From ac17ab87e8d93dabe93dfdc134e9853d47de7af1 Mon Sep 17 00:00:00 2001
From: dandan20031031 <2425926169@qq.com>
Date: Wed, 28 Oct 2026 17:12:03 +0800
Subject: [PATCH 1/5] feat(rope): add Qwen-Image multi-axis rotation with WS1
 validation

Provide stride-aware native CUDA and FlashAttention-derived Triton paths with fixed FP32 arithmetic, dispatcher/fake/autograd registration, standard operator validation, package build checks, and mature-baseline benchmark evidence. Refs #386.

Signed-off-by: dandan20031031 <2425926169@qq.com>
---
 MANIFEST.in                                   |    1 +
 .../rope/benchmark_multi_axis_rope.py         |  290 ++
 build_tools/extensions.py                     |   29 +
 csrc/bindings/rope_module.cpp                 |    8 +
 csrc/cuda/rope/multi_axis_rope.cu             |   82 +
 docs/.nav.yml                                 |    1 +
 docs/operators/README.md                      |    1 +
 docs/operators/multi-axis-rope.md             |  199 ++
 pyproject.toml                                |    2 +-
 .../README.md                                 |  152 +
 .../baseline-layout.txt                       |  138 +
 .../benchmark-bfloat16.json                   | 2781 +++++++++++++++++
 .../installed-wheel.json                      |   37 +
 .../installed-wheel.txt                       |   38 +
 .../layout.txt                                |  138 +
 .../manifest.json                             |   65 +
 .../memcheck.txt                              |    8 +
 .../operator-cuda.json                        |   70 +
 .../operator-triton.json                      |   70 +
 .../package-build.txt                         |   40 +
 .../package.json                              |   31 +
 .../precommit.txt                             |    7 +
 .../quality.txt                               |   58 +
 .../regression.txt                            |   10 +
 .../tests.txt                                 |    5 +
 .../upstream.json                             |  335 ++
 .../ws1-cuda.json                             | 1962 ++++++++++++
 .../ws1-cuda.txt                              |    3 +
 .../ws1-triton.json                           | 1962 ++++++++++++
 .../ws1-triton.txt                            |    3 +
 .../backends/cuda/rope/multi_axis_rope.py     |   49 +
 .../triton/rope/FLASH_ATTENTION_LICENSE       |   29 +
 .../shared/triton/rope/_flash_attn_rotary.py  |  141 +
 .../shared/triton/rope/multi_axis_rope.py     |   68 +
 rl_engine/ops/rope/_registration.py           |   41 +
 rl_engine/ops/rope/multi_axis_rope.py         |  132 +
 rl_engine/reference/rope/multi_axis_rope.py   |   25 +
 rl_engine/runtime/registry.py                 |   14 +
 .../validation/operators/gradient_adapters.py |   29 +
 .../validation/operators/operator_inputs.py   |   19 +
 .../validation/operators/operator_specs.py    |   14 +
 .../validation/operators/qwen_image_rope.py   |   80 +
 tests/build/test_multi_axis_rope_build.py     |   46 +
 tests/ops/rope/test_multi_axis_rope.py        |  443 +++
 .../operators/test_operator_inputs.py         |   18 +
 .../check_multi_axis_rope_install.py          |   67 +
 .../check_qwenimage_rope_upstream.py          |   99 +
 .../operators/validate_multi_axis_rope.py     |   85 +
 48 files changed, 9924 insertions(+), 1 deletion(-)
 create mode 100644 benchmarks/operators/rope/benchmark_multi_axis_rope.py
 create mode 100644 csrc/bindings/rope_module.cpp
 create mode 100644 csrc/cuda/rope/multi_axis_rope.cu
 create mode 100644 docs/operators/multi-axis-rope.md
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/README.md
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/baseline-layout.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/benchmark-bfloat16.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/layout.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/manifest.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/memcheck.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-cuda.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-triton.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package-build.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/precommit.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/quality.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/regression.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/tests.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/upstream.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.txt
 create mode 100644 rl_engine/backends/cuda/rope/multi_axis_rope.py
 create mode 100644 rl_engine/backends/shared/triton/rope/FLASH_ATTENTION_LICENSE
 create mode 100644 rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py
 create mode 100644 rl_engine/backends/shared/triton/rope/multi_axis_rope.py
 create mode 100644 rl_engine/ops/rope/_registration.py
 create mode 100644 rl_engine/ops/rope/multi_axis_rope.py
 create mode 100644 rl_engine/reference/rope/multi_axis_rope.py
 create mode 100644 rl_engine/validation/operators/qwen_image_rope.py
 create mode 100644 tests/build/test_multi_axis_rope_build.py
 create mode 100644 tests/ops/rope/test_multi_axis_rope.py
 create mode 100644 tools/validation/operators/check_multi_axis_rope_install.py
 create mode 100644 tools/validation/operators/check_qwenimage_rope_upstream.py
 create mode 100644 tools/validation/operators/validate_multi_axis_rope.py

diff --git a/MANIFEST.in b/MANIFEST.in
index 5ce1ee23f..67165e195 100644
--- a/MANIFEST.in
+++ b/MANIFEST.in
@@ -2,3 +2,4 @@ include pyproject.toml setup.py README.md LICENSE
 recursive-include build_tools *.py
 recursive-include csrc *.cpp *.cu *.cuh *.h *.hip *.asc
 recursive-include rl_engine *.json *.jsonl *.sh *.patch *.pyi
+recursive-include rl_engine FLASH_ATTENTION_LICENSE
diff --git a/benchmarks/operators/rope/benchmark_multi_axis_rope.py b/benchmarks/operators/rope/benchmark_multi_axis_rope.py
new file mode 100644
index 000000000..316a75ba7
--- /dev/null
+++ b/benchmarks/operators/rope/benchmark_multi_axis_rope.py
@@ -0,0 +1,290 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Standalone RoPE: contiguous/packed QKV, eager/graph, optional pinned FlashAttention."""
+
+from __future__ import annotations
+
+import argparse
+import ast
+import gc
+import hashlib
+import importlib.util
+import json
+import platform
+import statistics
+import subprocess
+import sys
+import time
+from pathlib import Path
+from typing import Optional, Union
+
+import torch
+
+ROOT = Path(__file__).resolve().parents[3]
+if str(ROOT) not in sys.path:
+    sys.path.insert(0, str(ROOT))
+
+from rl_engine.ops.rope.multi_axis_rope import (  # noqa: E402
+    build_multi_axis_cos_sin,
+    get_multi_axis_rope_op,
+    qwen_image_positions,
+)
+
+FLASH_REVISION = "98c1cee0ed59307833841cdbf567fe336ceec13a"
+FLASH_SHA256 = "d2037e623854c702fb075784fe4292d4a89d6b9ac1ae33cab3d6788f8f94f240"
+FLASH_LAYER_SHA256 = "a79cb42dfc33e126227bdba27ae7a024d7d5cc5c8fd7f7a5f26c68d501024e0a"
+
+
+def upstream_complex(x, cis):
+    """Diffusers use_real=False arithmetic; table assembly is outside timing."""
+    z = torch.view_as_complex(x.float().reshape(*x.shape[:-1], 64, 2))
+    return torch.view_as_real(z * cis.unsqueeze(-2)).flatten(-2).to(x.dtype)
+
+
+def load_flash_baselines(source_path, layer_path):
+    """Load checksum-pinned upstream kernel and autograd class without installing flash-attn."""
+    source = source_path.read_text(encoding="utf-8")
+    layer = layer_path.read_text(encoding="utf-8")
+    for text, expected in ((source, FLASH_SHA256), (layer, FLASH_LAYER_SHA256)):
+        if hashlib.sha256(text.encode()).hexdigest() != expected:
+            raise ValueError("FlashAttention source checksum mismatch (normalized UTF-8 text)")
+    node = next(n for n in ast.parse(layer).body if getattr(n, "name", None) == "ApplyRotaryEmb")
+    directory = ROOT / "artifacts/flash_rope_baseline"
+    directory.mkdir(parents=True, exist_ok=True)
+    result = {}
+    assert source.count("BLOCK_H=2,") == 1
+    for variant in ("default", "no_fma"):
+        code = source
+        if variant == "no_fma":
+            code = code.replace("BLOCK_H=2,", "BLOCK_H=2,\n            enable_fp_fusion=False,")
+        path = directory / f"flash_rotary_{variant}.py"
+        path.write_text(code, encoding="utf-8")
+        name = f"rlk_flash_baseline_{variant}"
+        spec = importlib.util.spec_from_file_location(name, path)
+        module = importlib.util.module_from_spec(spec)
+        sys.modules[name] = module
+        spec.loader.exec_module(module)
+        namespace = {
+            "__name__": name + "_autograd",
+            "torch": torch,
+            "Tensor": torch.Tensor,
+            "Optional": Optional,
+            "Union": Union,
+            "apply_rotary": module.apply_rotary,
+        }
+        exec(compile(ast.Module(body=[node], type_ignores=[]), str(layer_path), "exec"), namespace)
+        cls = namespace["ApplyRotaryEmb"]
+        result[f"flash_{variant}"] = lambda x, c, s, cls=cls: cls.apply(x, c, s, True)
+    return result
+
+
+def measure(fn, args):
+    for _ in range(args.warmup):
+        result = fn()
+    torch.cuda.synchronize()
+    del result
+    wall = []
+    for _ in range(args.trials):
+        start = time.perf_counter()
+        for _ in range(args.iterations):
+            result = fn()
+        torch.cuda.synchronize()
+        wall.append((time.perf_counter() - start) * 1000 / args.iterations)
+    del result
+
+    # Collect graph owners before capture, so GC cannot destroy another graph
+    # during stream capture. Preserve the caller's GC state.
+    gc.collect()
+    enabled = gc.isenabled()
+    gc.disable()
+    graph = torch.cuda.CUDAGraph()
+    try:
+        with torch.cuda.graph(graph, stream=torch.cuda.Stream()):
+            for _ in range(args.graph_calls):
+                result = fn()
+        for _ in range(args.warmup):
+            graph.replay()
+        events = [
+            (torch.cuda.Event(enable_timing=True), torch.cuda.Event(enable_timing=True))
+            for _ in range(args.iterations)
+        ]
+        for start, end in events:
+            start.record()
+            graph.replay()
+            end.record()
+        torch.cuda.synchronize()
+        times = [start.elapsed_time(end) / args.graph_calls for start, end in events]
+        del graph, result
+    finally:
+        if enabled:
+            gc.enable()
+
+    before = torch.cuda.memory_allocated()
+    torch.cuda.reset_peak_memory_stats()
+    result = fn()
+    torch.cuda.synchronize()
+    extra = torch.cuda.max_memory_allocated() - before
+    del result
+    return {
+        "graph_median_ms": statistics.median(times),
+        "graph_min_ms": min(times),
+        "graph_max_ms": max(times),
+        "eager_wall_median_ms": statistics.median(wall),
+        "eager_wall_trials_ms": wall,
+        "extra_peak_bytes": extra,
+    }
+
+
+def byte_differences(a, b):
+    return int((a.contiguous().view(torch.uint8) != b.contiguous().view(torch.uint8)).sum())
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--output", type=Path, required=True)
+    parser.add_argument("--iterations", type=int, default=30)
+    parser.add_argument("--warmup", type=int, default=5)
+    parser.add_argument("--trials", type=int, default=3)
+    parser.add_argument("--graph-calls", type=int, default=20)
+    parser.add_argument("--dtype", choices=["float32", "float16", "bfloat16"], default="bfloat16")
+    parser.add_argument(
+        "--layouts",
+        nargs="+",
+        choices=["contiguous", "packed_qkv"],
+        default=["contiguous", "packed_qkv"],
+    )
+    parser.add_argument("--flash-source", type=Path)
+    parser.add_argument("--flash-layer-source", type=Path)
+    args = parser.parse_args()
+    if min(args.iterations, args.warmup, args.trials, args.graph_calls) < 1:
+        parser.error("timing counts must be positive")
+    if bool(args.flash_source) != bool(args.flash_layer_source):
+        parser.error("both FlashAttention source files are required")
+    dtype = getattr(torch, args.dtype)
+    torch.manual_seed(386)
+    ours = {name: get_multi_axis_rope_op(name, device="cuda") for name in ("cuda", "triton")}
+    flash = (
+        load_flash_baselines(args.flash_source, args.flash_layer_source)
+        if args.flash_source
+        else {}
+    )
+    reference = get_multi_axis_rope_op("reference")
+    rows = []
+    correctness = []
+    for height, width in ((2, 3), (64, 64), (83, 83), (104, 58)):
+        positions, _ = qwen_image_positions(height, width, 0)
+        cos, sin = build_multi_axis_cos_sin(positions, device="cuda")
+        cis = torch.complex(cos, sin)
+
+        def upstream_op(x, _cos, _sin, cis=cis):
+            return upstream_complex(x, cis)
+
+        ops = {"upstream_complex": upstream_op, **ours, **flash}
+        for layout in args.layouts:
+            if layout == "packed_qkv":
+                packed = torch.randn(1, height * width, 3, 24, 128, device="cuda", dtype=dtype)
+                x = packed[:, :, 0].detach().requires_grad_()
+            else:
+                x = torch.randn(
+                    1, height * width, 24, 128, device="cuda", dtype=dtype, requires_grad=True
+                )
+            grad = torch.randn_like(x)
+            gold = reference(x, cos, sin)
+            gold_dx = torch.autograd.grad(gold, x, grad)[0]
+            for name, op in ops.items():
+                result = op(x, cos, sin)
+                dx = torch.autograd.grad(result, x, grad)[0]
+                diff = [byte_differences(result, gold), byte_differences(dx, gold_dx)]
+                if name in ("cuda", "triton", "flash_no_fma"):
+                    assert diff == [0, 0], (name, layout, x.shape, diff)
+                else:
+                    tol = {torch.float32: 2e-6, torch.float16: 0.004, torch.bfloat16: 0.04}[dtype]
+                    torch.testing.assert_close(result, gold, atol=tol, rtol=tol)
+                    torch.testing.assert_close(dx, gold_dx, atol=tol, rtol=tol)
+                correctness.append(
+                    {
+                        "backend": name,
+                        "layout": layout,
+                        "shape": list(x.shape),
+                        "different_bytes_y_dx": diff,
+                    }
+                )
+            del result, dx, gold, gold_dx
+            for name, op in ops.items():
+                for phase in ("forward", "forward_backward"):
+
+                    def call(op=op, phase=phase, x=x, cos=cos, sin=sin, grad=grad):
+                        if phase == "forward":
+                            with torch.no_grad():
+                                return op(x, cos, sin)
+                        return torch.autograd.grad(op(x, cos, sin), x, grad)[0]
+
+                    stats = measure(call, args)
+                    rows.append(
+                        {
+                            "backend": name,
+                            "phase": phase,
+                            "layout": layout,
+                            "shape": list(x.shape),
+                            "input_stride": list(x.stride()),
+                            **stats,
+                        }
+                    )
+                    print(
+                        f"{height}x{width} {layout:11s} {name:16s} {phase:16s} "
+                        f"graph={stats['graph_median_ms']:.5f}ms "
+                        f"eager={stats['eager_wall_median_ms']:.5f}ms",
+                        flush=True,
+                    )
+    sources = [
+        "csrc/cuda/rope/multi_axis_rope.cu",
+        "rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py",
+        "rl_engine/backends/shared/triton/rope/multi_axis_rope.py",
+        "rl_engine/backends/cuda/rope/multi_axis_rope.py",
+        "rl_engine/ops/rope/_registration.py",
+        "rl_engine/ops/rope/multi_axis_rope.py",
+        "rl_engine/reference/rope/multi_axis_rope.py",
+        "benchmarks/operators/rope/benchmark_multi_axis_rope.py",
+    ]
+    import triton
+
+    evidence = {
+        "gpu": torch.cuda.get_device_name(),
+        "capability": torch.cuda.get_device_capability(),
+        "os": platform.platform(),
+        "python": sys.version,
+        "torch": torch.__version__,
+        "cuda_runtime": torch.version.cuda,
+        "triton": triton.__version__,
+        "dtype": args.dtype,
+        "base_commit": subprocess.check_output(
+            ["git", "rev-parse", "HEAD"], cwd=ROOT, text=True
+        ).strip(),
+        "source_sha256": {p: hashlib.sha256((ROOT / p).read_bytes()).hexdigest() for p in sources},
+        "timing": "CUDA graph ms/call and synchronized eager wall ms/call; warm allocation; "
+        "excludes tables, compilation, and initial allocation of inputs",
+        "memory": "extra peak allocated bytes excluding inputs/tables/grad; includes returned "
+        "output or gradient; packed QKV backing storage allocated before timing",
+        "iterations": args.iterations,
+        "warmup": args.warmup,
+        "trials": args.trials,
+        "graph_calls": args.graph_calls,
+        "flash_baseline": {
+            "included": bool(flash),
+            "revision": FLASH_REVISION,
+            "kernel_sha256_normalized": FLASH_SHA256,
+            "layer_sha256_normalized": FLASH_LAYER_SHA256,
+            "mode": "upstream rotary_kernel, apply_rotary and ApplyRotaryEmb, interleaved=True",
+            "no_fma_change": "only add enable_fp_fusion=False at the original kernel launch",
+        },
+        "trace": {name: op.trace() for name, op in ours.items()},
+        "qualification": "Exact bytes vs unfused FP32 gold for CUDA, Triton and flash_no_fma; "
+        "declared tolerance for upstream complex and default FlashAttention",
+        "correctness": correctness,
+        "results": rows,
+    }
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(json.dumps(evidence, indent=2) + "\n", encoding="utf-8")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/build_tools/extensions.py b/build_tools/extensions.py
index db079e828..2f04f5cdd 100644
--- a/build_tools/extensions.py
+++ b/build_tools/extensions.py
@@ -105,6 +105,7 @@ def _native_extension_required() -> bool:
     """Whether the caller explicitly requested a native extension build."""
     return (
         envs.env_flag(envs.RL_KERNEL_REQUIRE_EXT)
+        or envs.env_flag("RL_KERNEL_ROPE_ONLY")
         or bool(os.environ.get("PYTORCH_ROCM_ARCH", "").strip())
         or bool(os.environ.get("TORCH_CUDA_ARCH_LIST", "").strip())
         or envs.env_flag("FORCE_CUDA")
@@ -182,6 +183,33 @@ def get_extensions():
         torch_rpath.append(f"-Wl,-rpath,{torch_lib_dir}")
     is_rocm = getattr(torch.version, "hip", None) is not None
 
+    # A normal wheel build for the independently usable RoPE library. This
+    # explicit profile avoids compiling unrelated Hopper/IPC operators and is
+    # also useful on Windows. Default builds still include RoPE in the full _C.
+    if envs.env_flag("RL_KERNEL_ROPE_ONLY"):
+        if is_rocm or getattr(torch.version, "cuda", None) is None:
+            raise RuntimeError("RL_KERNEL_ROPE_ONLY requires CUDA-enabled PyTorch")
+        from torch.utils.cpp_extension import CUDA_HOME
+
+        library_dirs = []
+        if os.name == "nt" and CUDA_HOME and (Path(CUDA_HOME) / "lib/cudart.lib").is_file():
+            library_dirs.append(str(Path(CUDA_HOME) / "lib"))
+        return [
+            GpuExtension(
+                name="rl_engine._rope_cuda",
+                sources=[
+                    "csrc/bindings/rope_module.cpp",
+                    "csrc/cuda/rope/multi_axis_rope.cu",
+                ],
+                library_dirs=library_dirs,
+                extra_compile_args={
+                    "cxx": ["/O2", "/std:c++17"] if os.name == "nt" else ["-O3", "-std=c++17"],
+                    "nvcc": ["-O3", "-std=c++17", "--fmad=false"],
+                },
+                extra_link_args=[] if os.name == "nt" else list(torch_rpath),
+            )
+        ]
+
     if _musa_build_available(torch):
         extensions.append(
             GpuExtension(
@@ -233,6 +261,7 @@ def get_extensions():
                 ]
             )
         else:
+            cuda_sources.append("csrc/cuda/rope/multi_axis_rope.cu")
             # CUDA IPC and the fixed-tree collective implementation are not
             # part of the ROCm extension.
             cuda_sources.append("csrc/cuda/collectives/deterministic_collective.cu")
diff --git a/csrc/bindings/rope_module.cpp b/csrc/bindings/rope_module.cpp
new file mode 100644
index 000000000..c12e5b90d
--- /dev/null
+++ b/csrc/bindings/rope_module.cpp
@@ -0,0 +1,8 @@
+// SPDX-License-Identifier: Apache-2.0
+// Python module entry point for the optional RoPE-only wheel. Operators are
+// registered by multi_axis_rope.cu through the PyTorch dispatcher.
+#include <pybind11/pybind11.h>
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.doc() = "Qwen-Image RoPE dispatcher library";
+}
diff --git a/csrc/cuda/rope/multi_axis_rope.cu b/csrc/cuda/rope/multi_axis_rope.cu
new file mode 100644
index 000000000..3de49caba
--- /dev/null
+++ b/csrc/cuda/rope/multi_axis_rope.cu
@@ -0,0 +1,82 @@
+// SPDX-License-Identifier: Apache-2.0
+// Qwen-Image adjacent-pair rotation. Reads strided BSHD inputs/tables directly.
+#include <ATen/ATen.h>
+#include <torch/library.h>
+#include <cuda_runtime.h>
+#include <c10/cuda/CUDAStream.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <c10/cuda/CUDAException.h>
+#include <algorithm>
+
+namespace {
+struct Strides {
+  int64_t xb, xs, xh, xd;
+  int64_t cb, cs, cd, sb, ss, sd;
+};
+
+template <typename scalar_t>
+__global__ void rotate_pairs(const scalar_t* x, const float* cos, const float* sin,
+                             scalar_t* y, int64_t n, int64_t seq, int64_t heads,
+                             Strides strides, bool transpose) {
+  const int64_t step = static_cast<int64_t>(blockDim.x) * gridDim.x;
+  for (int64_t i = static_cast<int64_t>(blockIdx.x) * blockDim.x + threadIdx.x;
+       i < n; i += step) {
+    const int64_t pair = i % 64;
+    const int64_t row = i / 64;
+    const int64_t head = row % heads;
+    const int64_t token = row / heads;
+    const int64_t batch = token / seq;
+    const int64_t position = token % seq;
+    const int64_t xi = batch * strides.xb + position * strides.xs +
+                       head * strides.xh + 2 * pair * strides.xd;
+    const float c = cos[batch * strides.cb + position * strides.cs + pair * strides.cd];
+    const float sn = sin[batch * strides.sb + position * strides.ss + pair * strides.sd];
+    const float s = transpose ? -sn : sn;
+    const float a = static_cast<float>(x[xi]);
+    const float b = static_cast<float>(x[xi + strides.xd]);
+    y[2 * i] = static_cast<scalar_t>(__fsub_rn(__fmul_rn(a, c), __fmul_rn(b, s)));
+    y[2 * i + 1] = static_cast<scalar_t>(__fadd_rn(__fmul_rn(b, c), __fmul_rn(a, s)));
+  }
+}
+
+at::Tensor multi_axis_rope_apply(const at::Tensor& x, const at::Tensor& cos,
+                                const at::Tensor& sin, bool transpose) {
+  TORCH_CHECK(x.is_cuda() && cos.is_cuda() && sin.is_cuda(), "multi_axis_rope: CUDA tensors required");
+  TORCH_CHECK(x.device() == cos.device() && x.device() == sin.device(), "multi_axis_rope: device mismatch");
+  TORCH_CHECK(x.dim() == 4 && x.size(3) == 128 && x.size(2) > 0,
+              "multi_axis_rope: x must be [B, S, H, 128], H > 0");
+  TORCH_CHECK(x.scalar_type() == at::kFloat || x.scalar_type() == at::kHalf ||
+              x.scalar_type() == at::kBFloat16, "multi_axis_rope: unsupported input dtype");
+  TORCH_CHECK(cos.scalar_type() == at::kFloat && sin.scalar_type() == at::kFloat,
+              "multi_axis_rope: tables must be fp32");
+  TORCH_CHECK(cos.sizes() == sin.sizes(), "multi_axis_rope: table shape mismatch");
+  TORCH_CHECK((cos.dim() == 2 && cos.size(0) == x.size(1) && cos.size(1) == 64) ||
+              (cos.dim() == 3 && cos.size(0) == x.size(0) && cos.size(1) == x.size(1) && cos.size(2) == 64),
+              "multi_axis_rope: tables must be [S, 64] or [B, S, 64]");
+  TORCH_CHECK(!cos.requires_grad() && !sin.requires_grad(), "multi_axis_rope: table gradients unsupported");
+  const c10::cuda::CUDAGuard guard(x.device());
+  auto y = at::empty(x.sizes(), x.options());
+  const int64_t n = x.numel() / 2;
+  if (n == 0) return y;
+  const Strides strides{x.stride(0), x.stride(1), x.stride(2), x.stride(3),
+      cos.dim() == 3 ? cos.stride(0) : 0, cos.stride(-2), cos.stride(-1),
+      sin.dim() == 3 ? sin.stride(0) : 0, sin.stride(-2), sin.stride(-1)};
+  const int blocks = static_cast<int>(std::min<int64_t>((n + 255) / 256, 65535));
+  auto stream = c10::cuda::getCurrentCUDAStream(x.get_device());
+  AT_DISPATCH_FLOATING_TYPES_AND2(at::ScalarType::Half, at::ScalarType::BFloat16,
+      x.scalar_type(), "multi_axis_rope", [&] {
+    rotate_pairs<scalar_t><<<blocks, 256, 0, stream>>>(x.data_ptr<scalar_t>(), cos.data_ptr<float>(),
+        sin.data_ptr<float>(), y.data_ptr<scalar_t>(), n, x.size(1), x.size(2), strides, transpose);
+  });
+  C10_CUDA_KERNEL_LAUNCH_CHECK();
+  return y;
+}
+}  // namespace
+
+// The focused validator and normal package build load this same registration.
+TORCH_LIBRARY_FRAGMENT(rl_kernel_native, m) {
+  m.def("multi_axis_rope(Tensor x, Tensor cos, Tensor sin, bool transpose) -> Tensor");
+}
+TORCH_LIBRARY_IMPL(rl_kernel_native, CUDA, m) {
+  m.impl("multi_axis_rope", &multi_axis_rope_apply);
+}
diff --git a/docs/.nav.yml b/docs/.nav.yml
index 29b5ead37..d03ade331 100644
--- a/docs/.nav.yml
+++ b/docs/.nav.yml
@@ -29,6 +29,7 @@ nav:
       - operators/sampling.md
       - operators/det-gemm.md
       - operators/embedding.md
+      - operators/multi-axis-rope.md
   - Developer Guide:
     - contributing/README.md
     - Contributor Guide: contributing/contributor-guide.md
diff --git a/docs/operators/README.md b/docs/operators/README.md
index 00f4cbb45..2d3b27d5d 100644
--- a/docs/operators/README.md
+++ b/docs/operators/README.md
@@ -26,6 +26,7 @@ Every operator page should include:
 - [Fused Linear LogP TP Test Runbook](linear-logp-tp-test.md)
 - [GRPO Loss](grpo-loss.md)
 - [RoPE](rope.md)
+- [Qwen-Image Multi-Axis RoPE](multi-axis-rope.md)
 - [LM Head](lm_head.md)
 - [Policy Ratio + KL Penalty](ratio-kl.md)
 - [Pack and Pad](pack-and-pad.md)
diff --git a/docs/operators/multi-axis-rope.md b/docs/operators/multi-axis-rope.md
new file mode 100644
index 000000000..fa908a03e
--- /dev/null
+++ b/docs/operators/multi-axis-rope.md
@@ -0,0 +1,199 @@
+# Qwen-Image Multi-Axis RoPE
+
+This operator implements the `multi_axis_rope` CUDA row of
+[RFC #386](https://github.com/RL-Align/RL-Kernel/issues/386). It rotates **adjacent
+channel pairs** with three independently indexed axes of widths `[16, 56, 56]`.
+The existing GPT-NeoX rotate-half operator has different semantics.
+
+## Reference and scope
+
+Semantics are pinned to Diffusers
+[`40f63690da440e8c382ab53e67508835deb1ed2f`](https://github.com/huggingface/diffusers/blob/40f63690da440e8c382ab53e67508835deb1ed2f/src/diffusers/models/transformers/transformer_qwenimage.py):
+`QwenEmbedRope(scale_rope=True)` and `apply_rotary_emb_qwen(use_real=False)`.
+The position helper covers single-frame text-to-image at packed latent resolution.
+For a 1024×1024 image the packed grid is 64×64. Multi-image editing, video,
+QwenEmbedLayer3DRope, distributed execution and full-model integration are outside
+this operator's qualification.
+
+Image coordinates are `(0, h, w)` in row-major order. The centered coordinate
+range is `[-ceil(size/2), floor(size/2))`, so an odd grid has one extra negative
+coordinate. Text starts at `max(height//2, width//2)` on all three axes. The helper
+returns image and text positions separately: the caller determines stream order.
+
+## API
+
+```python
+import torch
+from rl_engine.ops.rope.multi_axis_rope import (
+    qwen_image_positions, build_multi_axis_cos_sin, get_multi_axis_rope_op,
+)
+
+image_pos, text_pos = qwen_image_positions(64, 64, 128)
+cos, sin = build_multi_axis_cos_sin(image_pos, device="cuda")
+x = torch.randn(1, 4096, 24, 128, device="cuda", dtype=torch.bfloat16,
+                requires_grad=True)
+op = get_multi_axis_rope_op("cuda", device=x.device)
+y = op(x, cos, sin)
+y.float().square().mean().backward()
+print(op.trace())
+```
+
+| Argument | Contract |
+| --- | --- |
+| `x` | `[B, S, H, 128]`, H > 0; FP16, BF16 or FP32 |
+| `cos`, `sin` | Matching FP32 `[S, 64]` shared tables or `[B, S, 64]` per-sample tables |
+| Devices | Input and tables on the same device |
+| Layout | BSHD, adjacent pairs; input and independent table strides read directly |
+| Output | Input shape/dtype/device; contiguous |
+| Empty input | B=0 and/or S=0 supported, including backward |
+| Gradients | Input gradients and higher-order input VJPs; table gradients rejected |
+| Position builder | int32/int64 `[S, 3]` or `[B, S, 3]`, within `[-4096, 4095]` |
+
+Table construction follows the upstream CPU FP32 `pow` and `polar` operations,
+then copies the results to the target device. Build and retain tables before CUDA
+Graph capture. Reusing these tables also prevents per-batch transcendental drift.
+The CPU factories select their device explicitly, including inside a CUDA or
+meta default-device context. Packed QKV views require no materializing copy.
+
+## Numerical contract
+
+Contract ID: `qwenimage-interleaved-rope-fp32-v1`.
+
+For each adjacent pair `(a, b)`, compute `(a*c - b*s, b*c + a*s)` in FP32 with
+separate rounded multiplications and additions, followed by one output cast.
+There is no cross-element reduction, split-K, atomic accumulation or TF32.
+CUDA uses round-to-nearest arithmetic intrinsics; Triton disables FP contraction.
+Backward uses the transpose rotation (negate sine), including recursive VJPs.
+
+CUDA and Triton must agree in raw bytes within a pinned environment. Compare
+batch placement, token chunking, padding and graph replay the same way.
+Independent FP64 complex arithmetic provides an accuracy oracle. Upstream's GPU
+complex multiplication may contract arithmetic, so its comparison uses declared
+tolerances, not an unverified bitwise claim. Representative test tolerances are
+`8e-7` for FP32, `0.002` for FP16 and `0.02` for BF16 (absolute and relative).
+Non-finite propagation follows the explicit arithmetic; NaN payload equality is
+not a cross-backend guarantee.
+
+## Backends and dispatch
+
+- `cuda`: production function in `csrc/cuda/rope/multi_axis_rope.cu`, registered
+  in `rl_engine._C` by ordinary CUDA builds. Does not require Hopper-specific PTX.
+- `triton`: adapts FlashAttention's existing interleaved rotary kernel, with
+  explicit shared/per-sample table strides and `enable_fp_fusion=False`.
+  The upstream tile geometry is retained; this profile is limited to NVIDIA CUDA.
+- `reference`: PyTorch FP32 arithmetic, including CPU.
+- `auto`: CUDA → Triton → reference on NVIDIA, reference on other platforms.
+
+Explicit `cuda`/`triton` requests fail when unavailable; they never silently switch
+implementations. Strict callers select an explicit backend and retain `op.trace()`.
+`auto` is for diagnostic use and is not a strict-route certification. The legacy
+registry also exposes `kernel_registry.get_op("multi_axis_rope", device=...)`.
+This PR does not add an Ascend backend or certify ROCm. Existing Ascend work in
+PR #411 can share semantics once its reference contract is aligned.
+
+Both GPU paths expose a PyTorch custom operator with fake-tensor and autograd
+registration. `torch.compile(op, fullgraph=True)` supports forward and backward.
+The native implementation is registered through `TORCH_LIBRARY`; Python does
+not attach symbols to an unrelated extension at validation time.
+
+## Related work and reuse
+
+The Triton kernel comes from FlashAttention
+[`98c1cee`](https://github.com/Dao-AILab/flash-attention/blob/98c1cee0ed59307833841cdbf567fe336ceec13a/flash_attn/ops/triton/rotary.py),
+with its BSD-3-Clause copyright and full license retained beside the adapted
+source and included in wheels/source distributions. Its default fused arithmetic
+can differ in low bits; disabling contraction gives the declared profile.
+The benchmark loads the original kernel and autograd wrapper from checksum-pinned
+source files to compare the default and contraction-disabled variants.
+
+[Ascend PR #411](https://github.com/RL-Align/RL-Kernel/pull/411) covers the same
+roadmap row on another platform. Its current half-split reference and uncentered
+positions differ from the pinned Diffusers contract here. Combining the two
+references requires semantic alignment. Existing Qwen-Image implementations in
+[SGLang](https://github.com/sgl-project/sglang/tree/a1966e1b5d8b750b494d7e53cca1d2464f960038/python/sglang/multimodal_gen)
+and [vLLM-Omni](https://github.com/vllm-project/vllm-omni/tree/c9c288e3f5828f2a2a9b07acb16b3b709906059a)
+also motivate stride support and future QK-norm/RoPE fusion. This contribution
+qualifies the unfused operator; it makes no new-algorithm or fused-model claim.
+
+## Validation and benchmarks
+
+```bash
+python -m pytest tests/ops/rope/test_multi_axis_rope.py -q -rs
+python tools/validation/operators/check_operator.py --op multi_axis_rope \
+  --candidate cuda --device cuda --dtype bf16 --batch 2 --seq 7 --n-heads 3 --check-grad
+python -m benchmarks.operators.rope.benchmark_multi_axis_rope --output rope.json
+python tools/validation/operators/validate_multi_axis_rope.py --ws1 cuda --output ws1-cuda.json
+python tools/validation/operators/validate_multi_axis_rope.py --ws1 triton --output ws1-triton.json
+mkdocs build --strict -f mkdocs.yaml
+```
+
+For operator-only native compilation, without building every unrelated extension:
+
+```bash
+python tools/validation/operators/validate_multi_axis_rope.py -q -rs
+python tools/validation/operators/validate_multi_axis_rope.py --benchmark --output rope.json
+```
+
+To audit against the actual pinned upstream symbols without installing Diffusers,
+download the source file linked above and run:
+
+```bash
+python tools/validation/operators/check_qwenimage_rope_upstream.py \
+  --source transformer_qwenimage.py --output upstream.json
+```
+
+The audit checks the source checksum before loading the two relevant definitions,
+then compares image/text tables and forward/backward across three dtypes.
+
+The focused runner loads the **same production CUDA source and dispatcher
+registration**, without a generated binding or extension monkeypatch. It reuses
+an installed native library when available. Set a matching `CUDA_HOME`, CUDA
+architecture and host compiler before invoking it.
+
+Ordinary CUDA builds include the source in `rl_engine._C`. An explicit operator
+wheel profile builds `rl_engine._rope_cuda` with the same dispatcher registration:
+
+```bash
+RL_KERNEL_ROPE_ONLY=1 python -m pip wheel --no-build-isolation --no-deps . -w dist
+python -m pip install --force-reinstall --no-deps dist/rl_kernel-*.whl
+# Change to a directory outside the source checkout before this command:
+python /path/to/RL-Kernel/tools/validation/operators/check_multi_axis_rope_install.py \
+  --package-root /path/to/site-packages --output installed-wheel.json
+```
+
+On Windows set `$env:RL_KERNEL_ROPE_ONLY='1'` in PowerShell and enter the matching
+MSVC build environment. Select a CUDA toolkit matching `torch.version.cuda`.
+The operator profile does not certify the complete collection of unrelated
+native operators. CPU-only/ROCm builds reject this explicit CUDA wheel profile.
+
+`multi_axis_rope` has an `OP_SPECS` entry, standard input builder and gradient
+adapter. `--ws1` runs the existing forward and gradient accuracy/invariance APIs
+with C1's elementwise tolerance profile and C2's logical configuration transforms.
+This gate qualifies the Qwen-Image operator and does not certify the Dense Qwen3
+chain, distributed execution, or model assembly.
+
+Tests cover independent forward/VJP accuracy, adjacent-pair layout, shared and
+batched positions, odd and rectangular grids, batch/padding/chunk invariance,
+non-contiguous tensors, packed QKV views without copies, default-device contexts,
+empty inputs, boundary rejection, higher derivatives, CUDA/Triton byte equality,
+graph replay, the current CUDA stream, `torch.library.opcheck` and full-graph
+compilation. A two-device guard test skips when only one GPU is present.
+
+The benchmark covers a short synthetic case and 64×64, 83×83 and 104×58 grids,
+with 24 heads and both contiguous/packed-QKV layouts. It checks forward and input
+gradients before timing and records CUDA Graph and synchronized eager wall time,
+extra peak allocation, source hashes and actual-backend traces. Table generation
+and compilation are excluded. These measurements establish no model speedup.
+
+To include the mature FlashAttention baselines, download `flash_attn/ops/triton/rotary.py`
+and `flash_attn/layers/rotary.py` at the linked revision, then pass both paths:
+
+```bash
+python tools/validation/operators/validate_multi_axis_rope.py --benchmark \
+  --flash-source /path/to/ops/rotary.py --flash-layer-source /path/to/layers/rotary.py \
+  --output rope-with-flash.json
+```
+
+The normalized source checksums are enforced before loading the audited symbols.
+The default variant retains upstream arithmetic; `flash_no_fma` changes only the
+kernel launch contraction setting and must match the FP32 gold's output bytes.
diff --git a/pyproject.toml b/pyproject.toml
index 207606282..f4f8b8ae9 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -41,7 +41,7 @@ where = ["."]
 include = ["rl_engine*"]
 
 [tool.setuptools.package-data]
-rl_engine = ["**/*.json", "**/*.jsonl", "**/*.sh", "**/*.patch", "*.pyi"]
+rl_engine = ["**/*.json", "**/*.jsonl", "**/*.sh", "**/*.patch", "*.pyi", "**/FLASH_ATTENTION_LICENSE"]
 
 [tool.ruff]
 line-length = 100
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/README.md b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/README.md
new file mode 100644
index 000000000..64232c731
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/README.md
@@ -0,0 +1,152 @@
+# Qwen-Image multi-axis RoPE validation — 2026-10-28
+
+Evidence for the CUDA `multi_axis_rope` row of
+[RFC #386](https://github.com/RL-Align/RL-Kernel/issues/386), based on
+`test-qwenimage@16b81175874ecf11d5ff05c00de3b69e9457591b`.
+The source fingerprints in `manifest.json` identify the final implementation.
+This is single-GPU operator evidence; it does not qualify a model or trajectory.
+
+## Implementation
+
+- Adjacent pairs, axes `[16, 56, 56]`, canonical centered image positions and
+  diagonal text offsets, pinned to Diffusers `40f63690da440e8c382ab53e67508835deb1ed2f`.
+- Explicit CPU table factories remain CPU factories inside CUDA/meta default-device contexts.
+- Native CUDA reads all input/table strides directly, including packed QKV views.
+  The shared Triton implementation adapts FlashAttention's existing rotary kernel
+  at `98c1cee0ed59307833841cdbf567fe336ceec13a`, retaining attribution and BSD-3 terms.
+- Both GPU paths have dispatcher, fake and input-autograd registration. The
+  arithmetic is separate FP32 multiply/add followed by one output cast.
+- OP_SPECS, the ordinary input builder and a gradient adapter expose the operator
+  to existing validation APIs. The operator gate runs all four forward/gradient
+  accuracy/invariance judgments with explicit backend/kernel provenance.
+- Default CUDA builds include the native source in `_C`. The explicit
+  `RL_KERNEL_ROPE_ONLY=1` build profile also produces an installable companion
+  library, without compiling unrelated native operators.
+
+## Environment
+
+| Item | Value |
+| --- | --- |
+| GPU | NVIDIA GeForce RTX 5080, SM120, 16 GB |
+| Driver | 617.42 |
+| OS / Python | Windows / 3.13.9 |
+| PyTorch / CUDA toolkit | 2.9.1+cu128 / 12.8 |
+| Triton | triton-windows 3.5.1.post24; runtime 3.5.1 |
+| Native compiler | MSVC 2022 x64, C++17; CUDA target 12.0 |
+| Sanitizer | NVIDIA Compute Sanitizer 13.3, memcheck |
+| Benchmark dtype | BF16, FP32 position tables |
+
+The installed wheel is CPython 3.13/Windows-specific. Toolkit/driver versions
+have different roles; the native code was compiled with CUDA 12.8, not 13.3.
+
+## Validation
+
+| Check | Result | Evidence |
+| --- | --- | --- |
+| Public operator suite using the built companion library | 81 passed, 1 two-GPU skip | `tests.txt` |
+| Operator/build/registry/validation regressions | 195 passed, 7 Ascend skips | `regression.txt` |
+| CUDA WS1 four judgments | Passed | `ws1-cuda.json` |
+| Triton WS1 four judgments | Passed | `ws1-triton.json` |
+| Standard CLI, BF16 forward and input gradient, CUDA/Triton | Both passed | `operator-{cuda,triton}.json` |
+| Pinned actual Diffusers symbols | 30 cases passed; image/text tables exact | `upstream.json` |
+| Memory safety on selected accuracy/strided/empty/stream cases | 34 passed, zero memcheck errors | `memcheck.txt` |
+| Source distribution + operator wheel build | Passed; native sources and BSD license present | `package-build.txt`, `package.json` |
+| Fresh interpreter outside checkout after wheel installation | CUDA/Triton eager and fullgraph forward/input-gradient bytes exact | `installed-wheel.json` |
+| Changed Python formatting/static checks; strict docs build | Passed | `quality.txt` |
+| Repository layout check | 5 failures, 31 passed, 1 skipped | `layout.txt` |
+| Same layout check on clean base | Same 5 failures, 27 passed, 1 skipped | `baseline-layout.txt` |
+
+Operator tests cover FP32/FP16/BF16, shared/per-sample tables, odd and rectangular
+grids, batch/chunk/padding invariance, higher input VJPs, strided gradients,
+independent table strides, empty tensors, invalid-input rejection, non-default
+streams, graph replay with changed static buffers, `torch.library.opcheck`, and
+`torch.compile(..., fullgraph=True)` through backward. A CPU profiler assertion
+rejects `contiguous`, `clone` or `copy_` inside either GPU call for a packed view.
+
+The layout failures reproduce on the clean base: the Windows driver-link flag
+assertion, symlink creation without Windows privilege, and three checks affected
+by symlink-disabled JSON/source placeholders. The added build tests pass. This
+report does not describe the entire repository test suite as green.
+
+## Mature baseline and performance
+
+`benchmark-bfloat16.json` contains 40 pre-timing correctness comparisons and 80
+measurements: four geometries, two layouts, five implementations, two phases.
+The CUDA, adapted Triton and contraction-disabled upstream FlashAttention paths
+all have zero differing output/gradient bytes against the unfused FP32 gold.
+Original FlashAttention and Diffusers complex arithmetic use declared tolerances.
+
+The optional FlashAttention comparison loads the checksum-pinned original
+`rotary_kernel`, `apply_rotary` and `ApplyRotaryEmb` class without installing
+flash-attn. `flash_no_fma` adds only `enable_fp_fusion=False` to the original
+launch; the default baseline retains its fused arithmetic.
+
+CUDA Graph medians in ms/call; each cell is **forward / forward+backward**.
+Batch 1, 24 heads, D=128:
+
+| Grid / layout | Native CUDA | Adapted Triton | FlashAttention, no FMA | Diffusers complex |
+| --- | --- | --- | --- | --- |
+| 64×64 contiguous | 0.05771 / 0.12953 | 0.05191 / 0.12463 | 0.05202 / 0.12443 | 0.24277 / 0.50118 |
+| 64×64 packed QKV | 0.05866 / 0.13089 | 0.04835 / 0.12314 | 0.04896 / 0.12374 | 0.24599 / 0.50083 |
+| 83×83 contiguous | 0.11203 / 0.22478 | 0.10685 / 0.21339 | 0.10693 / 0.21351 | 0.51009 / 1.05054 |
+| 83×83 packed QKV | 0.11280 / 0.22514 | 0.10758 / 0.21415 | 0.10746 / 0.21431 | 0.51117 / 1.05461 |
+| 104×58 contiguous | 0.09809 / 0.19646 | 0.09356 / 0.18708 | 0.09362 / 0.18709 | 0.43411 / 0.89356 |
+| 104×58 packed QKV | 0.09596 / 0.19664 | 0.09341 / 0.18776 | 0.09344 / 0.18786 | 0.43668 / 0.89561 |
+
+The adapted Triton path is close to its mature upstream baseline. The native
+CUDA path avoids requiring Triton but is slower in this graph comparison.
+No speed advantage over FlashAttention is established. At 4096 tokens, either
+GPU path's forward extra peak is 24 MiB, exactly the returned BF16 output size,
+for both layouts. There is no extra packed-input materialization.
+
+Five warm calls/replays, 20 calls per captured graph, median of 30 replay
+measurements. Synchronized eager wall time is also recorded as three trials of
+30 calls. It includes Python/dispatcher/autograd overhead and can differ
+substantially from graph time. Tables, compilation, input allocation and packed
+backing storage are outside timing. Extra peak allocation includes the returned
+output or gradient. These are one-machine, one-process microbenchmarks, without
+clock pinning; close differences do not establish a stable winner.
+
+## Reproduction
+
+Set a toolkit matching PyTorch, the target architecture and a matching host
+compiler. On the recorded Windows host, run from an x64 MSVC environment with
+`CUDA_HOME`/`CUDA_PATH` selecting CUDA 12.8 and `TORCH_CUDA_ARCH_LIST=12.0`.
+Use `$env:RL_KERNEL_ROPE_ONLY='1'` for the operator package profile.
+
+```bash
+python setup.py sdist bdist_wheel
+python setup.py build_ext --inplace
+python -m pytest tests/ops/rope/test_multi_axis_rope.py -q -rs
+python tools/validation/operators/check_operator.py --op multi_axis_rope \
+  --candidate cuda --device cuda --dtype bf16 --batch 2 --seq 7 --n-heads 3 --check-grad --json
+python tools/validation/operators/validate_multi_axis_rope.py --ws1 cuda --output ws1-cuda.json
+python tools/validation/operators/validate_multi_axis_rope.py --ws1 triton --output ws1-triton.json
+python tools/validation/operators/validate_multi_axis_rope.py --benchmark \
+  --flash-source /path/to/flash_attn/ops/triton/rotary.py \
+  --flash-layer-source /path/to/flash_attn/layers/rotary.py --output benchmark.json
+python tools/validation/operators/check_qwenimage_rope_upstream.py \
+  --source /path/to/transformer_qwenimage.py --output upstream.json
+compute-sanitizer --tool memcheck --error-exitcode 1 python \
+  tools/validation/operators/validate_multi_axis_rope.py -q \
+  -k "forward_backward_accuracy or noncontiguous_input_tables_and_gradient or empty_forward_backward or packed_qkv_and_independent_table_strides or native_entrypoint_checks"
+mkdocs build --strict -f mkdocs.yaml
+```
+
+Install the built wheel, change to a directory outside the checkout, then run
+`check_multi_axis_rope_install.py --package-root /path/to/site-packages --output installed-wheel.json`.
+This probe neither changes sys.path nor invokes the focused JIT loader.
+See `quality.txt` for the regression command and static-check scope.
+
+## Related work and remaining qualification
+
+The [operator page](../../../docs/operators/multi-axis-rope.md) documents
+FlashAttention reuse and the related Ascend #411, SGLang and vLLM-Omni paths.
+The live repository search on 2026-10-28 found #411 for the exact row name;
+its head remains `285c4dccab7e9048956584caf431946d82405c09`. Its half-split and
+uncentered reference differs from the pinned adjacent-pair/centered contract.
+
+Linux, A100/H100, multi-GPU guards, the complete monolithic `_C` build, QK-norm
+fusion, model/block assembly and trajectories remain unqualified. There is no
+Dense Qwen3 or distributed certification implied by the operator WS1 adapter.
+Cross-version/device byte equality and fast-math builds are not claimed.
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/baseline-layout.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/baseline-layout.txt
new file mode 100644
index 000000000..16a5c97e8
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/baseline-layout.txt
@@ -0,0 +1,138 @@
+.FF.............s........F.FF....                                        [100%]
+================================== FAILURES ===================================
+________________ test_cuda_build_keeps_existing_ipc_collective ________________
+
+monkeypatch = <_pytest.monkeypatch.MonkeyPatch object at 0x0000020FF72168B0>
+
+    def test_cuda_build_keeps_existing_ipc_collective(monkeypatch) -> None:
+        extension = _load_extension_config(monkeypatch, hip=None)
+
+        assert "csrc/cuda/collectives/deterministic_collective.cu" in extension["sources"]
+        assert "-DKERNEL_ALIGN_WITH_CUDA" in extension["extra_compile_args"]["cxx"]
+        assert "-DKERNEL_ALIGN_WITH_ROCM" not in extension["extra_compile_args"]["cxx"]
+>       assert "-lcuda" in extension["extra_link_args"]
+E       AssertionError: assert '-lcuda' in ['-Wl,-rpath,$ORIGIN/../torch/lib']
+
+tests\build\test_build_platform_collectives.py:61: AssertionError
+____________ test_package_data_json_links_resolve_to_real_content _____________
+
+tmp_path = WindowsPath('C:/Users/123/AppData/Local/Temp/pytest-of-123/pytest-4/test_package_data_json_links_r0')
+
+    def test_package_data_json_links_resolve_to_real_content(tmp_path) -> None:
+        target = tmp_path / "configs" / "profile.json"
+        target.parent.mkdir()
+        target.write_text('{"name": "profile"}\n', encoding="utf-8")
+        source = tmp_path / "rl_engine" / "profile.json"
+        source.parent.mkdir()
+>       source.symlink_to("../configs/profile.json")
+
+tests\build\test_build_platform_collectives.py:70:
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
+
+self = WindowsPath('C:/Users/123/AppData/Local/Temp/pytest-of-123/pytest-4/test_package_data_json_links_r0/rl_engine/profile.json')
+target = '../configs/profile.json', target_is_directory = False
+
+    def symlink_to(self, target, target_is_directory=False):
+        """
+        Make this path a symlink pointing to the target path.
+        Note the order of arguments (link, target) is the reverse of os.symlink.
+        """
+>       os.symlink(target, self, target_is_directory)
+E       OSError: [WinError 1314] �ͻ���û���������Ȩ��: '../configs/profile.json' -> 'C:\\Users\\123\\AppData\\Local\\Temp\\pytest-of-123\\pytest-4\\test_package_data_json_links_r0\\rl_engine\\profile.json'
+
+D:\anaconda\Lib\pathlib\_local.py:789: OSError
+_______ test_recorded_algorithm_sources_resolve_to_real_implementations _______
+
+    def test_recorded_algorithm_sources_resolve_to_real_implementations():
+        from rl_engine.config.workload import load_manifest
+
+        moves = json.loads((ROOT / "tools/migration/layout.json").read_text())["moves"]
+        for case in load_manifest().representative_cases:
+            source, symbol = case["provenance_evidence"]["algorithm_source"].rsplit(":", 1)
+            implementation = ROOT / moves.get(source, source)
+>           text = implementation.read_text()
+                   ^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+tests\runtime\test_layout_compatibility.py:87:
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
+D:\anaconda\Lib\pathlib\_local.py:546: in read_text
+    return PathBase.read_text(self, encoding, errors, newline)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
+
+self = WindowsPath('D:/dandan123/dandan/issue/investigation-rl-kernel/baseline-validation/csrc/cuda/attention/deterministic_attention.cu')
+encoding = 'locale', errors = None, newline = None
+
+    def read_text(self, encoding=None, errors=None, newline=None):
+        """
+        Open the file in text mode, read it, and close the file.
+        """
+        with self.open(mode='r', encoding=encoding, errors=errors, newline=newline) as f:
+>           return f.read()
+                   ^^^^^^^^
+E           UnicodeDecodeError: 'gbk' codec can't decode byte 0x92 in position 173: illegal multibyte sequence
+
+D:\anaconda\Lib\pathlib\_abc.py:633: UnicodeDecodeError
+_ test_packaged_contract_bytes_match_main[rl_engine.config-workloads/qwen3_8b.json-80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6] _
+
+package = 'rl_engine.config', resource = 'workloads/qwen3_8b.json'
+digest = '80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6'
+
+    @pytest.mark.parametrize(
+        ("package", "resource", "digest"),
+        [
+            (
+                "rl_engine.config",
+                "workloads/qwen3_8b.json",
+                "80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6",
+            ),
+            (
+                "rl_engine.contracts",
+                "profiles/precision/ws1.json",
+                "044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532",
+            ),
+        ],
+    )
+    def test_packaged_contract_bytes_match_main(package, resource, digest):
+>       assert hashlib.sha256(files(package).joinpath(resource).read_bytes()).hexdigest() == digest
+E       AssertionError: assert 'c97e966fb075...75e111e052d7d' == '80395644b5e8...f19635d3df0f6'
+E
+E         - 80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6
+E         + c97e966fb075dad8ce7dcd7a23a58ffb856e7474dfb3c50510675e111e052d7d
+
+tests\runtime\test_layout_compatibility.py:118: AssertionError
+_ test_packaged_contract_bytes_match_main[rl_engine.contracts-profiles/precision/ws1.json-044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532] _
+
+package = 'rl_engine.contracts', resource = 'profiles/precision/ws1.json'
+digest = '044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532'
+
+    @pytest.mark.parametrize(
+        ("package", "resource", "digest"),
+        [
+            (
+                "rl_engine.config",
+                "workloads/qwen3_8b.json",
+                "80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6",
+            ),
+            (
+                "rl_engine.contracts",
+                "profiles/precision/ws1.json",
+                "044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532",
+            ),
+        ],
+    )
+    def test_packaged_contract_bytes_match_main(package, resource, digest):
+>       assert hashlib.sha256(files(package).joinpath(resource).read_bytes()).hexdigest() == digest
+E       AssertionError: assert '2020c02e7af4...313d6752cd34e' == '044167ec21e3...7628cbd101532'
+E
+E         - 044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532
+E         + 2020c02e7af4e2ca4163e1ecfb9c491baef07b73deba39b0f44313d6752cd34e
+
+tests\runtime\test_layout_compatibility.py:118: AssertionError
+=========================== short test summary info ===========================
+FAILED tests/build/test_build_platform_collectives.py::test_cuda_build_keeps_existing_ipc_collective
+FAILED tests/build/test_build_platform_collectives.py::test_package_data_json_links_resolve_to_real_content
+FAILED tests/runtime/test_layout_compatibility.py::test_recorded_algorithm_sources_resolve_to_real_implementations
+FAILED tests/runtime/test_layout_compatibility.py::test_packaged_contract_bytes_match_main[rl_engine.config-workloads/qwen3_8b.json-80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6]
+FAILED tests/runtime/test_layout_compatibility.py::test_packaged_contract_bytes_match_main[rl_engine.contracts-profiles/precision/ws1.json-044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532]
+5 failed, 27 passed, 1 skipped in 5.70s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/benchmark-bfloat16.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/benchmark-bfloat16.json
new file mode 100644
index 000000000..72fb42bbe
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/benchmark-bfloat16.json
@@ -0,0 +1,2781 @@
+{
+  "gpu": "NVIDIA GeForce RTX 5080",
+  "capability": [
+    12,
+    0
+  ],
+  "os": "Windows-11-10.0.26200-SP0",
+  "python": "3.13.9 | packaged by Anaconda, Inc. | (main, Oct 21 2025, 19:09:58) [MSC v.1929 64 bit (AMD64)]",
+  "torch": "2.9.1+cu128",
+  "cuda_runtime": "12.8",
+  "triton": "3.5.1",
+  "dtype": "bfloat16",
+  "base_commit": "16b81175874ecf11d5ff05c00de3b69e9457591b",
+  "source_sha256": {
+    "csrc/cuda/rope/multi_axis_rope.cu": "f5db5cf5d4962053194455d9cde8cf91cf4e824bad7abf72e8410a404bcb0e44",
+    "rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py": "5c6984da351e5228f7549629d4ff53e5772b9a6ef2a01ba2a38493e769b15ab5",
+    "rl_engine/backends/shared/triton/rope/multi_axis_rope.py": "5b989bd63efd77913267b48654d10204bc563768c6c5d4e1cde9fb55e1354547",
+    "rl_engine/backends/cuda/rope/multi_axis_rope.py": "d779d9b60ce5d1d6651925f723049930d898024fdf0548ce29a2019b84753d65",
+    "rl_engine/ops/rope/_registration.py": "add0375d1639444908091d01e8d29781f7a975beffec64b77bfaa81a5b9a0ae3",
+    "rl_engine/ops/rope/multi_axis_rope.py": "07bcff1160f22c1ff6730be731206b2f4a153e2fada26afc93fc884e780568cb",
+    "rl_engine/reference/rope/multi_axis_rope.py": "c9a94730047f51659f7aa0a443e88d6ce7d03d00d6813a9aed548cafbf6347bf",
+    "benchmarks/operators/rope/benchmark_multi_axis_rope.py": "6961a55803fbd4e566bd060558fb59e658e7f1f704e61ca97f912e5d16b55884"
+  },
+  "timing": "CUDA graph ms/call and synchronized eager wall ms/call; warm allocation; excludes tables, compilation, and initial allocation of inputs",
+  "memory": "extra peak allocated bytes excluding inputs/tables/grad; includes returned output or gradient; packed QKV backing storage allocated before timing",
+  "iterations": 30,
+  "warmup": 5,
+  "trials": 3,
+  "graph_calls": 20,
+  "flash_baseline": {
+    "included": true,
+    "revision": "98c1cee0ed59307833841cdbf567fe336ceec13a",
+    "kernel_sha256_normalized": "d2037e623854c702fb075784fe4292d4a89d6b9ac1ae33cab3d6788f8f94f240",
+    "layer_sha256_normalized": "a79cb42dfc33e126227bdba27ae7a024d7d5cc5c8fd7f7a5f26c68d501024e0a",
+    "mode": "upstream rotary_kernel, apply_rotary and ApplyRotaryEmb, interleaved=True",
+    "no_fma_change": "only add enable_fp_fusion=False at the original kernel launch"
+  },
+  "trace": {
+    "cuda": {
+      "contract": "qwenimage-interleaved-rope-fp32-v1",
+      "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
+      "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
+      "layout": "BSHD, adjacent channel pairs",
+      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+    },
+    "triton": {
+      "contract": "qwenimage-interleaved-rope-fp32-v1",
+      "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
+      "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
+      "layout": "BSHD, adjacent channel pairs",
+      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+    }
+  },
+  "qualification": "Exact bytes vs unfused FP32 gold for CUDA, Triton and flash_no_fma; declared tolerance for upstream complex and default FlashAttention",
+  "correctness": [
+    {
+      "backend": "upstream_complex",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        130,
+        139
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        130,
+        139
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        139,
+        140
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        139,
+        140
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        234,
+        225
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        234,
+        225
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        220,
+        226
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        220,
+        226
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        196,
+        203
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        196,
+        203
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        230,
+        218
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        230,
+        218
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    }
+  ],
+  "results": [
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.003558399900794029,
+      "graph_min_ms": 0.0035264000296592714,
+      "graph_max_ms": 0.015438400208950043,
+      "eager_wall_median_ms": 0.0339666667059646,
+      "eager_wall_trials_ms": [
+        0.0356399999873247,
+        0.0339666667059646,
+        0.03332666662269427
+      ],
+      "extra_peak_bytes": 184320
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.007688799872994423,
+      "graph_min_ms": 0.007651200145483017,
+      "graph_max_ms": 0.007790400087833405,
+      "eager_wall_median_ms": 0.21265999997316007,
+      "eager_wall_trials_ms": [
+        0.21187666667780528,
+        0.21265999997316007,
+        0.22587333332921844
+      ],
+      "extra_peak_bytes": 187392
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0012016000226140022,
+      "graph_min_ms": 0.0011567999608814716,
+      "graph_max_ms": 0.0012384000234305858,
+      "eager_wall_median_ms": 0.02289333333465038,
+      "eager_wall_trials_ms": [
+        0.02289333333465038,
+        0.02245666667780218,
+        0.02458333328831941
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0022280000150203703,
+      "graph_min_ms": 0.0021888000890612604,
+      "graph_max_ms": 0.0022655999287962913,
+      "eager_wall_median_ms": 0.1764300000407578,
+      "eager_wall_trials_ms": [
+        0.20031666669334905,
+        0.1764300000407578,
+        0.15246333338533682
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.001142400037497282,
+      "graph_min_ms": 0.0010591999627649785,
+      "graph_max_ms": 0.005691200122237205,
+      "eager_wall_median_ms": 0.057766666698929235,
+      "eager_wall_trials_ms": [
+        0.07126333330234047,
+        0.057766666698929235,
+        0.05640333335274287
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.002125599980354309,
+      "graph_min_ms": 0.002091200090944767,
+      "graph_max_ms": 0.002160000056028366,
+      "eager_wall_median_ms": 0.21872666663208898,
+      "eager_wall_trials_ms": [
+        0.22702333335473668,
+        0.21872666663208898,
+        0.21611666670651175
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0010935999918729067,
+      "graph_min_ms": 0.001033599954098463,
+      "graph_max_ms": 0.0011536000296473504,
+      "eager_wall_median_ms": 0.10024000002886169,
+      "eager_wall_trials_ms": [
+        0.10126666669142044,
+        0.09858333329854456,
+        0.10024000002886169
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0020527999848127367,
+      "graph_min_ms": 0.002019199915230274,
+      "graph_max_ms": 0.0021023999899625776,
+      "eager_wall_median_ms": 0.3111766666431019,
+      "eager_wall_trials_ms": [
+        0.30222666664485587,
+        0.3111766666431019,
+        0.333893333360417
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0010991999879479408,
+      "graph_min_ms": 0.001033599954098463,
+      "graph_max_ms": 0.0011392000131309032,
+      "eager_wall_median_ms": 0.10421333336125826,
+      "eager_wall_trials_ms": [
+        0.10421333336125826,
+        0.10106999998242827,
+        0.12048333337588701
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0020503999665379526,
+      "graph_min_ms": 0.002001599967479706,
+      "graph_max_ms": 0.00886240005493164,
+      "eager_wall_median_ms": 0.31822000000829576,
+      "eager_wall_trials_ms": [
+        0.28550666665978497,
+        0.31822000000829576,
+        0.33777000001767493
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.00379519984126091,
+      "graph_min_ms": 0.003745599836111069,
+      "graph_max_ms": 0.003859199956059456,
+      "eager_wall_median_ms": 0.10653666668076767,
+      "eager_wall_trials_ms": [
+        0.10653666668076767,
+        0.10879000001295935,
+        0.09587666669782872
+      ],
+      "extra_peak_bytes": 184320
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.008064000308513642,
+      "graph_min_ms": 0.008011200278997422,
+      "graph_max_ms": 0.008102399855852127,
+      "eager_wall_median_ms": 0.2260500000071867,
+      "eager_wall_trials_ms": [
+        0.23646333332483968,
+        0.2187999999781217,
+        0.2260500000071867
+      ],
+      "extra_peak_bytes": 187392
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0012024000287055969,
+      "graph_min_ms": 0.0011392000131309032,
+      "graph_max_ms": 0.0025040000677108766,
+      "eager_wall_median_ms": 0.023706666615908034,
+      "eager_wall_trials_ms": [
+        0.023706666615908034,
+        0.024339999981748406,
+        0.023019999995691858
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.002226400002837181,
+      "graph_min_ms": 0.002174399979412556,
+      "graph_max_ms": 0.0022671999409794807,
+      "eager_wall_median_ms": 0.157756666643157,
+      "eager_wall_trials_ms": [
+        0.165013333329019,
+        0.15389999995628992,
+        0.157756666643157
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0011008000001311302,
+      "graph_min_ms": 0.0010688000358641148,
+      "graph_max_ms": 0.0011359999887645244,
+      "eager_wall_median_ms": 0.059236666675133165,
+      "eager_wall_trials_ms": [
+        0.053669999942940194,
+        0.06250666665437166,
+        0.059236666675133165
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0021215999498963357,
+      "graph_min_ms": 0.002065600082278252,
+      "graph_max_ms": 0.0021679999306797983,
+      "eager_wall_median_ms": 0.25444666662224336,
+      "eager_wall_trials_ms": [
+        0.25444666662224336,
+        0.27043333332888625,
+        0.2240899999984928
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.001101600006222725,
+      "graph_min_ms": 0.001065600011497736,
+      "graph_max_ms": 0.0011376000009477138,
+      "eager_wall_median_ms": 0.12566333334689261,
+      "eager_wall_trials_ms": [
+        0.12566333334689261,
+        0.12918333328949907,
+        0.10983666664590903
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0020527999848127367,
+      "graph_min_ms": 0.0020160000771284104,
+      "graph_max_ms": 0.0021023999899625776,
+      "eager_wall_median_ms": 0.28767333333235,
+      "eager_wall_trials_ms": [
+        0.2609133333559536,
+        0.28767333333235,
+        0.3289566666353494
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.001101600006222725,
+      "graph_min_ms": 0.0010223999619483949,
+      "graph_max_ms": 0.0022975999861955644,
+      "eager_wall_median_ms": 0.0995366666757036,
+      "eager_wall_trials_ms": [
+        0.10132333336514421,
+        0.0995366666757036,
+        0.0990000000456348
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0020527999848127367,
+      "graph_min_ms": 0.0020128000527620316,
+      "graph_max_ms": 0.0021007999777793883,
+      "eager_wall_median_ms": 0.3198099999887442,
+      "eager_wall_trials_ms": [
+        0.33375666668386356,
+        0.30899333335886087,
+        0.3198099999887442
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.2427680015563965,
+      "graph_min_ms": 0.24123198986053468,
+      "graph_max_ms": 0.3049216032028198,
+      "eager_wall_median_ms": 0.25121666670505266,
+      "eager_wall_trials_ms": [
+        0.24960999999166233,
+        0.2669566666857766,
+        0.25121666670505266
+      ],
+      "extra_peak_bytes": 125829120
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.5011807918548584,
+      "graph_min_ms": 0.4943295955657959,
+      "graph_max_ms": 0.5467391967773437,
+      "eager_wall_median_ms": 0.5078799999561548,
+      "eager_wall_trials_ms": [
+        0.5254200000005463,
+        0.5078799999561548,
+        0.5070699999729792
+      ],
+      "extra_peak_bytes": 127926272
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.05770559906959534,
+      "graph_min_ms": 0.05762240290641785,
+      "graph_max_ms": 0.07225120067596436,
+      "eager_wall_median_ms": 0.061780000032740645,
+      "eager_wall_trials_ms": [
+        0.06511333334249987,
+        0.06060333331940152,
+        0.061780000032740645
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.12953280210494994,
+      "graph_min_ms": 0.12918239831924438,
+      "graph_max_ms": 0.14451359510421752,
+      "eager_wall_median_ms": 0.15664999997776854,
+      "eager_wall_trials_ms": [
+        0.2333100000517637,
+        0.15664999997776854,
+        0.14973666669296412
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.05190559923648834,
+      "graph_min_ms": 0.05078719854354859,
+      "graph_max_ms": 0.06408159732818604,
+      "eager_wall_median_ms": 0.0845666666161075,
+      "eager_wall_trials_ms": [
+        0.08885000000494377,
+        0.08201333330362104,
+        0.0845666666161075
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.12463040351867677,
+      "graph_min_ms": 0.12331520318984986,
+      "graph_max_ms": 0.1529088020324707,
+      "eager_wall_median_ms": 0.2539966667124342,
+      "eager_wall_trials_ms": [
+        0.2757299999454214,
+        0.25390333333537757,
+        0.2539966667124342
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.05208719968795776,
+      "graph_min_ms": 0.05123199820518494,
+      "graph_max_ms": 0.06861280202865601,
+      "eager_wall_median_ms": 0.13019999999717888,
+      "eager_wall_trials_ms": [
+        0.1370399999965836,
+        0.12465666665472479,
+        0.13019999999717888
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.12441120147705079,
+      "graph_min_ms": 0.12314239740371705,
+      "graph_max_ms": 0.139137601852417,
+      "eager_wall_median_ms": 0.340176666698729,
+      "eager_wall_trials_ms": [
+        0.35652999995363643,
+        0.340176666698729,
+        0.33921999996285496
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.052018401026725766,
+      "graph_min_ms": 0.05074719786643982,
+      "graph_max_ms": 0.07086880207061767,
+      "eager_wall_median_ms": 0.10117999997116083,
+      "eager_wall_trials_ms": [
+        0.10250666667464732,
+        0.10117999997116083,
+        0.10012000002461718
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.12442959547042846,
+      "graph_min_ms": 0.12299360036849975,
+      "graph_max_ms": 0.14664160013198851,
+      "eager_wall_median_ms": 0.34945666666317265,
+      "eager_wall_trials_ms": [
+        0.36080999998375773,
+        0.3163533333160255,
+        0.34945666666317265
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.2459879994392395,
+      "graph_min_ms": 0.24467840194702148,
+      "graph_max_ms": 0.272545599937439,
+      "eager_wall_median_ms": 0.2536533333720096,
+      "eager_wall_trials_ms": [
+        0.27815666665749933,
+        0.2529566666756485,
+        0.2536533333720096
+      ],
+      "extra_peak_bytes": 125829120
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.5008344173431396,
+      "graph_min_ms": 0.4996367931365967,
+      "graph_max_ms": 0.5373487949371338,
+      "eager_wall_median_ms": 0.5262199999682101,
+      "eager_wall_trials_ms": [
+        0.5269766666970099,
+        0.5070566666593853,
+        0.5262199999682101
+      ],
+      "extra_peak_bytes": 127926272
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.05865840017795562,
+      "graph_min_ms": 0.058550399541854856,
+      "graph_max_ms": 0.08192480206489564,
+      "eager_wall_median_ms": 0.06282000000889336,
+      "eager_wall_trials_ms": [
+        0.06282000000889336,
+        0.06640333334265354,
+        0.06277333335068154
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.1308944046497345,
+      "graph_min_ms": 0.13030719757080078,
+      "graph_max_ms": 0.145305597782135,
+      "eager_wall_median_ms": 0.1688300000144712,
+      "eager_wall_trials_ms": [
+        0.1688300000144712,
+        0.15922666667999388,
+        0.17570666665657578
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.04834639877080918,
+      "graph_min_ms": 0.047383999824523924,
+      "graph_max_ms": 0.06508319973945617,
+      "eager_wall_median_ms": 0.08053666667062014,
+      "eager_wall_trials_ms": [
+        0.09511666667094687,
+        0.07707999999790142,
+        0.08053666667062014
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.1231384038925171,
+      "graph_min_ms": 0.12216000556945801,
+      "graph_max_ms": 0.1478927969932556,
+      "eager_wall_median_ms": 0.2741499999804849,
+      "eager_wall_trials_ms": [
+        0.28788666665301815,
+        0.2741499999804849,
+        0.24507333334137607
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.048105600476264956,
+      "graph_min_ms": 0.0474047988653183,
+      "graph_max_ms": 0.06945440173149109,
+      "eager_wall_median_ms": 0.12940666668631215,
+      "eager_wall_trials_ms": [
+        0.14179000002817096,
+        0.12940666668631215,
+        0.12683000004471978
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.12347760200500488,
+      "graph_min_ms": 0.12205439805984497,
+      "graph_max_ms": 0.14951200485229493,
+      "eager_wall_median_ms": 0.36926333335334977,
+      "eager_wall_trials_ms": [
+        0.36926333335334977,
+        0.3322200000184239,
+        0.36990666667406913
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.04896320104598999,
+      "graph_min_ms": 0.047470399737358095,
+      "graph_max_ms": 0.06636160016059875,
+      "eager_wall_median_ms": 0.14505666664869446,
+      "eager_wall_trials_ms": [
+        0.14700000004571243,
+        0.14505666664869446,
+        0.13565333332129134
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.12373999953269958,
+      "graph_min_ms": 0.1219599962234497,
+      "graph_max_ms": 0.1542415976524353,
+      "eager_wall_median_ms": 0.33643999995547347,
+      "eager_wall_trials_ms": [
+        0.33643999995547347,
+        0.3219899999749032,
+        0.3583933333478247
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.51009202003479,
+      "graph_min_ms": 0.5082543849945068,
+      "graph_max_ms": 0.5431327819824219,
+      "eager_wall_median_ms": 0.526480000007723,
+      "eager_wall_trials_ms": [
+        0.5092500000197712,
+        0.5466633333223095,
+        0.526480000007723
+      ],
+      "extra_peak_bytes": 211630080
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 1.0505383968353272,
+      "graph_min_ms": 1.0236895561218262,
+      "graph_max_ms": 1.129139232635498,
+      "eager_wall_median_ms": 1.044759999967937,
+      "eager_wall_trials_ms": [
+        1.044759999967937,
+        1.0434166666527744,
+        1.0474333333453008
+      ],
+      "extra_peak_bytes": 215157248
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.11203439831733704,
+      "graph_min_ms": 0.1114351987838745,
+      "graph_max_ms": 0.14540959596633912,
+      "eager_wall_median_ms": 0.11692666666931473,
+      "eager_wall_trials_ms": [
+        0.11580999998841435,
+        0.11692666666931473,
+        0.12043333335896023
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.22477760314941406,
+      "graph_min_ms": 0.22440481185913086,
+      "graph_max_ms": 0.24935200214385986,
+      "eager_wall_median_ms": 0.23845333331943644,
+      "eager_wall_trials_ms": [
+        0.23845333331943644,
+        0.2526333333056148,
+        0.230336666633472
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.10685279965400696,
+      "graph_min_ms": 0.10539519786834717,
+      "graph_max_ms": 0.12008960247039795,
+      "eager_wall_median_ms": 0.11027333330275724,
+      "eager_wall_trials_ms": [
+        0.11027333330275724,
+        0.10863333330538201,
+        0.11599999997997656
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.21339280605316163,
+      "graph_min_ms": 0.2126607894897461,
+      "graph_max_ms": 0.23967199325561522,
+      "eager_wall_median_ms": 0.27515000001585577,
+      "eager_wall_trials_ms": [
+        0.2870199999961187,
+        0.2669133333256468,
+        0.27515000001585577
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.10692960023880005,
+      "graph_min_ms": 0.10526080131530761,
+      "graph_max_ms": 0.12183040380477905,
+      "eager_wall_median_ms": 0.10955999999472017,
+      "eager_wall_trials_ms": [
+        0.1204166666866513,
+        0.10955999999472017,
+        0.10893666670502473
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.21346559524536132,
+      "graph_min_ms": 0.21225121021270751,
+      "graph_max_ms": 0.2430032014846802,
+      "eager_wall_median_ms": 0.338726666692916,
+      "eager_wall_trials_ms": [
+        0.33888666663794237,
+        0.3213533333716138,
+        0.338726666692916
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.10693199634552003,
+      "graph_min_ms": 0.10515999794006348,
+      "graph_max_ms": 0.12142080068588257,
+      "eager_wall_median_ms": 0.1100600000427221,
+      "eager_wall_trials_ms": [
+        0.11902333335456206,
+        0.1100600000427221,
+        0.10893666664439176
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.21350640058517456,
+      "graph_min_ms": 0.2123823881149292,
+      "graph_max_ms": 0.23954401016235352,
+      "eager_wall_median_ms": 0.3223733333167426,
+      "eager_wall_trials_ms": [
+        0.29901333333934116,
+        0.3223733333167426,
+        0.326143333343983
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.5111671924591064,
+      "graph_min_ms": 0.5095888137817383,
+      "graph_max_ms": 0.5389599800109863,
+      "eager_wall_median_ms": 0.5120633333717706,
+      "eager_wall_trials_ms": [
+        0.5099033333560025,
+        0.532273333374178,
+        0.5120633333717706
+      ],
+      "extra_peak_bytes": 211630080
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 1.0546072006225586,
+      "graph_min_ms": 1.0272704124450684,
+      "graph_max_ms": 1.1091728210449219,
+      "eager_wall_median_ms": 1.0453999999905743,
+      "eager_wall_trials_ms": [
+        1.0565066666458733,
+        1.0453999999905743,
+        1.042836666662576
+      ],
+      "extra_peak_bytes": 215157248
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.11280480027198792,
+      "graph_min_ms": 0.11215200424194335,
+      "graph_max_ms": 0.13529759645462036,
+      "eager_wall_median_ms": 0.12308333334658528,
+      "eager_wall_trials_ms": [
+        0.11563333331044608,
+        0.12308333334658528,
+        0.12365000002318993
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.2251416087150574,
+      "graph_min_ms": 0.22450559139251708,
+      "graph_max_ms": 0.2644239902496338,
+      "eager_wall_median_ms": 0.2328933333653064,
+      "eager_wall_trials_ms": [
+        0.23086333330866182,
+        0.2328933333653064,
+        0.25025666667109664
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.1075808048248291,
+      "graph_min_ms": 0.1061360001564026,
+      "graph_max_ms": 0.12249759435653687,
+      "eager_wall_median_ms": 0.10984000000462402,
+      "eager_wall_trials_ms": [
+        0.1298533332980393,
+        0.10860666661756113,
+        0.10984000000462402
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.21414560079574585,
+      "graph_min_ms": 0.21337599754333497,
+      "graph_max_ms": 0.23953759670257568,
+      "eager_wall_median_ms": 0.26501666670810664,
+      "eager_wall_trials_ms": [
+        0.2692166666747653,
+        0.23041333333821967,
+        0.26501666670810664
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.10766559839248657,
+      "graph_min_ms": 0.10606240034103394,
+      "graph_max_ms": 0.13348640203475953,
+      "eager_wall_median_ms": 0.11105333330003002,
+      "eager_wall_trials_ms": [
+        0.11928999999023897,
+        0.11105333330003002,
+        0.11004666666849516
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.21448720693588258,
+      "graph_min_ms": 0.21337599754333497,
+      "graph_max_ms": 0.25474560260772705,
+      "eager_wall_median_ms": 0.30118666666870314,
+      "eager_wall_trials_ms": [
+        0.32142000000021653,
+        0.30118666666870314,
+        0.2931233333583805
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.1074567973613739,
+      "graph_min_ms": 0.10611200332641602,
+      "graph_max_ms": 0.1223423957824707,
+      "eager_wall_median_ms": 0.11078333336627111,
+      "eager_wall_trials_ms": [
+        0.12015666664713838,
+        0.10989999997642978,
+        0.11078333336627111
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.21430879831314087,
+      "graph_min_ms": 0.2134079933166504,
+      "graph_max_ms": 0.23155200481414795,
+      "eager_wall_median_ms": 0.3170500000123866,
+      "eager_wall_trials_ms": [
+        0.3170500000123866,
+        0.32710333331730607,
+        0.31006333332091646
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.43411200046539306,
+      "graph_min_ms": 0.4321375846862793,
+      "graph_max_ms": 0.46126718521118165,
+      "eager_wall_median_ms": 0.4498766666680846,
+      "eager_wall_trials_ms": [
+        0.43220333336648764,
+        0.4498766666680846,
+        0.46697666663627996
+      ],
+      "extra_peak_bytes": 185303040
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.8935583591461181,
+      "graph_min_ms": 0.870751953125,
+      "graph_max_ms": 0.9448016166687012,
+      "eager_wall_median_ms": 0.8892600000156866,
+      "eager_wall_trials_ms": [
+        0.8870666666553007,
+        0.8892600000156866,
+        0.8902699999453034
+      ],
+      "extra_peak_bytes": 188391424
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09808879792690278,
+      "graph_min_ms": 0.09759039878845215,
+      "graph_max_ms": 0.11292320489883423,
+      "eager_wall_median_ms": 0.11404999998679462,
+      "eager_wall_trials_ms": [
+        0.11404999998679462,
+        0.11419666667886001,
+        0.102829999984048
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.19646480083465576,
+      "graph_min_ms": 0.19599839448928832,
+      "graph_max_ms": 0.21091840267181397,
+      "eager_wall_median_ms": 0.20405333331533862,
+      "eager_wall_trials_ms": [
+        0.20405333331533862,
+        0.20323666661473302,
+        0.2081700000417186
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09356399774551391,
+      "graph_min_ms": 0.09257439970970154,
+      "graph_max_ms": 0.10800800323486329,
+      "eager_wall_median_ms": 0.10138999999374694,
+      "eager_wall_trials_ms": [
+        0.10716000003109609,
+        0.10138999999374694,
+        0.10003000000627556
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.18708080053329468,
+      "graph_min_ms": 0.1862496018409729,
+      "graph_max_ms": 0.21909279823303224,
+      "eager_wall_median_ms": 0.26604333331003244,
+      "eager_wall_trials_ms": [
+        0.30756000002535683,
+        0.2481433333135404,
+        0.26604333331003244
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09362079799175263,
+      "graph_min_ms": 0.09246399998664856,
+      "graph_max_ms": 0.10841599702835084,
+      "eager_wall_median_ms": 0.10133000002194119,
+      "eager_wall_trials_ms": [
+        0.14010333331195093,
+        0.10133000002194119,
+        0.10039666667580605
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.18697920441627502,
+      "graph_min_ms": 0.18585760593414308,
+      "graph_max_ms": 0.21307199001312255,
+      "eager_wall_median_ms": 0.3559766667118917,
+      "eager_wall_trials_ms": [
+        0.3559766667118917,
+        0.3666866666511244,
+        0.35564333332634607
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09362399876117705,
+      "graph_min_ms": 0.09233919978141784,
+      "graph_max_ms": 0.10795680284500123,
+      "eager_wall_median_ms": 0.1356800000091122,
+      "eager_wall_trials_ms": [
+        0.14346666666824603,
+        0.1356800000091122,
+        0.1269966666465431
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.18709120154380798,
+      "graph_min_ms": 0.18642239570617675,
+      "graph_max_ms": 0.22162079811096191,
+      "eager_wall_median_ms": 0.3122899999652873,
+      "eager_wall_trials_ms": [
+        0.29407333331619157,
+        0.3122899999652873,
+        0.36742666664698237
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.43668241500854493,
+      "graph_min_ms": 0.43459200859069824,
+      "graph_max_ms": 0.4815072059631348,
+      "eager_wall_median_ms": 0.4541133333380761,
+      "eager_wall_trials_ms": [
+        0.4636933333434475,
+        0.43558999999125564,
+        0.4541133333380761
+      ],
+      "extra_peak_bytes": 185303040
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.8956063747406007,
+      "graph_min_ms": 0.8733648300170899,
+      "graph_max_ms": 0.9540863990783691,
+      "eager_wall_median_ms": 0.8916500000244317,
+      "eager_wall_trials_ms": [
+        0.8923599999737538,
+        0.8916500000244317,
+        0.8882866666681366
+      ],
+      "extra_peak_bytes": 188391424
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09595520198345184,
+      "graph_min_ms": 0.09553279876708984,
+      "graph_max_ms": 0.11051679849624634,
+      "eager_wall_median_ms": 0.1001133333678202,
+      "eager_wall_trials_ms": [
+        0.1001133333678202,
+        0.09971333335367187,
+        0.10664000001270324
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.19663520455360411,
+      "graph_min_ms": 0.19614880084991454,
+      "graph_max_ms": 0.22351360321044922,
+      "eager_wall_median_ms": 0.20444999997077198,
+      "eager_wall_trials_ms": [
+        0.22424333334735516,
+        0.20414666669239523,
+        0.20444999997077198
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0934136003255844,
+      "graph_min_ms": 0.0916208028793335,
+      "graph_max_ms": 0.10537919998168946,
+      "eager_wall_median_ms": 0.09456333336856915,
+      "eager_wall_trials_ms": [
+        0.09251333334153362,
+        0.09570666667665743,
+        0.09456333336856915
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.1877624034881592,
+      "graph_min_ms": 0.18668639659881592,
+      "graph_max_ms": 0.21486239433288573,
+      "eager_wall_median_ms": 0.24588000002646973,
+      "eager_wall_trials_ms": [
+        0.26271666671770316,
+        0.24588000002646973,
+        0.23198333334827717
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09341759979724884,
+      "graph_min_ms": 0.0921504020690918,
+      "graph_max_ms": 0.11279360055923462,
+      "eager_wall_median_ms": 0.14909333337224476,
+      "eager_wall_trials_ms": [
+        0.15717333335487638,
+        0.12570666664638944,
+        0.14909333337224476
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.1880079984664917,
+      "graph_min_ms": 0.18690240383148193,
+      "graph_max_ms": 0.2143280029296875,
+      "eager_wall_median_ms": 0.3311766666835562,
+      "eager_wall_trials_ms": [
+        0.3571433333490859,
+        0.3311766666835562,
+        0.32030999997611315
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0934448003768921,
+      "graph_min_ms": 0.09198240041732789,
+      "graph_max_ms": 0.10836479663848878,
+      "eager_wall_median_ms": 0.132580000051045,
+      "eager_wall_trials_ms": [
+        0.132580000051045,
+        0.14571666670235572,
+        0.12959000002107737
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.18786399960517886,
+      "graph_min_ms": 0.18675040006637572,
+      "graph_max_ms": 0.2141263961791992,
+      "eager_wall_median_ms": 0.34321999998307245,
+      "eager_wall_trials_ms": [
+        0.36083333331286366,
+        0.3284899999925983,
+        0.34321999998307245
+      ],
+      "extra_peak_bytes": 74809344
+    }
+  ]
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.json
new file mode 100644
index 000000000..468ce07af
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.json
@@ -0,0 +1,37 @@
+{
+  "package": "D:\\dandan123\\dandan\\issue\\investigation-rl-kernel\\RL-Kernel\\.venv\\Lib\\site-packages\\rl_engine",
+  "companion_library": "D:\\dandan123\\dandan\\issue\\investigation-rl-kernel\\RL-Kernel\\.venv\\Lib\\site-packages\\rl_engine\\_rope_cuda.cp313-win_amd64.pyd",
+  "native_dispatcher_registered": true,
+  "torch": "2.9.1+cu128",
+  "gpu": "NVIDIA GeForce RTX 5080",
+  "license_present": true,
+  "packed_qkv_stride": [
+    17280,
+    1152,
+    128,
+    1
+  ],
+  "checks": {
+    "cuda_eager": "passed",
+    "cuda_compiled": "passed",
+    "cuda_trace": {
+      "contract": "qwenimage-interleaved-rope-fp32-v1",
+      "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
+      "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
+      "layout": "BSHD, adjacent channel pairs",
+      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+    },
+    "triton_eager": "passed",
+    "triton_compiled": "passed",
+    "triton_trace": {
+      "contract": "qwenimage-interleaved-rope-fp32-v1",
+      "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
+      "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
+      "layout": "BSHD, adjacent channel pairs",
+      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+    }
+  },
+  "passed": true
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.txt
new file mode 100644
index 000000000..01c1dd377
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.txt
@@ -0,0 +1,38 @@
+WARNING 10-28 17:08:06 [RL-Kernel]: Core binary extension (_C) unavailable: cannot import name '_C' from 'rl_engine' (D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\.venv\Lib\site-packages\rl_engine\__init__.py). Falling back to native code.
+{
+  "package": "D:\\dandan123\\dandan\\issue\\investigation-rl-kernel\\RL-Kernel\\.venv\\Lib\\site-packages\\rl_engine",
+  "companion_library": "D:\\dandan123\\dandan\\issue\\investigation-rl-kernel\\RL-Kernel\\.venv\\Lib\\site-packages\\rl_engine\\_rope_cuda.cp313-win_amd64.pyd",
+  "native_dispatcher_registered": true,
+  "torch": "2.9.1+cu128",
+  "gpu": "NVIDIA GeForce RTX 5080",
+  "license_present": true,
+  "packed_qkv_stride": [
+    17280,
+    1152,
+    128,
+    1
+  ],
+  "checks": {
+    "cuda_eager": "passed",
+    "cuda_compiled": "passed",
+    "cuda_trace": {
+      "contract": "qwenimage-interleaved-rope-fp32-v1",
+      "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
+      "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
+      "layout": "BSHD, adjacent channel pairs",
+      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+    },
+    "triton_eager": "passed",
+    "triton_compiled": "passed",
+    "triton_trace": {
+      "contract": "qwenimage-interleaved-rope-fp32-v1",
+      "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
+      "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
+      "layout": "BSHD, adjacent channel pairs",
+      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+    }
+  },
+  "passed": true
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/layout.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/layout.txt
new file mode 100644
index 000000000..0086d019b
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/layout.txt
@@ -0,0 +1,138 @@
+.FF.............s............F.FF....                                    [100%]
+================================== FAILURES ===================================
+________________ test_cuda_build_keeps_existing_ipc_collective ________________
+
+monkeypatch = <_pytest.monkeypatch.MonkeyPatch object at 0x000001D4B0E668B0>
+
+    def test_cuda_build_keeps_existing_ipc_collective(monkeypatch) -> None:
+        extension = _load_extension_config(monkeypatch, hip=None)
+
+        assert "csrc/cuda/collectives/deterministic_collective.cu" in extension["sources"]
+        assert "-DKERNEL_ALIGN_WITH_CUDA" in extension["extra_compile_args"]["cxx"]
+        assert "-DKERNEL_ALIGN_WITH_ROCM" not in extension["extra_compile_args"]["cxx"]
+>       assert "-lcuda" in extension["extra_link_args"]
+E       AssertionError: assert '-lcuda' in ['-Wl,-rpath,$ORIGIN/../torch/lib']
+
+tests\build\test_build_platform_collectives.py:61: AssertionError
+____________ test_package_data_json_links_resolve_to_real_content _____________
+
+tmp_path = WindowsPath('C:/Users/123/AppData/Local/Temp/pytest-of-123/pytest-3/test_package_data_json_links_r0')
+
+    def test_package_data_json_links_resolve_to_real_content(tmp_path) -> None:
+        target = tmp_path / "configs" / "profile.json"
+        target.parent.mkdir()
+        target.write_text('{"name": "profile"}\n', encoding="utf-8")
+        source = tmp_path / "rl_engine" / "profile.json"
+        source.parent.mkdir()
+>       source.symlink_to("../configs/profile.json")
+
+tests\build\test_build_platform_collectives.py:70:
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
+
+self = WindowsPath('C:/Users/123/AppData/Local/Temp/pytest-of-123/pytest-3/test_package_data_json_links_r0/rl_engine/profile.json')
+target = '../configs/profile.json', target_is_directory = False
+
+    def symlink_to(self, target, target_is_directory=False):
+        """
+        Make this path a symlink pointing to the target path.
+        Note the order of arguments (link, target) is the reverse of os.symlink.
+        """
+>       os.symlink(target, self, target_is_directory)
+E       OSError: [WinError 1314] �ͻ���û���������Ȩ��: '../configs/profile.json' -> 'C:\\Users\\123\\AppData\\Local\\Temp\\pytest-of-123\\pytest-3\\test_package_data_json_links_r0\\rl_engine\\profile.json'
+
+D:\anaconda\Lib\pathlib\_local.py:789: OSError
+_______ test_recorded_algorithm_sources_resolve_to_real_implementations _______
+
+    def test_recorded_algorithm_sources_resolve_to_real_implementations():
+        from rl_engine.config.workload import load_manifest
+
+        moves = json.loads((ROOT / "tools/migration/layout.json").read_text())["moves"]
+        for case in load_manifest().representative_cases:
+            source, symbol = case["provenance_evidence"]["algorithm_source"].rsplit(":", 1)
+            implementation = ROOT / moves.get(source, source)
+>           text = implementation.read_text()
+                   ^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+tests\runtime\test_layout_compatibility.py:87:
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
+D:\anaconda\Lib\pathlib\_local.py:546: in read_text
+    return PathBase.read_text(self, encoding, errors, newline)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
+
+self = WindowsPath('D:/dandan123/dandan/issue/investigation-rl-kernel/RL-Kernel/csrc/cuda/attention/deterministic_attention.cu')
+encoding = 'locale', errors = None, newline = None
+
+    def read_text(self, encoding=None, errors=None, newline=None):
+        """
+        Open the file in text mode, read it, and close the file.
+        """
+        with self.open(mode='r', encoding=encoding, errors=errors, newline=newline) as f:
+>           return f.read()
+                   ^^^^^^^^
+E           UnicodeDecodeError: 'gbk' codec can't decode byte 0x92 in position 173: illegal multibyte sequence
+
+D:\anaconda\Lib\pathlib\_abc.py:633: UnicodeDecodeError
+_ test_packaged_contract_bytes_match_main[rl_engine.config-workloads/qwen3_8b.json-80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6] _
+
+package = 'rl_engine.config', resource = 'workloads/qwen3_8b.json'
+digest = '80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6'
+
+    @pytest.mark.parametrize(
+        ("package", "resource", "digest"),
+        [
+            (
+                "rl_engine.config",
+                "workloads/qwen3_8b.json",
+                "80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6",
+            ),
+            (
+                "rl_engine.contracts",
+                "profiles/precision/ws1.json",
+                "044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532",
+            ),
+        ],
+    )
+    def test_packaged_contract_bytes_match_main(package, resource, digest):
+>       assert hashlib.sha256(files(package).joinpath(resource).read_bytes()).hexdigest() == digest
+E       AssertionError: assert 'c97e966fb075...75e111e052d7d' == '80395644b5e8...f19635d3df0f6'
+E
+E         - 80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6
+E         + c97e966fb075dad8ce7dcd7a23a58ffb856e7474dfb3c50510675e111e052d7d
+
+tests\runtime\test_layout_compatibility.py:118: AssertionError
+_ test_packaged_contract_bytes_match_main[rl_engine.contracts-profiles/precision/ws1.json-044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532] _
+
+package = 'rl_engine.contracts', resource = 'profiles/precision/ws1.json'
+digest = '044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532'
+
+    @pytest.mark.parametrize(
+        ("package", "resource", "digest"),
+        [
+            (
+                "rl_engine.config",
+                "workloads/qwen3_8b.json",
+                "80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6",
+            ),
+            (
+                "rl_engine.contracts",
+                "profiles/precision/ws1.json",
+                "044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532",
+            ),
+        ],
+    )
+    def test_packaged_contract_bytes_match_main(package, resource, digest):
+>       assert hashlib.sha256(files(package).joinpath(resource).read_bytes()).hexdigest() == digest
+E       AssertionError: assert '2020c02e7af4...313d6752cd34e' == '044167ec21e3...7628cbd101532'
+E
+E         - 044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532
+E         + 2020c02e7af4e2ca4163e1ecfb9c491baef07b73deba39b0f44313d6752cd34e
+
+tests\runtime\test_layout_compatibility.py:118: AssertionError
+=========================== short test summary info ===========================
+FAILED tests/build/test_build_platform_collectives.py::test_cuda_build_keeps_existing_ipc_collective
+FAILED tests/build/test_build_platform_collectives.py::test_package_data_json_links_resolve_to_real_content
+FAILED tests/runtime/test_layout_compatibility.py::test_recorded_algorithm_sources_resolve_to_real_implementations
+FAILED tests/runtime/test_layout_compatibility.py::test_packaged_contract_bytes_match_main[rl_engine.config-workloads/qwen3_8b.json-80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6]
+FAILED tests/runtime/test_layout_compatibility.py::test_packaged_contract_bytes_match_main[rl_engine.contracts-profiles/precision/ws1.json-044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532]
+5 failed, 31 passed, 1 skipped in 5.67s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/manifest.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/manifest.json
new file mode 100644
index 000000000..190cf7834
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/manifest.json
@@ -0,0 +1,65 @@
+{
+  "recorded_at_utc": "2026-10-28T09:12:00.361759+00:00",
+  "base_commit": "16b81175874ecf11d5ff05c00de3b69e9457591b",
+  "source_sha256": {
+    "MANIFEST.in": "706ec113547f7012cfa6d15d062817d5d2d9e0c0e18e0fb706eca5ba2335bcc5",
+    "benchmarks/operators/rope/benchmark_multi_axis_rope.py": "6961a55803fbd4e566bd060558fb59e658e7f1f704e61ca97f912e5d16b55884",
+    "build_tools/extensions.py": "cf07cd49cba0ba7b4da459c26546380b7b288232fcc032e7641863e09ae3014a",
+    "csrc/bindings/rope_module.cpp": "c69eb3f200de306ca5885fd7c4c7ccf49c822ed1f2f1f639223ec36f0c51b524",
+    "csrc/cuda/rope/multi_axis_rope.cu": "f5db5cf5d4962053194455d9cde8cf91cf4e824bad7abf72e8410a404bcb0e44",
+    "docs/.nav.yml": "814a6d60839465e8f9405a9a26306dbc203b0fcb6aa36db264fd7651624d7ef3",
+    "docs/operators/README.md": "ab926bf5a3fb719e4a57cb65f0784a52019f6ed4bd49c0b5f1b15a1765fd6d7a",
+    "docs/operators/multi-axis-rope.md": "79f94810ee15ead5f1ce24715456f65f12b383728f2a3571104614b1a33b47d8",
+    "pyproject.toml": "a132159ae2a2c077d9fe001194fa2b9ab382acbcd58a9eccdea9387286bfb258",
+    "rl_engine/backends/cuda/rope/multi_axis_rope.py": "d779d9b60ce5d1d6651925f723049930d898024fdf0548ce29a2019b84753d65",
+    "rl_engine/backends/shared/triton/rope/FLASH_ATTENTION_LICENSE": "8c9ccb96c065e706135b6cbad279b721da6156e51f3a5f27c6b3329af9416d73",
+    "rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py": "5c6984da351e5228f7549629d4ff53e5772b9a6ef2a01ba2a38493e769b15ab5",
+    "rl_engine/backends/shared/triton/rope/multi_axis_rope.py": "5b989bd63efd77913267b48654d10204bc563768c6c5d4e1cde9fb55e1354547",
+    "rl_engine/ops/rope/_registration.py": "add0375d1639444908091d01e8d29781f7a975beffec64b77bfaa81a5b9a0ae3",
+    "rl_engine/ops/rope/multi_axis_rope.py": "07bcff1160f22c1ff6730be731206b2f4a153e2fada26afc93fc884e780568cb",
+    "rl_engine/reference/rope/multi_axis_rope.py": "c9a94730047f51659f7aa0a443e88d6ce7d03d00d6813a9aed548cafbf6347bf",
+    "rl_engine/runtime/registry.py": "9f9c69bb6c2e941b057a66ac2d388be021cbb4b55cb3c602976844aabc6106e1",
+    "rl_engine/validation/operators/gradient_adapters.py": "5ee2e9ad619a4cad6e81d8f4a6ccf81ecf45e63a7764d0fcb4c204206b37c16b",
+    "rl_engine/validation/operators/operator_inputs.py": "55c8d9e7cbde2070662dd1d0ebe9888aef106b45fb05902057aec032d29a9836",
+    "rl_engine/validation/operators/operator_specs.py": "cef2e5bed8075509d5d5a538b378477b09f1adfe5fd01829156357cefa99f983",
+    "rl_engine/validation/operators/qwen_image_rope.py": "c3ddfc732f32de3c633de16064810d67202254ea78fd8107d558b471bb15a37f",
+    "tests/build/test_multi_axis_rope_build.py": "90c3740b2d4946acc6206d952622c19afaa0089cca904919088563427555b956",
+    "tests/ops/rope/test_multi_axis_rope.py": "1a0950b09eec407e8ffd490b40dda4118f8d9b876f4db4af8aac648af1a76687",
+    "tests/validation/operators/test_operator_inputs.py": "36365f662194f158276b3c07f6447471a0cf17253ec420ab63c8b9149c511b49",
+    "tools/validation/operators/check_multi_axis_rope_install.py": "ff8198ff55779ea2ede88d4ecfe07caa627631b4b08bce38139ccfa147abf538",
+    "tools/validation/operators/check_qwenimage_rope_upstream.py": "cc368d5bebf0f0acb21b7455348592d30442b6a8d349de0f49d5a09c5f20ce4e",
+    "tools/validation/operators/validate_multi_axis_rope.py": "def4747da9ea9194c761bef24479017ac74ac10de46a3fbe1a9528ec6ee07731"
+  },
+  "evidence_sha256": {
+    "baseline-layout.txt": "b050d798a5e3a6c2d43c7c22e4e1140529ebf05ebab982df17f501f5b69cff12",
+    "benchmark-bfloat16.json": "0679a422e8129a8a36ada938b024794334f784cdea28013574f99c70673e13a0",
+    "installed-wheel.json": "60f6a1a48d35fa25a639a1ff168ce0c37659e0258737422cafda3741d8ef0200",
+    "installed-wheel.txt": "2be6479a8e30938410095dbcc7724ad114cb61cb35ef86c7f86b8d507a3193f6",
+    "layout.txt": "83169d7995bdadc14744a0ed0fff69b334f2ed6ff0b92cda005a9cf0f733116d",
+    "memcheck.txt": "f6f05d4e5b772909f29baf2a1d6be04b4456660ca024e5067c3f64c7efa4b81f",
+    "operator-cuda.json": "1a231d26f2411a38cffe70fdb59b388227fc0e7da9625370b165d7d590e1e011",
+    "operator-triton.json": "af2be1f4df1bf9d7afb4802ca02c81826e64acacae891436f83ad6b3342109f5",
+    "package-build.txt": "34b11d626dc4988f3d862e4122599a0b1a949b1d799436981dd74f3ee5c1b07e",
+    "package.json": "e267d82707cdeb0c53d120928cd32a544a44190e8e41c43f3b6c1d4668275a7e",
+    "precommit.txt": "04f9db16780e0ad0b41e6450d362e6f0a7bdb07128596c8ea13f8ea8bb2cc674",
+    "quality.txt": "d4728355fe488102ca644f510795c474b54f8d42bb92ddfda34557ba521c6afd",
+    "README.md": "aa2cabc1e8423f4d737d0db90786142c4a919041dbf97335b17991475a4cbaca",
+    "regression.txt": "5d058ac93516d9ce1f486fbd4f35b84a254a32d8ddf26eedf41cf2388153e004",
+    "tests.txt": "f6af2230f6a0b5d993cb6bca2724a82f772c6b33df456edce6e728996f2a76d5",
+    "upstream.json": "8c087d60013f0266303c6d60237d87e5f1ff67b03e8e8c1900de4b20349e2e93",
+    "ws1-cuda.json": "531124223ae5fe43eca890fb1ccf80e09f33affe34bab132c7c17742081d85a3",
+    "ws1-cuda.txt": "b4cf09ef5165e535dba9ed19eeec08f15eccaf86b549bb1a253fd08e87be35ab",
+    "ws1-triton.json": "29eedee8e50b8750d0ce9be95deb39fc6aa9ab47b0867fe20ce2e7ad5cb47548",
+    "ws1-triton.txt": "a1daefba35f7c1a06d222750ad1ce9c4bdc97cc03410969e7b4f4ce011a6770a"
+  },
+  "layout_failures_equal_clean_base": true,
+  "layout_failure_ids": [
+    "tests/build/test_build_platform_collectives.py::test_cuda_build_keeps_existing_ipc_collective",
+    "tests/build/test_build_platform_collectives.py::test_package_data_json_links_resolve_to_real_content",
+    "tests/runtime/test_layout_compatibility.py::test_recorded_algorithm_sources_resolve_to_real_implementations",
+    "tests/runtime/test_layout_compatibility.py::test_packaged_contract_bytes_match_main[rl_engine.config-workloads/qwen3_8b.json-80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6]",
+    "tests/runtime/test_layout_compatibility.py::test_packaged_contract_bytes_match_main[rl_engine.contracts-profiles/precision/ws1.json-044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532]"
+  ],
+  "package_source_and_license_checks": "passed",
+  "hash_mode": "SHA256 of recorded working-tree bytes; upstream baselines separately normalize newlines"
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/memcheck.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/memcheck.txt
new file mode 100644
index 000000000..dd59e40b0
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/memcheck.txt
@@ -0,0 +1,8 @@
+========= COMPUTE-SANITIZER
+WARNING 10-28 16:46:09 [RL-Kernel]: Core binary extension (_C) unavailable: cannot import name '_C' from 'rl_engine' (D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\rl_engine\__init__.py). Falling back to native code.
+W1028 16:46:10.075000 16608 site-packages\torch\utils\cpp_extension.py:480] Error checking compiler version for cl: 'cp1' codec can't decode bytes in position 0--1: No mapping for the Unicode character exists in the target code page.
+ninja: no work to do.
+Focused CUDA source build; torch=2.9.1+cu128, GPU=NVIDIA GeForce RTX 5080
+..................................                                       [100%]
+34 passed, 48 deselected in 3.96s
+========= ERROR SUMMARY: 0 errors
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-cuda.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-cuda.json
new file mode 100644
index 000000000..03d59bb4b
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-cuda.json
@@ -0,0 +1,70 @@
+{
+  "suite_name": "multi_axis_rope",
+  "total_candidates": 1,
+  "passed_candidates": 1,
+  "pass_rate": 1.0,
+  "passed": true,
+  "candidates": [
+    {
+      "candidate_name": "cuda-multi_axis_rope",
+      "backend": "cuda",
+      "total_outputs": 2,
+      "passed_outputs": 2,
+      "pass_rate": 1.0,
+      "passed": true,
+      "cases": [
+        {
+          "case_name": "multi_axis_rope-torch.bfloat16-2x7x3x128",
+          "dtype": "torch.bfloat16",
+          "op_class": "elementwise",
+          "passed": true,
+          "outputs": [
+            {
+              "output_index": 0,
+              "shape": [
+                2,
+                7,
+                3,
+                128
+              ],
+              "candidate_dtype": "torch.bfloat16",
+              "gold_dtype": "torch.float32",
+              "atol": 0.02,
+              "rtol": 0.016,
+              "max_abs_error": 0.007704973220825195,
+              "mean_abs_error": 0.0006018931162543595,
+              "max_rel_error": 0.003828622167930007,
+              "passed": true,
+              "judgment": "forward_accuracy",
+              "comparison_lhs_role": "bf16_candidate",
+              "comparison_rhs_role": "fp32_reference",
+              "message": ""
+            },
+            {
+              "output_index": 1,
+              "shape": [
+                2,
+                7,
+                3,
+                128
+              ],
+              "candidate_dtype": "torch.bfloat16",
+              "gold_dtype": "torch.bfloat16",
+              "atol": 0.02,
+              "rtol": 0.016,
+              "max_abs_error": 0.0,
+              "mean_abs_error": 0.0,
+              "max_rel_error": 0.0,
+              "passed": true,
+              "judgment": "gradient_accuracy",
+              "comparison_lhs_role": "bf16_candidate",
+              "comparison_rhs_role": "fp32_reference",
+              "message": "gradient:x"
+            }
+          ]
+        }
+      ],
+      "backend_provenance": null
+    }
+  ]
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-triton.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-triton.json
new file mode 100644
index 000000000..6f086edac
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-triton.json
@@ -0,0 +1,70 @@
+{
+  "suite_name": "multi_axis_rope",
+  "total_candidates": 1,
+  "passed_candidates": 1,
+  "pass_rate": 1.0,
+  "passed": true,
+  "candidates": [
+    {
+      "candidate_name": "triton-multi_axis_rope",
+      "backend": "triton",
+      "total_outputs": 2,
+      "passed_outputs": 2,
+      "pass_rate": 1.0,
+      "passed": true,
+      "cases": [
+        {
+          "case_name": "multi_axis_rope-torch.bfloat16-2x7x3x128",
+          "dtype": "torch.bfloat16",
+          "op_class": "elementwise",
+          "passed": true,
+          "outputs": [
+            {
+              "output_index": 0,
+              "shape": [
+                2,
+                7,
+                3,
+                128
+              ],
+              "candidate_dtype": "torch.bfloat16",
+              "gold_dtype": "torch.float32",
+              "atol": 0.02,
+              "rtol": 0.016,
+              "max_abs_error": 0.007704973220825195,
+              "mean_abs_error": 0.0006018931162543595,
+              "max_rel_error": 0.003828622167930007,
+              "passed": true,
+              "judgment": "forward_accuracy",
+              "comparison_lhs_role": "bf16_candidate",
+              "comparison_rhs_role": "fp32_reference",
+              "message": ""
+            },
+            {
+              "output_index": 1,
+              "shape": [
+                2,
+                7,
+                3,
+                128
+              ],
+              "candidate_dtype": "torch.bfloat16",
+              "gold_dtype": "torch.bfloat16",
+              "atol": 0.02,
+              "rtol": 0.016,
+              "max_abs_error": 0.0,
+              "mean_abs_error": 0.0,
+              "max_rel_error": 0.0,
+              "passed": true,
+              "judgment": "gradient_accuracy",
+              "comparison_lhs_role": "bf16_candidate",
+              "comparison_rhs_role": "fp32_reference",
+              "message": "gradient:x"
+            }
+          ]
+        }
+      ],
+      "backend_provenance": null
+    }
+  ]
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package-build.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package-build.txt
new file mode 100644
index 000000000..d63d79d52
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package-build.txt
@@ -0,0 +1,40 @@
+Selected lines from the completed sdist/wheel build log.
+Full local log SHA256: bfbcd342efaf8caa36adcd208acee385db1ba6fbc4326fe8e0325f9523f199b0
+running sdist
+copying rl_engine\backends\shared\triton\rope\FLASH_ATTENTION_LICENSE -> rl_kernel-0.1.0\rl_engine\backends\shared\triton\rope
+running bdist_wheel
+copying rl_engine\backends\shared\triton\rope\FLASH_ATTENTION_LICENSE -> build\lib.win-amd64-cpython-313\rl_engine\backends\shared\triton\rope
+copying rl_engine\backends\shared\triton\rope\FLASH_ATTENTION_LICENSE -> build\lib.win-amd64-cpython-313\rl_engine\backends\shared\triton\rope
+running build_ext
+ninja: no work to do.
+"C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.44.35207\bin\HostX64\x64\link.exe" /nologo /INCREMENTAL:NO /LTCG /DLL /MANIFEST:EMBED,ID=2 /MANIFESTUAC:NO /LIBPATH:D:\dandan123\dandan\myvllm\.cuda-12.8\Library\lib /LIBPATH:D:\anaconda\Lib\site-packages\torch\lib /LIBPATH:D:\dandan123\dandan\myvllm\.cuda-12.8\Library\lib\x64 /LIBPATH:D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\.venv\libs /LIBPATH:D:\anaconda\libs /LIBPATH:D:\anaconda /LIBPATH:D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\.venv\PCbuild\amd64 "/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.44.35207\lib\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\10\lib\10.0.26100.0\ucrt\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\10\\lib\10.0.26100.0\\um\x64" c10.lib torch.lib torch_cpu.lib torch_python.lib cudart.lib c10_cuda.lib torch_cuda.lib /EXPORT:PyInit__rope_cuda D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\bindings\rope_module.obj D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\cuda\rope\multi_axis_rope.obj /OUT:build\lib.win-amd64-cpython-313\rl_engine\_rope_cuda.cp313-win_amd64.pyd /IMPLIB:D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\bindings\_rope_cuda.cp313-win_amd64.lib
+copying build\lib.win-amd64-cpython-313\rl_engine\backends\shared\triton\rope\FLASH_ATTENTION_LICENSE -> build\bdist.win-amd64\wheel\.\rl_engine\backends\shared\triton\rope
+copying build\lib.win-amd64-cpython-313\rl_engine\_rope_cuda.cp313-win_amd64.pyd -> build\bdist.win-amd64\wheel\.\rl_engine
+adding 'rl_engine/_rope_cuda.cp313-win_amd64.pyd'
+adding 'rl_engine/backends/shared/triton/rope/FLASH_ATTENTION_LICENSE'
+
+In-place build log:
+D:\anaconda\Lib\site-packages\setuptools\config\_apply_pyprojecttoml.py:82: SetuptoolsDeprecationWarning: `project.license` as a TOML table is deprecated
+!!
+
+        ********************************************************************************
+        Please use a simple string containing a SPDX expression for `project.license`. You can also use `project.license-files`. (Both options available on setuptools>=77.0.0).
+
+        By 2027-Feb-18, you need to update your project and remove deprecated calls
+        or your builds will no longer be supported.
+
+        See https://packaging.python.org/en/latest/guides/writing-pyproject-toml/#license for details.
+        ********************************************************************************
+
+!!
+  corresp(dist, value, root_dir)
+running build_ext
+W1028 16:52:54.107000 9384 site-packages\torch\utils\cpp_extension.py:480] Error checking compiler version for cl: 'cp1' codec can't decode bytes in position 0--1: No mapping for the Unicode character exists in the target code page.
+building 'rl_engine._rope_cuda' extension
+W1028 16:52:54.150000 9384 site-packages\torch\utils\cpp_extension.py:480] Error checking compiler version for cl: 'cp1' codec can't decode bytes in position 0--1: No mapping for the Unicode character exists in the target code page.
+ninja: no work to do.
+"C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.44.35207\bin\HostX64\x64\link.exe" /nologo /INCREMENTAL:NO /LTCG /DLL /MANIFEST:EMBED,ID=2 /MANIFESTUAC:NO /LIBPATH:D:\dandan123\dandan\myvllm\.cuda-12.8\Library\lib /LIBPATH:D:\anaconda\Lib\site-packages\torch\lib /LIBPATH:D:\dandan123\dandan\myvllm\.cuda-12.8\Library\lib\x64 /LIBPATH:D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\.venv\libs /LIBPATH:D:\anaconda\libs /LIBPATH:D:\anaconda /LIBPATH:D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\.venv\PCbuild\amd64 "/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.44.35207\lib\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\10\lib\10.0.26100.0\ucrt\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\10\\lib\10.0.26100.0\\um\x64" c10.lib torch.lib torch_cpu.lib torch_python.lib cudart.lib c10_cuda.lib torch_cuda.lib /EXPORT:PyInit__rope_cuda D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\bindings\rope_module.obj D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\cuda\rope\multi_axis_rope.obj /OUT:build\lib.win-amd64-cpython-313\rl_engine\_rope_cuda.cp313-win_amd64.pyd /IMPLIB:D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\bindings\_rope_cuda.cp313-win_amd64.lib
+  正在创建库 D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\bindings\_rope_cuda.cp313-win_amd64.lib 和对象 D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\bindings\_rope_cuda.cp313-win_amd64.exp
+正在生成代码
+已完成代码的生成
+copying build\lib.win-amd64-cpython-313\rl_engine\_rope_cuda.cp313-win_amd64.pyd -> rl_engine
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package.json
new file mode 100644
index 000000000..51e102428
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package.json
@@ -0,0 +1,31 @@
+{
+  "wheel": "rl_kernel-0.1.0-cp313-cp313-win_amd64.whl",
+  "wheel_sha256": "bc5418bbc4be779ec6d5b4fa99d54b3244f2570d60a29adc3e6a1c38be25a864",
+  "sdist": "rl_kernel-0.1.0.tar.gz",
+  "sdist_sha256": "56026574fa3307e85e6e278298a09deee13b43d4ec5c61899698ac5e55163a12",
+  "native_library_member": "rl_engine/_rope_cuda.cp313-win_amd64.pyd",
+  "native_library_sha256": "c0878eeac05b6b400ed7c3935561de525ce3af8b2a050b5d8bfe5cdd66650940",
+  "inplace_library_sha256": "aab5bf0777da43609e067e6a41bc37c46fdf6b1dd0c60d1071c15bbc4f707242",
+  "binary_hash_scope": "Wheel and in-place libraries were linked separately; record both hashes",
+  "license_in_both_archives": true,
+  "license_sha256": "8c9ccb96c065e706135b6cbad279b721da6156e51f3a5f27c6b3329af9416d73",
+  "wheel_sources_match_checkout": [
+    "rl_engine/backends/cuda/rope/multi_axis_rope.py",
+    "rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py",
+    "rl_engine/backends/shared/triton/rope/multi_axis_rope.py",
+    "rl_engine/ops/rope/_registration.py",
+    "rl_engine/ops/rope/multi_axis_rope.py",
+    "rl_engine/reference/rope/multi_axis_rope.py",
+    "rl_engine/runtime/registry.py",
+    "rl_engine/validation/operators/gradient_adapters.py",
+    "rl_engine/validation/operators/operator_inputs.py",
+    "rl_engine/validation/operators/operator_specs.py",
+    "rl_engine/validation/operators/qwen_image_rope.py"
+  ],
+  "sdist_required_sources_match_checkout": [
+    "csrc/bindings/rope_module.cpp",
+    "csrc/cuda/rope/multi_axis_rope.cu",
+    "rl_engine/backends/shared/triton/rope/FLASH_ATTENTION_LICENSE"
+  ],
+  "profile": "RL_KERNEL_ROPE_ONLY=1; full monolithic _C build not qualified"
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/precommit.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/precommit.txt
new file mode 100644
index 000000000..475a8c0d8
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/precommit.txt
@@ -0,0 +1,7 @@
+trim trailing whitespace.................................................Passed
+fix end of files.........................................................Passed
+check yaml...............................................................Passed
+check for added large files..............................................Passed
+black....................................................................Passed
+isort....................................................................Passed
+flake8...................................................................Passed
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/quality.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/quality.txt
new file mode 100644
index 000000000..2756f4004
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/quality.txt
@@ -0,0 +1,58 @@
+Changed-file checks; repository-pinned Black 24.4.2, isort 5.13.2 and Flake8 7.0.0.
+Test commands are separate from the static checks below.
+Regression command: python -m pytest tests/build/test_multi_axis_rope_build.py tests/validation/operators/test_operator_inputs.py tests/validation/operators/test_gradient_invariance.py tests/validation/operators/test_forward_invariance.py tests/validation/operators/test_elementwise_inventory.py tests/validation/operators/test_four_judgment_matrix.py tests/validation/operators/test_ws1_workload.py tests/runtime/test_kernel_registry.py tests/runtime/test_dispatch.py tests/ops/rope/test_rope.py -q -rs
+Layout output copies have only trailing whitespace normalized by pre-commit.
+
+Command: python -m isort --check-only --profile black --line-length 100 benchmarks/operators/rope/benchmark_multi_axis_rope.py build_tools/extensions.py rl_engine/backends/cuda/rope/multi_axis_rope.py rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py rl_engine/backends/shared/triton/rope/multi_axis_rope.py rl_engine/ops/rope/_registration.py rl_engine/ops/rope/multi_axis_rope.py rl_engine/reference/rope/multi_axis_rope.py rl_engine/runtime/registry.py rl_engine/validation/operators/gradient_adapters.py rl_engine/validation/operators/operator_inputs.py rl_engine/validation/operators/operator_specs.py rl_engine/validation/operators/qwen_image_rope.py tests/build/test_multi_axis_rope_build.py tests/ops/rope/test_multi_axis_rope.py tests/validation/operators/test_operator_inputs.py tools/validation/operators/check_multi_axis_rope_install.py tools/validation/operators/check_qwenimage_rope_upstream.py tools/validation/operators/validate_multi_axis_rope.py
+Exit code: 0
+
+
+Command: python -m black --check --line-length 100 benchmarks/operators/rope/benchmark_multi_axis_rope.py build_tools/extensions.py rl_engine/backends/cuda/rope/multi_axis_rope.py rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py rl_engine/backends/shared/triton/rope/multi_axis_rope.py rl_engine/ops/rope/_registration.py rl_engine/ops/rope/multi_axis_rope.py rl_engine/reference/rope/multi_axis_rope.py rl_engine/runtime/registry.py rl_engine/validation/operators/gradient_adapters.py rl_engine/validation/operators/operator_inputs.py rl_engine/validation/operators/operator_specs.py rl_engine/validation/operators/qwen_image_rope.py tests/build/test_multi_axis_rope_build.py tests/ops/rope/test_multi_axis_rope.py tests/validation/operators/test_operator_inputs.py tools/validation/operators/check_multi_axis_rope_install.py tools/validation/operators/check_qwenimage_rope_upstream.py tools/validation/operators/validate_multi_axis_rope.py
+Exit code: 0
+All done! ✨ 🍰 ✨
+
+19 files would be left unchanged.
+
+Command: python -m flake8 --max-line-length=100 --extend-ignore=E203,E704 benchmarks/operators/rope/benchmark_multi_axis_rope.py build_tools/extensions.py rl_engine/backends/cuda/rope/multi_axis_rope.py rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py rl_engine/backends/shared/triton/rope/multi_axis_rope.py rl_engine/ops/rope/_registration.py rl_engine/ops/rope/multi_axis_rope.py rl_engine/reference/rope/multi_axis_rope.py rl_engine/runtime/registry.py rl_engine/validation/operators/gradient_adapters.py rl_engine/validation/operators/operator_inputs.py rl_engine/validation/operators/operator_specs.py rl_engine/validation/operators/qwen_image_rope.py tests/build/test_multi_axis_rope_build.py tests/ops/rope/test_multi_axis_rope.py tests/validation/operators/test_operator_inputs.py tools/validation/operators/check_multi_axis_rope_install.py tools/validation/operators/check_qwenimage_rope_upstream.py tools/validation/operators/validate_multi_axis_rope.py
+Exit code: 0
+
+
+Command: python -m ruff check benchmarks/operators/rope/benchmark_multi_axis_rope.py build_tools/extensions.py rl_engine/backends/cuda/rope/multi_axis_rope.py rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py rl_engine/backends/shared/triton/rope/multi_axis_rope.py rl_engine/ops/rope/_registration.py rl_engine/ops/rope/multi_axis_rope.py rl_engine/reference/rope/multi_axis_rope.py rl_engine/runtime/registry.py rl_engine/validation/operators/gradient_adapters.py rl_engine/validation/operators/operator_inputs.py rl_engine/validation/operators/operator_specs.py rl_engine/validation/operators/qwen_image_rope.py tests/build/test_multi_axis_rope_build.py tests/ops/rope/test_multi_axis_rope.py tests/validation/operators/test_operator_inputs.py tools/validation/operators/check_multi_axis_rope_install.py tools/validation/operators/check_qwenimage_rope_upstream.py tools/validation/operators/validate_multi_axis_rope.py
+Exit code: 0
+All checks passed!
+
+Command: python -m mkdocs build --strict -f mkdocs.yaml
+Exit code: 0
+INFO    -  Cleaning site directory
+
+INFO    -  Building documentation to directory: D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\site
+
+mkdocs_git_revision_date_localized_plugin: 'D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\docs\operators\multi-axis-rope.md' has no git logs, using current timestamp
+
+INFO    -  The following pages exist in the docs directory, but are not included in the "nav" configuration:
+
+  - blog\2026-07-08-announcing-rl-kernel-linear-logp-for-vime-zh.md
+
+  - blog\2026-08-09-dsv4-flash-moe-consistency-roadmap.md
+
+  - operators\matmul.md
+
+  - operators\rope.md
+
+  - usage\cross-config-drift-report.md
+
+  - usage\cuda-rocm-consistency-audit.md
+
+  - usage\h100-configurable.md
+
+  - usage\h100-cp-gradient-validation.md
+
+  - usage\h100-matrix-validation.md
+
+  - usage\h100-pcp-validation.md
+
+  - usage\rocm-sparse-performance.md
+
+INFO    -  Documentation built in 4.01 seconds
+
+Repository pre-commit: passed; see precommit.txt.
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/regression.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/regression.txt
new file mode 100644
index 000000000..066e09cec
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/regression.txt
@@ -0,0 +1,10 @@
+........................................................................ [ 35%]
+........................................................................ [ 71%]
+...................................................sssssss               [100%]
+=========================== short test summary info ===========================
+SKIPPED [3] tests\ops\rope\test_rope.py:390: rope_apply_ascend is not compiled (requires an Ascend NPU build).
+SKIPPED [1] tests\ops\rope\test_rope.py:406: rope_apply_ascend is not compiled (requires an Ascend NPU build).
+SKIPPED [1] tests\ops\rope\test_rope.py:413: rope_apply_ascend is not compiled (requires an Ascend NPU build).
+SKIPPED [1] tests\ops\rope\test_rope.py:420: rope_apply_ascend is not compiled (requires an Ascend NPU build).
+SKIPPED [1] tests\ops\rope\test_rope.py:433: rope_apply_ascend is not compiled (requires an Ascend NPU build).
+195 passed, 7 skipped in 4.31s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/tests.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/tests.txt
new file mode 100644
index 000000000..38e8e21e6
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/tests.txt
@@ -0,0 +1,5 @@
+........................................................................ [ 87%]
+s.........                                                               [100%]
+=========================== short test summary info ===========================
+SKIPPED [1] tests\ops\rope\test_multi_axis_rope.py:336: requires two NVIDIA GPUs
+81 passed, 1 skipped in 5.55s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/upstream.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/upstream.json
new file mode 100644
index 000000000..1f89880c1
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/upstream.json
@@ -0,0 +1,335 @@
+{
+  "source_sha256": "0bad11901deae69f05aec01fd464e51e38582d455068f2851ee665036179321f",
+  "cases": [
+    {
+      "grid": [
+        2,
+        3
+      ],
+      "stream": "image",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        2,
+        3
+      ],
+      "stream": "image",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        2,
+        3
+      ],
+      "stream": "image",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        2,
+        3
+      ],
+      "stream": "text",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        2,
+        3
+      ],
+      "stream": "text",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        2,
+        3
+      ],
+      "stream": "text",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        3,
+        5
+      ],
+      "stream": "image",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        3,
+        5
+      ],
+      "stream": "image",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        3,
+        5
+      ],
+      "stream": "image",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        3,
+        5
+      ],
+      "stream": "text",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        3,
+        5
+      ],
+      "stream": "text",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        3,
+        5
+      ],
+      "stream": "text",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        64,
+        64
+      ],
+      "stream": "image",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        64,
+        64
+      ],
+      "stream": "image",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        64,
+        64
+      ],
+      "stream": "image",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        64,
+        64
+      ],
+      "stream": "text",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        64,
+        64
+      ],
+      "stream": "text",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        64,
+        64
+      ],
+      "stream": "text",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        83,
+        83
+      ],
+      "stream": "image",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        83,
+        83
+      ],
+      "stream": "image",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        83,
+        83
+      ],
+      "stream": "image",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        83,
+        83
+      ],
+      "stream": "text",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        83,
+        83
+      ],
+      "stream": "text",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        83,
+        83
+      ],
+      "stream": "text",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        104,
+        58
+      ],
+      "stream": "image",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        104,
+        58
+      ],
+      "stream": "image",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        104,
+        58
+      ],
+      "stream": "image",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        104,
+        58
+      ],
+      "stream": "text",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        104,
+        58
+      ],
+      "stream": "text",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        104,
+        58
+      ],
+      "stream": "text",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    }
+  ]
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.json
new file mode 100644
index 000000000..20ad4dd78
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.json
@@ -0,0 +1,1962 @@
+{
+  "scope": "Qwen-Image multi_axis_rope operator; C1 elementwise and C2 config transforms",
+  "trace": {
+    "contract": "qwenimage-interleaved-rope-fp32-v1",
+    "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
+    "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
+    "layout": "BSHD, adjacent channel pairs",
+    "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+    "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+  },
+  "passed": true,
+  "forward": {
+    "op_name": "multi_axis_rope",
+    "backend_profile": "cuda_bf16",
+    "accuracy_reports": [
+      {
+        "config_id": "BN/full",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/chunked",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/chunked",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004673302173614502,
+            "mean_abs_error": 6.917401333339512e-05,
+            "max_rel_error": 0.003753266530111432,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.00048516690731048584,
+            "mean_abs_error": 7.802647451171651e-05,
+            "max_rel_error": 0.0037793784867972136,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004864335060119629,
+            "mean_abs_error": 7.697804539930075e-05,
+            "max_rel_error": 0.003821910358965397,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.710371912457049e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004673302173614502,
+            "mean_abs_error": 6.917401333339512e-05,
+            "max_rel_error": 0.003753266530111432,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.00048516690731048584,
+            "mean_abs_error": 7.802647451171651e-05,
+            "max_rel_error": 0.0037793784867972136,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004864335060119629,
+            "mean_abs_error": 7.697804539930075e-05,
+            "max_rel_error": 0.003821910358965397,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.710371912457049e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/permuted",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/permuted",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_right",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/padded_right",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_left",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/padded_left",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      }
+    ],
+    "invariance_reports": [
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/chunked",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/chunked"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s0",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s1",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s2",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s3",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/permuted",
+        "transform_kind": "permutation",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/permuted"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_right",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_right"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_left",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_left"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      }
+    ],
+    "logprob_smoke": null,
+    "backend_provenance": {
+      "backend_profile": "cuda_bf16",
+      "requested_backend": "cuda",
+      "actual_backend": "cuda",
+      "execution_dtype": "bfloat16",
+      "accumulation_dtype": "float32",
+      "output_dtype": "bfloat16",
+      "reference_dtype": "float32",
+      "candidate_tf32_enabled": false,
+      "reference_tf32_enabled": false
+    },
+    "candidate_id": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
+    "device": "cuda",
+    "compute_capability": "sm120",
+    "seed": 20260812,
+    "fallback_reason": null,
+    "passed": true,
+    "provenance_valid": true,
+    "metadata_valid": true,
+    "observed_kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope"
+  },
+  "gradient": {
+    "op_name": "multi_axis_rope",
+    "backend_profile": "cuda_bf16",
+    "accuracy_reports": [
+      {
+        "config_id": "BN/full",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/chunked",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/chunked",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.7887988835573196e-05,
+            "mean_abs_error": 3.6821968478761846e-06,
+            "max_rel_error": 1.8164952993392944,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.7113780965446495e-06,
+            "max_rel_error": 0.4316565990447998,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.995925024151802e-05,
+            "mean_abs_error": 4.7168209675874095e-06,
+            "max_rel_error": 0.7439075112342834,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.0211638659238815e-05,
+            "mean_abs_error": 4.9605891945248e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.7887988835573196e-05,
+            "mean_abs_error": 3.6821968478761846e-06,
+            "max_rel_error": 1.8164952993392944,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.7113780965446495e-06,
+            "max_rel_error": 0.4316565990447998,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.995925024151802e-05,
+            "mean_abs_error": 4.7168209675874095e-06,
+            "max_rel_error": 0.7439075112342834,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.0211638659238815e-05,
+            "mean_abs_error": 4.9605891945248e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/permuted",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/permuted",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_right",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/padded_right",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_left",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/padded_left",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      }
+    ],
+    "invariance_reports": [
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/chunked",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/chunked"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s0",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s1",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s2",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s3",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/permuted",
+        "transform_kind": "permutation",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/permuted"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_right",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_right"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_left",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_left"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      }
+    ],
+    "singleton_aggregate_reports": [],
+    "backend_provenance": {
+      "backend_profile": "cuda_bf16",
+      "requested_backend": "cuda",
+      "actual_backend": "cuda",
+      "execution_dtype": "bfloat16",
+      "accumulation_dtype": "float32",
+      "output_dtype": "bfloat16",
+      "reference_dtype": "float32",
+      "candidate_tf32_enabled": false,
+      "reference_tf32_enabled": false
+    },
+    "candidate_id": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
+    "device": "cuda",
+    "compute_capability": "sm120",
+    "seed": 20260812,
+    "fallback_reason": null,
+    "passed": true,
+    "provenance_valid": true,
+    "metadata_valid": true,
+    "loss_reduction": "sum_over_active_tokens_then_optional_mean_by_active_count",
+    "active_token_denominator": 27,
+    "grad_tensor_names": [
+      "dx"
+    ],
+    "first_failing_op": null,
+    "first_failing_tensor": null,
+    "first_failing_config_pair": null,
+    "observed_kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope"
+  }
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.txt
new file mode 100644
index 000000000..e69d60986
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.txt
@@ -0,0 +1,3 @@
+WARNING 10-28 16:54:40 [RL-Kernel]: Core binary extension (_C) unavailable: cannot import name '_C' from 'rl_engine' (D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\rl_engine\__init__.py). Falling back to native code.
+Using the installed production RoPE dispatcher library
+WS1 four judgments backend=cuda: passed=True
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.json
new file mode 100644
index 000000000..d81adfea1
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.json
@@ -0,0 +1,1962 @@
+{
+  "scope": "Qwen-Image multi_axis_rope operator; C1 elementwise and C2 config transforms",
+  "trace": {
+    "contract": "qwenimage-interleaved-rope-fp32-v1",
+    "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
+    "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
+    "layout": "BSHD, adjacent channel pairs",
+    "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+    "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+  },
+  "passed": true,
+  "forward": {
+    "op_name": "multi_axis_rope",
+    "backend_profile": "triton_cuda_bf16",
+    "accuracy_reports": [
+      {
+        "config_id": "BN/full",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/chunked",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/chunked",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004673302173614502,
+            "mean_abs_error": 6.917401333339512e-05,
+            "max_rel_error": 0.003753266530111432,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.00048516690731048584,
+            "mean_abs_error": 7.802647451171651e-05,
+            "max_rel_error": 0.0037793784867972136,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004864335060119629,
+            "mean_abs_error": 7.697804539930075e-05,
+            "max_rel_error": 0.003821910358965397,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.710371912457049e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004673302173614502,
+            "mean_abs_error": 6.917401333339512e-05,
+            "max_rel_error": 0.003753266530111432,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.00048516690731048584,
+            "mean_abs_error": 7.802647451171651e-05,
+            "max_rel_error": 0.0037793784867972136,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004864335060119629,
+            "mean_abs_error": 7.697804539930075e-05,
+            "max_rel_error": 0.003821910358965397,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.710371912457049e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/permuted",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/permuted",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_right",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/padded_right",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_left",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/padded_left",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      }
+    ],
+    "invariance_reports": [
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/chunked",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/chunked"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s0",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s1",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s2",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s3",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/permuted",
+        "transform_kind": "permutation",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/permuted"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_right",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_right"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_left",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_left"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      }
+    ],
+    "logprob_smoke": null,
+    "backend_provenance": {
+      "backend_profile": "triton_cuda_bf16",
+      "requested_backend": "triton",
+      "actual_backend": "triton",
+      "execution_dtype": "bfloat16",
+      "accumulation_dtype": "float32",
+      "output_dtype": "bfloat16",
+      "reference_dtype": "float32",
+      "candidate_tf32_enabled": false,
+      "reference_tf32_enabled": false
+    },
+    "candidate_id": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
+    "device": "cuda",
+    "compute_capability": "sm120",
+    "seed": 20260812,
+    "fallback_reason": null,
+    "passed": true,
+    "provenance_valid": true,
+    "metadata_valid": true,
+    "observed_kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel"
+  },
+  "gradient": {
+    "op_name": "multi_axis_rope",
+    "backend_profile": "triton_cuda_bf16",
+    "accuracy_reports": [
+      {
+        "config_id": "BN/full",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/chunked",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/chunked",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.7887988835573196e-05,
+            "mean_abs_error": 3.6821968478761846e-06,
+            "max_rel_error": 1.8164952993392944,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.7113780965446495e-06,
+            "max_rel_error": 0.4316565990447998,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.995925024151802e-05,
+            "mean_abs_error": 4.7168209675874095e-06,
+            "max_rel_error": 0.7439075112342834,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.0211638659238815e-05,
+            "mean_abs_error": 4.9605891945248e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.7887988835573196e-05,
+            "mean_abs_error": 3.6821968478761846e-06,
+            "max_rel_error": 1.8164952993392944,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.7113780965446495e-06,
+            "max_rel_error": 0.4316565990447998,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.995925024151802e-05,
+            "mean_abs_error": 4.7168209675874095e-06,
+            "max_rel_error": 0.7439075112342834,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.0211638659238815e-05,
+            "mean_abs_error": 4.9605891945248e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/permuted",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/permuted",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_right",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/padded_right",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_left",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/padded_left",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      }
+    ],
+    "invariance_reports": [
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/chunked",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/chunked"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s0",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s1",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s2",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s3",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/permuted",
+        "transform_kind": "permutation",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/permuted"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_right",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_right"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_left",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_left"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      }
+    ],
+    "singleton_aggregate_reports": [],
+    "backend_provenance": {
+      "backend_profile": "triton_cuda_bf16",
+      "requested_backend": "triton",
+      "actual_backend": "triton",
+      "execution_dtype": "bfloat16",
+      "accumulation_dtype": "float32",
+      "output_dtype": "bfloat16",
+      "reference_dtype": "float32",
+      "candidate_tf32_enabled": false,
+      "reference_tf32_enabled": false
+    },
+    "candidate_id": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
+    "device": "cuda",
+    "compute_capability": "sm120",
+    "seed": 20260812,
+    "fallback_reason": null,
+    "passed": true,
+    "provenance_valid": true,
+    "metadata_valid": true,
+    "loss_reduction": "sum_over_active_tokens_then_optional_mean_by_active_count",
+    "active_token_denominator": 27,
+    "grad_tensor_names": [
+      "dx"
+    ],
+    "first_failing_op": null,
+    "first_failing_tensor": null,
+    "first_failing_config_pair": null,
+    "observed_kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel"
+  }
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.txt
new file mode 100644
index 000000000..283d1933a
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.txt
@@ -0,0 +1,3 @@
+WARNING 10-28 16:54:47 [RL-Kernel]: Core binary extension (_C) unavailable: cannot import name '_C' from 'rl_engine' (D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\rl_engine\__init__.py). Falling back to native code.
+Using the installed production RoPE dispatcher library
+WS1 four judgments backend=triton: passed=True
diff --git a/rl_engine/backends/cuda/rope/multi_axis_rope.py b/rl_engine/backends/cuda/rope/multi_axis_rope.py
new file mode 100644
index 000000000..83ef741ca
--- /dev/null
+++ b/rl_engine/backends/cuda/rope/multi_axis_rope.py
@@ -0,0 +1,49 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Architecture-independent CUDA implementation; no Hopper-only instructions."""
+
+import importlib.util
+
+import torch
+
+from rl_engine.backends import extension  # noqa: F401 - loads the normal native library
+from rl_engine.ops.rope._registration import register_rotation
+from rl_engine.ops.rope.multi_axis_rope import MultiAxisRoPEMetadata, validate_inputs
+
+
+def _apply(x, cos, sin, transpose):
+    return torch.ops.rl_kernel_native.multi_axis_rope(x, cos, sin, transpose)
+
+
+def native_available():
+    if torch.version.hip is not None:
+        return False
+    if not hasattr(torch.ops.rl_kernel_native, "multi_axis_rope"):
+        library = importlib.util.find_spec("rl_engine._rope_cuda")
+        if library is not None:
+            torch.ops.load_library(library.origin)
+    return hasattr(torch.ops.rl_kernel_native, "multi_axis_rope")
+
+
+_rotation = register_rotation(
+    "qwen_image_rope_cuda",
+    _apply,
+    kernel_id="torch.ops.rl_kernel_native.multi_axis_rope",
+    family="cuda",
+)
+
+
+class MultiAxisRoPECudaOp(MultiAxisRoPEMetadata):
+    backend_id = "rlkernel.cuda.qwenimage_multi_axis_rope.v1"
+    kernel_id = "torch.ops.rl_kernel_native.multi_axis_rope"
+
+    def __init__(self):
+        if not native_available():
+            raise RuntimeError(
+                "CUDA multi_axis_rope extension is unavailable; rebuild rl_engine._C"
+            )
+
+    def __call__(self, x, cos, sin):
+        validate_inputs(x, cos, sin)
+        if x.device.type != "cuda":
+            raise ValueError("CUDA multi_axis_rope requires CUDA tensors")
+        return _rotation(x, cos, sin, False)
diff --git a/rl_engine/backends/shared/triton/rope/FLASH_ATTENTION_LICENSE b/rl_engine/backends/shared/triton/rope/FLASH_ATTENTION_LICENSE
new file mode 100644
index 000000000..5860e4b33
--- /dev/null
+++ b/rl_engine/backends/shared/triton/rope/FLASH_ATTENTION_LICENSE
@@ -0,0 +1,29 @@
+BSD 3-Clause License
+
+Copyright (c) 2022, the respective contributors, as shown by the AUTHORS file.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+* Redistributions of source code must retain the above copyright notice, this
+  list of conditions and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright notice,
+  this list of conditions and the following disclaimer in the documentation
+  and/or other materials provided with the distribution.
+
+* Neither the name of the copyright holder nor the names of its
+  contributors may be used to endorse or promote products derived from
+  this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py b/rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py
new file mode 100644
index 000000000..bd617c8c1
--- /dev/null
+++ b/rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py
@@ -0,0 +1,141 @@
+# SPDX-License-Identifier: BSD-3-Clause
+# Copyright (c) 2025, Tri Dao.
+# Adapted from Dao-AILab/flash-attention, revision
+# 98c1cee0ed59307833841cdbf567fe336ceec13a, flash_attn/ops/triton/rotary.py.
+# Changes: explicit shared/per-sample table strides. Launch policy is in the
+# sibling wrapper. See FLASH_ATTENTION_LICENSE for the retained BSD-3 terms.
+
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def rotary_kernel(
+    OUT,  # Pointers to matrices
+    X,
+    COS,
+    SIN,
+    CU_SEQLENS,
+    SEQLEN_OFFSETS,  # this could be int or a pointer
+    # Matrix dimensions
+    seqlen,
+    nheads,
+    seqlen_ro,
+    # strides
+    stride_out_batch,
+    stride_out_seqlen,
+    stride_out_nheads,
+    stride_out_headdim,
+    stride_x_batch,
+    stride_x_seqlen,
+    stride_x_nheads,
+    stride_x_headdim,
+    stride_cos_batch,
+    stride_cos_seqlen,
+    stride_cos_dim,
+    stride_sin_batch,
+    stride_sin_seqlen,
+    stride_sin_dim,
+    # Meta-parameters
+    # We want ROTARY_DIM to be constexpr, otherwise the triton compiler doesn't know that
+    # the mask is constant every 8 elements, and it will generate LDG.16 instead of LDG.128
+    ROTARY_DIM: tl.constexpr,
+    IS_SEQLEN_OFFSETS_TENSOR: tl.constexpr,
+    IS_VARLEN: tl.constexpr,
+    INTERLEAVED: tl.constexpr,
+    CONJUGATE: tl.constexpr,
+    BLOCK_H: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+):
+    BLOCK_K: tl.constexpr = triton.next_power_of_2(ROTARY_DIM)
+    ROTARY_DIM_HALF = ROTARY_DIM // 2
+    pid_head = tl.program_id(axis=0)
+    pid_m = tl.program_id(axis=1)
+    pid_batch = tl.program_id(axis=2)
+
+    if not IS_VARLEN:
+        X = X + pid_batch * stride_x_batch
+        OUT = OUT + pid_batch * stride_out_batch
+    else:
+        start_idx = tl.load(CU_SEQLENS + pid_batch)
+        seqlen = tl.load(CU_SEQLENS + pid_batch + 1) - start_idx
+        X = X + start_idx * stride_x_seqlen
+        OUT = OUT + start_idx * stride_out_seqlen
+
+    if pid_m * BLOCK_M >= seqlen:
+        return
+
+    rh = pid_head * BLOCK_H + tl.arange(0, BLOCK_H)
+    rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    if not IS_SEQLEN_OFFSETS_TENSOR:
+        rm_cs = rm + SEQLEN_OFFSETS
+    else:
+        rm_cs = rm + tl.load(SEQLEN_OFFSETS + pid_batch)
+
+    rk_half = tl.arange(0, BLOCK_K // 2)
+    COS = (
+        COS
+        + pid_batch * stride_cos_batch
+        + (rm_cs[:, None] * stride_cos_seqlen + rk_half[None, :] * stride_cos_dim)
+    )
+    SIN = (
+        SIN
+        + pid_batch * stride_sin_batch
+        + (rm_cs[:, None] * stride_sin_seqlen + rk_half[None, :] * stride_sin_dim)
+    )
+    mask_cs = (rm_cs[:, None] < seqlen_ro) & (rk_half[None, :] < ROTARY_DIM_HALF)
+    cos = tl.load(COS, mask=mask_cs, other=1.0).to(tl.float32)
+    sin = tl.load(SIN, mask=mask_cs, other=0.0).to(tl.float32)
+    if CONJUGATE:
+        sin = -sin
+
+    if not INTERLEAVED:
+        # Load the 1st and 2nd halves of X, do calculation, then store to 1st and 2nd halves of OUT
+        X = X + (
+            rh[:, None, None] * stride_x_nheads
+            + rm[None, :, None] * stride_x_seqlen
+            + rk_half[None, None, :] * stride_x_headdim
+        )
+        OUT = OUT + (
+            rh[:, None, None] * stride_out_nheads
+            + rm[None, :, None] * stride_out_seqlen
+            + rk_half[None, None, :] * stride_out_headdim
+        )
+        mask = (
+            (rh[:, None, None] < nheads)
+            & (rm[None, :, None] < seqlen)
+            & (rk_half[None, None, :] < ROTARY_DIM_HALF)
+        )
+        x0 = tl.load(X, mask=mask, other=0.0).to(tl.float32)
+        x1 = tl.load(
+            X + ROTARY_DIM_HALF * stride_x_headdim,
+            mask=mask,
+            other=0.0,
+        ).to(tl.float32)
+        o0 = x0 * cos - x1 * sin
+        o1 = x0 * sin + x1 * cos
+        tl.store(OUT, o0, mask=mask)
+        tl.store(OUT + ROTARY_DIM_HALF * stride_out_headdim, o1, mask=mask)
+    else:
+        rk = tl.arange(0, BLOCK_K)
+        X = X + (
+            rh[:, None, None] * stride_x_nheads
+            + rm[None, :, None] * stride_x_seqlen
+            + rk[None, None, :] * stride_x_headdim
+        )
+        OUT = OUT + (
+            rh[:, None, None] * stride_out_nheads
+            + rm[None, :, None] * stride_out_seqlen
+            + rk[None, None, :] * stride_out_headdim
+        )
+        mask = (
+            (rh[:, None, None] < nheads)
+            & (rm[None, :, None] < seqlen)
+            & (rk[None, None, :] < ROTARY_DIM)
+        )
+        x = tl.load(X, mask=mask, other=0.0).to(tl.float32)
+        x0, x1 = tl.split(tl.reshape(x, [BLOCK_H, BLOCK_M, BLOCK_K // 2, 2]))
+        o0 = x0 * cos - x1 * sin
+        o1 = x0 * sin + x1 * cos
+        o = tl.reshape(tl.join(o0, o1), [BLOCK_H, BLOCK_M, BLOCK_K])
+        tl.store(OUT, o, mask=mask)
diff --git a/rl_engine/backends/shared/triton/rope/multi_axis_rope.py b/rl_engine/backends/shared/triton/rope/multi_axis_rope.py
new file mode 100644
index 000000000..1a37836fe
--- /dev/null
+++ b/rl_engine/backends/shared/triton/rope/multi_axis_rope.py
@@ -0,0 +1,68 @@
+# SPDX-License-Identifier: Apache-2.0
+"""FlashAttention-derived, stride-aware rotation under a frozen FP32 profile."""
+
+import torch
+import triton
+
+from rl_engine.ops.rope._registration import register_rotation
+from rl_engine.ops.rope.multi_axis_rope import MultiAxisRoPEMetadata, validate_inputs
+
+from ._flash_attn_rotary import rotary_kernel
+
+
+def _apply(x, cos, sin, transpose):
+    output = torch.empty(x.shape, device=x.device, dtype=x.dtype)
+    batch, seq, heads, _ = x.shape
+    if batch and seq:
+        with torch.cuda.device(x.device):
+            rotary_kernel[(triton.cdiv(heads, 2), triton.cdiv(seq, 8), batch)](
+                output,
+                x,
+                cos,
+                sin,
+                None,
+                0,
+                seq,
+                heads,
+                seq,
+                *output.stride(),
+                *x.stride(),
+                cos.stride(0) if cos.ndim == 3 else 0,
+                cos.stride(-2),
+                cos.stride(-1),
+                sin.stride(0) if sin.ndim == 3 else 0,
+                sin.stride(-2),
+                sin.stride(-1),
+                128,
+                False,
+                False,
+                True,
+                transpose,
+                BLOCK_H=2,
+                BLOCK_M=8,
+                enable_fp_fusion=False,
+            )
+    return output
+
+
+_rotation = register_rotation(
+    "qwen_image_rope_triton",
+    _apply,
+    kernel_id="rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
+    family="triton",
+)
+
+
+class MultiAxisRoPETritonOp(MultiAxisRoPEMetadata):
+    backend_id = "rlkernel.triton.qwenimage_multi_axis_rope.v1"
+    kernel_id = "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel"
+
+    def __init__(self):
+        if torch.version.hip is not None or not torch.cuda.is_available():
+            raise RuntimeError("this multi_axis_rope Triton profile requires NVIDIA CUDA")
+
+    def __call__(self, x, cos, sin):
+        validate_inputs(x, cos, sin)
+        if x.device.type != "cuda":
+            raise ValueError("Triton multi_axis_rope requires CUDA tensors")
+        return _rotation(x, cos, sin, False)
diff --git a/rl_engine/ops/rope/_registration.py b/rl_engine/ops/rope/_registration.py
new file mode 100644
index 000000000..343d41a9c
--- /dev/null
+++ b/rl_engine/ops/rope/_registration.py
@@ -0,0 +1,41 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Dispatcher, fake and recursive input-VJP registration for the RoPE profile."""
+
+import torch
+
+from rl_engine.ops.autograd.backward_runtime import record_backward
+from rl_engine.ops.rope.multi_axis_rope import validate_inputs
+
+
+def register_rotation(name, implementation, *, kernel_id, family):
+    @torch.library.custom_op(f"rl_kernel::{name}", mutates_args=(), device_types="cuda")
+    def rotation(
+        x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor, transpose: bool
+    ) -> torch.Tensor:
+        validate_inputs(x, cos, sin)
+        result = implementation(x, cos, sin, transpose)
+        if transpose:
+            record_backward(
+                "multi_axis_rope",
+                kernel_id=kernel_id,
+                impl="interleaved_fp32_transpose",
+                family=family,
+            )
+        return result
+
+    @rotation.register_fake
+    def fake(x, cos, sin, transpose):
+        validate_inputs(x, cos, sin)
+        return torch.empty(x.shape, dtype=x.dtype, device=x.device)
+
+    def setup_context(ctx, inputs, output):
+        _, cos, sin, transpose = inputs
+        ctx.save_for_backward(cos, sin)
+        ctx.transpose = transpose
+
+    def backward(ctx, grad):
+        cos, sin = ctx.saved_tensors
+        return rotation(grad, cos, sin, not ctx.transpose), None, None, None
+
+    rotation.register_autograd(backward, setup_context=setup_context)
+    return rotation
diff --git a/rl_engine/ops/rope/multi_axis_rope.py b/rl_engine/ops/rope/multi_axis_rope.py
new file mode 100644
index 000000000..ff95adff2
--- /dev/null
+++ b/rl_engine/ops/rope/multi_axis_rope.py
@@ -0,0 +1,132 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Qwen-Image interleaved three-axis RoPE (RFC #386).
+
+Tables are immutable inputs, built before capture/replay. The rotation accepts
+BSHD tensors, unlike the legacy BHSD, rotate-half RoPE operator.
+"""
+
+from __future__ import annotations
+
+import torch
+
+CONTRACT = "qwenimage-interleaved-rope-fp32-v1"
+AXES_DIMS = (16, 56, 56)
+UPSTREAM_REVISION = "40f63690da440e8c382ab53e67508835deb1ed2f"
+
+
+def qwen_image_positions(
+    height: int, width: int, text_length: int
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """CPU int64 positions for one image's *packed latent* H/W, then text.
+
+    Matches QwenEmbedRope(scale_rope=True), single-frame text-to-image. Odd
+    grids have the extra negative coordinate: H=3 gives [-2, -1, 0].
+    Return separate tables so callers explicitly choose their stream order.
+    """
+    for name, value in (
+        ("height", height),
+        ("width", width),
+        ("text_length", text_length),
+    ):
+        if isinstance(value, bool) or not isinstance(value, int):
+            raise TypeError(f"{name} must be an integer")
+    if not 1 <= height <= 8192 or not 1 <= width <= 8192 or text_length < 0:
+        raise ValueError("height/width must be in [1, 8192], text_length must be nonnegative")
+    start = max(height // 2, width // 2)
+    if start + text_length > 4096:
+        raise ValueError("text positions exceed the pinned upstream 4096-entry table")
+    h = torch.arange(-(height - height // 2), height // 2, device="cpu")
+    w = torch.arange(-(width - width // 2), width // 2, device="cpu")
+    hh, ww = torch.meshgrid(h, w, indexing="ij")
+    image = torch.stack((torch.zeros_like(hh), hh, ww), dim=-1).reshape(-1, 3)
+    text = (
+        torch.arange(start, start + text_length, device="cpu")[:, None].expand(-1, 3).contiguous()
+    )
+    return image, text
+
+
+def build_multi_axis_cos_sin(
+    positions: torch.Tensor, *, device=None
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Build canonical FP32 tables on CPU, then copy to the requested device.
+
+    positions: int32/int64 [S, 3] or [B, S, 3], range [-4096, 4095].
+    The CPU torch.polar construction follows the pinned Diffusers reference;
+    no GPU transcendental or per-batch table recomputation enters the hot path.
+    """
+    if positions.dtype not in (torch.int32, torch.int64):
+        raise TypeError("positions must be int32 or int64")
+    if positions.ndim not in (2, 3) or positions.shape[-1] != 3:
+        raise ValueError("positions must be [S, 3] or [B, S, 3]")
+    cpu = positions.detach().to(device="cpu", dtype=torch.float32)
+    if torch.any((cpu < -4096) | (cpu >= 4096)):
+        raise ValueError("positions exceed the pinned upstream [-4096, 4095] range")
+    phases = []
+    for axis, dim in enumerate(AXES_DIMS):
+        inv_freq = 1.0 / torch.pow(10000, torch.arange(0, dim, 2, device="cpu").float().div(dim))
+        phases.append(cpu[..., axis, None] * inv_freq)
+    phase = torch.cat(phases, dim=-1)
+    cis = torch.polar(torch.ones_like(phase), phase)
+    target = positions.device if device is None else device
+    return cis.real.contiguous().to(target), cis.imag.contiguous().to(target)
+
+
+def validate_inputs(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> None:
+    if x.ndim != 4 or x.shape[-1] != 128 or x.shape[2] <= 0:
+        raise ValueError("x must be [B, S, H, 128] with H > 0")
+    if x.dtype not in (torch.float16, torch.bfloat16, torch.float32):
+        raise TypeError("x must be float16, bfloat16 or float32")
+    shared = (x.shape[1], 64)
+    batched = (x.shape[0], x.shape[1], 64)
+    if cos.shape not in (shared, batched) or sin.shape != cos.shape:
+        raise ValueError("cos/sin must have matching [S, 64] or [B, S, 64] shapes")
+    if cos.dtype != torch.float32 or sin.dtype != torch.float32:
+        raise TypeError("cos/sin must be float32")
+    if x.device != cos.device or x.device != sin.device:
+        raise ValueError("x, cos and sin must be on the same device")
+    if cos.requires_grad or sin.requires_grad:
+        raise ValueError("cos/sin are fixed tables; table gradients are unsupported")
+
+
+def get_multi_axis_rope_op(backend: str = "auto", *, device="cpu"):
+    """Resolve an op exposing trace(); explicit backends never silently fall back.
+
+    Strict callers must request cuda or triton and retain the returned trace.
+    auto is a diagnostic/convenience mode with a PyTorch reference fallback.
+    """
+    if backend == "auto":
+        from rl_engine.runtime.registry import kernel_registry
+
+        return kernel_registry.get_op("multi_axis_rope", device=device)
+    if backend == "cuda":
+        from rl_engine.backends.cuda.rope.multi_axis_rope import MultiAxisRoPECudaOp
+
+        return MultiAxisRoPECudaOp()
+    if backend == "triton":
+        from rl_engine.backends.shared.triton.rope.multi_axis_rope import MultiAxisRoPETritonOp
+
+        return MultiAxisRoPETritonOp()
+    if backend == "reference":
+        from rl_engine.reference.rope.multi_axis_rope import NativeMultiAxisRoPEOp
+
+        return NativeMultiAxisRoPEOp()
+    raise ValueError(f"unknown multi_axis_rope backend: {backend!r}")
+
+
+def multi_axis_rope(x, cos, sin, *, backend="auto"):
+    return get_multi_axis_rope_op(backend, device=x.device)(x, cos, sin)
+
+
+class MultiAxisRoPEMetadata:
+    backend_id: str
+    kernel_id: str
+
+    def trace(self) -> dict:
+        return {
+            "contract": CONTRACT,
+            "actual_backend": self.backend_id,
+            "kernel_id": self.kernel_id,
+            "layout": "BSHD, adjacent channel pairs",
+            "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+            "table_revision": UPSTREAM_REVISION,
+        }
diff --git a/rl_engine/reference/rope/multi_axis_rope.py b/rl_engine/reference/rope/multi_axis_rope.py
new file mode 100644
index 000000000..353eb6a00
--- /dev/null
+++ b/rl_engine/reference/rope/multi_axis_rope.py
@@ -0,0 +1,25 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Explicit FP32 reference for the Qwen-Image interleaved rotation."""
+
+import torch
+
+from rl_engine.ops.rope.multi_axis_rope import MultiAxisRoPEMetadata, validate_inputs
+
+
+class NativeMultiAxisRoPEOp(MultiAxisRoPEMetadata):
+    op_class = "elementwise"
+    backend_id = "rlkernel.reference.qwenimage_multi_axis_rope.v1"
+    kernel_id = "rl_engine.reference.rope.multi_axis_rope.NativeMultiAxisRoPEOp"
+
+    def __call__(self, x, cos, sin):
+        return self.forward(x, cos, sin)
+
+    def forward(self, x, cos, sin):
+        return self.forward_fp32(x, cos, sin).to(x.dtype)
+
+    def forward_fp32(self, x, cos, sin):
+        validate_inputs(x, cos, sin)
+        c, s = cos.unsqueeze(-2), sin.unsqueeze(-2)
+        even, odd = x[..., 0::2].float(), x[..., 1::2].float()
+        out = torch.stack((even * c - odd * s, odd * c + even * s), dim=-1)
+        return out.flatten(-2)
diff --git a/rl_engine/runtime/registry.py b/rl_engine/runtime/registry.py
index 7751d5901..326081f44 100644
--- a/rl_engine/runtime/registry.py
+++ b/rl_engine/runtime/registry.py
@@ -167,6 +167,11 @@ class OpBackend(Enum, metaclass=_KernelEnumMeta):
     PYTORCH_NATIVE_ROPE = "rl_engine.kernels.ops.pytorch.rotary_embedding.rope.NativeRoPEOp"
     TRITON_ROPE = "rl_engine.kernels.ops.triton.rotary_embedding.rope.TritonRoPEOp"
     CUDA_ROPE_SM90 = "rl_engine.kernels.ops.cuda.rotary_embedding.rope.RoPESM90Op"
+    CUDA_MULTI_AXIS_ROPE = "rl_engine.backends.cuda.rope.multi_axis_rope.MultiAxisRoPECudaOp"
+    TRITON_MULTI_AXIS_ROPE = (
+        "rl_engine.backends.shared.triton.rope.multi_axis_rope.MultiAxisRoPETritonOp"
+    )
+    PYTORCH_MULTI_AXIS_ROPE = "rl_engine.reference.rope.multi_axis_rope.NativeMultiAxisRoPEOp"
     ASCEND_ROPE = "rl_engine.kernels.ops.ascend.rotary_embedding.rope.RoPEAscendOp"
     PYTORCH_NATIVE_SILU = "rl_engine.kernels.ops.pytorch.activation.swiglu.NativeSiLUOp"
     PYTORCH_NATIVE_SWIGLU = "rl_engine.kernels.ops.pytorch.activation.swiglu.NativeSwiGLUOp"
@@ -737,6 +742,15 @@ def __init__(self):
                 ],
             },
         }
+        # CUDA qualification for the new interleaved Qwen-Image contract.
+        # Other platforms retain an explicit diagnostic reference, not logp.
+        for candidates in self._priority_map.values():
+            candidates["multi_axis_rope"] = [OpBackend.PYTORCH_MULTI_AXIS_ROPE]
+        self._priority_map["cuda"]["multi_axis_rope"] = [
+            OpBackend.CUDA_MULTI_AXIS_ROPE,
+            OpBackend.TRITON_MULTI_AXIS_ROPE,
+            OpBackend.PYTORCH_MULTI_AXIS_ROPE,
+        ]
         # Preserve the former CPU fallback behavior for every operator on NPU,
         # then override only the operators with an Ascend-specific backend.
         self._priority_map["npu"] = {
diff --git a/rl_engine/validation/operators/gradient_adapters.py b/rl_engine/validation/operators/gradient_adapters.py
index f22d80285..6bc6e8104 100644
--- a/rl_engine/validation/operators/gradient_adapters.py
+++ b/rl_engine/validation/operators/gradient_adapters.py
@@ -220,6 +220,21 @@ def to_dict(self) -> dict[str, Any]:
             "csrc/cuda/logprob/fused_linear_logp_sm90.cu",
         ),
     ),
+    # Optional Qwen-Image extension: do not certify it as a Dense Qwen3 chain node.
+    "multi_axis_rope": GradientAdapterSpec(
+        op_name="multi_axis_rope",
+        chain_node="multi_axis_rope",
+        op_class="elementwise",
+        spec_name="multi_axis_rope",
+        tensors=(_DX,),
+        requirement="optional_fused",
+        source_files=(
+            "rl_engine/backends/cuda/rope/multi_axis_rope.py",
+            "rl_engine/backends/shared/triton/rope/multi_axis_rope.py",
+            "rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py",
+            "csrc/cuda/rope/multi_axis_rope.cu",
+        ),
+    ),
     "rope": GradientAdapterSpec(
         op_name="rope",
         chain_node="rope",
@@ -712,6 +727,16 @@ def _row_inputs(
             "weight": params["weight"],
             "bias": None,
         }
+    if op_name == "multi_axis_rope":
+        from rl_engine.ops.rope.multi_axis_rope import build_multi_axis_cos_sin
+
+        if head_dim != 128:
+            raise ValueError("multi_axis_rope gradient adapter requires head_dim=128")
+        rows = _stack_rows(keys, leading, (n_heads, head_dim), device=device, dtype=dtype)
+        positions = _row_positions(keys, device=torch.device("cpu"))
+        coords = torch.stack((positions % 7, positions % 127 - 63, positions % 61 - 30), dim=-1)
+        cos, sin = build_multi_axis_cos_sin(coords, device=device)
+        return {"x": rows.unsqueeze(0), "cos": cos, "sin": sin}
     if op_name == "rope":
         rows = _stack_rows(keys, leading, (n_heads, head_dim), device=device, dtype=dtype)
         return {
@@ -724,6 +749,10 @@ def _row_inputs(
 
 def _to_rows(op_name: str, value: torch.Tensor, n_rows: int) -> torch.Tensor:
     """Normalize an operator output / input-grad back to (n_rows, *tail)."""
+    if op_name == "multi_axis_rope":
+        if value.shape[0] != 1 or value.shape[1] != n_rows:
+            raise ValueError("multi_axis_rope row adapter expects [1, tokens, heads, 128]")
+        return value.squeeze(0)
     if op_name == "rope":
         # RoPE runs as (1, heads, tokens, head_dim); tokens is the row axis.
         permuted = value.permute(0, 2, 1, 3)
diff --git a/rl_engine/validation/operators/operator_inputs.py b/rl_engine/validation/operators/operator_inputs.py
index ca3b7c120..137d8d69b 100644
--- a/rl_engine/validation/operators/operator_inputs.py
+++ b/rl_engine/validation/operators/operator_inputs.py
@@ -37,6 +37,7 @@ def make_operator_inputs(
         "linear_logp": _make_linear_logp_inputs,
         "batch_invariant_logp": _make_batch_invariant_logp_inputs,
         "rope": _make_rope_inputs,
+        "multi_axis_rope": _make_multi_axis_rope_inputs,
         "silu": _make_silu_inputs,
         "swiglu": _make_swiglu_inputs,
         "embedding": _make_embedding_inputs,
@@ -67,6 +68,7 @@ def operator_shape_name(op_name: str, args: argparse.Namespace) -> str:
         "linear_logp": f"{batch}x{seq}x{_normalized_dim(args)}x{vocab}",
         "batch_invariant_logp": f"{batch}x{seq}x{vocab}",
         "rope": f"{batch}x{DEFAULT_N_HEADS}x{seq}x{DEFAULT_HEAD_DIM}",
+        "multi_axis_rope": f"{batch}x{seq}x{_arg_int(args, 'n_heads', 24)}x128",
         "silu": f"{batch}x{seq}x{DEFAULT_INTERMEDIATE}",
         "swiglu": f"{batch}x{seq}x{DEFAULT_INTERMEDIATE}",
         "embedding": f"{batch}x{seq}x{vocab}x{_normalized_dim(args)}",
@@ -250,6 +252,23 @@ def _make_batch_invariant_logp_inputs(
     }
 
 
+def _make_multi_axis_rope_inputs(args, dtype, device):
+    from rl_engine.ops.rope.multi_axis_rope import build_multi_axis_cos_sin
+
+    batch, seq = _batch_seq(args)
+    if _arg_int(args, "head_dim", 128) != 128:
+        raise ValueError("multi_axis_rope requires head_dim=128")
+    heads = _arg_int(args, "n_heads", 24)
+    row = torch.arange(seq, device="cpu", dtype=torch.int64)
+    positions = torch.stack((row % 7, (row // 127) % 4096, row % 127 - 63), dim=-1)
+    cos, sin = build_multi_axis_cos_sin(positions, device=device)
+    return {
+        "x": _floating_tensor((batch, seq, heads, 128), args, dtype, device, offset=0),
+        "cos": cos,
+        "sin": sin,
+    }
+
+
 def _make_rope_inputs(
     args: argparse.Namespace, dtype: torch.dtype, device: torch.device
 ) -> dict[str, Any]:
diff --git a/rl_engine/validation/operators/operator_specs.py b/rl_engine/validation/operators/operator_specs.py
index 2323f4787..bc23ab28a 100644
--- a/rl_engine/validation/operators/operator_specs.py
+++ b/rl_engine/validation/operators/operator_specs.py
@@ -197,6 +197,20 @@ def _load_object(path: str) -> Any:
         },
         grad_input_names=("a", "b"),
     ),
+    "multi_axis_rope": OperatorSpec(
+        name="multi_axis_rope",
+        op_class="elementwise",
+        gold_path="rl_engine.reference.rope.multi_axis_rope.NativeMultiAxisRoPEOp",
+        gold_method="forward_fp32",
+        candidate_paths={
+            "pytorch": "rl_engine.reference.rope.multi_axis_rope.NativeMultiAxisRoPEOp",
+            "cuda": "rl_engine.backends.cuda.rope.multi_axis_rope.MultiAxisRoPECudaOp",
+            "triton": (
+                "rl_engine.backends.shared.triton.rope.multi_axis_rope.MultiAxisRoPETritonOp"
+            ),
+        },
+        grad_input_names=("x",),
+    ),
     "rope": OperatorSpec(
         name="rope",
         op_class="elementwise",
diff --git a/rl_engine/validation/operators/qwen_image_rope.py b/rl_engine/validation/operators/qwen_image_rope.py
new file mode 100644
index 000000000..bc605bd32
--- /dev/null
+++ b/rl_engine/validation/operators/qwen_image_rope.py
@@ -0,0 +1,80 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Qwen-Image operator-only four-judgment gate using the standard C3/C4 APIs.
+
+Reuses C2's logical sample/config transforms and C1's elementwise tolerances.
+It does not certify the Dense Qwen3 chain, distributed execution or Qwen-Image
+model assembly. The operator has its own OP_SPECS entry and gradient adapter.
+"""
+
+import torch
+
+from rl_engine.contracts.numerical import BackendProvenance
+from rl_engine.ops.rope.multi_axis_rope import get_multi_axis_rope_op
+from rl_engine.reference.rope.multi_axis_rope import NativeMultiAxisRoPEOp
+from rl_engine.validation.operators.forward_invariance import assert_forward_batch_invariant
+from rl_engine.validation.operators.gradient_adapters import (
+    get_adapter,
+    make_forward_runner,
+    make_gradient_runner,
+)
+from rl_engine.validation.operators.gradient_invariance import assert_gradient_batch_invariant
+
+
+def run_multi_axis_rope_ws1(backend):
+    if backend not in ("cuda", "triton"):
+        raise ValueError("WS1 GPU evidence requires an explicit cuda or triton backend")
+    device = torch.device("cuda")
+    profile = "cuda_bf16" if backend == "cuda" else "triton_cuda_bf16"
+    op = get_multi_axis_rope_op(backend, device=device)
+    gold = NativeMultiAxisRoPEOp().forward_fp32
+    trace = op.trace()
+    dtype = torch.bfloat16
+    provenance = BackendProvenance(
+        backend_profile=profile,
+        requested_backend=backend,
+        actual_backend=backend,
+        execution_dtype="bfloat16",
+        accumulation_dtype="float32",
+        output_dtype="bfloat16",
+        reference_dtype="float32",
+        candidate_tf32_enabled=False,
+        reference_tf32_enabled=False,
+    )
+    metadata = {
+        "backend_profile": profile,
+        "provenance": provenance,
+        "op_class": "elementwise",
+        "dtype": dtype,
+        "op_name": "multi_axis_rope",
+        "candidate_id": trace["actual_backend"],
+        "device": str(device),
+        "compute_capability": "sm%d%d" % torch.cuda.get_device_capability(),
+        "observed_actual_backend": backend,
+        "observed_kernel_id": trace["kernel_id"],
+        "observed_output_dtype": "bfloat16",
+    }
+    runner_options = {"device": device, "dtype": dtype, "n_heads": 3, "head_dim": 128}
+    candidate_options = {"backend_family": backend, "kernel_id": trace["kernel_id"]}
+    forward = assert_forward_batch_invariant(
+        make_forward_runner(
+            "multi_axis_rope", op, reference=False, **runner_options, **candidate_options
+        ),
+        gold_fn=make_forward_runner("multi_axis_rope", gold, reference=True, **runner_options),
+        include_logprob_smoke=False,
+        **metadata,
+    )
+    gradient = assert_gradient_batch_invariant(
+        make_gradient_runner(
+            "multi_axis_rope", op, reference=False, **runner_options, **candidate_options
+        ),
+        gold_fn=make_gradient_runner("multi_axis_rope", gold, reference=True, **runner_options),
+        grad_tensors=get_adapter("multi_axis_rope").tensors,
+        **metadata,
+    )
+    return {
+        "scope": "Qwen-Image multi_axis_rope operator; C1 elementwise and C2 config transforms",
+        "trace": trace,
+        "passed": forward.passed and gradient.passed,
+        "forward": forward.to_dict(),
+        "gradient": gradient.to_dict(),
+    }
diff --git a/tests/build/test_multi_axis_rope_build.py b/tests/build/test_multi_axis_rope_build.py
new file mode 100644
index 000000000..c9d06791e
--- /dev/null
+++ b/tests/build/test_multi_axis_rope_build.py
@@ -0,0 +1,46 @@
+# SPDX-License-Identifier: Apache-2.0
+"""The explicit operator wheel profile must build native code or fail closed."""
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+
+from build_tools import extensions
+
+
+def test_rope_only_wheel_selects_production_sources(monkeypatch):
+    monkeypatch.setenv("RL_KERNEL_ROPE_ONLY", "1")
+    fake_torch = SimpleNamespace(
+        __file__=torch.__file__, version=SimpleNamespace(cuda="12.8", hip=None)
+    )
+    monkeypatch.setattr(
+        extensions,
+        "_load_torch_extension_tools",
+        lambda: (fake_torch, None, lambda **kwargs: kwargs),
+    )
+    (extension,) = extensions.get_extensions()
+    assert extension["name"] == "rl_engine._rope_cuda"
+    assert set(extension["sources"]) == {
+        "csrc/bindings/rope_module.cpp",
+        "csrc/cuda/rope/multi_axis_rope.cu",
+    }
+    assert "--fmad=false" in extension["extra_compile_args"]["nvcc"]
+
+
+@pytest.mark.parametrize("hip", [None, "6.2"])
+def test_rope_only_wheel_rejects_non_cuda_builds(monkeypatch, hip):
+    monkeypatch.setenv("RL_KERNEL_ROPE_ONLY", "1")
+    fake_torch = SimpleNamespace(
+        __file__=torch.__file__, version=SimpleNamespace(cuda=None, hip=hip)
+    )
+    monkeypatch.setattr(extensions, "_load_torch_extension_tools", lambda: (fake_torch, None, None))
+    with pytest.raises(RuntimeError, match="requires CUDA-enabled PyTorch"):
+        extensions.get_extensions()
+
+
+def test_rope_only_wheel_rejects_missing_torch(monkeypatch):
+    monkeypatch.setenv("RL_KERNEL_ROPE_ONLY", "1")
+    monkeypatch.setattr(extensions, "_load_torch_extension_tools", lambda: (None, None, None))
+    with pytest.raises(RuntimeError, match="PyTorch is unavailable"):
+        extensions.get_extensions()
diff --git a/tests/ops/rope/test_multi_axis_rope.py b/tests/ops/rope/test_multi_axis_rope.py
new file mode 100644
index 000000000..6cbcb2b85
--- /dev/null
+++ b/tests/ops/rope/test_multi_axis_rope.py
@@ -0,0 +1,443 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Accuracy is independent FP64 complex algebra; invariance compares raw bytes."""
+
+import gc
+import math
+
+import pytest
+import torch
+
+from rl_engine.ops.rope.multi_axis_rope import (
+    build_multi_axis_cos_sin,
+    get_multi_axis_rope_op,
+    multi_axis_rope,
+    qwen_image_positions,
+)
+
+
+def assert_bits(a, b):
+    assert a.dtype == b.dtype and a.shape == b.shape
+    assert torch.equal(a.contiguous().view(torch.uint8), b.contiguous().view(torch.uint8))
+
+
+def complex_reference(x, cos, sin, *, transpose=False):
+    pairs = x.double().reshape(*x.shape[:-1], 64, 2)
+    z = torch.view_as_complex(pairs.contiguous())
+    phase = torch.complex(cos.double(), sin.double()).unsqueeze(-2)
+    if transpose:
+        phase = phase.conj()
+    return torch.view_as_real(z * phase).flatten(-2)
+
+
+@pytest.fixture(params=["reference", "cuda", "triton"])
+def implementation(request):
+    backend = request.param
+    device = "cpu" if backend == "reference" else "cuda"
+    if device == "cuda" and (not torch.cuda.is_available() or torch.version.hip is not None):
+        pytest.skip("NVIDIA CUDA required")
+    if backend == "cuda":
+        from rl_engine.backends.cuda.rope.multi_axis_rope import native_available
+
+        if not native_available():
+            pytest.skip("build native multi_axis_rope or use the focused CUDA validation runner")
+    if backend == "triton":
+        pytest.importorskip("triton")
+    return get_multi_axis_rope_op(backend, device=device), device
+
+
+def inputs(device, dtype=torch.float32, *, batch=2, seq=7, heads=3, batched=False):
+    generator = torch.Generator().manual_seed(386)
+    x = torch.randn(batch, seq, heads, 128, generator=generator).to(device, dtype)
+    shape = (batch, seq, 3) if batched else (seq, 3)
+    pos = torch.randint(-42, 53, shape, generator=generator)
+    cos, sin = build_multi_axis_cos_sin(pos, device=device)
+    return x, cos, sin
+
+
+@pytest.mark.parametrize("height,width", [(2, 4), (3, 5), (64, 64), (83, 83), (104, 58)])
+def test_positions_and_axis_frequencies(height, width):
+    image, text = qwen_image_positions(height, width, 3)
+    # Explicit flattened endpoints and a row boundary catch swapped H/W and odd-grid shifts.
+    assert image[0].tolist() == [0, -((height + 1) // 2), -((width + 1) // 2)]
+    assert image[-1].tolist() == [0, height // 2 - 1, width // 2 - 1]
+    assert image[width].tolist() == [0, -((height + 1) // 2) + 1, -((width + 1) // 2)]
+    assert text[:, 0].tolist() == list(
+        range(max(height // 2, width // 2), max(height // 2, width // 2) + 3)
+    )
+    cos, sin = build_multi_axis_cos_sin(image[[0, -1]])
+    # Independent scalar FP64 trig oracle, per-axis exponents reset at 8 and 36 pairs.
+    for row, position in enumerate(image[[0, -1]].tolist()):
+        offset = 0
+        for axis, width_axis in enumerate((16, 56, 56)):
+            for pair in range(width_axis // 2):
+                angle = position[axis] / (10000.0 ** (2 * pair / width_axis))
+                assert abs(cos[row, offset + pair].item() - math.cos(angle)) < 4e-6
+                assert abs(sin[row, offset + pair].item() - math.sin(angle)) < 4e-6
+            offset += width_axis // 2
+
+
+@pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16])
+@pytest.mark.parametrize("batched", [False, True])
+def test_forward_backward_accuracy(implementation, dtype, batched):
+    op, device = implementation
+    x, cos, sin = inputs(device, dtype, batched=batched)
+    x.requires_grad_()
+    grad = torch.randn_like(x)
+    result = op(x, cos, sin)
+    dx = torch.autograd.grad(result, x, grad)[0]
+    tolerance = {torch.float32: 8e-7, torch.float16: 0.002, torch.bfloat16: 0.02}[dtype]
+    torch.testing.assert_close(
+        result.double(), complex_reference(x, cos, sin), atol=tolerance, rtol=tolerance
+    )
+    torch.testing.assert_close(
+        dx.double(),
+        complex_reference(grad, cos, sin, transpose=True),
+        atol=tolerance,
+        rtol=tolerance,
+    )
+
+
+def test_adjacent_pair_layout_not_rotate_half(implementation):
+    op, device = implementation
+    x = torch.arange(128, dtype=torch.float32, device=device).reshape(1, 1, 1, 128)
+    cos = torch.zeros(1, 64, device=device)
+    sin = torch.ones_like(cos)
+    y = op(x, cos, sin)
+    assert_bits(y[..., 0::2], -x[..., 1::2])
+    assert_bits(y[..., 1::2], x[..., 0::2])
+
+
+def test_batch_position_padding_and_chunk_invariance(implementation):
+    op, device = implementation
+    x, cos, sin = inputs(device, batched=True, batch=3, seq=19)
+    x.requires_grad_()
+    grad = torch.randn_like(x)
+    y = op(x, cos, sin)
+    dx = torch.autograd.grad(y, x, grad)[0]
+    # Reordering samples with different positions must not reuse sample zero's table.
+    permutation = torch.tensor([2, 0, 1], device=device)
+    assert_bits(op(x[permutation], cos[permutation], sin[permutation]), y[permutation])
+    for b in range(3):
+        for lo, hi in [(0, 1), (1, 8), (8, 19)]:
+            xx = x[b : b + 1, lo:hi].detach().requires_grad_()
+            yy = op(xx, cos[b, lo:hi], sin[b, lo:hi])
+            dd = torch.autograd.grad(yy, xx, grad[b : b + 1, lo:hi])[0]
+            assert_bits(yy, y[b : b + 1, lo:hi])
+            assert_bits(dd, dx[b : b + 1, lo:hi])
+    padded = torch.cat((x.detach(), torch.randn(3, 5, 3, 128, device=device)), dim=1)
+    pc = torch.cat((cos, torch.ones(3, 5, 64, device=device)), dim=1)
+    ps = torch.cat((sin, torch.zeros(3, 5, 64, device=device)), dim=1)
+    assert_bits(op(padded, pc, ps)[:, :19], y)
+
+
+def test_noncontiguous_input_tables_and_gradient(implementation):
+    op, device = implementation
+    base, cos, sin = inputs(device, batched=True, seq=14, heads=4)
+    x = base[:, ::2, ::2].detach().requires_grad_()
+    c, s = cos[:, ::2], sin[:, ::2]
+    grad = torch.randn_like(base)[:, ::2, ::2]
+    result = op(x, c, s)
+    dx = torch.autograd.grad(result, x, grad)[0]
+    xx = x.detach().contiguous().requires_grad_()
+    yy = op(xx, c.contiguous(), s.contiguous())
+    dd = torch.autograd.grad(yy, xx, grad.contiguous())[0]
+    assert_bits(result, yy)
+    assert_bits(dx, dd)
+
+
+@pytest.mark.parametrize("batch,seq", [(0, 7), (2, 0), (0, 0)])
+def test_empty_forward_backward(implementation, batch, seq):
+    op, device = implementation
+    x, cos, sin = inputs(device, batch=batch, seq=seq)
+    x.requires_grad_()
+    result = op(x, cos, sin)
+    result.sum().backward()
+    assert result.shape == x.shape and x.grad.shape == x.shape
+
+
+def test_second_derivative(implementation):
+    op, device = implementation
+    x, cos, sin = inputs(device, batch=1, seq=2)
+    x.requires_grad_()
+    y = op(x, cos, sin)
+    first = torch.autograd.grad(y.square().sum(), x, create_graph=True)[0]
+    second = torch.autograd.grad(first.sum(), x)[0]
+    # R^T R = (cos^2 + sin^2) I, including the small table rounding error.
+    expected = (
+        (2 * (cos.double().square() + sin.double().square()))
+        .repeat_interleave(2, -1)[None, :, None, :]
+        .expand_as(x)
+    )
+    torch.testing.assert_close(second.double(), expected, atol=8e-7, rtol=8e-7)
+
+
+@pytest.mark.parametrize(
+    "problem", ["shape", "dtype", "table_shape", "table_dtype", "table_grad", "device"]
+)
+def test_reject_invalid_inputs(implementation, problem):
+    op, device = implementation
+    x, cos, sin = inputs(device)
+    if problem == "shape":
+        x = x[..., :64]
+    elif problem == "dtype":
+        x = x.double()
+    elif problem == "table_shape":
+        sin = sin[:-1]
+    elif problem == "table_dtype":
+        cos = cos.half()
+    elif problem == "table_grad":
+        cos.requires_grad_()
+    elif problem == "device":
+        sin = torch.empty_like(sin, device="meta")
+    with pytest.raises((ValueError, TypeError)):
+        op(x, cos, sin)
+
+
+def test_explicit_backend_and_reference_dispatch(monkeypatch):
+    from rl_engine.runtime.registry import KernelRegistry
+
+    op = KernelRegistry().get_op("multi_axis_rope", device="cpu")
+    assert op.trace()["actual_backend"].startswith("rlkernel.reference.")
+    from rl_engine.backends.cuda.rope import multi_axis_rope as cuda_module
+
+    monkeypatch.setattr(cuda_module, "native_available", lambda: False)
+    with pytest.raises(RuntimeError, match="unavailable"):
+        get_multi_axis_rope_op("cuda")
+    with pytest.raises(ValueError, match="unknown"):
+        get_multi_axis_rope_op("typo")
+    x, cos, sin = inputs("cpu")
+    assert_bits(multi_axis_rope(x, cos, sin), op(x, cos, sin))
+
+
+def test_cuda_registry_uses_native_when_present():
+    from rl_engine.backends.cuda.rope.multi_axis_rope import native_available
+    from rl_engine.runtime.registry import KernelRegistry
+
+    if not torch.cuda.is_available() or not native_available():
+        pytest.skip("native CUDA multi_axis_rope extension required")
+    op = KernelRegistry().get_op("multi_axis_rope", device="cuda")
+    assert op.trace()["actual_backend"] == "rlkernel.cuda.qwenimage_multi_axis_rope.v1"
+    x, cos, sin = inputs("cuda")
+    torch.testing.assert_close(
+        op(x, cos, sin).double(), complex_reference(x, cos, sin), atol=8e-7, rtol=8e-7
+    )
+
+
+def test_cuda_registry_falls_back_to_triton(monkeypatch):
+    from rl_engine.runtime.registry import KernelRegistry
+
+    if not torch.cuda.is_available() or torch.version.hip is not None:
+        pytest.skip("NVIDIA CUDA required")
+    pytest.importorskip("triton")
+    from rl_engine.backends.cuda.rope import multi_axis_rope as cuda_module
+
+    monkeypatch.setattr(cuda_module, "native_available", lambda: False)
+    op = KernelRegistry().get_op("multi_axis_rope", device="cuda")
+    assert op.trace()["actual_backend"] == "rlkernel.triton.qwenimage_multi_axis_rope.v1"
+    x, cos, sin = inputs("cuda")
+    torch.testing.assert_close(
+        op(x, cos, sin).double(), complex_reference(x, cos, sin), atol=8e-7, rtol=8e-7
+    )
+
+
+@pytest.mark.parametrize(
+    "bad",
+    [torch.zeros(2, 3), torch.zeros(2, 4, dtype=torch.int64), torch.full((2, 3), 4096)],
+)
+def test_reject_invalid_positions(bad):
+    with pytest.raises((TypeError, ValueError)):
+        build_multi_axis_cos_sin(bad)
+
+
+@pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16])
+def test_cuda_triton_byte_parity_and_graph(dtype):
+    if not torch.cuda.is_available() or torch.version.hip is not None:
+        pytest.skip("NVIDIA CUDA required")
+    from rl_engine.backends.cuda.rope.multi_axis_rope import native_available
+
+    if not native_available():
+        pytest.skip("native multi_axis_rope extension required")
+    pytest.importorskip("triton")
+    cuda, triton = get_multi_axis_rope_op("cuda"), get_multi_axis_rope_op("triton")
+    x, cos, sin = inputs("cuda", dtype, batched=True, seq=17, heads=24)
+    x.requires_grad_()
+    grad = torch.randn_like(x)
+    yc, yt = cuda(x, cos, sin), triton(x, cos, sin)
+    dc = torch.autograd.grad(yc, x, grad)[0]
+    dt = torch.autograd.grad(yt, x, grad)[0]
+    assert_bits(yc, yt)
+    assert_bits(dc, dt)
+    reference = get_multi_axis_rope_op("reference")
+    yr = reference(x, cos, sin)
+    dr = torch.autograd.grad(yr, x, grad)[0]
+    assert_bits(yc, yr)
+    assert_bits(dc, dr)
+    # Capture forward + autograd after warmup on a side stream.
+    for op in (cuda, triton):
+        stream = torch.cuda.Stream()
+        stream.wait_stream(torch.cuda.current_stream())
+        with torch.cuda.stream(stream):
+            static_x = x.detach().clone().requires_grad_()
+            static_grad = grad.clone()
+            for _ in range(3):
+                torch.autograd.grad(op(static_x, cos, sin), static_x, static_grad)
+        torch.cuda.current_stream().wait_stream(stream)
+        graph = torch.cuda.CUDAGraph()
+        gc.collect()
+        was_enabled = gc.isenabled()
+        gc.disable()
+        try:
+            with torch.cuda.graph(graph, stream=stream):
+                result = op(static_x, cos, sin)
+                derivative = torch.autograd.grad(result, static_x, static_grad)[0]
+        finally:
+            if was_enabled:
+                gc.enable()
+        for _ in range(3):
+            graph.replay()
+        torch.cuda.synchronize()
+        assert_bits(result, yc)
+        assert_bits(derivative, dc)
+        # Replay must consume updated static buffers, not just retain captured outputs.
+        with torch.no_grad():
+            static_x.copy_(x * 2)
+            static_grad.copy_(grad * 0.5)
+        graph.replay()
+        torch.cuda.synchronize()
+        expected = op(static_x, cos, sin)
+        expected_dx = torch.autograd.grad(expected, static_x, static_grad)[0]
+        assert_bits(result, expected)
+        assert_bits(derivative, expected_dx)
+
+
+def test_native_entrypoint_checks_and_nondefault_stream():
+    if not torch.cuda.is_available() or torch.version.hip is not None:
+        pytest.skip("NVIDIA CUDA required")
+    from rl_engine.backends.cuda.rope.multi_axis_rope import native_available
+
+    if not native_available():
+        pytest.skip("native multi_axis_rope extension required")
+    x, cos, sin = inputs("cuda")
+    native = torch.ops.rl_kernel_native.multi_axis_rope
+    for c, s in [(cos.half(), sin), (cos, sin[:-1]), (cos.cpu(), sin)]:
+        with pytest.raises(RuntimeError):
+            native(x, c, s, False)
+    stream = torch.cuda.Stream()
+    stream.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(stream):
+        xx = x * 3
+        y = native(xx, cos, sin, False)
+    torch.cuda.current_stream().wait_stream(stream)
+    torch.testing.assert_close(y.double(), complex_reference(xx, cos, sin), atol=2e-6, rtol=2e-6)
+
+
+def test_device_guard():
+    if torch.cuda.device_count() < 2 or torch.version.hip is not None:
+        pytest.skip("requires two NVIDIA GPUs")
+    from rl_engine.backends.cuda.rope.multi_axis_rope import native_available
+
+    if not native_available():
+        pytest.skip("native multi_axis_rope extension required")
+    with torch.cuda.device(0):
+        x, cos, sin = inputs("cuda:1")
+        for backend in ("cuda", "triton"):
+            y = get_multi_axis_rope_op(backend)(x, cos, sin)
+            assert y.device == x.device
+            assert torch.cuda.current_device() == 0
+            torch.testing.assert_close(
+                y.double(), complex_reference(x, cos, sin), atol=8e-7, rtol=8e-7
+            )
+
+
+@pytest.mark.parametrize("default_device", ["meta", "cuda"])
+def test_canonical_tables_ignore_default_device(default_device):
+    if default_device == "cuda" and not torch.cuda.is_available():
+        pytest.skip("CUDA context requires CUDA")
+    positions, text = qwen_image_positions(3, 5, 7)
+    expected = build_multi_axis_cos_sin(positions)
+    with torch.device(default_device):
+        image_again, text_again = qwen_image_positions(3, 5, 7)
+        actual = build_multi_axis_cos_sin(positions, device="cpu")
+    assert image_again.device.type == "cpu" and text_again.device.type == "cpu"
+    assert_bits(image_again, positions)
+    assert_bits(text_again, text)
+    for a, b in zip(actual, expected, strict=True):
+        assert_bits(a, b)
+
+
+def test_packed_qkv_and_independent_table_strides(implementation):
+    op, device = implementation
+    packed = torch.randn(2, 17, 3, 3, 256, device=device)
+    x = packed[:, :, 0, :, ::2].detach().requires_grad_()
+    cos_base = torch.randn(2, 17, 128, device=device)
+    cos = cos_base[..., ::2]
+    sin = torch.randn(17, 2, 64, device=device).transpose(0, 1)
+    grad = torch.randn_like(x)
+    result = op(x, cos, sin)
+    dx = torch.autograd.grad(result, x, grad)[0]
+    assert result.is_contiguous()
+    torch.testing.assert_close(
+        result.double(), complex_reference(x, cos, sin), atol=2e-6, rtol=2e-6
+    )
+    torch.testing.assert_close(
+        dx.double(), complex_reference(grad, cos, sin, transpose=True), atol=2e-6, rtol=2e-6
+    )
+    if device == "cuda":
+        with torch.profiler.profile(activities=[torch.profiler.ProfilerActivity.CPU]) as profile:
+            op(x, cos, sin)
+        copies = {event.key for event in profile.key_averages()} & {
+            "aten::contiguous",
+            "aten::clone",
+            "aten::copy_",
+        }
+        assert not copies, f"RoPE materialized its strided inputs: {copies}"
+
+
+@pytest.mark.parametrize("backend", ["cuda", "triton"])
+def test_dispatcher_fake_autograd_and_fullgraph_compile(backend):
+    if not torch.cuda.is_available() or torch.version.hip is not None:
+        pytest.skip("NVIDIA CUDA required")
+    if backend == "cuda":
+        from rl_engine.backends.cuda.rope.multi_axis_rope import _rotation, native_available
+
+        if not native_available():
+            pytest.skip("build native RoPE library")
+    else:
+        pytest.importorskip("triton")
+        from rl_engine.backends.shared.triton.rope.multi_axis_rope import _rotation
+    op = get_multi_axis_rope_op(backend)
+    x, cos, sin = inputs("cuda", batched=True, seq=5, heads=2)
+    x.requires_grad_()
+    checks = torch.library.opcheck(_rotation, (x, cos, sin, False))
+    assert all(value == "SUCCESS" for value in checks.values())
+    compiled = torch.compile(op, fullgraph=True)
+    grad = torch.randn_like(x)
+    expected = op(x, cos, sin)
+    expected_dx = torch.autograd.grad(expected, x, grad)[0]
+    result = compiled(x, cos, sin)
+    dx = torch.autograd.grad(result, x, grad)[0]
+    assert_bits(result, expected)
+    assert_bits(dx, expected_dx)
+
+
+@pytest.mark.parametrize("backend", ["cuda", "triton"])
+@pytest.mark.ws1
+def test_standard_ws1_four_judgments(backend):
+    if not torch.cuda.is_available() or torch.version.hip is not None:
+        pytest.skip("NVIDIA CUDA required")
+    if backend == "cuda":
+        from rl_engine.backends.cuda.rope.multi_axis_rope import native_available
+
+        if not native_available():
+            pytest.skip("build native RoPE library")
+    else:
+        pytest.importorskip("triton")
+    from rl_engine.validation.operators.qwen_image_rope import run_multi_axis_rope_ws1
+
+    result = run_multi_axis_rope_ws1(backend)
+    assert result["passed"]
+    assert result["trace"]["actual_backend"].startswith(f"rlkernel.{backend}.")
+    for phase in ("forward", "gradient"):
+        assert result[phase]["accuracy_reports"]
+        assert result[phase]["invariance_reports"]
+        assert all(report["passed"] for report in result[phase]["invariance_reports"])
diff --git a/tests/validation/operators/test_operator_inputs.py b/tests/validation/operators/test_operator_inputs.py
index 9e5015bf4..b2c054b18 100644
--- a/tests/validation/operators/test_operator_inputs.py
+++ b/tests/validation/operators/test_operator_inputs.py
@@ -103,6 +103,24 @@ def test_cp_attention_operator_spec_registers_backward_grad_inputs():
     assert candidate.name == "pytorch-cp_attention"
 
 
+def test_multi_axis_rope_operator_spec_and_seeded_inputs():
+    args = _args(op="multi_axis_rope", batch=2, seq=7, n_heads=3, input_mode="random")
+    case = make_operator_case(args, torch.float32, torch.device("cpu"))
+    inputs = make_operator_inputs("multi_axis_rope", args, torch.float32, torch.device("cpu"))
+    repeated = make_operator_inputs("multi_axis_rope", args, torch.float32, torch.device("cpu"))
+    assert case.op_class == "elementwise"
+    assert case.grad_input_names == ("x",)
+    assert inputs["x"].shape == (2, 7, 3, 128)
+    assert inputs["cos"].shape == inputs["sin"].shape == (7, 64)
+    for name in inputs:
+        assert torch.equal(inputs[name], repeated[name])
+    assert operator_shape_name("multi_axis_rope", args) == "2x7x3x128"
+    with pytest.raises(ValueError, match="128"):
+        make_operator_inputs(
+            "multi_axis_rope", _args(head_dim=64), torch.float32, torch.device("cpu")
+        )
+
+
 def test_constant_linear_logp_inputs_match_operator_contract():
     args = _args(input_mode="constant", constant_value=0.5, token_value=3)
     inputs = make_operator_inputs("linear_logp", args, torch.float32, torch.device("cpu"))
diff --git a/tools/validation/operators/check_multi_axis_rope_install.py b/tools/validation/operators/check_multi_axis_rope_install.py
new file mode 100644
index 000000000..56cfe2ccb
--- /dev/null
+++ b/tools/validation/operators/check_multi_axis_rope_install.py
@@ -0,0 +1,67 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Run from outside the checkout after installing a RoPE-enabled wheel.
+
+Does not change sys.path or compile/load a JIT extension. --package-root makes
+the expected installation location an explicit assertion in package validation.
+"""
+
+import argparse
+import importlib.util
+import json
+from pathlib import Path
+
+import torch
+
+import rl_engine
+from rl_engine.ops.rope.multi_axis_rope import (
+    build_multi_axis_cos_sin,
+    get_multi_axis_rope_op,
+    qwen_image_positions,
+)
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--package-root", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+    package = Path(rl_engine.__file__).resolve().parent
+    assert package.is_relative_to(args.package_root.resolve()), package
+    license_path = package / "backends/shared/triton/rope/FLASH_ATTENTION_LICENSE"
+    assert "Redistribution and use" in license_path.read_text(encoding="utf-8")
+    positions, _ = qwen_image_positions(3, 5, 0)
+    cos, sin = build_multi_axis_cos_sin(positions, device="cuda")
+    packed = torch.randn(2, 15, 3, 3, 128, device="cuda", dtype=torch.bfloat16)
+    x = packed[:, :, 0].detach().requires_grad_()
+    grad = torch.randn_like(x)
+    gold = get_multi_axis_rope_op("reference")(x, cos, sin)
+    gold_dx = torch.autograd.grad(gold, x, grad)[0]
+    outcomes = {}
+    for backend in ("cuda", "triton"):
+        op = get_multi_axis_rope_op(backend, device=x.device)
+        for mode, fn in (("eager", op), ("compiled", torch.compile(op, fullgraph=True))):
+            result = fn(x, cos, sin)
+            dx = torch.autograd.grad(result, x, grad)[0]
+            assert torch.equal(result.view(torch.uint8), gold.contiguous().view(torch.uint8))
+            assert torch.equal(dx.view(torch.uint8), gold_dx.contiguous().view(torch.uint8))
+            outcomes[f"{backend}_{mode}"] = "passed"
+        outcomes[f"{backend}_trace"] = op.trace()
+    native = importlib.util.find_spec("rl_engine._rope_cuda")
+    report = {
+        "package": str(package),
+        "companion_library": native.origin if native else None,
+        "native_dispatcher_registered": hasattr(torch.ops.rl_kernel_native, "multi_axis_rope"),
+        "torch": torch.__version__,
+        "gpu": torch.cuda.get_device_name(),
+        "license_present": True,
+        "packed_qkv_stride": list(x.stride()),
+        "checks": outcomes,
+        "passed": True,
+    }
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
+    print(json.dumps(report, indent=2))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tools/validation/operators/check_qwenimage_rope_upstream.py b/tools/validation/operators/check_qwenimage_rope_upstream.py
new file mode 100644
index 000000000..153000904
--- /dev/null
+++ b/tools/validation/operators/check_qwenimage_rope_upstream.py
@@ -0,0 +1,99 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Audit table and rotation semantics against a checksum-pinned upstream file.
+
+Supply transformer_qwenimage.py from Diffusers revision
+40f63690da440e8c382ab53e67508835deb1ed2f. Only the two audited symbol definitions
+are executed, after verifying the source hash. No Diffusers installation or model
+checkpoint is required. This checks CPU semantics, not full-model integration.
+"""
+
+from __future__ import annotations
+
+import argparse
+import ast
+import functools
+import hashlib
+import json
+import logging
+import sys
+from pathlib import Path
+
+import torch
+
+ROOT = Path(__file__).resolve().parents[3]
+sys.path.insert(0, str(ROOT))
+SOURCE_SHA256 = "0bad11901deae69f05aec01fd464e51e38582d455068f2851ee665036179321f"
+
+
+def main():
+    from rl_engine.ops.rope.multi_axis_rope import (
+        build_multi_axis_cos_sin,
+        get_multi_axis_rope_op,
+        qwen_image_positions,
+    )
+
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--source", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+    source = args.source.read_text(encoding="utf-8")
+    digest = hashlib.sha256(source.encode()).hexdigest()
+    if digest != SOURCE_SHA256:
+        raise ValueError("upstream source does not match the pinned revision checksum")
+    symbols = {"QwenEmbedRope", "apply_rotary_emb_qwen"}
+    nodes = [n for n in ast.parse(source).body if getattr(n, "name", None) in symbols]
+    if len(nodes) != 2:
+        raise ValueError("missing audited upstream definitions")
+    namespace = {
+        "torch": torch,
+        "nn": torch.nn,
+        "lru_cache_unless_export": functools.lru_cache,
+        "logger": logging.getLogger("upstream_audit"),
+    }
+    exec(
+        compile(ast.Module(body=nodes, type_ignores=[]), str(args.source), "exec"),
+        namespace,
+    )
+    embed = namespace["QwenEmbedRope"](theta=10000, axes_dim=[16, 56, 56], scale_rope=True)
+    apply_upstream = namespace["apply_rotary_emb_qwen"]
+    op = get_multi_axis_rope_op("reference")
+    torch.manual_seed(386)
+    results = []
+    for height, width in ((2, 3), (3, 5), (64, 64), (83, 83), (104, 58)):
+        positions = qwen_image_positions(height, width, 7)
+        upstream_tables = embed((1, height, width), device=torch.device("cpu"), max_txt_seq_len=7)
+        for stream, pos, upstream in zip(
+            ("image", "text"), positions, upstream_tables, strict=True
+        ):
+            cos, sin = build_multi_axis_cos_sin(pos)
+            assert torch.equal(cos, upstream.real) and torch.equal(sin, upstream.imag)
+            for dtype in (torch.float32, torch.float16, torch.bfloat16):
+                x = torch.randn(1, len(pos), 2, 128, dtype=dtype, requires_grad=True)
+                grad = torch.randn_like(x)
+                expected = apply_upstream(x, upstream, use_real=False)
+                actual = op(x, cos, sin)
+                eg = torch.autograd.grad(expected, x, grad)[0]
+                ag = torch.autograd.grad(actual, x, grad)[0]
+                tol = {torch.float32: 8e-7, torch.float16: 0.002, torch.bfloat16: 0.02}[dtype]
+                torch.testing.assert_close(actual, expected, atol=tol, rtol=tol)
+                torch.testing.assert_close(ag, eg, atol=tol, rtol=tol)
+                results.append(
+                    {
+                        "grid": [height, width],
+                        "stream": stream,
+                        "dtype": str(dtype),
+                        "table_bitwise": True,
+                        "forward_max_abs": (actual.float() - expected.float()).abs().max().item(),
+                        "dx_max_abs": (ag.float() - eg.float()).abs().max().item(),
+                    }
+                )
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(json.dumps({"source_sha256": digest, "cases": results}, indent=2) + "\n")
+    print(
+        f"{len(results)} upstream forward/backward cases passed; "
+        "all image/text tables exactly equal"
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tools/validation/operators/validate_multi_axis_rope.py b/tools/validation/operators/validate_multi_axis_rope.py
new file mode 100644
index 000000000..d18ef7cbb
--- /dev/null
+++ b/tools/validation/operators/validate_multi_axis_rope.py
@@ -0,0 +1,85 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Compile only the production RoPE CUDA source for focused GPU validation.
+
+Loads the production TORCH_LIBRARY registration directly. Normal package builds
+include the same source in rl_engine._C. No generated binding or monkeypatch is
+used. This validates the operator library, not every unrelated native operator.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import runpy
+import sys
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[3]
+sys.path.insert(0, str(ROOT))
+
+
+def load_focused_extension():
+    import torch
+    from torch.utils.cpp_extension import CUDA_HOME, load
+
+    from rl_engine.backends.cuda.rope.multi_axis_rope import native_available
+
+    if native_available():
+        print("Using the installed production RoPE dispatcher library")
+        return
+
+    build = ROOT / "artifacts" / "multi_axis_rope_native"
+    build.mkdir(parents=True, exist_ok=True)
+    # Conda CUDA toolkits use lib/, whereas the Windows SDK convention is lib/x64/.
+    link_flags = []
+    if os.name == "nt" and CUDA_HOME and (Path(CUDA_HOME) / "lib/cudart.lib").is_file():
+        link_flags.append(f'/LIBPATH:"{Path(CUDA_HOME) / "lib"}"')
+    os.environ["PATH"] = str(Path(sys.executable).parent) + os.pathsep + os.environ["PATH"]
+    load(
+        name="rlk_multi_axis_rope_native",
+        sources=[str(ROOT / "csrc/cuda/rope/multi_axis_rope.cu")],
+        build_directory=str(build),
+        extra_cflags=(["/O2", "/std:c++17"] if os.name == "nt" else ["-O3", "-std=c++17"]),
+        extra_cuda_cflags=["-O3", "-std=c++17", "--fmad=false"],
+        extra_ldflags=link_flags,
+        is_python_module=False,
+        verbose=True,
+    )
+    assert hasattr(torch.ops.rl_kernel_native, "multi_axis_rope")
+    print(
+        f"Focused CUDA source build; torch={torch.__version__}, GPU={torch.cuda.get_device_name()}"
+    )
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--benchmark", action="store_true")
+    parser.add_argument("--ws1", choices=("cuda", "triton"))
+    parser.add_argument("--output", type=Path)
+    args, remaining = parser.parse_known_args()
+    load_focused_extension()
+    if args.ws1:
+        from rl_engine.validation.operators.qwen_image_rope import run_multi_axis_rope_ws1
+
+        result = run_multi_axis_rope_ws1(args.ws1)
+        if args.output:
+            args.output.parent.mkdir(parents=True, exist_ok=True)
+            args.output.write_text(json.dumps(result, indent=2) + "\n", encoding="utf-8")
+        print(f"WS1 four judgments backend={args.ws1}: passed={result['passed']}")
+        raise SystemExit(0 if result["passed"] else 1)
+    elif args.benchmark:
+        sys.argv = ["benchmark_multi_axis_rope", *remaining]
+        if args.output:
+            sys.argv.extend(["--output", str(args.output)])
+        runpy.run_module("benchmarks.operators.rope.benchmark_multi_axis_rope", run_name="__main__")
+    else:
+        import pytest
+
+        raise SystemExit(
+            pytest.main([str(ROOT / "tests/ops/rope/test_multi_axis_rope.py"), *remaining])
+        )
+
+
+if __name__ == "__main__":
+    main()

From a0f9276195b020681ea01c5ca47bc33fa260277a Mon Sep 17 00:00:00 2001
From: dandan20031031 <2425926169@qq.com>
Date: Wed, 28 Oct 2026 21:49:14 +0800
Subject: [PATCH 2/5] test(rope): qualify the complete CUDA package on WSL2

Add an explicit full-extension GPU gate and CPU CI for test-qwenimage. Record WSL2 wheel/import, WS1, layout, compiler, CPU-only, memory-check and mature-baseline evidence; expose the loaded full native library in the installation probe. Refs #386.

Signed-off-by: dandan20031031 <2425926169@qq.com>
---
 .github/workflows/qwenimage-rope.yml          |   82 +
 ci/scripts/run_qwenimage_rope_ws1.sh          |   89 +
 docs/operators/multi-axis-rope.md             |   23 +
 .../README.md                                 |  132 +
 .../benchmark-bfloat16.json                   | 2781 +++++++++++++++++
 .../cpu-contracts.txt                         |   20 +
 .../cpu-environment.json                      |    7 +
 .../full-extension.json                       |   22 +
 .../full-inplace-build.txt                    |   20 +
 .../full-wheel-build.txt                      |   10 +
 .../gpu-environment.json                      |   15 +
 .../installed-wheel.json                      |   38 +
 .../layout.txt                                |    2 +
 .../manifest.json                             |   58 +
 .../memcheck.txt                              |    4 +
 .../operator-cuda.json                        |   70 +
 .../operator-tests.xml                        |    1 +
 .../operator-triton.json                      |   70 +
 .../package.json                              |   11 +
 .../precommit.txt                             |    7 +
 .../python-packages.json                      |  114 +
 .../quality.txt                               |   34 +
 .../regression.txt                            |   10 +
 .../tests-and-ws1.txt                         |   11 +
 .../upstream.json                             |  335 ++
 .../ws1-cuda.json                             | 1962 ++++++++++++
 .../ws1-triton.json                           | 1962 ++++++++++++
 .../check_multi_axis_rope_install.py          |    2 +
 28 files changed, 7892 insertions(+)
 create mode 100644 .github/workflows/qwenimage-rope.yml
 create mode 100644 ci/scripts/run_qwenimage_rope_ws1.sh
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/README.md
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/benchmark-bfloat16.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-contracts.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-environment.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-extension.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-inplace-build.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-wheel-build.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/gpu-environment.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/installed-wheel.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/layout.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/manifest.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/memcheck.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-cuda.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-tests.xml
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-triton.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/package.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/precommit.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/python-packages.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/quality.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/regression.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/tests-and-ws1.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/upstream.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-cuda.json
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-triton.json

diff --git a/.github/workflows/qwenimage-rope.yml b/.github/workflows/qwenimage-rope.yml
new file mode 100644
index 000000000..7819acb9d
--- /dev/null
+++ b/.github/workflows/qwenimage-rope.yml
@@ -0,0 +1,82 @@
+# SPDX-License-Identifier: Apache-2.0
+name: Qwen-Image RoPE
+
+on:
+  pull_request:
+    branches: [test-qwenimage]
+    types: [opened, synchronize, reopened, ready_for_review]
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  cpu-contracts:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ['3.10', '3.12']
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: pip
+      - name: Install CPU reference dependencies
+        run: |
+          python -m pip install pip==25.3 setuptools==80.9.0 wheel==0.45.1
+          python -m pip install torch==2.9.1 --index-url https://download.pytorch.org/whl/cpu
+          python -m pip install --no-build-isolation -e . pytest
+      - name: Run CPU contracts, input registration and build-profile checks
+        env:
+          OMP_NUM_THREADS: '1'
+          MKL_NUM_THREADS: '1'
+        run: |
+          python -m pytest -q -rs \
+            tests/ops/rope/test_multi_axis_rope.py \
+            tests/validation/operators/test_operator_inputs.py \
+            tests/build/test_multi_axis_rope_build.py \
+            --junitxml=artifacts/rope-cpu.xml
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: rope-cpu-${{ matrix.python-version }}
+          path: artifacts/rope-cpu.xml
+
+  static-and-docs:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+          cache: pip
+      - run: python -m pip install pre-commit==4.3.0 -r requirements/docs.txt
+      - name: Check changed files
+        run: |
+          python - <<'PY'
+          import json
+          import os
+          import re
+          import subprocess
+          import sys
+
+          with open(os.environ['GITHUB_EVENT_PATH'], encoding='utf-8') as stream:
+              event = json.load(stream)
+          base = event.get('pull_request', {}).get('base', {}).get('sha')
+          if base:
+              assert re.fullmatch(r'[0-9a-f]{40}', base)
+          else:
+              base = 'HEAD^'
+          changed = subprocess.check_output([
+              'git', 'diff', '--name-only', '--diff-filter=ACMR', '-z', base, 'HEAD'
+          ]).decode('utf-8').split('\0')
+          files = [('./' + path if path.startswith('-') else path) for path in changed if path]
+          if files:
+              subprocess.check_call([sys.executable, '-m', 'pre_commit', 'run', '--files', *files])
+          PY
+      - name: Build documentation
+        run: mkdocs build --strict -f mkdocs.yaml
diff --git a/ci/scripts/run_qwenimage_rope_ws1.sh b/ci/scripts/run_qwenimage_rope_ws1.sh
new file mode 100644
index 000000000..447e93f14
--- /dev/null
+++ b/ci/scripts/run_qwenimage_rope_ws1.sh
@@ -0,0 +1,89 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: Apache-2.0
+# Operator-only gate: requires an ordinary, complete rl_engine._C installation.
+set -euo pipefail
+
+ROPE_ROOT="$(cd "$(dirname "$0")/../.." && pwd)"
+cd "$ROPE_ROOT"
+ROPE_PYTHON="${ROPE_PYTHON:-python3}"
+ROPE_WS1_OUTPUT="${ROPE_WS1_OUTPUT:-${TMPDIR:-/tmp}/rlk-qwenimage-rope}"
+mkdir -p "$ROPE_WS1_OUTPUT"
+export RL_KERNEL_REQUIRE_EXT=1
+
+"$ROPE_PYTHON" - "$ROPE_WS1_OUTPUT/environment.json" <<'PY'
+import hashlib
+import json
+import platform
+import sys
+from pathlib import Path
+
+import torch
+import triton
+
+from build_tools.envs import env_flag
+from rl_engine import _C
+
+if env_flag("RL_KERNEL_ROPE_ONLY"):
+    raise SystemExit("The complete _C build is required; unset RL_KERNEL_ROPE_ONLY")
+if env_flag("KERNEL_ALIGN_USE_FAST_MATH"):
+    raise SystemExit("This numerical profile does not qualify fast-math builds")
+if torch.version.hip is not None or not torch.cuda.is_available():
+    raise SystemExit("A real NVIDIA CUDA device is required")
+if not torch.cuda.is_bf16_supported():
+    raise SystemExit("The WS1 BF16 profile requires BF16-capable hardware")
+if not hasattr(torch.ops.rl_kernel_native, "multi_axis_rope"):
+    raise SystemExit("The imported _C library does not register multi_axis_rope")
+library = Path(_C.__file__).resolve()
+report = {
+    "scope": "Qwen-Image multi_axis_rope operator; ordinary full _C build",
+    "python": platform.python_version(),
+    "platform": platform.platform(),
+    "torch": torch.__version__,
+    "cuda": torch.version.cuda,
+    "triton": triton.__version__,
+    "gpu": torch.cuda.get_device_name(),
+    "capability": list(torch.cuda.get_device_capability()),
+    "native_library": str(library),
+    "native_library_sha256": hashlib.sha256(library.read_bytes()).hexdigest(),
+}
+Path(sys.argv[1]).write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
+print(json.dumps(report))
+PY
+
+"$ROPE_PYTHON" -m pytest tests/ops/rope/test_multi_axis_rope.py -q -rs \
+    --junitxml="$ROPE_WS1_OUTPUT/operator-tests.xml"
+
+# GPU/backward/compile cases must execute. Only the two-device guard may skip
+# on a single-GPU runner; missing CUDA/Triton support is not acceptance evidence.
+"$ROPE_PYTHON" - "$ROPE_WS1_OUTPUT/operator-tests.xml" <<'PY'
+import sys
+import xml.etree.ElementTree as ET
+
+root = ET.parse(sys.argv[1]).getroot()
+unexpected = [case.get("name") for case in root.iter("testcase")
+              if case.find("skipped") is not None and case.get("name") != "test_device_guard"]
+if unexpected:
+    raise SystemExit(f"Required RoPE tests skipped: {unexpected}")
+PY
+
+for backend in cuda triton; do
+    "$ROPE_PYTHON" tools/validation/operators/check_operator.py \
+        --op multi_axis_rope --candidate "$backend" --device cuda --dtype bf16 \
+        --batch 2 --seq 7 --n-heads 3 --check-grad --json \
+        > "$ROPE_WS1_OUTPUT/operator-$backend.json"
+    "$ROPE_PYTHON" tools/validation/operators/validate_multi_axis_rope.py \
+        --ws1 "$backend" --output "$ROPE_WS1_OUTPUT/ws1-$backend.json"
+done
+
+"$ROPE_PYTHON" - "$ROPE_WS1_OUTPUT" <<'PY'
+import json
+import sys
+from pathlib import Path
+
+root = Path(sys.argv[1])
+for backend in ("cuda", "triton"):
+    report = json.loads((root / f"ws1-{backend}.json").read_text(encoding="utf-8"))
+    if not report["passed"] or report["trace"]["actual_backend"] != f"rlkernel.{backend}.qwenimage_multi_axis_rope.v1":
+        raise SystemExit(f"WS1 failed or selected an unexpected backend: {backend}")
+print("Qwen-Image RoPE full-extension operator gate passed")
+PY
diff --git a/docs/operators/multi-axis-rope.md b/docs/operators/multi-axis-rope.md
index fa908a03e..16899a755 100644
--- a/docs/operators/multi-axis-rope.md
+++ b/docs/operators/multi-axis-rope.md
@@ -172,6 +172,29 @@ with C1's elementwise tolerance profile and C2's logical configuration transform
 This gate qualifies the Qwen-Image operator and does not certify the Dense Qwen3
 chain, distributed execution, or model assembly.
 
+For a complete native package installation on Linux, run the dedicated gate:
+
+```bash
+export CUDA_HOME=/usr/local/cuda-12.8
+export PATH="$CUDA_HOME/bin:$PATH"
+RL_KERNEL_REQUIRE_EXT=1 python -m pip install --no-build-isolation -e .
+ROPE_PYTHON=python bash ci/scripts/run_qwenimage_rope_ws1.sh
+```
+
+The gate requires a real NVIDIA GPU, Triton and an ordinary `rl_engine._C`
+containing the native RoPE registration. It rejects the operator-only profile
+and fast-math mode, checks required tests actually execute, and permits only
+the two-device guard to skip on a single-GPU host. Results include the loaded
+binary fingerprint, ordinary operator CLI checks and both WS1 backend reports.
+Set `ROPE_WS1_OUTPUT` to select the artifact directory.
+
+On WSL2, keep the checkout and Python environment in the Linux filesystem.
+GNU linking of the package's CUDA driver calls may need
+`LIBRARY_PATH=/usr/lib/wsl/lib`. Record the WSL kernel and physical GPU in the
+evidence; a WSL2 profile does not qualify a bare-metal Linux server or another
+GPU architecture. The `Qwen-Image RoPE` workflow runs CPU contracts and changed-file
+checks for PRs targeting `test-qwenimage`; GPU qualification uses the gate above.
+
 Tests cover independent forward/VJP accuracy, adjacent-pair layout, shared and
 batched positions, odd and rectangular grids, batch/padding/chunk invariance,
 non-contiguous tensors, packed QKV views without copies, default-device contexts,
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/README.md b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/README.md
new file mode 100644
index 000000000..ddc879318
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/README.md
@@ -0,0 +1,132 @@
+# Qwen-Image RoPE: full native package qualification on WSL2
+
+Recorded 2026-10-28 for [PR #623](https://github.com/RL-Align/RL-Kernel/pull/623).
+Numerical/native sources are based on `ac17ab87e8d93dabe93dfdc134e9853d47de7af1`;
+this follow-up adds the reproducible gate, CPU CI routing and package provenance.
+Source fingerprints are in `manifest.json`. Historical Windows measurements stay
+in the separate `qwenimage_multi_axis_rope_v2_sm120` directory.
+
+## Environment and build
+
+| Item | Recorded value |
+| --- | --- |
+| Linux distribution | Ubuntu 24.04.5 LTS, x86_64, glibc 2.39 |
+| Virtualization | WSL 3.0.3 on Windows 11 |
+| Linux kernel | 6.18.54.1-microsoft-standard-WSL2 |
+| GPU / driver | RTX 5080, SM120, 16 GB / Windows driver 617.42 |
+| Python / PyTorch / Triton | 3.12.3 / 2.9.1+cu128 / 3.5.1 |
+| CUDA compiler / host compiler | CUDA 12.8.93 / GCC 13.3.0 |
+| Compute Sanitizer | 2025.1.0.0, CUDA 12.8 |
+| Build profile | Ordinary complete `rl_engine._C`; 10 source translation units |
+| CUDA architecture / parallelism | `TORCH_CUDA_ARCH_LIST=12.0`, `MAX_JOBS=2` |
+
+CUDA compiler, development libraries and command-line tools were installed from
+NVIDIA's WSL-Ubuntu repository; the existing Windows GPU driver provides CUDA
+driver access. The WSL driver directory `/usr/lib/wsl/lib` was added to
+`LIBRARY_PATH` for linking the package's existing CUDA driver API calls.
+`RL_KERNEL_ROPE_ONLY`, forced SM90 and fast-math flags were unset.
+
+The normal PEP 517 wheel build compiled all ten configured translation units,
+including the existing logprob/GEMM/norm/activation/attention/collective code and
+the new RoPE source. The complete `_C` loaded and registered the native RoPE
+dispatcher. `full-extension.json` records sources, flags, binary path and hash;
+`package.json` records the built wheel and installed library fingerprints.
+
+The checkout, virtual environments, build outputs and caches are inside the
+Linux filesystem, backed by the D: virtual disk. A fresh Git clone preserved all
+27 tracked symlinks. This is a WSL2 Linux userspace/compiler qualification on
+SM120; bare-metal Linux and A100/H100 remain separate hardware profiles.
+
+## Results
+
+| Check | Result | Evidence |
+| --- | --- | --- |
+| Ordinary complete `_C` wheel and in-place build/import | Passed | `full-extension.json`, `full-wheel-build.txt`, `package.json` |
+| Operator accuracy/invariance/autograd/graphs/opcheck/fullgraph | 81 passed, 1 two-GPU guard skip | `tests-and-ws1.txt`, `operator-tests.xml` |
+| CUDA and Triton WS1 four judgments | Both passed with explicit provenance | `ws1-cuda.json`, `ws1-triton.json` |
+| Standard operator CLI forward/input gradient | Both passed | `operator-cuda.json`, `operator-triton.json` |
+| Related regressions | 195 passed, 7 Ascend skips | `regression.txt` |
+| Repository layout with extension enforcement enabled | 37 passed, zero skips/failures | `layout.txt` |
+| Fresh installed full wheel, outside source checkout | Eager/compiled forward and input-gradient bytes exact for both backends | `installed-wheel.json` |
+| Actual CPU-only PyTorch wheel/reference installation | 57 passed, 52 expected GPU skips; reference forward/backward passed | `cpu-contracts.txt`, `cpu-environment.json` |
+| Compute Sanitizer accuracy/stride/empty/stream subset | 34 passed; zero errors | `memcheck.txt` |
+| Pinned actual Diffusers symbols | 30 forward/backward cases passed; image/text tables exact | `upstream.json` |
+| Reference-grid benchmark correctness | 40 comparisons passed | `benchmark-bfloat16.json` |
+| Shell/workflow syntax, changed-file checks, strict docs | Passed locally | `quality.txt` |
+
+The full-extension GPU gate rejects missing native/Triton execution. Its only
+allowed operator-test skip is the two-device guard on a one-GPU runner.
+The CPU-only result is separate: it uses `torch==2.9.1+cpu`, and its installed
+package contains no native CUDA binary. CPU skips are not GPU acceptance evidence.
+The 37 passing Linux layout tests contrast with the five Windows baseline
+failures recorded earlier; those historical logs were not rewritten.
+
+The installed-wheel probe starts in a fresh process outside the checkout,
+loads `_C` from site-packages and reports that exact path. It tests packed-QKV
+forward/input-gradient equality in eager and `torch.compile(fullgraph=True)`
+for both CUDA and Triton. The FlashAttention BSD license is present in the wheel.
+
+## Standalone benchmark
+
+BF16, batch 1, 24 heads, head dimension 128. CUDA Graph medians below are
+**forward / forward+backward milliseconds per call**:
+
+| Grid / layout | Native CUDA | Adapted Triton | FlashAttention, no FMA | Diffusers complex |
+| --- | --- | --- | --- | --- |
+| 64x64 contiguous | 0.05780 / 0.12967 | 0.04929 / 0.12414 | 0.04898 / 0.12407 | 0.24446 / 0.51281 |
+| 64x64 packed QKV | 0.05868 / 0.13079 | 0.04731 / 0.11841 | 0.04722 / 0.11847 | 0.24754 / 0.51583 |
+| 83x83 contiguous | 0.11198 / 0.22392 | 0.10720 / 0.22538 | 0.10739 / 0.21420 | 0.52579 / 1.06721 |
+| 83x83 packed QKV | 0.11233 / 0.22486 | 0.10745 / 0.21478 | 0.10744 / 0.21469 | 0.52820 / 1.06947 |
+| 104x58 contiguous | 0.09797 / 0.19708 | 0.09366 / 0.18497 | 0.09362 / 0.18458 | 0.44992 / 0.90110 |
+| 104x58 packed QKV | 0.09610 / 0.19612 | 0.09310 / 0.18828 | 0.09323 / 0.18817 | 0.44979 / 0.90276 |
+
+These are 20 calls per graph, five warm calls/replays and 30 replay measurements.
+Three synchronized eager-wall trials of 30 calls are also in the JSON. Tables,
+compilation and input/backing allocations are outside timing. Extra peak includes
+the returned output or gradient. The short synthetic case is included in raw data.
+At 4096 tokens, both GPU paths' forward extra peak is 24 MiB for either layout,
+the returned BF16 output size. Packed inputs are read without materialization.
+
+Original FlashAttention and its autograd wrapper are checksum-pinned to
+`98c1cee0ed59307833841cdbf567fe336ceec13a`. The no-FMA baseline changes only the
+launch contraction option. CUDA, adapted Triton and that baseline have zero
+differing forward/input-gradient bytes against the unfused FP32 gold in every
+benchmark case. Default FlashAttention and complex arithmetic use declared tolerances.
+The adapted Triton path is close to the mature baseline; no performance advantage
+over FlashAttention is established. Timings are one-process measurements on a
+shared Windows/WSL GPU without clock pinning, not a model-throughput result or
+a controlled comparison against the earlier Windows process.
+
+## Reproduction and CI
+
+Use the matching CUDA 12.8 toolkit and Linux PyTorch 2.9.1+cu128 in a fresh
+Linux virtual environment. Other Python package versions are recorded in
+`python-packages.json`. From the Linux checkout:
+
+```bash
+export CUDA_HOME=/usr/local/cuda-12.8
+export PATH="$CUDA_HOME/bin:$PATH"
+export LIBRARY_PATH="/usr/lib/wsl/lib${LIBRARY_PATH:+:$LIBRARY_PATH}"
+export MAX_JOBS=2 TORCH_CUDA_ARCH_LIST=12.0 RL_KERNEL_REQUIRE_EXT=1
+unset RL_KERNEL_ROPE_ONLY KERNEL_ALIGN_FORCE_SM90 KERNEL_ALIGN_USE_FAST_MATH
+python -m pip wheel --no-build-isolation --no-deps . -w /tmp/rope-full-wheel
+python -m pip install --force-reinstall --no-deps /tmp/rope-full-wheel/rl_kernel-*.whl
+python setup.py build_ext --inplace
+ROPE_PYTHON=python ROPE_WS1_OUTPUT=/tmp/rope-ws1 bash ci/scripts/run_qwenimage_rope_ws1.sh
+python ci/run.py layout -q -rs
+compute-sanitizer --tool memcheck --error-exitcode 1 python -m pytest \
+  tests/ops/rope/test_multi_axis_rope.py -q \
+  -k "forward_backward_accuracy or noncontiguous_input_tables_and_gradient or empty_forward_backward or packed_qkv_and_independent_table_strides or native_entrypoint_checks"
+```
+
+The new [operator gate](../../../ci/scripts/run_qwenimage_rope_ws1.sh) requires
+the complete `_C` registration, records actual library/backend identity and
+enforces the GPU skip policy. The [CPU workflow](../../../.github/workflows/qwenimage-rope.yml)
+targets `test-qwenimage`, runs Python 3.10/3.12 CPU contracts, checks changed files
+and builds documentation. This report's actual CPU-only run used Python 3.12;
+hosted CI status is tracked on the PR. GPU runners use the explicit gate.
+
+Bare-metal Linux, A100/H100, two-GPU guards, QK-norm fusion, model/block assembly
+and trajectories remain unqualified. Building the entire library does not
+certify every native operator. No Dense Qwen3/distributed certification or
+cross-device/version byte-equality claim is made.
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/benchmark-bfloat16.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/benchmark-bfloat16.json
new file mode 100644
index 000000000..ca80105b0
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/benchmark-bfloat16.json
@@ -0,0 +1,2781 @@
+{
+  "gpu": "NVIDIA GeForce RTX 5080",
+  "capability": [
+    12,
+    0
+  ],
+  "os": "Linux-6.18.54.1-microsoft-standard-WSL2-x86_64-with-glibc2.39",
+  "python": "3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0]",
+  "torch": "2.9.1+cu128",
+  "cuda_runtime": "12.8",
+  "triton": "3.5.1",
+  "dtype": "bfloat16",
+  "base_commit": "ac17ab87e8d93dabe93dfdc134e9853d47de7af1",
+  "source_sha256": {
+    "csrc/cuda/rope/multi_axis_rope.cu": "f5db5cf5d4962053194455d9cde8cf91cf4e824bad7abf72e8410a404bcb0e44",
+    "rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py": "cf0eebf315830a4178053a71a77e650b58aa85bba1d1a08aff7677350af755d9",
+    "rl_engine/backends/shared/triton/rope/multi_axis_rope.py": "f08ff90627c66e15e9c558f8a68f132434a741a76997b3800bfbf73dc8940e9e",
+    "rl_engine/backends/cuda/rope/multi_axis_rope.py": "d779d9b60ce5d1d6651925f723049930d898024fdf0548ce29a2019b84753d65",
+    "rl_engine/ops/rope/_registration.py": "add0375d1639444908091d01e8d29781f7a975beffec64b77bfaa81a5b9a0ae3",
+    "rl_engine/ops/rope/multi_axis_rope.py": "07bcff1160f22c1ff6730be731206b2f4a153e2fada26afc93fc884e780568cb",
+    "rl_engine/reference/rope/multi_axis_rope.py": "c9a94730047f51659f7aa0a443e88d6ce7d03d00d6813a9aed548cafbf6347bf",
+    "benchmarks/operators/rope/benchmark_multi_axis_rope.py": "6961a55803fbd4e566bd060558fb59e658e7f1f704e61ca97f912e5d16b55884"
+  },
+  "timing": "CUDA graph ms/call and synchronized eager wall ms/call; warm allocation; excludes tables, compilation, and initial allocation of inputs",
+  "memory": "extra peak allocated bytes excluding inputs/tables/grad; includes returned output or gradient; packed QKV backing storage allocated before timing",
+  "iterations": 30,
+  "warmup": 5,
+  "trials": 3,
+  "graph_calls": 20,
+  "flash_baseline": {
+    "included": true,
+    "revision": "98c1cee0ed59307833841cdbf567fe336ceec13a",
+    "kernel_sha256_normalized": "d2037e623854c702fb075784fe4292d4a89d6b9ac1ae33cab3d6788f8f94f240",
+    "layer_sha256_normalized": "a79cb42dfc33e126227bdba27ae7a024d7d5cc5c8fd7f7a5f26c68d501024e0a",
+    "mode": "upstream rotary_kernel, apply_rotary and ApplyRotaryEmb, interleaved=True",
+    "no_fma_change": "only add enable_fp_fusion=False at the original kernel launch"
+  },
+  "trace": {
+    "cuda": {
+      "contract": "qwenimage-interleaved-rope-fp32-v1",
+      "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
+      "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
+      "layout": "BSHD, adjacent channel pairs",
+      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+    },
+    "triton": {
+      "contract": "qwenimage-interleaved-rope-fp32-v1",
+      "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
+      "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
+      "layout": "BSHD, adjacent channel pairs",
+      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+    }
+  },
+  "qualification": "Exact bytes vs unfused FP32 gold for CUDA, Triton and flash_no_fma; declared tolerance for upstream complex and default FlashAttention",
+  "correctness": [
+    {
+      "backend": "upstream_complex",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        130,
+        139
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        130,
+        139
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        141,
+        139
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        141,
+        139
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        235,
+        224
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        235,
+        224
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        221,
+        226
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        221,
+        226
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        195,
+        199
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        195,
+        199
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "upstream_complex",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        230,
+        220
+      ]
+    },
+    {
+      "backend": "cuda",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "triton",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    },
+    {
+      "backend": "flash_default",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        230,
+        220
+      ]
+    },
+    {
+      "backend": "flash_no_fma",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "different_bytes_y_dx": [
+        0,
+        0
+      ]
+    }
+  ],
+  "results": [
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0034856000915169717,
+      "graph_min_ms": 0.0034543998539447785,
+      "graph_max_ms": 0.0035312000662088396,
+      "eager_wall_median_ms": 0.08385686669498682,
+      "eager_wall_trials_ms": [
+        0.09771363335554877,
+        0.08163450002030004,
+        0.08385686669498682
+      ],
+      "extra_peak_bytes": 184320
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.007651200145483017,
+      "graph_min_ms": 0.007601600140333176,
+      "graph_max_ms": 0.007692799717187881,
+      "eager_wall_median_ms": 0.19571573332844613,
+      "eager_wall_trials_ms": [
+        0.2535254000273805,
+        0.19571573332844613,
+        0.18590186670432254
+      ],
+      "extra_peak_bytes": 187392
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0012047999538481236,
+      "graph_min_ms": 0.0011695999652147292,
+      "graph_max_ms": 0.0012400000356137752,
+      "eager_wall_median_ms": 0.05762376664885475,
+      "eager_wall_trials_ms": [
+        0.07130193328824437,
+        0.05762376664885475,
+        0.057149166665719044
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.002329600043594837,
+      "graph_min_ms": 0.002223999984562397,
+      "graph_max_ms": 0.0026383999735116958,
+      "eager_wall_median_ms": 0.17063139997238372,
+      "eager_wall_trials_ms": [
+        0.17063139997238372,
+        0.16431806670880178,
+        0.24000986668397672
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0013432000298053027,
+      "graph_min_ms": 0.0010928000323474408,
+      "graph_max_ms": 0.006166400015354156,
+      "eager_wall_median_ms": 0.08490003328915918,
+      "eager_wall_trials_ms": [
+        0.08490003328915918,
+        0.0852208666401566,
+        0.08217023332690587
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.002125599980354309,
+      "graph_min_ms": 0.0020735999569296837,
+      "graph_max_ms": 0.0021679999306797983,
+      "eager_wall_median_ms": 0.3603733666144156,
+      "eager_wall_trials_ms": [
+        0.3862667332820517,
+        0.31255590001819655,
+        0.3603733666144156
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0011024000123143196,
+      "graph_min_ms": 0.0010544000193476676,
+      "graph_max_ms": 0.01971679925918579,
+      "eager_wall_median_ms": 0.07505266664035541,
+      "eager_wall_trials_ms": [
+        0.09395336667997374,
+        0.07483469998987857,
+        0.07505266664035541
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0020584000274538996,
+      "graph_min_ms": 0.002022399939596653,
+      "graph_max_ms": 0.002151999995112419,
+      "eager_wall_median_ms": 0.3825640333161573,
+      "eager_wall_trials_ms": [
+        0.3916477333405055,
+        0.38042896667320747,
+        0.3825640333161573
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0011288000270724297,
+      "graph_min_ms": 0.0010703999549150466,
+      "graph_max_ms": 0.0025040000677108766,
+      "eager_wall_median_ms": 0.07896646663236122,
+      "eager_wall_trials_ms": [
+        0.07896646663236122,
+        0.08368196668016026,
+        0.07650009999148703
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        18432,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.002057600021362305,
+      "graph_min_ms": 0.002022399939596653,
+      "graph_max_ms": 0.0021487999707460403,
+      "eager_wall_median_ms": 0.3408336999806731,
+      "eager_wall_trials_ms": [
+        0.3408336999806731,
+        0.32808933331883355,
+        0.35737479999321903
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0037903999909758568,
+      "graph_min_ms": 0.003734400123357773,
+      "graph_max_ms": 0.0038384001702070234,
+      "eager_wall_median_ms": 0.026099166704322368,
+      "eager_wall_trials_ms": [
+        0.027518266688275617,
+        0.02535739998468974,
+        0.026099166704322368
+      ],
+      "extra_peak_bytes": 184320
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.008167200163006781,
+      "graph_min_ms": 0.00809599980711937,
+      "graph_max_ms": 0.008207999914884568,
+      "eager_wall_median_ms": 0.22357730000900725,
+      "eager_wall_trials_ms": [
+        0.255850233346185,
+        0.22357730000900725,
+        0.2018136666568656
+      ],
+      "extra_peak_bytes": 187392
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.001232000021263957,
+      "graph_min_ms": 0.0011552000418305398,
+      "graph_max_ms": 0.0026464000344276427,
+      "eager_wall_median_ms": 0.05505463335187718,
+      "eager_wall_trials_ms": [
+        0.05505463335187718,
+        0.052852566659566946,
+        0.05531913332864254
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.002279200032353401,
+      "graph_min_ms": 0.00220320001244545,
+      "graph_max_ms": 0.005976000055670738,
+      "eager_wall_median_ms": 0.2586098333267728,
+      "eager_wall_trials_ms": [
+        0.2587837333218583,
+        0.25146820001585485,
+        0.2586098333267728
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0011359999887645244,
+      "graph_min_ms": 0.001071999967098236,
+      "graph_max_ms": 0.002603200078010559,
+      "eager_wall_median_ms": 0.03804060003555302,
+      "eager_wall_trials_ms": [
+        0.03804060003555302,
+        0.04575726664673615,
+        0.03741419999035619
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0021271999925374986,
+      "graph_min_ms": 0.0020896000787615776,
+      "graph_max_ms": 0.016212800145149232,
+      "eager_wall_median_ms": 0.24443720000514682,
+      "eager_wall_trials_ms": [
+        0.281011633342132,
+        0.24443720000514682,
+        0.23313996668245332
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.001104000024497509,
+      "graph_min_ms": 0.001065600011497736,
+      "graph_max_ms": 0.0011983999982476234,
+      "eager_wall_median_ms": 0.078476833308135,
+      "eager_wall_trials_ms": [
+        0.07911106664929927,
+        0.07519836666081876,
+        0.078476833308135
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0020527999848127367,
+      "graph_min_ms": 0.0020207999274134635,
+      "graph_max_ms": 0.0021023999899625776,
+      "eager_wall_median_ms": 0.30830893338134047,
+      "eager_wall_trials_ms": [
+        0.3673079666744646,
+        0.29416493331761256,
+        0.30830893338134047
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0011272000148892403,
+      "graph_min_ms": 0.0010511999949812888,
+      "graph_max_ms": 0.002239999920129776,
+      "eager_wall_median_ms": 0.07592803334167304,
+      "eager_wall_trials_ms": [
+        0.0775905666766145,
+        0.07592803334167304,
+        0.0742380666755101
+      ],
+      "extra_peak_bytes": 36864
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6,
+        24,
+        128
+      ],
+      "input_stride": [
+        55296,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0020535999909043314,
+      "graph_min_ms": 0.002014400064945221,
+      "graph_max_ms": 0.0020975999534130095,
+      "eager_wall_median_ms": 0.31871896662778454,
+      "eager_wall_trials_ms": [
+        0.31871896662778454,
+        0.33167873334605247,
+        0.2915063333299865
+      ],
+      "extra_peak_bytes": 73728
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.24445840120315554,
+      "graph_min_ms": 0.24224960803985596,
+      "graph_max_ms": 0.29114880561828616,
+      "eager_wall_median_ms": 0.247559633377629,
+      "eager_wall_trials_ms": [
+        0.25793806665509084,
+        0.2471523333345734,
+        0.247559633377629
+      ],
+      "extra_peak_bytes": 125829120
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.5128072023391723,
+      "graph_min_ms": 0.49546399116516116,
+      "graph_max_ms": 0.5571055889129639,
+      "eager_wall_median_ms": 0.5245368333514003,
+      "eager_wall_trials_ms": [
+        0.5245368333514003,
+        0.518849266639639,
+        0.5350402333230401
+      ],
+      "extra_peak_bytes": 127926272
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.05780480206012726,
+      "graph_min_ms": 0.057633602619171144,
+      "graph_max_ms": 0.07439360022544861,
+      "eager_wall_median_ms": 0.06806933330760027,
+      "eager_wall_trials_ms": [
+        0.07017140002668991,
+        0.06806933330760027,
+        0.06168783335548748
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.1296720027923584,
+      "graph_min_ms": 0.12902239561080933,
+      "graph_max_ms": 0.15688639879226685,
+      "eager_wall_median_ms": 0.2333034999537631,
+      "eager_wall_trials_ms": [
+        0.23402639999403618,
+        0.19631726663646987,
+        0.2333034999537631
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.049288800358772276,
+      "graph_min_ms": 0.04815520048141479,
+      "graph_max_ms": 0.06759200096130372,
+      "eager_wall_median_ms": 0.08061226665934858,
+      "eager_wall_trials_ms": [
+        0.07100329997532147,
+        0.0930295666573026,
+        0.08061226665934858
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.12414479851722718,
+      "graph_min_ms": 0.12152320146560669,
+      "graph_max_ms": 0.16048799753189086,
+      "eager_wall_median_ms": 0.35214599999259616,
+      "eager_wall_trials_ms": [
+        0.35569400000288925,
+        0.35214599999259616,
+        0.32247826663175755
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.04924079924821853,
+      "graph_min_ms": 0.04797120094299316,
+      "graph_max_ms": 0.068995201587677,
+      "eager_wall_median_ms": 0.13288023334704727,
+      "eager_wall_trials_ms": [
+        0.14359453337723002,
+        0.12000013339275029,
+        0.13288023334704727
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.12360240221023559,
+      "graph_min_ms": 0.1215999960899353,
+      "graph_max_ms": 0.16038399934768677,
+      "eager_wall_median_ms": 0.35429976663484314,
+      "eager_wall_trials_ms": [
+        0.34954116666388774,
+        0.3772011999899405,
+        0.35429976663484314
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0489767998456955,
+      "graph_min_ms": 0.0479856014251709,
+      "graph_max_ms": 0.06645600199699402,
+      "eager_wall_median_ms": 0.1330274333061728,
+      "eager_wall_trials_ms": [
+        0.1330274333061728,
+        0.1200082000044252,
+        0.13366413331823424
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        12582912,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.12407039999961852,
+      "graph_min_ms": 0.12193280458450317,
+      "graph_max_ms": 0.1605712056159973,
+      "eager_wall_median_ms": 0.36529683329717955,
+      "eager_wall_trials_ms": [
+        0.3895335000076254,
+        0.36529683329717955,
+        0.3503133667133322
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.2475391983985901,
+      "graph_min_ms": 0.24529919624328614,
+      "graph_max_ms": 0.27714240550994873,
+      "eager_wall_median_ms": 0.25804310001452296,
+      "eager_wall_trials_ms": [
+        0.25804310001452296,
+        0.26510979999632883,
+        0.2503190333300154
+      ],
+      "extra_peak_bytes": 125829120
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.5158272027969361,
+      "graph_min_ms": 0.49741439819335936,
+      "graph_max_ms": 0.5595920085906982,
+      "eager_wall_median_ms": 0.5221092666639985,
+      "eager_wall_trials_ms": [
+        0.5362037999778598,
+        0.5221092666639985,
+        0.5111432000073061
+      ],
+      "extra_peak_bytes": 127926272
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.05868320167064667,
+      "graph_min_ms": 0.058520001173019406,
+      "graph_max_ms": 0.0754975974559784,
+      "eager_wall_median_ms": 0.06486223331497361,
+      "eager_wall_trials_ms": [
+        0.06433113333817649,
+        0.08489736665069358,
+        0.06486223331497361
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.13079119920730592,
+      "graph_min_ms": 0.13011519908905028,
+      "graph_max_ms": 0.1711359977722168,
+      "eager_wall_median_ms": 0.21280260001124893,
+      "eager_wall_trials_ms": [
+        0.2440493333779159,
+        0.20012976668416135,
+        0.21280260001124893
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.04731200039386749,
+      "graph_min_ms": 0.04663040041923523,
+      "graph_max_ms": 0.06482719779014587,
+      "eager_wall_median_ms": 0.08378989999376547,
+      "eager_wall_trials_ms": [
+        0.08378989999376547,
+        0.07920856666411662,
+        0.09377416663483018
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.11841280460357666,
+      "graph_min_ms": 0.11643359661102295,
+      "graph_max_ms": 0.14446239471435546,
+      "eager_wall_median_ms": 0.3498463667104564,
+      "eager_wall_trials_ms": [
+        0.3498463667104564,
+        0.35651856663510745,
+        0.3451830999741408
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.047358399629592894,
+      "graph_min_ms": 0.04637599885463715,
+      "graph_max_ms": 0.06467199921607972,
+      "eager_wall_median_ms": 0.12952533331069085,
+      "eager_wall_trials_ms": [
+        0.12952533331069085,
+        0.1346659000167468,
+        0.12932516668418734
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.11804000139236451,
+      "graph_min_ms": 0.11632959842681885,
+      "graph_max_ms": 0.13619199991226197,
+      "eager_wall_median_ms": 0.3809861666619933,
+      "eager_wall_trials_ms": [
+        0.3809861666619933,
+        0.39690620002753957,
+        0.3660936666468236
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.04721840023994446,
+      "graph_min_ms": 0.04616160094738007,
+      "graph_max_ms": 0.06441760063171387,
+      "eager_wall_median_ms": 0.1290158666355031,
+      "eager_wall_trials_ms": [
+        0.13406613331123177,
+        0.1290158666355031,
+        0.12222770001244498
+      ],
+      "extra_peak_bytes": 25165824
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        4096,
+        24,
+        128
+      ],
+      "input_stride": [
+        37748736,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.11847440004348755,
+      "graph_min_ms": 0.11651359796524048,
+      "graph_max_ms": 0.1468176007270813,
+      "eager_wall_median_ms": 0.4179154666417162,
+      "eager_wall_trials_ms": [
+        0.4434371333142432,
+        0.4179154666417162,
+        0.4101795332947707
+      ],
+      "extra_peak_bytes": 50331648
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.5257872104644776,
+      "graph_min_ms": 0.5080063819885254,
+      "graph_max_ms": 0.5525023937225342,
+      "eager_wall_median_ms": 0.5246295667044857,
+      "eager_wall_trials_ms": [
+        0.5097412666752158,
+        0.528316466625256,
+        0.5246295667044857
+      ],
+      "extra_peak_bytes": 211630080
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 1.067212772369385,
+      "graph_min_ms": 1.0389776229858398,
+      "graph_max_ms": 1.0976528167724608,
+      "eager_wall_median_ms": 1.0610819333426964,
+      "eager_wall_trials_ms": [
+        1.0702722999970622,
+        1.0479776333644015,
+        1.0610819333426964
+      ],
+      "extra_peak_bytes": 215157248
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.11197760105133056,
+      "graph_min_ms": 0.11102559566497802,
+      "graph_max_ms": 0.13988800048828126,
+      "eager_wall_median_ms": 0.12250546666715915,
+      "eager_wall_trials_ms": [
+        0.11638536664880424,
+        0.13100786666958206,
+        0.12250546666715915
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.22392239570617678,
+      "graph_min_ms": 0.22292959690093994,
+      "graph_max_ms": 0.26030399799346926,
+      "eager_wall_median_ms": 0.24186149997452353,
+      "eager_wall_trials_ms": [
+        0.24518829995940905,
+        0.23021603331775017,
+        0.24186149997452353
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.10719599723815917,
+      "graph_min_ms": 0.10527039766311645,
+      "graph_max_ms": 0.123745596408844,
+      "eager_wall_median_ms": 0.10772686667526916,
+      "eager_wall_trials_ms": [
+        0.10711883332987782,
+        0.10772686667526916,
+        0.10832119999880281
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.2253783941268921,
+      "graph_min_ms": 0.21142399311065674,
+      "graph_max_ms": 0.25633440017700193,
+      "eager_wall_median_ms": 0.29183916667534504,
+      "eager_wall_trials_ms": [
+        0.29183916667534504,
+        0.31016709999676095,
+        0.27477109997562366
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.10732319951057434,
+      "graph_min_ms": 0.10543199777603149,
+      "graph_max_ms": 0.12232480049133301,
+      "eager_wall_median_ms": 0.11885433329249888,
+      "eager_wall_trials_ms": [
+        0.11160320003303544,
+        0.11885433329249888,
+        0.148421633275575
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.22528079748153684,
+      "graph_min_ms": 0.21176319122314452,
+      "graph_max_ms": 0.24576001167297362,
+      "eager_wall_median_ms": 0.41978080001475365,
+      "eager_wall_trials_ms": [
+        0.4051902666712219,
+        0.42735956667456776,
+        0.41978080001475365
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.10738959908485413,
+      "graph_min_ms": 0.10557440519332886,
+      "graph_max_ms": 0.1310464024543762,
+      "eager_wall_median_ms": 0.1100779333379857,
+      "eager_wall_trials_ms": [
+        0.11928496666466042,
+        0.10885563333431492,
+        0.1100779333379857
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        21163008,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.21419600248336793,
+      "graph_min_ms": 0.2123471975326538,
+      "graph_max_ms": 0.24199841022491456,
+      "eager_wall_median_ms": 0.42007340004298993,
+      "eager_wall_trials_ms": [
+        0.4409272999813159,
+        0.42007340004298993,
+        0.40205200002674246
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.5282040119171143,
+      "graph_min_ms": 0.5110288143157959,
+      "graph_max_ms": 0.5885712146759033,
+      "eager_wall_median_ms": 0.547415499992591,
+      "eager_wall_trials_ms": [
+        0.547415499992591,
+        0.5160773666527044,
+        0.5478402666388623
+      ],
+      "extra_peak_bytes": 211630080
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 1.0694655895233156,
+      "graph_min_ms": 1.0273263931274415,
+      "graph_max_ms": 1.0971407890319824,
+      "eager_wall_median_ms": 1.0825770000034634,
+      "eager_wall_trials_ms": [
+        1.0599153666286536,
+        1.0825770000034634,
+        1.0843000333503976
+      ],
+      "extra_peak_bytes": 215157248
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.11232720017433166,
+      "graph_min_ms": 0.11165920495986939,
+      "graph_max_ms": 0.13819040060043336,
+      "eager_wall_median_ms": 0.12610396667393312,
+      "eager_wall_trials_ms": [
+        0.12610396667393312,
+        0.11783170005704353,
+        0.1361641000282058
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.22485519647598268,
+      "graph_min_ms": 0.2238991975784302,
+      "graph_max_ms": 0.2728208065032959,
+      "eager_wall_median_ms": 0.2468356666213367,
+      "eager_wall_trials_ms": [
+        0.2468356666213367,
+        0.25081796666199807,
+        0.23804076669572774
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.10744640231132507,
+      "graph_min_ms": 0.10558079481124878,
+      "graph_max_ms": 0.12357280254364014,
+      "eager_wall_median_ms": 0.10917900005248764,
+      "eager_wall_trials_ms": [
+        0.108484466666899,
+        0.11029286670236615,
+        0.10917900005248764
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.21478239297866822,
+      "graph_min_ms": 0.21232800483703612,
+      "graph_max_ms": 0.2458656072616577,
+      "eager_wall_median_ms": 0.3064794333719571,
+      "eager_wall_trials_ms": [
+        0.3064794333719571,
+        0.29039823330094805,
+        0.3166810000038822
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.10760959982872009,
+      "graph_min_ms": 0.10598560571670532,
+      "graph_max_ms": 0.13379520177841187,
+      "eager_wall_median_ms": 0.11580663334219328,
+      "eager_wall_trials_ms": [
+        0.11580663334219328,
+        0.12037279999882837,
+        0.10920460002428929
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.2147576093673706,
+      "graph_min_ms": 0.21242880821228027,
+      "graph_max_ms": 0.2415616035461426,
+      "eager_wall_median_ms": 0.42939333337320323,
+      "eager_wall_trials_ms": [
+        0.4258981332895928,
+        0.45155456670424127,
+        0.42939333337320323
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.1074351966381073,
+      "graph_min_ms": 0.10590399503707885,
+      "graph_max_ms": 0.13916159868240358,
+      "eager_wall_median_ms": 0.11420323335187277,
+      "eager_wall_trials_ms": [
+        0.11669516667704254,
+        0.11420323335187277,
+        0.11166136664542137
+      ],
+      "extra_peak_bytes": 42326016
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6889,
+        24,
+        128
+      ],
+      "input_stride": [
+        63489024,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.21468960046768187,
+      "graph_min_ms": 0.21285600662231446,
+      "graph_max_ms": 0.24125280380249023,
+      "eager_wall_median_ms": 0.4036813999846345,
+      "eager_wall_trials_ms": [
+        0.40494703334843507,
+        0.4027272333284297,
+        0.4036813999846345
+      ],
+      "extra_peak_bytes": 84652032
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.44991838932037354,
+      "graph_min_ms": 0.4312863826751709,
+      "graph_max_ms": 0.4923136234283447,
+      "eager_wall_median_ms": 0.4565672999888193,
+      "eager_wall_trials_ms": [
+        0.46419443333434174,
+        0.4565672999888193,
+        0.43665636664324364
+      ],
+      "extra_peak_bytes": 185303040
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.9010960102081299,
+      "graph_min_ms": 0.8724479675292969,
+      "graph_max_ms": 0.9627167701721191,
+      "eager_wall_median_ms": 0.9057802333093908,
+      "eager_wall_trials_ms": [
+        0.9057802333093908,
+        0.9399917666087276,
+        0.8861856666650662
+      ],
+      "extra_peak_bytes": 188391424
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09797199964523315,
+      "graph_min_ms": 0.0971664011478424,
+      "graph_max_ms": 0.12116639614105225,
+      "eager_wall_median_ms": 0.10149109997049284,
+      "eager_wall_trials_ms": [
+        0.10149109997049284,
+        0.10145610000715048,
+        0.10150603332779913
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.19708160161972046,
+      "graph_min_ms": 0.19586080312728882,
+      "graph_max_ms": 0.2340928077697754,
+      "eager_wall_median_ms": 0.23985833334639514,
+      "eager_wall_trials_ms": [
+        0.2599421666673152,
+        0.21093993333731,
+        0.23985833334639514
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09365599751472473,
+      "graph_min_ms": 0.09192960262298584,
+      "graph_max_ms": 0.1098512053489685,
+      "eager_wall_median_ms": 0.09574563337082509,
+      "eager_wall_trials_ms": [
+        0.10015153335795428,
+        0.09574563337082509,
+        0.09447829997952795
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.18496879935264587,
+      "graph_min_ms": 0.18224799633026123,
+      "graph_max_ms": 0.20439839363098145,
+      "eager_wall_median_ms": 0.27221376664480584,
+      "eager_wall_trials_ms": [
+        0.30083316663270426,
+        0.2585961666530541,
+        0.27221376664480584
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09344720244407653,
+      "graph_min_ms": 0.09223840236663819,
+      "graph_max_ms": 0.10959680080413818,
+      "eager_wall_median_ms": 0.10214689997762132,
+      "eager_wall_trials_ms": [
+        0.10214689997762132,
+        0.09693080003974804,
+        0.10462310001457809
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.18451839685440063,
+      "graph_min_ms": 0.1817072033882141,
+      "graph_max_ms": 0.2279695987701416,
+      "eager_wall_median_ms": 0.38619293330460397,
+      "eager_wall_trials_ms": [
+        0.4084413000479496,
+        0.38619293330460397,
+        0.3539249666573596
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09362159967422486,
+      "graph_min_ms": 0.09215999841690063,
+      "graph_max_ms": 0.1108240008354187,
+      "eager_wall_median_ms": 0.10443516669814319,
+      "eager_wall_trials_ms": [
+        0.10080796667656007,
+        0.10632786670612404,
+        0.10443516669814319
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "contiguous",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        18530304,
+        3072,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.18458240032196044,
+      "graph_min_ms": 0.18273439407348632,
+      "graph_max_ms": 0.21199359893798828,
+      "eager_wall_median_ms": 0.3521715333287527,
+      "eager_wall_trials_ms": [
+        0.3573932999643148,
+        0.32707516663019004,
+        0.3521715333287527
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.44979281425476075,
+      "graph_min_ms": 0.4314896106719971,
+      "graph_max_ms": 0.48181757926940916,
+      "eager_wall_median_ms": 0.4543782333106113,
+      "eager_wall_trials_ms": [
+        0.4543782333106113,
+        0.4393956000058097,
+        0.4658290666460137
+      ],
+      "extra_peak_bytes": 185303040
+    },
+    {
+      "backend": "upstream_complex",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.9027584075927735,
+      "graph_min_ms": 0.8707056045532227,
+      "graph_max_ms": 0.9478416442871094,
+      "eager_wall_median_ms": 0.9098065333091654,
+      "eager_wall_trials_ms": [
+        0.92959719998665,
+        0.9058386999943954,
+        0.9098065333091654
+      ],
+      "extra_peak_bytes": 188391424
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.0960999995470047,
+      "graph_min_ms": 0.09566079974174499,
+      "graph_max_ms": 0.1124384045600891,
+      "eager_wall_median_ms": 0.10085823329670045,
+      "eager_wall_trials_ms": [
+        0.10085823329670045,
+        0.1027275666880693,
+        0.10056819998377857
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "cuda",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.19612080454826356,
+      "graph_min_ms": 0.19543039798736572,
+      "graph_max_ms": 0.22353920936584473,
+      "eager_wall_median_ms": 0.21260113335301867,
+      "eager_wall_trials_ms": [
+        0.21260113335301867,
+        0.23863556671130937,
+        0.20749836670195995
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "triton",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09310480058193207,
+      "graph_min_ms": 0.09154719710350037,
+      "graph_max_ms": 0.1167631983757019,
+      "eager_wall_median_ms": 0.10304526670855314,
+      "eager_wall_trials_ms": [
+        0.09351696662633913,
+        0.10913006669094709,
+        0.10304526670855314
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "triton",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.18827759623527526,
+      "graph_min_ms": 0.18518240451812745,
+      "graph_max_ms": 0.21609120368957518,
+      "eager_wall_median_ms": 0.33319789999950444,
+      "eager_wall_trials_ms": [
+        0.33319789999950444,
+        0.3456709999833644,
+        0.3028138333320385
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09317199885845184,
+      "graph_min_ms": 0.09170079827308655,
+      "graph_max_ms": 0.11182240247726441,
+      "eager_wall_median_ms": 0.13384176666780453,
+      "eager_wall_trials_ms": [
+        0.12049876671274735,
+        0.15613870000379393,
+        0.13384176666780453
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "flash_default",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.18772000074386597,
+      "graph_min_ms": 0.18648799657821655,
+      "graph_max_ms": 0.2264847993850708,
+      "eager_wall_median_ms": 0.3818082333358082,
+      "eager_wall_trials_ms": [
+        0.44640976666414645,
+        0.3674994000296768,
+        0.3818082333358082
+      ],
+      "extra_peak_bytes": 74809344
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.09322640001773834,
+      "graph_min_ms": 0.09153280258178711,
+      "graph_max_ms": 0.11082240343093872,
+      "eager_wall_median_ms": 0.13263150000663396,
+      "eager_wall_trials_ms": [
+        0.13263150000663396,
+        0.13320910002221353,
+        0.12726086667195582
+      ],
+      "extra_peak_bytes": 37060608
+    },
+    {
+      "backend": "flash_no_fma",
+      "phase": "forward_backward",
+      "layout": "packed_qkv",
+      "shape": [
+        1,
+        6032,
+        24,
+        128
+      ],
+      "input_stride": [
+        55590912,
+        9216,
+        128,
+        1
+      ],
+      "graph_median_ms": 0.18817280530929564,
+      "graph_min_ms": 0.18564800024032593,
+      "graph_max_ms": 0.21453280448913575,
+      "eager_wall_median_ms": 0.3810788333794335,
+      "eager_wall_trials_ms": [
+        0.3668230666638313,
+        0.3885829333739821,
+        0.3810788333794335
+      ],
+      "extra_peak_bytes": 74809344
+    }
+  ]
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-contracts.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-contracts.txt
new file mode 100644
index 000000000..de9819e68
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-contracts.txt
@@ -0,0 +1,20 @@
+...........ssssssssssss.ss.ss.ss...ssssss.ss......ssssssssssss.ss...ssss [ 66%]
+s.s.ssssss...........................                                    [100%]
+=========================== short test summary info ============================
+SKIPPED [12] tests/ops/rope/test_multi_axis_rope.py:79: NVIDIA CUDA required
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:100: NVIDIA CUDA required
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:110: NVIDIA CUDA required
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:133: NVIDIA CUDA required
+SKIPPED [6] tests/ops/rope/test_multi_axis_rope.py:148: NVIDIA CUDA required
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:158: NVIDIA CUDA required
+SKIPPED [12] tests/ops/rope/test_multi_axis_rope.py:174: NVIDIA CUDA required
+SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:217: native CUDA multi_axis_rope extension required
+SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:230: NVIDIA CUDA required
+SKIPPED [3] tests/ops/rope/test_multi_axis_rope.py:255: NVIDIA CUDA required
+SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:315: NVIDIA CUDA required
+SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:336: requires two NVIDIA GPUs
+SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:355: CUDA context requires CUDA
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:368: NVIDIA CUDA required
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:399: NVIDIA CUDA required
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:427: NVIDIA CUDA required
+57 passed, 52 skipped in 1.70s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-environment.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-environment.json
new file mode 100644
index 000000000..e41fd057f
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-environment.json
@@ -0,0 +1,7 @@
+{
+  "torch": "2.9.1+cpu",
+  "cuda": null,
+  "package": "/home/dandan/venvs/rlk-rope-cpu/lib/python3.12/site-packages/rl_engine/__init__.py",
+  "reference_forward_backward": true,
+  "passed": true
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-extension.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-extension.json
new file mode 100644
index 000000000..420aa5da8
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-extension.json
@@ -0,0 +1,22 @@
+{
+  "profile": "ordinary complete _C; no RL_KERNEL_ROPE_ONLY",
+  "module": "/home/dandan/src/RL-Kernel/rl_engine/_C.cpython-312-x86_64-linux-gnu.so",
+  "binary_sha256": "2661e7188984c7d78acae1ef5186a3e90ec4aa8bb6f457522b23435a1f24f94a",
+  "sources": [
+    "csrc/bindings/ops.cpp",
+    "csrc/cuda/logprob/fused_logp_kernel.cu",
+    "csrc/cuda/logprob/deterministic_logp_kernel.cu",
+    "csrc/cuda/gemm/det_gemm_kernel.cu",
+    "csrc/cuda/norm/rmsnorm.cu",
+    "csrc/cuda/activation/activation.cu",
+    "csrc/cuda/attention/deterministic_attention.cu",
+    "csrc/cuda/rope/multi_axis_rope.cu",
+    "csrc/cuda/collectives/deterministic_collective.cu",
+    "csrc/cuda/attention/prefix_shared_attention.cu"
+  ],
+  "cuda_home": "/usr/local/cuda-12.8",
+  "driver_link_search": "/usr/lib/wsl/lib",
+  "torch": "2.9.1+cu128",
+  "architecture": "12.0",
+  "max_jobs": "2"
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-inplace-build.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-inplace-build.txt
new file mode 100644
index 000000000..d61fb7517
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-inplace-build.txt
@@ -0,0 +1,20 @@
+/home/dandan/src/RL-Kernel/.venv/lib/python3.12/site-packages/setuptools/config/_apply_pyprojecttoml.py:82: SetuptoolsDeprecationWarning: `project.license` as a TOML table is deprecated
+!!
+
+        ********************************************************************************
+        Please use a simple string containing a SPDX expression for `project.license`. You can also use `project.license-files`. (Both options available on setuptools>=77.0.0).
+
+        This deprecation is overdue, please update your project and remove deprecated
+        calls to avoid build errors in the future.
+
+        See https://packaging.python.org/en/latest/guides/writing-pyproject-toml/#license for details.
+        ********************************************************************************
+
+!!
+  corresp(dist, value, root_dir)
+running build_ext
+W1028 18:37:59.459000 9791 .venv/lib/python3.12/site-packages/torch/utils/cpp_extension.py:531] There are no x86_64-linux-gnu-g++ version bounds defined for CUDA version 12.8
+building 'rl_engine._C' extension
+ninja: no work to do.
+x86_64-linux-gnu-g++ -fno-strict-overflow -Wsign-compare -DNDEBUG -g -O2 -Wall -shared -Wl,-O1 -Wl,-Bsymbolic-functions /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/bindings/ops.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/activation/activation.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/attention/deterministic_attention.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/attention/prefix_shared_attention.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/collectives/deterministic_collective.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/gemm/det_gemm_kernel.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/logprob/deterministic_logp_kernel.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/logprob/fused_logp_kernel.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/norm/rmsnorm.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/rope/multi_axis_rope.o -L/home/dandan/src/RL-Kernel/.venv/lib/python3.12/site-packages/torch/lib -L/usr/local/cuda-12.8/lib64 -L/usr/lib/x86_64-linux-gnu -lc10 -ltorch -ltorch_cpu -ltorch_python -lcudart -lc10_cuda -ltorch_cuda -o build/lib.linux-x86_64-cpython-312/rl_engine/_C.cpython-312-x86_64-linux-gnu.so -Wl,-rpath,$ORIGIN/../torch/lib -lcuda
+copying build/lib.linux-x86_64-cpython-312/rl_engine/_C.cpython-312-x86_64-linux-gnu.so -> rl_engine
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-wheel-build.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-wheel-build.txt
new file mode 100644
index 000000000..592c254f2
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-wheel-build.txt
@@ -0,0 +1,10 @@
+Processing /home/dandan/src/RL-Kernel
+  Preparing metadata (pyproject.toml): started
+  Preparing metadata (pyproject.toml): finished with status 'done'
+Building wheels for collected packages: RL-Kernel
+  Building wheel for RL-Kernel (pyproject.toml): started
+  Building wheel for RL-Kernel (pyproject.toml): still running...
+  Building wheel for RL-Kernel (pyproject.toml): finished with status 'done'
+  Created wheel for RL-Kernel: filename=rl_kernel-0.1.0-cp312-cp312-linux_x86_64.whl size=5020592 sha256=32a39cd9cbb30b87816f6584096238fc2e4b0803eb387a2d3e1e80f3d1b40d8a
+  Stored in directory: /tmp/pip-ephem-wheel-cache-8gtts1_e/wheels/de/2c/1f/c224e9a9cc43d1b39ee96ba5703938f5f1d66da44d8695a43a
+Successfully built RL-Kernel
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/gpu-environment.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/gpu-environment.json
new file mode 100644
index 000000000..e40096583
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/gpu-environment.json
@@ -0,0 +1,15 @@
+{
+  "scope": "Qwen-Image multi_axis_rope operator; ordinary full _C build",
+  "python": "3.12.3",
+  "platform": "Linux-6.18.54.1-microsoft-standard-WSL2-x86_64-with-glibc2.39",
+  "torch": "2.9.1+cu128",
+  "cuda": "12.8",
+  "triton": "3.5.1",
+  "gpu": "NVIDIA GeForce RTX 5080",
+  "capability": [
+    12,
+    0
+  ],
+  "native_library": "/home/dandan/src/RL-Kernel/rl_engine/_C.cpython-312-x86_64-linux-gnu.so",
+  "native_library_sha256": "2661e7188984c7d78acae1ef5186a3e90ec4aa8bb6f457522b23435a1f24f94a"
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/installed-wheel.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/installed-wheel.json
new file mode 100644
index 000000000..28d3066fc
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/installed-wheel.json
@@ -0,0 +1,38 @@
+{
+  "package": "/home/dandan/src/RL-Kernel/.venv/lib/python3.12/site-packages/rl_engine",
+  "companion_library": null,
+  "full_extension_library": "/home/dandan/src/RL-Kernel/.venv/lib/python3.12/site-packages/rl_engine/_C.cpython-312-x86_64-linux-gnu.so",
+  "native_dispatcher_registered": true,
+  "torch": "2.9.1+cu128",
+  "gpu": "NVIDIA GeForce RTX 5080",
+  "license_present": true,
+  "packed_qkv_stride": [
+    17280,
+    1152,
+    128,
+    1
+  ],
+  "checks": {
+    "cuda_eager": "passed",
+    "cuda_compiled": "passed",
+    "cuda_trace": {
+      "contract": "qwenimage-interleaved-rope-fp32-v1",
+      "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
+      "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
+      "layout": "BSHD, adjacent channel pairs",
+      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+    },
+    "triton_eager": "passed",
+    "triton_compiled": "passed",
+    "triton_trace": {
+      "contract": "qwenimage-interleaved-rope-fp32-v1",
+      "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
+      "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
+      "layout": "BSHD, adjacent channel pairs",
+      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+    }
+  },
+  "passed": true
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/layout.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/layout.txt
new file mode 100644
index 000000000..8c81ecbf5
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/layout.txt
@@ -0,0 +1,2 @@
+.....................................                                    [100%]
+37 passed in 3.81s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/manifest.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/manifest.json
new file mode 100644
index 000000000..3acb62a79
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/manifest.json
@@ -0,0 +1,58 @@
+{
+  "profile": "Ubuntu 24.04.5 / WSL2 / RTX 5080 SM120 / ordinary complete _C",
+  "numerical_source_commit": "ac17ab87e8d93dabe93dfdc134e9853d47de7af1",
+  "source_sha256": {
+    ".github/workflows/qwenimage-rope.yml": "39e7e02a1f1f75603482e6acce9eb99ba8b5c5260ec6043271714061843651c3",
+    "MANIFEST.in": "87e59e6c40c67e54ae9bcd057957ac7d7823067cb455cbc4dff45b1ec88bdb3c",
+    "benchmarks/operators/rope/benchmark_multi_axis_rope.py": "6961a55803fbd4e566bd060558fb59e658e7f1f704e61ca97f912e5d16b55884",
+    "build_tools/extensions.py": "3723f14a3fb9307889790af15b5e3f360c4c46e12b8fef22423070e9660f70c5",
+    "ci/scripts/run_qwenimage_rope_ws1.sh": "8e1e12c1aad9231f215505a172b3a5416c54e7cbfb8856f5ff00d225be7bfa95",
+    "csrc/bindings/ops.cpp": "10b76a5e236c0dc3715f9703e28c057f404f21f2a9802245d5221c4997b2c3f4",
+    "csrc/cuda/activation/activation.cu": "35ef0cefcaa6167aec7a901fd19a2f6c90a8b9f01c2e786187aacdeee10c4b63",
+    "csrc/cuda/attention/deterministic_attention.cu": "99b44addd61b999a5fb6a04ca227ed8814d454368f361a8fd032e83d4d83a561",
+    "csrc/cuda/attention/prefix_shared_attention.cu": "218ab9647cd274f7a4d8246f8df8e875c5cf431a5a73f252e946b10c74a2d63f",
+    "csrc/cuda/collectives/deterministic_collective.cu": "9e114d5aae7a73a912e4e4afaa23a74126bef5557ba3b54832da0c0cdf9d97a2",
+    "csrc/cuda/gemm/det_gemm_kernel.cu": "08d08b0a32900fd9590d47f04819b34a2d37cbf53996aa5ed1373d88eae29f33",
+    "csrc/cuda/logprob/deterministic_logp_kernel.cu": "271399040a0489a3cabd9e0c06b64fa8b1bac92ab08710fce8922bcb1ff2b04f",
+    "csrc/cuda/logprob/fused_logp_kernel.cu": "ea67bcd307f0c63753397225543d333f58cf3d66a0ae69b643499d1e6729d8e7",
+    "csrc/cuda/norm/rmsnorm.cu": "fbb765563dbe9c063b35f38354321eaf522cf62ba03addf327154395db9683b9",
+    "csrc/cuda/rope/multi_axis_rope.cu": "f5db5cf5d4962053194455d9cde8cf91cf4e824bad7abf72e8410a404bcb0e44",
+    "docs/operators/multi-axis-rope.md": "1fb3859a9a2639aae809ece52a785737e9966098c24a08626b9240dad0d6cbf6",
+    "pyproject.toml": "c2d948784b379a0a828bbf3d8cfd9e0e7398a1ee5694fe8298509514684b2709",
+    "rl_engine/backends/cuda/rope/multi_axis_rope.py": "d779d9b60ce5d1d6651925f723049930d898024fdf0548ce29a2019b84753d65",
+    "rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py": "cf0eebf315830a4178053a71a77e650b58aa85bba1d1a08aff7677350af755d9",
+    "rl_engine/backends/shared/triton/rope/multi_axis_rope.py": "f08ff90627c66e15e9c558f8a68f132434a741a76997b3800bfbf73dc8940e9e",
+    "rl_engine/ops/rope/_registration.py": "add0375d1639444908091d01e8d29781f7a975beffec64b77bfaa81a5b9a0ae3",
+    "rl_engine/ops/rope/multi_axis_rope.py": "07bcff1160f22c1ff6730be731206b2f4a153e2fada26afc93fc884e780568cb",
+    "rl_engine/reference/rope/multi_axis_rope.py": "c9a94730047f51659f7aa0a443e88d6ce7d03d00d6813a9aed548cafbf6347bf",
+    "rl_engine/validation/operators/qwen_image_rope.py": "c3ddfc732f32de3c633de16064810d67202254ea78fd8107d558b471bb15a37f",
+    "tests/ops/rope/test_multi_axis_rope.py": "c9ff3eb6e6e013243fa741f853debb256fdac6c1d3a3820eb7de764a830e240c",
+    "tools/validation/operators/check_multi_axis_rope_install.py": "f295fc8ac5878ac13b6255c2f0be83b66d34ff2ad2f75111bb8b633508a865a7"
+  },
+  "evidence_sha256": {
+    "benchmark-bfloat16.json": "e9c1e4b1085623d84800656c179e4a0e86d70a0ef7fc9ed92f9e6ccc50f0f491",
+    "cpu-contracts.txt": "9518f53a6efc603bca1c52f160e5430f1006bb2abc9a7ef57694dae6feac5d78",
+    "cpu-environment.json": "4dfc68f3d76af05a050d830840b4a6b8c8fa0a4c12c29d11d632a56a57a2d4c6",
+    "full-extension.json": "8fae782e3c5e6aae59de6fe6b4e7ae185729e49d2282dbd0cb3b9ca08c0274cd",
+    "full-inplace-build.txt": "04d5ef122650ab305847d58850a6c3b44438eacf97920492e1eb1c5580e9a21e",
+    "full-wheel-build.txt": "a7883dfb1211798c869ede1ee0b6f1dfb32edd7b27b359ee18c6e5ab9fe449d5",
+    "gpu-environment.json": "bf808ab535d90b9b52dfcfb86543809e7d412cb48418677923cad22783bb9b8e",
+    "installed-wheel.json": "c726717106f30336e0a541e4a38d1f1b3858ed5dd262516c86618d4511a43554",
+    "layout.txt": "daef0733010cdad31123cb6b8cb041fe6c626319d527a5d97c88d1b958479039",
+    "memcheck.txt": "1a0c572f8f0d18e3699c847dbb69b345e4fd19200cabda3a39385d588cff0f10",
+    "operator-cuda.json": "fd95922bac2c5921744a8e601e9a267e51b7b7d7d44d94745f53964c2da2feb1",
+    "operator-tests.xml": "0ed98ddf85e34ba87f41465be1db4f417997e82b50cc5f8f5a83bd00f1f1543e",
+    "operator-triton.json": "7791aa1dfa68652dd66e3c332d8f036e9e21851e1df7fd41a090d594e22b06ff",
+    "package.json": "3f75c06fdee58afb1cdd9a2a611b432c8be1960fa640d96263c86b87423fe7bd",
+    "precommit.txt": "7cf90e06cf73194271c176ebc2d147e64cf8f5ed522d8e0e03093a4647f63759",
+    "python-packages.json": "617f8cef5a24f3b2a683e70e0b491e47db6d24a7ca848e5b3e461e87993a7201",
+    "quality.txt": "8826f0e406946a923400d792dd60d0f545b32f9fcffa3a148268c969e30e14c0",
+    "README.md": "af365e658fe4875df643704b2495e1f115c6c889f5d449880233d49e511040ac",
+    "regression.txt": "039cd9813a2ea54bfc5b7750fc0555b54d2bf30b09dbca2eba4a0403578ab9b0",
+    "tests-and-ws1.txt": "68397e27144c9244743dbcff39ffdca8a77a30eaf1e8180b2ca84dfc1216d633",
+    "upstream.json": "04e5472a82477a305b996ccbb420d0f3cc7b39af4053c867d37916a9bf5de51a",
+    "ws1-cuda.json": "7302e11f798816178622393095858871e2dc53ab925509eb130196a40b3c18ad",
+    "ws1-triton.json": "b21f985fbdff7e9b14f689d5eac25c9df5d6d3b0189900085837d5fc400f9d38"
+  },
+  "hash_mode": "SHA256 of recorded Linux file bytes; text is LF"
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/memcheck.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/memcheck.txt
new file mode 100644
index 000000000..915746141
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/memcheck.txt
@@ -0,0 +1,4 @@
+========= COMPUTE-SANITIZER
+..................................                                       [100%]
+34 passed, 48 deselected in 3.49s
+========= ERROR SUMMARY: 0 errors
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-cuda.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-cuda.json
new file mode 100644
index 000000000..801312e4a
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-cuda.json
@@ -0,0 +1,70 @@
+{
+  "suite_name": "multi_axis_rope",
+  "total_candidates": 1,
+  "passed_candidates": 1,
+  "pass_rate": 1.0,
+  "passed": true,
+  "candidates": [
+    {
+      "candidate_name": "cuda-multi_axis_rope",
+      "backend": "cuda",
+      "total_outputs": 2,
+      "passed_outputs": 2,
+      "pass_rate": 1.0,
+      "passed": true,
+      "cases": [
+        {
+          "case_name": "multi_axis_rope-torch.bfloat16-2x7x3x128",
+          "dtype": "torch.bfloat16",
+          "op_class": "elementwise",
+          "passed": true,
+          "outputs": [
+            {
+              "output_index": 0,
+              "shape": [
+                2,
+                7,
+                3,
+                128
+              ],
+              "candidate_dtype": "torch.bfloat16",
+              "gold_dtype": "torch.float32",
+              "atol": 0.02,
+              "rtol": 0.016,
+              "max_abs_error": 0.007704973220825195,
+              "mean_abs_error": 0.0006018930580466986,
+              "max_rel_error": 0.003828622167930007,
+              "passed": true,
+              "judgment": "forward_accuracy",
+              "comparison_lhs_role": "bf16_candidate",
+              "comparison_rhs_role": "fp32_reference",
+              "message": ""
+            },
+            {
+              "output_index": 1,
+              "shape": [
+                2,
+                7,
+                3,
+                128
+              ],
+              "candidate_dtype": "torch.bfloat16",
+              "gold_dtype": "torch.bfloat16",
+              "atol": 0.02,
+              "rtol": 0.016,
+              "max_abs_error": 0.0,
+              "mean_abs_error": 0.0,
+              "max_rel_error": 0.0,
+              "passed": true,
+              "judgment": "gradient_accuracy",
+              "comparison_lhs_role": "bf16_candidate",
+              "comparison_rhs_role": "fp32_reference",
+              "message": "gradient:x"
+            }
+          ]
+        }
+      ],
+      "backend_provenance": null
+    }
+  ]
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-tests.xml b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-tests.xml
new file mode 100644
index 000000000..9e26942b9
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-tests.xml
@@ -0,0 +1 @@
+<?xml version="1.0" encoding="utf-8"?><testsuites name="pytest tests"><testsuite name="pytest" errors="0" failures="0" skipped="1" tests="82" time="17.488" timestamp="2026-10-28T18:40:27.472020+08:00" hostname="DESKTOP-J18U1QU"><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_positions_and_axis_frequencies[2-4]" time="0.009" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_positions_and_axis_frequencies[3-5]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_positions_and_axis_frequencies[64-64]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_positions_and_axis_frequencies[83-83]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_positions_and_axis_frequencies[104-58]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[reference-False-dtype0]" time="1.137" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[reference-False-dtype1]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[reference-False-dtype2]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[reference-True-dtype0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[reference-True-dtype1]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[reference-True-dtype2]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[cuda-False-dtype0]" time="3.286" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[cuda-False-dtype1]" time="0.003" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[cuda-False-dtype2]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[cuda-True-dtype0]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[cuda-True-dtype1]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[cuda-True-dtype2]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[triton-False-dtype0]" time="0.900" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[triton-False-dtype1]" time="0.137" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[triton-False-dtype2]" time="0.144" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[triton-True-dtype0]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[triton-True-dtype1]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[triton-True-dtype2]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_adjacent_pair_layout_not_rotate_half[reference]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_adjacent_pair_layout_not_rotate_half[cuda]" time="0.767" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_adjacent_pair_layout_not_rotate_half[triton]" time="0.266" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_batch_position_padding_and_chunk_invariance[reference]" time="0.026" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_batch_position_padding_and_chunk_invariance[cuda]" time="0.097" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_batch_position_padding_and_chunk_invariance[triton]" time="0.352" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_noncontiguous_input_tables_and_gradient[reference]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_noncontiguous_input_tables_and_gradient[cuda]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_noncontiguous_input_tables_and_gradient[triton]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[reference-0-7]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[reference-2-0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[reference-0-0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[cuda-0-7]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[cuda-2-0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[cuda-0-0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[triton-0-7]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[triton-2-0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[triton-0-0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_second_derivative[reference]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_second_derivative[cuda]" time="0.931" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_second_derivative[triton]" time="0.292" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[reference-shape]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[reference-dtype]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[reference-table_shape]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[reference-table_dtype]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[reference-table_grad]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[reference-device]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[cuda-shape]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[cuda-dtype]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[cuda-table_shape]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[cuda-table_dtype]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[cuda-table_grad]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[cuda-device]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[triton-shape]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[triton-dtype]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[triton-table_shape]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[triton-table_dtype]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[triton-table_grad]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[triton-device]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_explicit_backend_and_reference_dispatch" time="0.022" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_cuda_registry_uses_native_when_present" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_cuda_registry_falls_back_to_triton" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_positions[bad0]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_positions[bad1]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_positions[bad2]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_cuda_triton_byte_parity_and_graph[dtype0]" time="0.104" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_cuda_triton_byte_parity_and_graph[dtype1]" time="0.107" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_cuda_triton_byte_parity_and_graph[dtype2]" time="0.099" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_native_entrypoint_checks_and_nondefault_stream" time="0.014" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_device_guard" time="0.000"><skipped type="pytest.skip" message="requires two NVIDIA GPUs">/home/dandan/src/RL-Kernel/tests/ops/rope/test_multi_axis_rope.py:336: requires two NVIDIA GPUs</skipped></testcase><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_canonical_tables_ignore_default_device[meta]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_canonical_tables_ignore_default_device[cuda]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_packed_qkv_and_independent_table_strides[reference]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_packed_qkv_and_independent_table_strides[cuda]" time="0.013" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_packed_qkv_and_independent_table_strides[triton]" time="0.596" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_dispatcher_fake_autograd_and_fullgraph_compile[cuda]" time="2.314" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_dispatcher_fake_autograd_and_fullgraph_compile[triton]" time="0.265" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_standard_ws1_four_judgments[cuda]" time="4.259" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_standard_ws1_four_judgments[triton]" time="0.596" /></testsuite></testsuites>
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-triton.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-triton.json
new file mode 100644
index 000000000..e105fb723
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-triton.json
@@ -0,0 +1,70 @@
+{
+  "suite_name": "multi_axis_rope",
+  "total_candidates": 1,
+  "passed_candidates": 1,
+  "pass_rate": 1.0,
+  "passed": true,
+  "candidates": [
+    {
+      "candidate_name": "triton-multi_axis_rope",
+      "backend": "triton",
+      "total_outputs": 2,
+      "passed_outputs": 2,
+      "pass_rate": 1.0,
+      "passed": true,
+      "cases": [
+        {
+          "case_name": "multi_axis_rope-torch.bfloat16-2x7x3x128",
+          "dtype": "torch.bfloat16",
+          "op_class": "elementwise",
+          "passed": true,
+          "outputs": [
+            {
+              "output_index": 0,
+              "shape": [
+                2,
+                7,
+                3,
+                128
+              ],
+              "candidate_dtype": "torch.bfloat16",
+              "gold_dtype": "torch.float32",
+              "atol": 0.02,
+              "rtol": 0.016,
+              "max_abs_error": 0.007704973220825195,
+              "mean_abs_error": 0.0006018930580466986,
+              "max_rel_error": 0.003828622167930007,
+              "passed": true,
+              "judgment": "forward_accuracy",
+              "comparison_lhs_role": "bf16_candidate",
+              "comparison_rhs_role": "fp32_reference",
+              "message": ""
+            },
+            {
+              "output_index": 1,
+              "shape": [
+                2,
+                7,
+                3,
+                128
+              ],
+              "candidate_dtype": "torch.bfloat16",
+              "gold_dtype": "torch.bfloat16",
+              "atol": 0.02,
+              "rtol": 0.016,
+              "max_abs_error": 0.0,
+              "mean_abs_error": 0.0,
+              "max_rel_error": 0.0,
+              "passed": true,
+              "judgment": "gradient_accuracy",
+              "comparison_lhs_role": "bf16_candidate",
+              "comparison_rhs_role": "fp32_reference",
+              "message": "gradient:x"
+            }
+          ]
+        }
+      ],
+      "backend_provenance": null
+    }
+  ]
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/package.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/package.json
new file mode 100644
index 000000000..1dce0c941
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/package.json
@@ -0,0 +1,11 @@
+{
+  "profile": "ordinary complete _C wheel; RL_KERNEL_ROPE_ONLY unset",
+  "wheel": "rl_kernel-0.1.0-cp312-cp312-linux_x86_64.whl",
+  "wheel_sha256": "32a39cd9cbb30b87816f6584096238fc2e4b0803eb387a2d3e1e80f3d1b40d8a",
+  "native_member": "rl_engine/_C.cpython-312-x86_64-linux-gnu.so",
+  "native_sha256": "2661e7188984c7d78acae1ef5186a3e90ec4aa8bb6f457522b23435a1f24f94a",
+  "license_present": true,
+  "installed_library": "/home/dandan/src/RL-Kernel/.venv/lib/python3.12/site-packages/rl_engine/_C.cpython-312-x86_64-linux-gnu.so",
+  "installed_library_sha256": "2661e7188984c7d78acae1ef5186a3e90ec4aa8bb6f457522b23435a1f24f94a",
+  "cpu_wheel_contains_no_cuda_binary": true
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/precommit.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/precommit.txt
new file mode 100644
index 000000000..475a8c0d8
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/precommit.txt
@@ -0,0 +1,7 @@
+trim trailing whitespace.................................................Passed
+fix end of files.........................................................Passed
+check yaml...............................................................Passed
+check for added large files..............................................Passed
+black....................................................................Passed
+isort....................................................................Passed
+flake8...................................................................Passed
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/python-packages.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/python-packages.json
new file mode 100644
index 000000000..41997f91d
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/python-packages.json
@@ -0,0 +1,114 @@
+{
+  "GitPython": "3.2.1",
+  "Jinja2": "3.1.6",
+  "Markdown": "3.11.2",
+  "MarkupSafe": "3.0.3",
+  "PyYAML": "6.0.3",
+  "Pygments": "2.21.0",
+  "RL-Kernel": "0.1.0",
+  "accelerate": "1.15.0",
+  "annotated-doc": "0.0.5",
+  "annotated-types": "0.8.0",
+  "anyio": "4.15.1",
+  "ast_serialize": "0.13.1",
+  "babel": "2.18.0",
+  "backrefs": "8.1",
+  "black": "24.4.2",
+  "bracex": "3.2",
+  "certifi": "2026.10.23",
+  "cfgv": "3.5.0",
+  "charset-normalizer": "3.5.2",
+  "click": "8.5.0",
+  "colorama": "0.4.6",
+  "distlib": "0.4.4",
+  "filelock": "3.32.3",
+  "flake8": "7.0.0",
+  "fsspec": "2026.7.0",
+  "ghp-import": "2.1.0",
+  "gitdb": "4.0.12",
+  "h11": "0.16.0",
+  "hf-xet": "1.7.0",
+  "httpcore": "1.0.9",
+  "httpx": "0.28.1",
+  "huggingface_hub": "1.33.0",
+  "identify": "2.6.20",
+  "idna": "3.20",
+  "iniconfig": "2.3.1",
+  "isort": "5.13.2",
+  "librt": "0.16.0",
+  "markdown-it-py": "4.2.0",
+  "mccabe": "0.7.0",
+  "mdurl": "0.1.2",
+  "mergedeep": "1.3.4",
+  "mkdocs": "1.6.1",
+  "mkdocs-awesome-nav": "3.3.0",
+  "mkdocs-get-deps": "0.2.2",
+  "mkdocs-git-revision-date-localized-plugin": "1.6.2",
+  "mkdocs-material": "9.8.0",
+  "mkdocs-material-extensions": "1.3.1",
+  "mpmath": "1.3.0",
+  "mypy": "2.4.0",
+  "mypy_extensions": "1.1.0",
+  "natsort": "8.4.0",
+  "networkx": "3.6.1",
+  "ninja": "1.13.2",
+  "nodeenv": "1.11.0",
+  "numpy": "2.5.4",
+  "nvidia-cublas-cu12": "12.8.4.1",
+  "nvidia-cuda-cupti-cu12": "12.8.90",
+  "nvidia-cuda-nvrtc-cu12": "12.8.93",
+  "nvidia-cuda-runtime-cu12": "12.8.90",
+  "nvidia-cudnn-cu12": "9.10.2.21",
+  "nvidia-cufft-cu12": "11.3.3.83",
+  "nvidia-cufile-cu12": "1.13.1.3",
+  "nvidia-curand-cu12": "10.3.9.90",
+  "nvidia-cusolver-cu12": "11.7.3.90",
+  "nvidia-cusparse-cu12": "12.5.8.93",
+  "nvidia-cusparselt-cu12": "0.7.1",
+  "nvidia-nccl-cu12": "2.27.5",
+  "nvidia-nvjitlink-cu12": "12.8.93",
+  "nvidia-nvshmem-cu12": "3.3.20",
+  "nvidia-nvtx-cu12": "12.8.90",
+  "packaging": "26.3",
+  "paginate": "0.5.7",
+  "pathspec": "1.1.1",
+  "pip": "25.3",
+  "platformdirs": "4.13.7",
+  "pluggy": "1.7.0",
+  "pre_commit": "4.7.0",
+  "psutil": "7.2.2",
+  "pycodestyle": "2.11.1",
+  "pydantic": "2.14.1",
+  "pydantic_core": "2.50.1",
+  "pyflakes": "3.2.0",
+  "pymdown-extensions": "12.3",
+  "pytest": "9.2.0",
+  "python-dateutil": "3.0.0",
+  "python-discovery": "1.6.7",
+  "pyyaml_env_tag": "1.1",
+  "regex": "2026.9.29",
+  "requests": "2.34.2",
+  "rich": "15.0.0",
+  "ruff": "0.17.3",
+  "safetensors": "0.8.0",
+  "setuptools": "80.9.0",
+  "shellingham": "1.5.4",
+  "six": "1.17.0",
+  "smmap": "5.0.3",
+  "sympy": "1.14.0",
+  "tabulate": "0.10.0",
+  "tokenizers": "0.22.2",
+  "torch": "2.9.1+cu128",
+  "tqdm": "4.70.1",
+  "transformers": "5.13.1",
+  "triton": "3.5.1",
+  "typer": "0.27.3",
+  "typing-inspection": "0.4.4",
+  "typing_extensions": "4.16.0",
+  "tzdata": "2026.6",
+  "urllib3": "2.8.0",
+  "virtualenv": "21.15.1",
+  "watchdog": "6.0.0",
+  "wcmatch": "11.2",
+  "wheel": "0.45.1"
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/quality.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/quality.txt
new file mode 100644
index 000000000..8e84b2473
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/quality.txt
@@ -0,0 +1,34 @@
+Local verification; no CUDA numerical/kernel sources changed for this follow-up.
+
+workflow-validation.log
+Workflow YAML, branch coverage and embedded Python syntax validated
+
+isort.log
+
+
+black.log
+All done! ✨ 🍰 ✨
+19 files would be left unchanged.
+
+flake8.log
+
+
+ruff.log
+All checks passed!
+
+docs.log
+INFO    -  Cleaning site directory
+INFO    -  Building documentation to directory: /home/dandan/src/RL-Kernel/site
+INFO    -  The following pages exist in the docs directory, but are not included in the "nav" configuration:
+  - blog/2026-07-08-announcing-rl-kernel-linear-logp-for-vime-zh.md
+  - blog/2026-08-09-dsv4-flash-moe-consistency-roadmap.md
+  - operators/matmul.md
+  - operators/rope.md
+  - usage/cross-config-drift-report.md
+  - usage/cuda-rocm-consistency-audit.md
+  - usage/h100-configurable.md
+  - usage/h100-cp-gradient-validation.md
+  - usage/h100-matrix-validation.md
+  - usage/h100-pcp-validation.md
+  - usage/rocm-sparse-performance.md
+INFO    -  Documentation built in 1.15 seconds
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/regression.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/regression.txt
new file mode 100644
index 000000000..95d27f812
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/regression.txt
@@ -0,0 +1,10 @@
+........................................................................ [ 35%]
+........................................................................ [ 71%]
+...................................................sssssss               [100%]
+=========================== short test summary info ============================
+SKIPPED [3] tests/ops/rope/test_rope.py:390: rope_apply_ascend is not compiled (requires an Ascend NPU build).
+SKIPPED [1] tests/ops/rope/test_rope.py:406: rope_apply_ascend is not compiled (requires an Ascend NPU build).
+SKIPPED [1] tests/ops/rope/test_rope.py:413: rope_apply_ascend is not compiled (requires an Ascend NPU build).
+SKIPPED [1] tests/ops/rope/test_rope.py:420: rope_apply_ascend is not compiled (requires an Ascend NPU build).
+SKIPPED [1] tests/ops/rope/test_rope.py:433: rope_apply_ascend is not compiled (requires an Ascend NPU build).
+195 passed, 7 skipped in 4.56s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/tests-and-ws1.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/tests-and-ws1.txt
new file mode 100644
index 000000000..0b8ce46dd
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/tests-and-ws1.txt
@@ -0,0 +1,11 @@
+{"scope": "Qwen-Image multi_axis_rope operator; ordinary full _C build", "python": "3.12.3", "platform": "Linux-6.18.54.1-microsoft-standard-WSL2-x86_64-with-glibc2.39", "torch": "2.9.1+cu128", "cuda": "12.8", "triton": "3.5.1", "gpu": "NVIDIA GeForce RTX 5080", "capability": [12, 0], "native_library": "/home/dandan/src/RL-Kernel/rl_engine/_C.cpython-312-x86_64-linux-gnu.so", "native_library_sha256": "2661e7188984c7d78acae1ef5186a3e90ec4aa8bb6f457522b23435a1f24f94a"}
+........................................................................ [ 87%]
+s.........                                                               [100%]
+=========================== short test summary info ============================
+SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:336: requires two NVIDIA GPUs
+81 passed, 1 skipped in 17.49s
+Using the installed production RoPE dispatcher library
+WS1 four judgments backend=cuda: passed=True
+Using the installed production RoPE dispatcher library
+WS1 four judgments backend=triton: passed=True
+Qwen-Image RoPE full-extension operator gate passed
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/upstream.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/upstream.json
new file mode 100644
index 000000000..032e23098
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/upstream.json
@@ -0,0 +1,335 @@
+{
+  "source_sha256": "0bad11901deae69f05aec01fd464e51e38582d455068f2851ee665036179321f",
+  "cases": [
+    {
+      "grid": [
+        2,
+        3
+      ],
+      "stream": "image",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        2,
+        3
+      ],
+      "stream": "image",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        2,
+        3
+      ],
+      "stream": "image",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        2,
+        3
+      ],
+      "stream": "text",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        2,
+        3
+      ],
+      "stream": "text",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        2,
+        3
+      ],
+      "stream": "text",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        3,
+        5
+      ],
+      "stream": "image",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        3,
+        5
+      ],
+      "stream": "image",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        3,
+        5
+      ],
+      "stream": "image",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        3,
+        5
+      ],
+      "stream": "text",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        3,
+        5
+      ],
+      "stream": "text",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        3,
+        5
+      ],
+      "stream": "text",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        64,
+        64
+      ],
+      "stream": "image",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        64,
+        64
+      ],
+      "stream": "image",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        64,
+        64
+      ],
+      "stream": "image",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        64,
+        64
+      ],
+      "stream": "text",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        64,
+        64
+      ],
+      "stream": "text",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        64,
+        64
+      ],
+      "stream": "text",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        83,
+        83
+      ],
+      "stream": "image",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 2.384185791015625e-07,
+      "dx_max_abs": 2.384185791015625e-07
+    },
+    {
+      "grid": [
+        83,
+        83
+      ],
+      "stream": "image",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        83,
+        83
+      ],
+      "stream": "image",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        83,
+        83
+      ],
+      "stream": "text",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        83,
+        83
+      ],
+      "stream": "text",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        83,
+        83
+      ],
+      "stream": "text",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        104,
+        58
+      ],
+      "stream": "image",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 2.384185791015625e-07,
+      "dx_max_abs": 2.384185791015625e-07
+    },
+    {
+      "grid": [
+        104,
+        58
+      ],
+      "stream": "image",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        104,
+        58
+      ],
+      "stream": "image",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        104,
+        58
+      ],
+      "stream": "text",
+      "dtype": "torch.float32",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        104,
+        58
+      ],
+      "stream": "text",
+      "dtype": "torch.float16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    },
+    {
+      "grid": [
+        104,
+        58
+      ],
+      "stream": "text",
+      "dtype": "torch.bfloat16",
+      "table_bitwise": true,
+      "forward_max_abs": 0.0,
+      "dx_max_abs": 0.0
+    }
+  ]
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-cuda.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-cuda.json
new file mode 100644
index 000000000..3a3e06daf
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-cuda.json
@@ -0,0 +1,1962 @@
+{
+  "scope": "Qwen-Image multi_axis_rope operator; C1 elementwise and C2 config transforms",
+  "trace": {
+    "contract": "qwenimage-interleaved-rope-fp32-v1",
+    "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
+    "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
+    "layout": "BSHD, adjacent channel pairs",
+    "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+    "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+  },
+  "passed": true,
+  "forward": {
+    "op_name": "multi_axis_rope",
+    "backend_profile": "cuda_bf16",
+    "accuracy_reports": [
+      {
+        "config_id": "BN/full",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/chunked",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/chunked",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004673302173614502,
+            "mean_abs_error": 6.917398422956467e-05,
+            "max_rel_error": 0.003753266530111432,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.00048516690731048584,
+            "mean_abs_error": 7.802648178767413e-05,
+            "max_rel_error": 0.0037793784867972136,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004864335060119629,
+            "mean_abs_error": 7.697803084738553e-05,
+            "max_rel_error": 0.003821910358965397,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.710371912457049e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004673302173614502,
+            "mean_abs_error": 6.917398422956467e-05,
+            "max_rel_error": 0.003753266530111432,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.00048516690731048584,
+            "mean_abs_error": 7.802648178767413e-05,
+            "max_rel_error": 0.0037793784867972136,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004864335060119629,
+            "mean_abs_error": 7.697803084738553e-05,
+            "max_rel_error": 0.003821910358965397,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.710371912457049e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/permuted",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/permuted",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_right",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/padded_right",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_left",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/padded_left",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      }
+    ],
+    "invariance_reports": [
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/chunked",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/chunked"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s0",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s1",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s2",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s3",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/permuted",
+        "transform_kind": "permutation",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/permuted"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_right",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_right"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_left",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_left"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      }
+    ],
+    "logprob_smoke": null,
+    "backend_provenance": {
+      "backend_profile": "cuda_bf16",
+      "requested_backend": "cuda",
+      "actual_backend": "cuda",
+      "execution_dtype": "bfloat16",
+      "accumulation_dtype": "float32",
+      "output_dtype": "bfloat16",
+      "reference_dtype": "float32",
+      "candidate_tf32_enabled": false,
+      "reference_tf32_enabled": false
+    },
+    "candidate_id": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
+    "device": "cuda",
+    "compute_capability": "sm120",
+    "seed": 20260812,
+    "fallback_reason": null,
+    "passed": true,
+    "provenance_valid": true,
+    "metadata_valid": true,
+    "observed_kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope"
+  },
+  "gradient": {
+    "op_name": "multi_axis_rope",
+    "backend_profile": "cuda_bf16",
+    "accuracy_reports": [
+      {
+        "config_id": "BN/full",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/chunked",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/chunked",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.7887988835573196e-05,
+            "mean_abs_error": 3.6821973026235355e-06,
+            "max_rel_error": 1.8164952993392944,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.7113780965446495e-06,
+            "max_rel_error": 0.4316565990447998,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.995925024151802e-05,
+            "mean_abs_error": 4.716822331829462e-06,
+            "max_rel_error": 0.7439075112342834,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.0211638659238815e-05,
+            "mean_abs_error": 4.9605891945248e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.7887988835573196e-05,
+            "mean_abs_error": 3.6821973026235355e-06,
+            "max_rel_error": 1.8164952993392944,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.7113780965446495e-06,
+            "max_rel_error": 0.4316565990447998,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.995925024151802e-05,
+            "mean_abs_error": 4.716822331829462e-06,
+            "max_rel_error": 0.7439075112342834,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.0211638659238815e-05,
+            "mean_abs_error": 4.9605891945248e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/permuted",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/permuted",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_right",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/padded_right",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_left",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/padded_left",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "cuda_bf16",
+          "requested_backend": "cuda",
+          "actual_backend": "cuda",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      }
+    ],
+    "invariance_reports": [
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/chunked",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/chunked"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s0",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s1",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s2",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s3",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/permuted",
+        "transform_kind": "permutation",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/permuted"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_right",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_right"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_left",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_left"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      }
+    ],
+    "singleton_aggregate_reports": [],
+    "backend_provenance": {
+      "backend_profile": "cuda_bf16",
+      "requested_backend": "cuda",
+      "actual_backend": "cuda",
+      "execution_dtype": "bfloat16",
+      "accumulation_dtype": "float32",
+      "output_dtype": "bfloat16",
+      "reference_dtype": "float32",
+      "candidate_tf32_enabled": false,
+      "reference_tf32_enabled": false
+    },
+    "candidate_id": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
+    "device": "cuda",
+    "compute_capability": "sm120",
+    "seed": 20260812,
+    "fallback_reason": null,
+    "passed": true,
+    "provenance_valid": true,
+    "metadata_valid": true,
+    "loss_reduction": "sum_over_active_tokens_then_optional_mean_by_active_count",
+    "active_token_denominator": 27,
+    "grad_tensor_names": [
+      "dx"
+    ],
+    "first_failing_op": null,
+    "first_failing_tensor": null,
+    "first_failing_config_pair": null,
+    "observed_kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope"
+  }
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-triton.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-triton.json
new file mode 100644
index 000000000..628fcbf70
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-triton.json
@@ -0,0 +1,1962 @@
+{
+  "scope": "Qwen-Image multi_axis_rope operator; C1 elementwise and C2 config transforms",
+  "trace": {
+    "contract": "qwenimage-interleaved-rope-fp32-v1",
+    "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
+    "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
+    "layout": "BSHD, adjacent channel pairs",
+    "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
+    "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
+  },
+  "passed": true,
+  "forward": {
+    "op_name": "multi_axis_rope",
+    "backend_profile": "triton_cuda_bf16",
+    "accuracy_reports": [
+      {
+        "config_id": "BN/full",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/chunked",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/chunked",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004673302173614502,
+            "mean_abs_error": 6.917398422956467e-05,
+            "max_rel_error": 0.003753266530111432,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.00048516690731048584,
+            "mean_abs_error": 7.802648178767413e-05,
+            "max_rel_error": 0.0037793784867972136,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004864335060119629,
+            "mean_abs_error": 7.697803084738553e-05,
+            "max_rel_error": 0.003821910358965397,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.710371912457049e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004673302173614502,
+            "mean_abs_error": 6.917398422956467e-05,
+            "max_rel_error": 0.003753266530111432,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.00048516690731048584,
+            "mean_abs_error": 7.802648178767413e-05,
+            "max_rel_error": 0.0037793784867972136,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004864335060119629,
+            "mean_abs_error": 7.697803084738553e-05,
+            "max_rel_error": 0.003821910358965397,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.710371912457049e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/permuted",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/permuted",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_right",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/padded_right",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_left",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/padded_left",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 0.0004878789186477661,
+            "mean_abs_error": 8.054684440139681e-05,
+            "max_rel_error": 0.0038475203327834606,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "forward_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      }
+    ],
+    "invariance_reports": [
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/chunked",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/chunked"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s0",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s1",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s2",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s3",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/permuted",
+        "transform_kind": "permutation",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/permuted"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_right",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_right"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_left",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "output",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_left"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "forward_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      }
+    ],
+    "logprob_smoke": null,
+    "backend_provenance": {
+      "backend_profile": "triton_cuda_bf16",
+      "requested_backend": "triton",
+      "actual_backend": "triton",
+      "execution_dtype": "bfloat16",
+      "accumulation_dtype": "float32",
+      "output_dtype": "bfloat16",
+      "reference_dtype": "float32",
+      "candidate_tf32_enabled": false,
+      "reference_tf32_enabled": false
+    },
+    "candidate_id": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
+    "device": "cuda",
+    "compute_capability": "sm120",
+    "seed": 20260812,
+    "fallback_reason": null,
+    "passed": true,
+    "provenance_valid": true,
+    "metadata_valid": true,
+    "observed_kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel"
+  },
+  "gradient": {
+    "op_name": "multi_axis_rope",
+    "backend_profile": "triton_cuda_bf16",
+    "accuracy_reports": [
+      {
+        "config_id": "BN/full",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/chunked",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/chunked",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.7887988835573196e-05,
+            "mean_abs_error": 3.6821973026235355e-06,
+            "max_rel_error": 1.8164952993392944,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.7113780965446495e-06,
+            "max_rel_error": 0.4316565990447998,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.995925024151802e-05,
+            "mean_abs_error": 4.716822331829462e-06,
+            "max_rel_error": 0.7439075112342834,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/full/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/full/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.0211638659238815e-05,
+            "mean_abs_error": 4.9605891945248e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s0",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s0",
+              "fp32_reference"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.7887988835573196e-05,
+            "mean_abs_error": 3.6821973026235355e-06,
+            "max_rel_error": 1.8164952993392944,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s1",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s1",
+              "fp32_reference"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.7113780965446495e-06,
+            "max_rel_error": 0.4316565990447998,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s2",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s2",
+              "fp32_reference"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 2.995925024151802e-05,
+            "mean_abs_error": 4.716822331829462e-06,
+            "max_rel_error": 0.7439075112342834,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "B1-singleton_aggregate/chunked/s3",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "B1-singleton_aggregate/chunked/s3",
+              "fp32_reference"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.0211638659238815e-05,
+            "mean_abs_error": 4.9605891945248e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/permuted",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/permuted",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_right",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/padded_right",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      },
+      {
+        "config_id": "BN/padded_left",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/padded_left",
+              "fp32_reference"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "float32",
+            "max_abs_error": 3.213249146938324e-05,
+            "mean_abs_error": 4.699562850873917e-06,
+            "max_rel_error": 5.814913272857666,
+            "atol": 0.02,
+            "rtol": 0.016,
+            "passed": true,
+            "judgment": "gradient_accuracy",
+            "comparison_lhs_role": "bf16_candidate",
+            "comparison_rhs_role": "fp32_reference"
+          }
+        ],
+        "passed": true,
+        "backend_provenance": {
+          "backend_profile": "triton_cuda_bf16",
+          "requested_backend": "triton",
+          "actual_backend": "triton",
+          "execution_dtype": "bfloat16",
+          "accumulation_dtype": "float32",
+          "output_dtype": "bfloat16",
+          "reference_dtype": "float32",
+          "candidate_tf32_enabled": false,
+          "reference_tf32_enabled": false
+        }
+      }
+    ],
+    "invariance_reports": [
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/chunked",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/chunked"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s0",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s1",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s2",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/full/s3",
+        "transform_kind": "batch_size",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/full/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s0"
+            ],
+            "shape": [
+              3,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s1"
+            ],
+            "shape": [
+              8,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s2"
+            ],
+            "shape": [
+              5,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
+        "transform_kind": "chunk",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "B1-singleton_aggregate/chunked/s3"
+            ],
+            "shape": [
+              11,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/permuted",
+        "transform_kind": "permutation",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/permuted"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_right",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_right"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      },
+      {
+        "canonical_config_id": "BN/full",
+        "transformed_config_id": "BN/padded_left",
+        "transform_kind": "padding",
+        "op_class": "elementwise",
+        "dtype": "bfloat16",
+        "backend_profile": "triton_cuda_bf16",
+        "details": [
+          {
+            "tensor_name": "dx",
+            "config_pair": [
+              "BN/full",
+              "BN/padded_left"
+            ],
+            "shape": [
+              27,
+              3,
+              128
+            ],
+            "dtype": "bfloat16",
+            "max_abs_error": 0.0,
+            "mean_abs_error": 0.0,
+            "max_rel_error": 0.0,
+            "atol": 0.0,
+            "rtol": 0.0,
+            "passed": true,
+            "judgment": "gradient_invariance",
+            "comparison_lhs_role": "transformed_config",
+            "comparison_rhs_role": "canonical_config"
+          }
+        ],
+        "passed": true
+      }
+    ],
+    "singleton_aggregate_reports": [],
+    "backend_provenance": {
+      "backend_profile": "triton_cuda_bf16",
+      "requested_backend": "triton",
+      "actual_backend": "triton",
+      "execution_dtype": "bfloat16",
+      "accumulation_dtype": "float32",
+      "output_dtype": "bfloat16",
+      "reference_dtype": "float32",
+      "candidate_tf32_enabled": false,
+      "reference_tf32_enabled": false
+    },
+    "candidate_id": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
+    "device": "cuda",
+    "compute_capability": "sm120",
+    "seed": 20260812,
+    "fallback_reason": null,
+    "passed": true,
+    "provenance_valid": true,
+    "metadata_valid": true,
+    "loss_reduction": "sum_over_active_tokens_then_optional_mean_by_active_count",
+    "active_token_denominator": 27,
+    "grad_tensor_names": [
+      "dx"
+    ],
+    "first_failing_op": null,
+    "first_failing_tensor": null,
+    "first_failing_config_pair": null,
+    "observed_kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel"
+  }
+}
diff --git a/tools/validation/operators/check_multi_axis_rope_install.py b/tools/validation/operators/check_multi_axis_rope_install.py
index 56cfe2ccb..4fa42d15f 100644
--- a/tools/validation/operators/check_multi_axis_rope_install.py
+++ b/tools/validation/operators/check_multi_axis_rope_install.py
@@ -13,6 +13,7 @@
 import torch
 
 import rl_engine
+from rl_engine.backends import extension
 from rl_engine.ops.rope.multi_axis_rope import (
     build_multi_axis_cos_sin,
     get_multi_axis_rope_op,
@@ -50,6 +51,7 @@ def main():
     report = {
         "package": str(package),
         "companion_library": native.origin if native else None,
+        "full_extension_library": getattr(extension._C, "__file__", None),
         "native_dispatcher_registered": hasattr(torch.ops.rl_kernel_native, "multi_axis_rope"),
         "torch": torch.__version__,
         "gpu": torch.cuda.get_device_name(),

From 84751473006401c19c1c8c17844e8ef8b1c44375 Mon Sep 17 00:00:00 2001
From: dandan20031031 <2425926169@qq.com>
Date: Wed, 28 Oct 2026 22:00:50 +0800
Subject: [PATCH 3/5] test(rope): record Python 3.10 CPU qualification

Signed-off-by: dandan20031031 <2425926169@qq.com>
---
 .../README.md                                 | 14 ++++++++-----
 .../cpu310-contracts.txt                      | 20 +++++++++++++++++++
 .../cpu310-environment.json                   |  9 +++++++++
 .../manifest.json                             |  4 +++-
 4 files changed, 41 insertions(+), 6 deletions(-)
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-contracts.txt
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-environment.json

diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/README.md b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/README.md
index ddc879318..9ef00fdb1 100644
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/README.md
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/README.md
@@ -48,7 +48,8 @@ SM120; bare-metal Linux and A100/H100 remain separate hardware profiles.
 | Related regressions | 195 passed, 7 Ascend skips | `regression.txt` |
 | Repository layout with extension enforcement enabled | 37 passed, zero skips/failures | `layout.txt` |
 | Fresh installed full wheel, outside source checkout | Eager/compiled forward and input-gradient bytes exact for both backends | `installed-wheel.json` |
-| Actual CPU-only PyTorch wheel/reference installation | 57 passed, 52 expected GPU skips; reference forward/backward passed | `cpu-contracts.txt`, `cpu-environment.json` |
+| Actual CPU-only PyTorch wheel/reference installation, Python 3.10.22 | 57 passed, 52 expected GPU skips; reference forward/backward passed | `cpu310-contracts.txt`, `cpu310-environment.json` |
+| Actual CPU-only PyTorch wheel/reference installation, Python 3.12.3 | 57 passed, 52 expected GPU skips; reference forward/backward passed | `cpu-contracts.txt`, `cpu-environment.json` |
 | Compute Sanitizer accuracy/stride/empty/stream subset | 34 passed; zero errors | `memcheck.txt` |
 | Pinned actual Diffusers symbols | 30 forward/backward cases passed; image/text tables exact | `upstream.json` |
 | Reference-grid benchmark correctness | 40 comparisons passed | `benchmark-bfloat16.json` |
@@ -56,8 +57,10 @@ SM120; bare-metal Linux and A100/H100 remain separate hardware profiles.
 
 The full-extension GPU gate rejects missing native/Triton execution. Its only
 allowed operator-test skip is the two-device guard on a one-GPU runner.
-The CPU-only result is separate: it uses `torch==2.9.1+cpu`, and its installed
-package contains no native CUDA binary. CPU skips are not GPU acceptance evidence.
+The CPU-only results are separate: both Python versions use `torch==2.9.1+cpu`,
+and their installed packages contain no native CUDA binary. CPU skips are not
+GPU acceptance evidence. Both runs install the normal CPU package and exercise
+the workflow's three test files locally; these are not hosted CI results.
 The 37 passing Linux layout tests contrast with the five Windows baseline
 failures recorded earlier; those historical logs were not rewritten.
 
@@ -123,8 +126,9 @@ The new [operator gate](../../../ci/scripts/run_qwenimage_rope_ws1.sh) requires
 the complete `_C` registration, records actual library/backend identity and
 enforces the GPU skip policy. The [CPU workflow](../../../.github/workflows/qwenimage-rope.yml)
 targets `test-qwenimage`, runs Python 3.10/3.12 CPU contracts, checks changed files
-and builds documentation. This report's actual CPU-only run used Python 3.12;
-hosted CI status is tracked on the PR. GPU runners use the explicit gate.
+and builds documentation. This report includes local CPU-only runs on both
+Python 3.10 and 3.12; hosted CI status is tracked on the PR. GPU runners use the
+explicit gate.
 
 Bare-metal Linux, A100/H100, two-GPU guards, QK-norm fusion, model/block assembly
 and trajectories remain unqualified. Building the entire library does not
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-contracts.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-contracts.txt
new file mode 100644
index 000000000..6f0703516
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-contracts.txt
@@ -0,0 +1,20 @@
+...........ssssssssssss.ss.ss.ss...ssssss.ss......ssssssssssss.ss...ssss [ 66%]
+s.s.ssssss...........................                                    [100%]
+=========================== short test summary info ============================
+SKIPPED [12] tests/ops/rope/test_multi_axis_rope.py:79: NVIDIA CUDA required
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:100: NVIDIA CUDA required
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:110: NVIDIA CUDA required
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:133: NVIDIA CUDA required
+SKIPPED [6] tests/ops/rope/test_multi_axis_rope.py:148: NVIDIA CUDA required
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:158: NVIDIA CUDA required
+SKIPPED [12] tests/ops/rope/test_multi_axis_rope.py:174: NVIDIA CUDA required
+SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:217: native CUDA multi_axis_rope extension required
+SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:230: NVIDIA CUDA required
+SKIPPED [3] tests/ops/rope/test_multi_axis_rope.py:255: NVIDIA CUDA required
+SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:315: NVIDIA CUDA required
+SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:336: requires two NVIDIA GPUs
+SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:355: CUDA context requires CUDA
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:368: NVIDIA CUDA required
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:399: NVIDIA CUDA required
+SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:427: NVIDIA CUDA required
+57 passed, 52 skipped in 1.72s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-environment.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-environment.json
new file mode 100644
index 000000000..833e69038
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-environment.json
@@ -0,0 +1,9 @@
+{
+  "python": "3.10.22",
+  "torch": "2.9.1+cpu",
+  "cuda": null,
+  "package": "/home/dandan/venvs/rlk-rope-cpu310/lib/python3.10/site-packages/rl_engine",
+  "reference_forward_backward": true,
+  "cpu_wheel_contains_no_cuda_binary": true,
+  "passed": true
+}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/manifest.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/manifest.json
index 3acb62a79..9db29dacc 100644
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/manifest.json
+++ b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/manifest.json
@@ -30,9 +30,12 @@
     "tools/validation/operators/check_multi_axis_rope_install.py": "f295fc8ac5878ac13b6255c2f0be83b66d34ff2ad2f75111bb8b633508a865a7"
   },
   "evidence_sha256": {
+    "README.md": "3ad6f8ef514b4fc61d51fb0ce8cb6d420d2180251c9ee4f84b8c931aa3a75094",
     "benchmark-bfloat16.json": "e9c1e4b1085623d84800656c179e4a0e86d70a0ef7fc9ed92f9e6ccc50f0f491",
     "cpu-contracts.txt": "9518f53a6efc603bca1c52f160e5430f1006bb2abc9a7ef57694dae6feac5d78",
     "cpu-environment.json": "4dfc68f3d76af05a050d830840b4a6b8c8fa0a4c12c29d11d632a56a57a2d4c6",
+    "cpu310-contracts.txt": "02f57808b9553cefe6f4dcec06d46a5163157b4290cc0c590b344a1616057852",
+    "cpu310-environment.json": "7130d4c164c98804284cc0f1e07ae6265c35e7498c2fb8285a5e6df87cb010b2",
     "full-extension.json": "8fae782e3c5e6aae59de6fe6b4e7ae185729e49d2282dbd0cb3b9ca08c0274cd",
     "full-inplace-build.txt": "04d5ef122650ab305847d58850a6c3b44438eacf97920492e1eb1c5580e9a21e",
     "full-wheel-build.txt": "a7883dfb1211798c869ede1ee0b6f1dfb32edd7b27b359ee18c6e5ab9fe449d5",
@@ -47,7 +50,6 @@
     "precommit.txt": "7cf90e06cf73194271c176ebc2d147e64cf8f5ed522d8e0e03093a4647f63759",
     "python-packages.json": "617f8cef5a24f3b2a683e70e0b491e47db6d24a7ca848e5b3e461e87993a7201",
     "quality.txt": "8826f0e406946a923400d792dd60d0f545b32f9fcffa3a148268c969e30e14c0",
-    "README.md": "af365e658fe4875df643704b2495e1f115c6c889f5d449880233d49e511040ac",
     "regression.txt": "039cd9813a2ea54bfc5b7750fc0555b54d2bf30b09dbca2eba4a0403578ab9b0",
     "tests-and-ws1.txt": "68397e27144c9244743dbcff39ffdca8a77a30eaf1e8180b2ca84dfc1216d633",
     "upstream.json": "04e5472a82477a305b996ccbb420d0f3cc7b39af4053c867d37916a9bf5de51a",

From 5931ce7b6d95799a6c19e0d3ae3eac958a768acf Mon Sep 17 00:00:00 2001
From: dandan20031031 <2425926169@qq.com>
Date: Wed, 28 Oct 2026 22:35:46 +0800
Subject: [PATCH 4/5] docs(rope): consolidate validation evidence for review

Signed-off-by: dandan20031031 <2425926169@qq.com>
---
 .../qwenimage_multi_axis_rope_sm120.md        |  122 +
 .../README.md                                 |  152 -
 .../baseline-layout.txt                       |  138 -
 .../benchmark-bfloat16.json                   | 2781 -----------------
 .../installed-wheel.json                      |   37 -
 .../installed-wheel.txt                       |   38 -
 .../layout.txt                                |  138 -
 .../manifest.json                             |   65 -
 .../memcheck.txt                              |    8 -
 .../operator-cuda.json                        |   70 -
 .../operator-triton.json                      |   70 -
 .../package-build.txt                         |   40 -
 .../package.json                              |   31 -
 .../precommit.txt                             |    7 -
 .../quality.txt                               |   58 -
 .../regression.txt                            |   10 -
 .../tests.txt                                 |    5 -
 .../upstream.json                             |  335 --
 .../ws1-cuda.json                             | 1962 ------------
 .../ws1-cuda.txt                              |    3 -
 .../ws1-triton.json                           | 1962 ------------
 .../ws1-triton.txt                            |    3 -
 .../README.md                                 |  136 -
 .../benchmark-bfloat16.json                   | 2781 -----------------
 .../cpu-contracts.txt                         |   20 -
 .../cpu-environment.json                      |    7 -
 .../cpu310-contracts.txt                      |   20 -
 .../cpu310-environment.json                   |    9 -
 .../full-extension.json                       |   22 -
 .../full-inplace-build.txt                    |   20 -
 .../full-wheel-build.txt                      |   10 -
 .../gpu-environment.json                      |   15 -
 .../installed-wheel.json                      |   38 -
 .../layout.txt                                |    2 -
 .../manifest.json                             |   60 -
 .../memcheck.txt                              |    4 -
 .../operator-cuda.json                        |   70 -
 .../operator-tests.xml                        |    1 -
 .../operator-triton.json                      |   70 -
 .../package.json                              |   11 -
 .../precommit.txt                             |    7 -
 .../python-packages.json                      |  114 -
 .../quality.txt                               |   34 -
 .../regression.txt                            |   10 -
 .../tests-and-ws1.txt                         |   11 -
 .../upstream.json                             |  335 --
 .../ws1-cuda.json                             | 1962 ------------
 .../ws1-triton.json                           | 1962 ------------
 48 files changed, 122 insertions(+), 15644 deletions(-)
 create mode 100644 reports/experiments/qwenimage_multi_axis_rope_sm120.md
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/README.md
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/baseline-layout.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/benchmark-bfloat16.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/layout.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/manifest.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/memcheck.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-cuda.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-triton.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package-build.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/precommit.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/quality.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/regression.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/tests.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/upstream.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/README.md
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/benchmark-bfloat16.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-contracts.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-environment.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-contracts.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-environment.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-extension.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-inplace-build.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-wheel-build.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/gpu-environment.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/installed-wheel.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/layout.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/manifest.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/memcheck.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-cuda.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-tests.xml
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-triton.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/package.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/precommit.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/python-packages.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/quality.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/regression.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/tests-and-ws1.txt
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/upstream.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-cuda.json
 delete mode 100644 reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-triton.json

diff --git a/reports/experiments/qwenimage_multi_axis_rope_sm120.md b/reports/experiments/qwenimage_multi_axis_rope_sm120.md
new file mode 100644
index 000000000..a7d82f65b
--- /dev/null
+++ b/reports/experiments/qwenimage_multi_axis_rope_sm120.md
@@ -0,0 +1,122 @@
+# Qwen-Image multi-axis RoPE: SM120 validation
+
+Recorded 2026-10-28 for [PR #623](https://github.com/RL-Align/RL-Kernel/pull/623).
+The numerical/native implementation is based on `ac17ab87e8d93dabe93dfdc134e9853d47de7af1`.
+The [operator documentation](../../docs/operators/multi-axis-rope.md) defines the
+contract, upstream reuse, API and detailed reproduction commands.
+
+## Environments
+
+Both profiles use one RTX 5080 (SM120, 16 GB), Windows driver 617.42,
+PyTorch 2.9.1+cu128 and CUDA Toolkit 12.8. They are separate recorded runs,
+not a controlled Windows-versus-WSL performance comparison.
+
+| Item | WSL2 Linux | Windows |
+| --- | --- | --- |
+| OS | Ubuntu 24.04.5, glibc 2.39; WSL 3.0.3 on Windows 11 | Windows |
+| Kernel | 6.18.54.1-microsoft-standard-WSL2 | Native Windows |
+| Python | 3.12.3 | 3.13.9 |
+| Compiler | CUDA 12.8.93, GCC 13.3 | CUDA 12.8, MSVC 2022 x64 |
+| Triton | 3.5.1 | triton-windows 3.5.1.post24 |
+| Native build | Ordinary complete `rl_engine._C`, all 10 configured translation units | Explicit operator-only `rl_engine._rope_cuda` |
+
+The WSL checkout, environments and caches are in the Linux filesystem.
+The normal full wheel was built with `RL_KERNEL_ROPE_ONLY`, forced SM90 and
+fast-math flags unset. Linking existing CUDA driver calls used
+`LIBRARY_PATH=/usr/lib/wsl/lib`; the Windows driver supplies CUDA driver access.
+
+## Recorded results
+
+| Check | WSL2 Linux | Windows |
+| --- | --- | --- |
+| Public operator suite | 81 passed, 1 two-GPU guard skip | 81 passed, 1 two-GPU guard skip |
+| Related regressions | 195 passed, 7 Ascend skips | 195 passed, 7 Ascend skips |
+| CUDA/Triton WS1 forward/gradient accuracy and invariance | Both passed with backend provenance | Both passed with backend provenance |
+| Standard operator CLI forward/input gradient | Both backends passed | Both backends passed |
+| Pinned actual Diffusers symbols | 30 cases passed; image/text tables exact | 30 cases passed; image/text tables exact |
+| Compute Sanitizer | 34 selected cases passed, zero errors | 34 selected cases passed, zero errors |
+| Fresh installed wheel outside checkout | Full wheel: both backends' eager/compiled forward and input-gradient bytes exact | Operator wheel: same checks passed |
+| CPU-only PyTorch 2.9.1+cpu, Python 3.10.22 and 3.12.3 | Each: 57 passed, 52 expected GPU skips; installed reference forward/backward passed | Not recorded |
+| Repository layout | 37 passed, zero skips/failures | 31 passed, 5 failures, 1 skip |
+| Changed-file static checks and strict docs | Passed locally | Passed locally |
+
+The operator suite covers FP32/FP16/BF16, independent accuracy, byte invariance,
+packed-QKV/no-copy and table strides, empty inputs, higher input VJPs, stream
+selection, graph replay, opcheck and fullgraph forward/backward. The installed
+wheels contain the FlashAttention BSD license; the CPU packages contain no CUDA
+binary. Local CPU checks are separate from hosted CI and GPU acceptance.
+
+The five historical Windows layout failures also reproduce on its clean base
+(27 passed, 5 failures, 1 skip): driver linking and Windows symlink privileges or
+placeholders. The later Linux run preserved all 27 tracked symlinks and passed.
+The earlier Windows logs remain unchanged in the archive.
+
+## Standalone benchmark
+
+WSL2 BF16, batch 1, 24 heads, D=128. CUDA Graph medians in ms/call;
+each cell is **forward / forward+backward**.
+
+| Grid / layout | Native CUDA | Adapted Triton | FlashAttention, no FMA | Diffusers complex |
+| --- | --- | --- | --- | --- |
+| 64x64 contiguous | 0.05780 / 0.12967 | 0.04929 / 0.12414 | 0.04898 / 0.12407 | 0.24446 / 0.51281 |
+| 64x64 packed QKV | 0.05868 / 0.13079 | 0.04731 / 0.11841 | 0.04722 / 0.11847 | 0.24754 / 0.51583 |
+| 83x83 contiguous | 0.11198 / 0.22392 | 0.10720 / 0.22538 | 0.10739 / 0.21420 | 0.52579 / 1.06721 |
+| 83x83 packed QKV | 0.11233 / 0.22486 | 0.10745 / 0.21478 | 0.10744 / 0.21469 | 0.52820 / 1.06947 |
+| 104x58 contiguous | 0.09797 / 0.19708 | 0.09366 / 0.18497 | 0.09362 / 0.18458 | 0.44992 / 0.90110 |
+| 104x58 packed QKV | 0.09610 / 0.19612 | 0.09310 / 0.18828 | 0.09323 / 0.18817 | 0.44979 / 0.90276 |
+
+All 40 pre-timing comparisons passed; CUDA, adapted Triton and the no-FMA
+baseline had zero differing output/input-gradient bytes against the FP32 gold.
+The archive includes 80 measurements, the short case, default FlashAttention,
+eager wall timings and peak allocation. At 4096 tokens, either GPU path's forward
+extra peak is 24 MiB, the returned output size, for both layouts.
+
+Each graph contains 20 calls, with five warm calls/replays and 30 measured
+replays. Eager wall timings use three synchronized trials of 30 calls.
+Tables, compilation and input/backing allocation are excluded; extra peak
+includes the returned output or gradient. The upstream source/autograd wrapper
+is pinned to `98c1cee0ed59307833841cdbf567fe336ceec13a`; the no-FMA variant changes
+only contraction. These single-process runs use a shared GPU without clock
+pinning and establish no advantage over FlashAttention or model speedup.
+
+## Reproduction
+
+Use a matching CUDA 12.8/PyTorch environment in a Linux checkout:
+
+```bash
+export CUDA_HOME=/usr/local/cuda-12.8
+export PATH="$CUDA_HOME/bin:$PATH"
+export LIBRARY_PATH="/usr/lib/wsl/lib${LIBRARY_PATH:+:$LIBRARY_PATH}"
+export MAX_JOBS=2 TORCH_CUDA_ARCH_LIST=12.0 RL_KERNEL_REQUIRE_EXT=1
+unset RL_KERNEL_ROPE_ONLY KERNEL_ALIGN_FORCE_SM90 KERNEL_ALIGN_USE_FAST_MATH
+python -m pip wheel --no-build-isolation --no-deps . -w /tmp/rope-full-wheel
+python -m pip install --force-reinstall --no-deps /tmp/rope-full-wheel/rl_kernel-*.whl
+python setup.py build_ext --inplace
+ROPE_PYTHON=python ROPE_WS1_OUTPUT=/tmp/rope-ws1 bash ci/scripts/run_qwenimage_rope_ws1.sh
+python ci/run.py layout -q -rs
+compute-sanitizer --tool memcheck --error-exitcode 1 python -m pytest \
+  tests/ops/rope/test_multi_axis_rope.py -q \
+  -k "forward_backward_accuracy or noncontiguous_input_tables_and_gradient or empty_forward_backward or packed_qkv_and_independent_table_strides or native_entrypoint_checks"
+```
+
+The [GPU gate](../../ci/scripts/run_qwenimage_rope_ws1.sh) requires the full
+native extension and Triton, records loaded binary/backend identity, and permits
+only the two-GPU guard to skip. The [CPU workflow](../../.github/workflows/qwenimage-rope.yml)
+contains the Python 3.10/3.12 contract commands. The operator documentation covers
+installed-wheel probes, pinned upstream audits and benchmark reproduction.
+Hosted CI status is tracked on the PR.
+
+## Archived evidence and qualification limits
+
+Detailed per-case JSON, JUnit, build/test logs, dependency versions, and
+source/binary SHA256 manifests remain available at immutable commit
+`84751473006401c19c1c8c17844e8ef8b1c44375`:
+
+- [WSL2 complete-extension evidence](https://github.com/dandan20031031/RL-Kernel/tree/84751473006401c19c1c8c17844e8ef8b1c44375/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120)
+- [Historical Windows operator-profile evidence](https://github.com/dandan20031031/RL-Kernel/tree/84751473006401c19c1c8c17844e8ef8b1c44375/reports/experiments/qwenimage_multi_axis_rope_v2_sm120)
+
+This summary preserves those recorded results; consolidating reports does not
+rerun the numerical suites. Bare-metal Linux, A100/H100, two-GPU guards,
+QK-norm fusion, model/block assembly and trajectories remain unqualified.
+Building the full library does not certify every native operator or Dense Qwen3.
+No distributed or cross-version/device byte-equality claim is made.
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/README.md b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/README.md
deleted file mode 100644
index 64232c731..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/README.md
+++ /dev/null
@@ -1,152 +0,0 @@
-# Qwen-Image multi-axis RoPE validation — 2026-10-28
-
-Evidence for the CUDA `multi_axis_rope` row of
-[RFC #386](https://github.com/RL-Align/RL-Kernel/issues/386), based on
-`test-qwenimage@16b81175874ecf11d5ff05c00de3b69e9457591b`.
-The source fingerprints in `manifest.json` identify the final implementation.
-This is single-GPU operator evidence; it does not qualify a model or trajectory.
-
-## Implementation
-
-- Adjacent pairs, axes `[16, 56, 56]`, canonical centered image positions and
-  diagonal text offsets, pinned to Diffusers `40f63690da440e8c382ab53e67508835deb1ed2f`.
-- Explicit CPU table factories remain CPU factories inside CUDA/meta default-device contexts.
-- Native CUDA reads all input/table strides directly, including packed QKV views.
-  The shared Triton implementation adapts FlashAttention's existing rotary kernel
-  at `98c1cee0ed59307833841cdbf567fe336ceec13a`, retaining attribution and BSD-3 terms.
-- Both GPU paths have dispatcher, fake and input-autograd registration. The
-  arithmetic is separate FP32 multiply/add followed by one output cast.
-- OP_SPECS, the ordinary input builder and a gradient adapter expose the operator
-  to existing validation APIs. The operator gate runs all four forward/gradient
-  accuracy/invariance judgments with explicit backend/kernel provenance.
-- Default CUDA builds include the native source in `_C`. The explicit
-  `RL_KERNEL_ROPE_ONLY=1` build profile also produces an installable companion
-  library, without compiling unrelated native operators.
-
-## Environment
-
-| Item | Value |
-| --- | --- |
-| GPU | NVIDIA GeForce RTX 5080, SM120, 16 GB |
-| Driver | 617.42 |
-| OS / Python | Windows / 3.13.9 |
-| PyTorch / CUDA toolkit | 2.9.1+cu128 / 12.8 |
-| Triton | triton-windows 3.5.1.post24; runtime 3.5.1 |
-| Native compiler | MSVC 2022 x64, C++17; CUDA target 12.0 |
-| Sanitizer | NVIDIA Compute Sanitizer 13.3, memcheck |
-| Benchmark dtype | BF16, FP32 position tables |
-
-The installed wheel is CPython 3.13/Windows-specific. Toolkit/driver versions
-have different roles; the native code was compiled with CUDA 12.8, not 13.3.
-
-## Validation
-
-| Check | Result | Evidence |
-| --- | --- | --- |
-| Public operator suite using the built companion library | 81 passed, 1 two-GPU skip | `tests.txt` |
-| Operator/build/registry/validation regressions | 195 passed, 7 Ascend skips | `regression.txt` |
-| CUDA WS1 four judgments | Passed | `ws1-cuda.json` |
-| Triton WS1 four judgments | Passed | `ws1-triton.json` |
-| Standard CLI, BF16 forward and input gradient, CUDA/Triton | Both passed | `operator-{cuda,triton}.json` |
-| Pinned actual Diffusers symbols | 30 cases passed; image/text tables exact | `upstream.json` |
-| Memory safety on selected accuracy/strided/empty/stream cases | 34 passed, zero memcheck errors | `memcheck.txt` |
-| Source distribution + operator wheel build | Passed; native sources and BSD license present | `package-build.txt`, `package.json` |
-| Fresh interpreter outside checkout after wheel installation | CUDA/Triton eager and fullgraph forward/input-gradient bytes exact | `installed-wheel.json` |
-| Changed Python formatting/static checks; strict docs build | Passed | `quality.txt` |
-| Repository layout check | 5 failures, 31 passed, 1 skipped | `layout.txt` |
-| Same layout check on clean base | Same 5 failures, 27 passed, 1 skipped | `baseline-layout.txt` |
-
-Operator tests cover FP32/FP16/BF16, shared/per-sample tables, odd and rectangular
-grids, batch/chunk/padding invariance, higher input VJPs, strided gradients,
-independent table strides, empty tensors, invalid-input rejection, non-default
-streams, graph replay with changed static buffers, `torch.library.opcheck`, and
-`torch.compile(..., fullgraph=True)` through backward. A CPU profiler assertion
-rejects `contiguous`, `clone` or `copy_` inside either GPU call for a packed view.
-
-The layout failures reproduce on the clean base: the Windows driver-link flag
-assertion, symlink creation without Windows privilege, and three checks affected
-by symlink-disabled JSON/source placeholders. The added build tests pass. This
-report does not describe the entire repository test suite as green.
-
-## Mature baseline and performance
-
-`benchmark-bfloat16.json` contains 40 pre-timing correctness comparisons and 80
-measurements: four geometries, two layouts, five implementations, two phases.
-The CUDA, adapted Triton and contraction-disabled upstream FlashAttention paths
-all have zero differing output/gradient bytes against the unfused FP32 gold.
-Original FlashAttention and Diffusers complex arithmetic use declared tolerances.
-
-The optional FlashAttention comparison loads the checksum-pinned original
-`rotary_kernel`, `apply_rotary` and `ApplyRotaryEmb` class without installing
-flash-attn. `flash_no_fma` adds only `enable_fp_fusion=False` to the original
-launch; the default baseline retains its fused arithmetic.
-
-CUDA Graph medians in ms/call; each cell is **forward / forward+backward**.
-Batch 1, 24 heads, D=128:
-
-| Grid / layout | Native CUDA | Adapted Triton | FlashAttention, no FMA | Diffusers complex |
-| --- | --- | --- | --- | --- |
-| 64×64 contiguous | 0.05771 / 0.12953 | 0.05191 / 0.12463 | 0.05202 / 0.12443 | 0.24277 / 0.50118 |
-| 64×64 packed QKV | 0.05866 / 0.13089 | 0.04835 / 0.12314 | 0.04896 / 0.12374 | 0.24599 / 0.50083 |
-| 83×83 contiguous | 0.11203 / 0.22478 | 0.10685 / 0.21339 | 0.10693 / 0.21351 | 0.51009 / 1.05054 |
-| 83×83 packed QKV | 0.11280 / 0.22514 | 0.10758 / 0.21415 | 0.10746 / 0.21431 | 0.51117 / 1.05461 |
-| 104×58 contiguous | 0.09809 / 0.19646 | 0.09356 / 0.18708 | 0.09362 / 0.18709 | 0.43411 / 0.89356 |
-| 104×58 packed QKV | 0.09596 / 0.19664 | 0.09341 / 0.18776 | 0.09344 / 0.18786 | 0.43668 / 0.89561 |
-
-The adapted Triton path is close to its mature upstream baseline. The native
-CUDA path avoids requiring Triton but is slower in this graph comparison.
-No speed advantage over FlashAttention is established. At 4096 tokens, either
-GPU path's forward extra peak is 24 MiB, exactly the returned BF16 output size,
-for both layouts. There is no extra packed-input materialization.
-
-Five warm calls/replays, 20 calls per captured graph, median of 30 replay
-measurements. Synchronized eager wall time is also recorded as three trials of
-30 calls. It includes Python/dispatcher/autograd overhead and can differ
-substantially from graph time. Tables, compilation, input allocation and packed
-backing storage are outside timing. Extra peak allocation includes the returned
-output or gradient. These are one-machine, one-process microbenchmarks, without
-clock pinning; close differences do not establish a stable winner.
-
-## Reproduction
-
-Set a toolkit matching PyTorch, the target architecture and a matching host
-compiler. On the recorded Windows host, run from an x64 MSVC environment with
-`CUDA_HOME`/`CUDA_PATH` selecting CUDA 12.8 and `TORCH_CUDA_ARCH_LIST=12.0`.
-Use `$env:RL_KERNEL_ROPE_ONLY='1'` for the operator package profile.
-
-```bash
-python setup.py sdist bdist_wheel
-python setup.py build_ext --inplace
-python -m pytest tests/ops/rope/test_multi_axis_rope.py -q -rs
-python tools/validation/operators/check_operator.py --op multi_axis_rope \
-  --candidate cuda --device cuda --dtype bf16 --batch 2 --seq 7 --n-heads 3 --check-grad --json
-python tools/validation/operators/validate_multi_axis_rope.py --ws1 cuda --output ws1-cuda.json
-python tools/validation/operators/validate_multi_axis_rope.py --ws1 triton --output ws1-triton.json
-python tools/validation/operators/validate_multi_axis_rope.py --benchmark \
-  --flash-source /path/to/flash_attn/ops/triton/rotary.py \
-  --flash-layer-source /path/to/flash_attn/layers/rotary.py --output benchmark.json
-python tools/validation/operators/check_qwenimage_rope_upstream.py \
-  --source /path/to/transformer_qwenimage.py --output upstream.json
-compute-sanitizer --tool memcheck --error-exitcode 1 python \
-  tools/validation/operators/validate_multi_axis_rope.py -q \
-  -k "forward_backward_accuracy or noncontiguous_input_tables_and_gradient or empty_forward_backward or packed_qkv_and_independent_table_strides or native_entrypoint_checks"
-mkdocs build --strict -f mkdocs.yaml
-```
-
-Install the built wheel, change to a directory outside the checkout, then run
-`check_multi_axis_rope_install.py --package-root /path/to/site-packages --output installed-wheel.json`.
-This probe neither changes sys.path nor invokes the focused JIT loader.
-See `quality.txt` for the regression command and static-check scope.
-
-## Related work and remaining qualification
-
-The [operator page](../../../docs/operators/multi-axis-rope.md) documents
-FlashAttention reuse and the related Ascend #411, SGLang and vLLM-Omni paths.
-The live repository search on 2026-10-28 found #411 for the exact row name;
-its head remains `285c4dccab7e9048956584caf431946d82405c09`. Its half-split and
-uncentered reference differs from the pinned adjacent-pair/centered contract.
-
-Linux, A100/H100, multi-GPU guards, the complete monolithic `_C` build, QK-norm
-fusion, model/block assembly and trajectories remain unqualified. There is no
-Dense Qwen3 or distributed certification implied by the operator WS1 adapter.
-Cross-version/device byte equality and fast-math builds are not claimed.
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/baseline-layout.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/baseline-layout.txt
deleted file mode 100644
index 16a5c97e8..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/baseline-layout.txt
+++ /dev/null
@@ -1,138 +0,0 @@
-.FF.............s........F.FF....                                        [100%]
-================================== FAILURES ===================================
-________________ test_cuda_build_keeps_existing_ipc_collective ________________
-
-monkeypatch = <_pytest.monkeypatch.MonkeyPatch object at 0x0000020FF72168B0>
-
-    def test_cuda_build_keeps_existing_ipc_collective(monkeypatch) -> None:
-        extension = _load_extension_config(monkeypatch, hip=None)
-
-        assert "csrc/cuda/collectives/deterministic_collective.cu" in extension["sources"]
-        assert "-DKERNEL_ALIGN_WITH_CUDA" in extension["extra_compile_args"]["cxx"]
-        assert "-DKERNEL_ALIGN_WITH_ROCM" not in extension["extra_compile_args"]["cxx"]
->       assert "-lcuda" in extension["extra_link_args"]
-E       AssertionError: assert '-lcuda' in ['-Wl,-rpath,$ORIGIN/../torch/lib']
-
-tests\build\test_build_platform_collectives.py:61: AssertionError
-____________ test_package_data_json_links_resolve_to_real_content _____________
-
-tmp_path = WindowsPath('C:/Users/123/AppData/Local/Temp/pytest-of-123/pytest-4/test_package_data_json_links_r0')
-
-    def test_package_data_json_links_resolve_to_real_content(tmp_path) -> None:
-        target = tmp_path / "configs" / "profile.json"
-        target.parent.mkdir()
-        target.write_text('{"name": "profile"}\n', encoding="utf-8")
-        source = tmp_path / "rl_engine" / "profile.json"
-        source.parent.mkdir()
->       source.symlink_to("../configs/profile.json")
-
-tests\build\test_build_platform_collectives.py:70:
-_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
-
-self = WindowsPath('C:/Users/123/AppData/Local/Temp/pytest-of-123/pytest-4/test_package_data_json_links_r0/rl_engine/profile.json')
-target = '../configs/profile.json', target_is_directory = False
-
-    def symlink_to(self, target, target_is_directory=False):
-        """
-        Make this path a symlink pointing to the target path.
-        Note the order of arguments (link, target) is the reverse of os.symlink.
-        """
->       os.symlink(target, self, target_is_directory)
-E       OSError: [WinError 1314] �ͻ���û���������Ȩ��: '../configs/profile.json' -> 'C:\\Users\\123\\AppData\\Local\\Temp\\pytest-of-123\\pytest-4\\test_package_data_json_links_r0\\rl_engine\\profile.json'
-
-D:\anaconda\Lib\pathlib\_local.py:789: OSError
-_______ test_recorded_algorithm_sources_resolve_to_real_implementations _______
-
-    def test_recorded_algorithm_sources_resolve_to_real_implementations():
-        from rl_engine.config.workload import load_manifest
-
-        moves = json.loads((ROOT / "tools/migration/layout.json").read_text())["moves"]
-        for case in load_manifest().representative_cases:
-            source, symbol = case["provenance_evidence"]["algorithm_source"].rsplit(":", 1)
-            implementation = ROOT / moves.get(source, source)
->           text = implementation.read_text()
-                   ^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-tests\runtime\test_layout_compatibility.py:87:
-_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
-D:\anaconda\Lib\pathlib\_local.py:546: in read_text
-    return PathBase.read_text(self, encoding, errors, newline)
-           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
-
-self = WindowsPath('D:/dandan123/dandan/issue/investigation-rl-kernel/baseline-validation/csrc/cuda/attention/deterministic_attention.cu')
-encoding = 'locale', errors = None, newline = None
-
-    def read_text(self, encoding=None, errors=None, newline=None):
-        """
-        Open the file in text mode, read it, and close the file.
-        """
-        with self.open(mode='r', encoding=encoding, errors=errors, newline=newline) as f:
->           return f.read()
-                   ^^^^^^^^
-E           UnicodeDecodeError: 'gbk' codec can't decode byte 0x92 in position 173: illegal multibyte sequence
-
-D:\anaconda\Lib\pathlib\_abc.py:633: UnicodeDecodeError
-_ test_packaged_contract_bytes_match_main[rl_engine.config-workloads/qwen3_8b.json-80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6] _
-
-package = 'rl_engine.config', resource = 'workloads/qwen3_8b.json'
-digest = '80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6'
-
-    @pytest.mark.parametrize(
-        ("package", "resource", "digest"),
-        [
-            (
-                "rl_engine.config",
-                "workloads/qwen3_8b.json",
-                "80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6",
-            ),
-            (
-                "rl_engine.contracts",
-                "profiles/precision/ws1.json",
-                "044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532",
-            ),
-        ],
-    )
-    def test_packaged_contract_bytes_match_main(package, resource, digest):
->       assert hashlib.sha256(files(package).joinpath(resource).read_bytes()).hexdigest() == digest
-E       AssertionError: assert 'c97e966fb075...75e111e052d7d' == '80395644b5e8...f19635d3df0f6'
-E
-E         - 80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6
-E         + c97e966fb075dad8ce7dcd7a23a58ffb856e7474dfb3c50510675e111e052d7d
-
-tests\runtime\test_layout_compatibility.py:118: AssertionError
-_ test_packaged_contract_bytes_match_main[rl_engine.contracts-profiles/precision/ws1.json-044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532] _
-
-package = 'rl_engine.contracts', resource = 'profiles/precision/ws1.json'
-digest = '044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532'
-
-    @pytest.mark.parametrize(
-        ("package", "resource", "digest"),
-        [
-            (
-                "rl_engine.config",
-                "workloads/qwen3_8b.json",
-                "80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6",
-            ),
-            (
-                "rl_engine.contracts",
-                "profiles/precision/ws1.json",
-                "044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532",
-            ),
-        ],
-    )
-    def test_packaged_contract_bytes_match_main(package, resource, digest):
->       assert hashlib.sha256(files(package).joinpath(resource).read_bytes()).hexdigest() == digest
-E       AssertionError: assert '2020c02e7af4...313d6752cd34e' == '044167ec21e3...7628cbd101532'
-E
-E         - 044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532
-E         + 2020c02e7af4e2ca4163e1ecfb9c491baef07b73deba39b0f44313d6752cd34e
-
-tests\runtime\test_layout_compatibility.py:118: AssertionError
-=========================== short test summary info ===========================
-FAILED tests/build/test_build_platform_collectives.py::test_cuda_build_keeps_existing_ipc_collective
-FAILED tests/build/test_build_platform_collectives.py::test_package_data_json_links_resolve_to_real_content
-FAILED tests/runtime/test_layout_compatibility.py::test_recorded_algorithm_sources_resolve_to_real_implementations
-FAILED tests/runtime/test_layout_compatibility.py::test_packaged_contract_bytes_match_main[rl_engine.config-workloads/qwen3_8b.json-80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6]
-FAILED tests/runtime/test_layout_compatibility.py::test_packaged_contract_bytes_match_main[rl_engine.contracts-profiles/precision/ws1.json-044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532]
-5 failed, 27 passed, 1 skipped in 5.70s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/benchmark-bfloat16.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/benchmark-bfloat16.json
deleted file mode 100644
index 72fb42bbe..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/benchmark-bfloat16.json
+++ /dev/null
@@ -1,2781 +0,0 @@
-{
-  "gpu": "NVIDIA GeForce RTX 5080",
-  "capability": [
-    12,
-    0
-  ],
-  "os": "Windows-11-10.0.26200-SP0",
-  "python": "3.13.9 | packaged by Anaconda, Inc. | (main, Oct 21 2025, 19:09:58) [MSC v.1929 64 bit (AMD64)]",
-  "torch": "2.9.1+cu128",
-  "cuda_runtime": "12.8",
-  "triton": "3.5.1",
-  "dtype": "bfloat16",
-  "base_commit": "16b81175874ecf11d5ff05c00de3b69e9457591b",
-  "source_sha256": {
-    "csrc/cuda/rope/multi_axis_rope.cu": "f5db5cf5d4962053194455d9cde8cf91cf4e824bad7abf72e8410a404bcb0e44",
-    "rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py": "5c6984da351e5228f7549629d4ff53e5772b9a6ef2a01ba2a38493e769b15ab5",
-    "rl_engine/backends/shared/triton/rope/multi_axis_rope.py": "5b989bd63efd77913267b48654d10204bc563768c6c5d4e1cde9fb55e1354547",
-    "rl_engine/backends/cuda/rope/multi_axis_rope.py": "d779d9b60ce5d1d6651925f723049930d898024fdf0548ce29a2019b84753d65",
-    "rl_engine/ops/rope/_registration.py": "add0375d1639444908091d01e8d29781f7a975beffec64b77bfaa81a5b9a0ae3",
-    "rl_engine/ops/rope/multi_axis_rope.py": "07bcff1160f22c1ff6730be731206b2f4a153e2fada26afc93fc884e780568cb",
-    "rl_engine/reference/rope/multi_axis_rope.py": "c9a94730047f51659f7aa0a443e88d6ce7d03d00d6813a9aed548cafbf6347bf",
-    "benchmarks/operators/rope/benchmark_multi_axis_rope.py": "6961a55803fbd4e566bd060558fb59e658e7f1f704e61ca97f912e5d16b55884"
-  },
-  "timing": "CUDA graph ms/call and synchronized eager wall ms/call; warm allocation; excludes tables, compilation, and initial allocation of inputs",
-  "memory": "extra peak allocated bytes excluding inputs/tables/grad; includes returned output or gradient; packed QKV backing storage allocated before timing",
-  "iterations": 30,
-  "warmup": 5,
-  "trials": 3,
-  "graph_calls": 20,
-  "flash_baseline": {
-    "included": true,
-    "revision": "98c1cee0ed59307833841cdbf567fe336ceec13a",
-    "kernel_sha256_normalized": "d2037e623854c702fb075784fe4292d4a89d6b9ac1ae33cab3d6788f8f94f240",
-    "layer_sha256_normalized": "a79cb42dfc33e126227bdba27ae7a024d7d5cc5c8fd7f7a5f26c68d501024e0a",
-    "mode": "upstream rotary_kernel, apply_rotary and ApplyRotaryEmb, interleaved=True",
-    "no_fma_change": "only add enable_fp_fusion=False at the original kernel launch"
-  },
-  "trace": {
-    "cuda": {
-      "contract": "qwenimage-interleaved-rope-fp32-v1",
-      "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
-      "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
-      "layout": "BSHD, adjacent channel pairs",
-      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-    },
-    "triton": {
-      "contract": "qwenimage-interleaved-rope-fp32-v1",
-      "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
-      "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
-      "layout": "BSHD, adjacent channel pairs",
-      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-    }
-  },
-  "qualification": "Exact bytes vs unfused FP32 gold for CUDA, Triton and flash_no_fma; declared tolerance for upstream complex and default FlashAttention",
-  "correctness": [
-    {
-      "backend": "upstream_complex",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        130,
-        139
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        130,
-        139
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        139,
-        140
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        139,
-        140
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        234,
-        225
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        234,
-        225
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        220,
-        226
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        220,
-        226
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        196,
-        203
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        196,
-        203
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        230,
-        218
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        230,
-        218
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    }
-  ],
-  "results": [
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.003558399900794029,
-      "graph_min_ms": 0.0035264000296592714,
-      "graph_max_ms": 0.015438400208950043,
-      "eager_wall_median_ms": 0.0339666667059646,
-      "eager_wall_trials_ms": [
-        0.0356399999873247,
-        0.0339666667059646,
-        0.03332666662269427
-      ],
-      "extra_peak_bytes": 184320
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.007688799872994423,
-      "graph_min_ms": 0.007651200145483017,
-      "graph_max_ms": 0.007790400087833405,
-      "eager_wall_median_ms": 0.21265999997316007,
-      "eager_wall_trials_ms": [
-        0.21187666667780528,
-        0.21265999997316007,
-        0.22587333332921844
-      ],
-      "extra_peak_bytes": 187392
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0012016000226140022,
-      "graph_min_ms": 0.0011567999608814716,
-      "graph_max_ms": 0.0012384000234305858,
-      "eager_wall_median_ms": 0.02289333333465038,
-      "eager_wall_trials_ms": [
-        0.02289333333465038,
-        0.02245666667780218,
-        0.02458333328831941
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0022280000150203703,
-      "graph_min_ms": 0.0021888000890612604,
-      "graph_max_ms": 0.0022655999287962913,
-      "eager_wall_median_ms": 0.1764300000407578,
-      "eager_wall_trials_ms": [
-        0.20031666669334905,
-        0.1764300000407578,
-        0.15246333338533682
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.001142400037497282,
-      "graph_min_ms": 0.0010591999627649785,
-      "graph_max_ms": 0.005691200122237205,
-      "eager_wall_median_ms": 0.057766666698929235,
-      "eager_wall_trials_ms": [
-        0.07126333330234047,
-        0.057766666698929235,
-        0.05640333335274287
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.002125599980354309,
-      "graph_min_ms": 0.002091200090944767,
-      "graph_max_ms": 0.002160000056028366,
-      "eager_wall_median_ms": 0.21872666663208898,
-      "eager_wall_trials_ms": [
-        0.22702333335473668,
-        0.21872666663208898,
-        0.21611666670651175
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0010935999918729067,
-      "graph_min_ms": 0.001033599954098463,
-      "graph_max_ms": 0.0011536000296473504,
-      "eager_wall_median_ms": 0.10024000002886169,
-      "eager_wall_trials_ms": [
-        0.10126666669142044,
-        0.09858333329854456,
-        0.10024000002886169
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0020527999848127367,
-      "graph_min_ms": 0.002019199915230274,
-      "graph_max_ms": 0.0021023999899625776,
-      "eager_wall_median_ms": 0.3111766666431019,
-      "eager_wall_trials_ms": [
-        0.30222666664485587,
-        0.3111766666431019,
-        0.333893333360417
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0010991999879479408,
-      "graph_min_ms": 0.001033599954098463,
-      "graph_max_ms": 0.0011392000131309032,
-      "eager_wall_median_ms": 0.10421333336125826,
-      "eager_wall_trials_ms": [
-        0.10421333336125826,
-        0.10106999998242827,
-        0.12048333337588701
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0020503999665379526,
-      "graph_min_ms": 0.002001599967479706,
-      "graph_max_ms": 0.00886240005493164,
-      "eager_wall_median_ms": 0.31822000000829576,
-      "eager_wall_trials_ms": [
-        0.28550666665978497,
-        0.31822000000829576,
-        0.33777000001767493
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.00379519984126091,
-      "graph_min_ms": 0.003745599836111069,
-      "graph_max_ms": 0.003859199956059456,
-      "eager_wall_median_ms": 0.10653666668076767,
-      "eager_wall_trials_ms": [
-        0.10653666668076767,
-        0.10879000001295935,
-        0.09587666669782872
-      ],
-      "extra_peak_bytes": 184320
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.008064000308513642,
-      "graph_min_ms": 0.008011200278997422,
-      "graph_max_ms": 0.008102399855852127,
-      "eager_wall_median_ms": 0.2260500000071867,
-      "eager_wall_trials_ms": [
-        0.23646333332483968,
-        0.2187999999781217,
-        0.2260500000071867
-      ],
-      "extra_peak_bytes": 187392
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0012024000287055969,
-      "graph_min_ms": 0.0011392000131309032,
-      "graph_max_ms": 0.0025040000677108766,
-      "eager_wall_median_ms": 0.023706666615908034,
-      "eager_wall_trials_ms": [
-        0.023706666615908034,
-        0.024339999981748406,
-        0.023019999995691858
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.002226400002837181,
-      "graph_min_ms": 0.002174399979412556,
-      "graph_max_ms": 0.0022671999409794807,
-      "eager_wall_median_ms": 0.157756666643157,
-      "eager_wall_trials_ms": [
-        0.165013333329019,
-        0.15389999995628992,
-        0.157756666643157
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0011008000001311302,
-      "graph_min_ms": 0.0010688000358641148,
-      "graph_max_ms": 0.0011359999887645244,
-      "eager_wall_median_ms": 0.059236666675133165,
-      "eager_wall_trials_ms": [
-        0.053669999942940194,
-        0.06250666665437166,
-        0.059236666675133165
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0021215999498963357,
-      "graph_min_ms": 0.002065600082278252,
-      "graph_max_ms": 0.0021679999306797983,
-      "eager_wall_median_ms": 0.25444666662224336,
-      "eager_wall_trials_ms": [
-        0.25444666662224336,
-        0.27043333332888625,
-        0.2240899999984928
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.001101600006222725,
-      "graph_min_ms": 0.001065600011497736,
-      "graph_max_ms": 0.0011376000009477138,
-      "eager_wall_median_ms": 0.12566333334689261,
-      "eager_wall_trials_ms": [
-        0.12566333334689261,
-        0.12918333328949907,
-        0.10983666664590903
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0020527999848127367,
-      "graph_min_ms": 0.0020160000771284104,
-      "graph_max_ms": 0.0021023999899625776,
-      "eager_wall_median_ms": 0.28767333333235,
-      "eager_wall_trials_ms": [
-        0.2609133333559536,
-        0.28767333333235,
-        0.3289566666353494
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.001101600006222725,
-      "graph_min_ms": 0.0010223999619483949,
-      "graph_max_ms": 0.0022975999861955644,
-      "eager_wall_median_ms": 0.0995366666757036,
-      "eager_wall_trials_ms": [
-        0.10132333336514421,
-        0.0995366666757036,
-        0.0990000000456348
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0020527999848127367,
-      "graph_min_ms": 0.0020128000527620316,
-      "graph_max_ms": 0.0021007999777793883,
-      "eager_wall_median_ms": 0.3198099999887442,
-      "eager_wall_trials_ms": [
-        0.33375666668386356,
-        0.30899333335886087,
-        0.3198099999887442
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.2427680015563965,
-      "graph_min_ms": 0.24123198986053468,
-      "graph_max_ms": 0.3049216032028198,
-      "eager_wall_median_ms": 0.25121666670505266,
-      "eager_wall_trials_ms": [
-        0.24960999999166233,
-        0.2669566666857766,
-        0.25121666670505266
-      ],
-      "extra_peak_bytes": 125829120
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.5011807918548584,
-      "graph_min_ms": 0.4943295955657959,
-      "graph_max_ms": 0.5467391967773437,
-      "eager_wall_median_ms": 0.5078799999561548,
-      "eager_wall_trials_ms": [
-        0.5254200000005463,
-        0.5078799999561548,
-        0.5070699999729792
-      ],
-      "extra_peak_bytes": 127926272
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.05770559906959534,
-      "graph_min_ms": 0.05762240290641785,
-      "graph_max_ms": 0.07225120067596436,
-      "eager_wall_median_ms": 0.061780000032740645,
-      "eager_wall_trials_ms": [
-        0.06511333334249987,
-        0.06060333331940152,
-        0.061780000032740645
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.12953280210494994,
-      "graph_min_ms": 0.12918239831924438,
-      "graph_max_ms": 0.14451359510421752,
-      "eager_wall_median_ms": 0.15664999997776854,
-      "eager_wall_trials_ms": [
-        0.2333100000517637,
-        0.15664999997776854,
-        0.14973666669296412
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.05190559923648834,
-      "graph_min_ms": 0.05078719854354859,
-      "graph_max_ms": 0.06408159732818604,
-      "eager_wall_median_ms": 0.0845666666161075,
-      "eager_wall_trials_ms": [
-        0.08885000000494377,
-        0.08201333330362104,
-        0.0845666666161075
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.12463040351867677,
-      "graph_min_ms": 0.12331520318984986,
-      "graph_max_ms": 0.1529088020324707,
-      "eager_wall_median_ms": 0.2539966667124342,
-      "eager_wall_trials_ms": [
-        0.2757299999454214,
-        0.25390333333537757,
-        0.2539966667124342
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.05208719968795776,
-      "graph_min_ms": 0.05123199820518494,
-      "graph_max_ms": 0.06861280202865601,
-      "eager_wall_median_ms": 0.13019999999717888,
-      "eager_wall_trials_ms": [
-        0.1370399999965836,
-        0.12465666665472479,
-        0.13019999999717888
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.12441120147705079,
-      "graph_min_ms": 0.12314239740371705,
-      "graph_max_ms": 0.139137601852417,
-      "eager_wall_median_ms": 0.340176666698729,
-      "eager_wall_trials_ms": [
-        0.35652999995363643,
-        0.340176666698729,
-        0.33921999996285496
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.052018401026725766,
-      "graph_min_ms": 0.05074719786643982,
-      "graph_max_ms": 0.07086880207061767,
-      "eager_wall_median_ms": 0.10117999997116083,
-      "eager_wall_trials_ms": [
-        0.10250666667464732,
-        0.10117999997116083,
-        0.10012000002461718
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.12442959547042846,
-      "graph_min_ms": 0.12299360036849975,
-      "graph_max_ms": 0.14664160013198851,
-      "eager_wall_median_ms": 0.34945666666317265,
-      "eager_wall_trials_ms": [
-        0.36080999998375773,
-        0.3163533333160255,
-        0.34945666666317265
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.2459879994392395,
-      "graph_min_ms": 0.24467840194702148,
-      "graph_max_ms": 0.272545599937439,
-      "eager_wall_median_ms": 0.2536533333720096,
-      "eager_wall_trials_ms": [
-        0.27815666665749933,
-        0.2529566666756485,
-        0.2536533333720096
-      ],
-      "extra_peak_bytes": 125829120
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.5008344173431396,
-      "graph_min_ms": 0.4996367931365967,
-      "graph_max_ms": 0.5373487949371338,
-      "eager_wall_median_ms": 0.5262199999682101,
-      "eager_wall_trials_ms": [
-        0.5269766666970099,
-        0.5070566666593853,
-        0.5262199999682101
-      ],
-      "extra_peak_bytes": 127926272
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.05865840017795562,
-      "graph_min_ms": 0.058550399541854856,
-      "graph_max_ms": 0.08192480206489564,
-      "eager_wall_median_ms": 0.06282000000889336,
-      "eager_wall_trials_ms": [
-        0.06282000000889336,
-        0.06640333334265354,
-        0.06277333335068154
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.1308944046497345,
-      "graph_min_ms": 0.13030719757080078,
-      "graph_max_ms": 0.145305597782135,
-      "eager_wall_median_ms": 0.1688300000144712,
-      "eager_wall_trials_ms": [
-        0.1688300000144712,
-        0.15922666667999388,
-        0.17570666665657578
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.04834639877080918,
-      "graph_min_ms": 0.047383999824523924,
-      "graph_max_ms": 0.06508319973945617,
-      "eager_wall_median_ms": 0.08053666667062014,
-      "eager_wall_trials_ms": [
-        0.09511666667094687,
-        0.07707999999790142,
-        0.08053666667062014
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.1231384038925171,
-      "graph_min_ms": 0.12216000556945801,
-      "graph_max_ms": 0.1478927969932556,
-      "eager_wall_median_ms": 0.2741499999804849,
-      "eager_wall_trials_ms": [
-        0.28788666665301815,
-        0.2741499999804849,
-        0.24507333334137607
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.048105600476264956,
-      "graph_min_ms": 0.0474047988653183,
-      "graph_max_ms": 0.06945440173149109,
-      "eager_wall_median_ms": 0.12940666668631215,
-      "eager_wall_trials_ms": [
-        0.14179000002817096,
-        0.12940666668631215,
-        0.12683000004471978
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.12347760200500488,
-      "graph_min_ms": 0.12205439805984497,
-      "graph_max_ms": 0.14951200485229493,
-      "eager_wall_median_ms": 0.36926333335334977,
-      "eager_wall_trials_ms": [
-        0.36926333335334977,
-        0.3322200000184239,
-        0.36990666667406913
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.04896320104598999,
-      "graph_min_ms": 0.047470399737358095,
-      "graph_max_ms": 0.06636160016059875,
-      "eager_wall_median_ms": 0.14505666664869446,
-      "eager_wall_trials_ms": [
-        0.14700000004571243,
-        0.14505666664869446,
-        0.13565333332129134
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.12373999953269958,
-      "graph_min_ms": 0.1219599962234497,
-      "graph_max_ms": 0.1542415976524353,
-      "eager_wall_median_ms": 0.33643999995547347,
-      "eager_wall_trials_ms": [
-        0.33643999995547347,
-        0.3219899999749032,
-        0.3583933333478247
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.51009202003479,
-      "graph_min_ms": 0.5082543849945068,
-      "graph_max_ms": 0.5431327819824219,
-      "eager_wall_median_ms": 0.526480000007723,
-      "eager_wall_trials_ms": [
-        0.5092500000197712,
-        0.5466633333223095,
-        0.526480000007723
-      ],
-      "extra_peak_bytes": 211630080
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 1.0505383968353272,
-      "graph_min_ms": 1.0236895561218262,
-      "graph_max_ms": 1.129139232635498,
-      "eager_wall_median_ms": 1.044759999967937,
-      "eager_wall_trials_ms": [
-        1.044759999967937,
-        1.0434166666527744,
-        1.0474333333453008
-      ],
-      "extra_peak_bytes": 215157248
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.11203439831733704,
-      "graph_min_ms": 0.1114351987838745,
-      "graph_max_ms": 0.14540959596633912,
-      "eager_wall_median_ms": 0.11692666666931473,
-      "eager_wall_trials_ms": [
-        0.11580999998841435,
-        0.11692666666931473,
-        0.12043333335896023
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.22477760314941406,
-      "graph_min_ms": 0.22440481185913086,
-      "graph_max_ms": 0.24935200214385986,
-      "eager_wall_median_ms": 0.23845333331943644,
-      "eager_wall_trials_ms": [
-        0.23845333331943644,
-        0.2526333333056148,
-        0.230336666633472
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.10685279965400696,
-      "graph_min_ms": 0.10539519786834717,
-      "graph_max_ms": 0.12008960247039795,
-      "eager_wall_median_ms": 0.11027333330275724,
-      "eager_wall_trials_ms": [
-        0.11027333330275724,
-        0.10863333330538201,
-        0.11599999997997656
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.21339280605316163,
-      "graph_min_ms": 0.2126607894897461,
-      "graph_max_ms": 0.23967199325561522,
-      "eager_wall_median_ms": 0.27515000001585577,
-      "eager_wall_trials_ms": [
-        0.2870199999961187,
-        0.2669133333256468,
-        0.27515000001585577
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.10692960023880005,
-      "graph_min_ms": 0.10526080131530761,
-      "graph_max_ms": 0.12183040380477905,
-      "eager_wall_median_ms": 0.10955999999472017,
-      "eager_wall_trials_ms": [
-        0.1204166666866513,
-        0.10955999999472017,
-        0.10893666670502473
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.21346559524536132,
-      "graph_min_ms": 0.21225121021270751,
-      "graph_max_ms": 0.2430032014846802,
-      "eager_wall_median_ms": 0.338726666692916,
-      "eager_wall_trials_ms": [
-        0.33888666663794237,
-        0.3213533333716138,
-        0.338726666692916
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.10693199634552003,
-      "graph_min_ms": 0.10515999794006348,
-      "graph_max_ms": 0.12142080068588257,
-      "eager_wall_median_ms": 0.1100600000427221,
-      "eager_wall_trials_ms": [
-        0.11902333335456206,
-        0.1100600000427221,
-        0.10893666664439176
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.21350640058517456,
-      "graph_min_ms": 0.2123823881149292,
-      "graph_max_ms": 0.23954401016235352,
-      "eager_wall_median_ms": 0.3223733333167426,
-      "eager_wall_trials_ms": [
-        0.29901333333934116,
-        0.3223733333167426,
-        0.326143333343983
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.5111671924591064,
-      "graph_min_ms": 0.5095888137817383,
-      "graph_max_ms": 0.5389599800109863,
-      "eager_wall_median_ms": 0.5120633333717706,
-      "eager_wall_trials_ms": [
-        0.5099033333560025,
-        0.532273333374178,
-        0.5120633333717706
-      ],
-      "extra_peak_bytes": 211630080
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 1.0546072006225586,
-      "graph_min_ms": 1.0272704124450684,
-      "graph_max_ms": 1.1091728210449219,
-      "eager_wall_median_ms": 1.0453999999905743,
-      "eager_wall_trials_ms": [
-        1.0565066666458733,
-        1.0453999999905743,
-        1.042836666662576
-      ],
-      "extra_peak_bytes": 215157248
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.11280480027198792,
-      "graph_min_ms": 0.11215200424194335,
-      "graph_max_ms": 0.13529759645462036,
-      "eager_wall_median_ms": 0.12308333334658528,
-      "eager_wall_trials_ms": [
-        0.11563333331044608,
-        0.12308333334658528,
-        0.12365000002318993
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.2251416087150574,
-      "graph_min_ms": 0.22450559139251708,
-      "graph_max_ms": 0.2644239902496338,
-      "eager_wall_median_ms": 0.2328933333653064,
-      "eager_wall_trials_ms": [
-        0.23086333330866182,
-        0.2328933333653064,
-        0.25025666667109664
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.1075808048248291,
-      "graph_min_ms": 0.1061360001564026,
-      "graph_max_ms": 0.12249759435653687,
-      "eager_wall_median_ms": 0.10984000000462402,
-      "eager_wall_trials_ms": [
-        0.1298533332980393,
-        0.10860666661756113,
-        0.10984000000462402
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.21414560079574585,
-      "graph_min_ms": 0.21337599754333497,
-      "graph_max_ms": 0.23953759670257568,
-      "eager_wall_median_ms": 0.26501666670810664,
-      "eager_wall_trials_ms": [
-        0.2692166666747653,
-        0.23041333333821967,
-        0.26501666670810664
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.10766559839248657,
-      "graph_min_ms": 0.10606240034103394,
-      "graph_max_ms": 0.13348640203475953,
-      "eager_wall_median_ms": 0.11105333330003002,
-      "eager_wall_trials_ms": [
-        0.11928999999023897,
-        0.11105333330003002,
-        0.11004666666849516
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.21448720693588258,
-      "graph_min_ms": 0.21337599754333497,
-      "graph_max_ms": 0.25474560260772705,
-      "eager_wall_median_ms": 0.30118666666870314,
-      "eager_wall_trials_ms": [
-        0.32142000000021653,
-        0.30118666666870314,
-        0.2931233333583805
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.1074567973613739,
-      "graph_min_ms": 0.10611200332641602,
-      "graph_max_ms": 0.1223423957824707,
-      "eager_wall_median_ms": 0.11078333336627111,
-      "eager_wall_trials_ms": [
-        0.12015666664713838,
-        0.10989999997642978,
-        0.11078333336627111
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.21430879831314087,
-      "graph_min_ms": 0.2134079933166504,
-      "graph_max_ms": 0.23155200481414795,
-      "eager_wall_median_ms": 0.3170500000123866,
-      "eager_wall_trials_ms": [
-        0.3170500000123866,
-        0.32710333331730607,
-        0.31006333332091646
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.43411200046539306,
-      "graph_min_ms": 0.4321375846862793,
-      "graph_max_ms": 0.46126718521118165,
-      "eager_wall_median_ms": 0.4498766666680846,
-      "eager_wall_trials_ms": [
-        0.43220333336648764,
-        0.4498766666680846,
-        0.46697666663627996
-      ],
-      "extra_peak_bytes": 185303040
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.8935583591461181,
-      "graph_min_ms": 0.870751953125,
-      "graph_max_ms": 0.9448016166687012,
-      "eager_wall_median_ms": 0.8892600000156866,
-      "eager_wall_trials_ms": [
-        0.8870666666553007,
-        0.8892600000156866,
-        0.8902699999453034
-      ],
-      "extra_peak_bytes": 188391424
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09808879792690278,
-      "graph_min_ms": 0.09759039878845215,
-      "graph_max_ms": 0.11292320489883423,
-      "eager_wall_median_ms": 0.11404999998679462,
-      "eager_wall_trials_ms": [
-        0.11404999998679462,
-        0.11419666667886001,
-        0.102829999984048
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.19646480083465576,
-      "graph_min_ms": 0.19599839448928832,
-      "graph_max_ms": 0.21091840267181397,
-      "eager_wall_median_ms": 0.20405333331533862,
-      "eager_wall_trials_ms": [
-        0.20405333331533862,
-        0.20323666661473302,
-        0.2081700000417186
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09356399774551391,
-      "graph_min_ms": 0.09257439970970154,
-      "graph_max_ms": 0.10800800323486329,
-      "eager_wall_median_ms": 0.10138999999374694,
-      "eager_wall_trials_ms": [
-        0.10716000003109609,
-        0.10138999999374694,
-        0.10003000000627556
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.18708080053329468,
-      "graph_min_ms": 0.1862496018409729,
-      "graph_max_ms": 0.21909279823303224,
-      "eager_wall_median_ms": 0.26604333331003244,
-      "eager_wall_trials_ms": [
-        0.30756000002535683,
-        0.2481433333135404,
-        0.26604333331003244
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09362079799175263,
-      "graph_min_ms": 0.09246399998664856,
-      "graph_max_ms": 0.10841599702835084,
-      "eager_wall_median_ms": 0.10133000002194119,
-      "eager_wall_trials_ms": [
-        0.14010333331195093,
-        0.10133000002194119,
-        0.10039666667580605
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.18697920441627502,
-      "graph_min_ms": 0.18585760593414308,
-      "graph_max_ms": 0.21307199001312255,
-      "eager_wall_median_ms": 0.3559766667118917,
-      "eager_wall_trials_ms": [
-        0.3559766667118917,
-        0.3666866666511244,
-        0.35564333332634607
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09362399876117705,
-      "graph_min_ms": 0.09233919978141784,
-      "graph_max_ms": 0.10795680284500123,
-      "eager_wall_median_ms": 0.1356800000091122,
-      "eager_wall_trials_ms": [
-        0.14346666666824603,
-        0.1356800000091122,
-        0.1269966666465431
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.18709120154380798,
-      "graph_min_ms": 0.18642239570617675,
-      "graph_max_ms": 0.22162079811096191,
-      "eager_wall_median_ms": 0.3122899999652873,
-      "eager_wall_trials_ms": [
-        0.29407333331619157,
-        0.3122899999652873,
-        0.36742666664698237
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.43668241500854493,
-      "graph_min_ms": 0.43459200859069824,
-      "graph_max_ms": 0.4815072059631348,
-      "eager_wall_median_ms": 0.4541133333380761,
-      "eager_wall_trials_ms": [
-        0.4636933333434475,
-        0.43558999999125564,
-        0.4541133333380761
-      ],
-      "extra_peak_bytes": 185303040
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.8956063747406007,
-      "graph_min_ms": 0.8733648300170899,
-      "graph_max_ms": 0.9540863990783691,
-      "eager_wall_median_ms": 0.8916500000244317,
-      "eager_wall_trials_ms": [
-        0.8923599999737538,
-        0.8916500000244317,
-        0.8882866666681366
-      ],
-      "extra_peak_bytes": 188391424
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09595520198345184,
-      "graph_min_ms": 0.09553279876708984,
-      "graph_max_ms": 0.11051679849624634,
-      "eager_wall_median_ms": 0.1001133333678202,
-      "eager_wall_trials_ms": [
-        0.1001133333678202,
-        0.09971333335367187,
-        0.10664000001270324
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.19663520455360411,
-      "graph_min_ms": 0.19614880084991454,
-      "graph_max_ms": 0.22351360321044922,
-      "eager_wall_median_ms": 0.20444999997077198,
-      "eager_wall_trials_ms": [
-        0.22424333334735516,
-        0.20414666669239523,
-        0.20444999997077198
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0934136003255844,
-      "graph_min_ms": 0.0916208028793335,
-      "graph_max_ms": 0.10537919998168946,
-      "eager_wall_median_ms": 0.09456333336856915,
-      "eager_wall_trials_ms": [
-        0.09251333334153362,
-        0.09570666667665743,
-        0.09456333336856915
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.1877624034881592,
-      "graph_min_ms": 0.18668639659881592,
-      "graph_max_ms": 0.21486239433288573,
-      "eager_wall_median_ms": 0.24588000002646973,
-      "eager_wall_trials_ms": [
-        0.26271666671770316,
-        0.24588000002646973,
-        0.23198333334827717
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09341759979724884,
-      "graph_min_ms": 0.0921504020690918,
-      "graph_max_ms": 0.11279360055923462,
-      "eager_wall_median_ms": 0.14909333337224476,
-      "eager_wall_trials_ms": [
-        0.15717333335487638,
-        0.12570666664638944,
-        0.14909333337224476
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.1880079984664917,
-      "graph_min_ms": 0.18690240383148193,
-      "graph_max_ms": 0.2143280029296875,
-      "eager_wall_median_ms": 0.3311766666835562,
-      "eager_wall_trials_ms": [
-        0.3571433333490859,
-        0.3311766666835562,
-        0.32030999997611315
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0934448003768921,
-      "graph_min_ms": 0.09198240041732789,
-      "graph_max_ms": 0.10836479663848878,
-      "eager_wall_median_ms": 0.132580000051045,
-      "eager_wall_trials_ms": [
-        0.132580000051045,
-        0.14571666670235572,
-        0.12959000002107737
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.18786399960517886,
-      "graph_min_ms": 0.18675040006637572,
-      "graph_max_ms": 0.2141263961791992,
-      "eager_wall_median_ms": 0.34321999998307245,
-      "eager_wall_trials_ms": [
-        0.36083333331286366,
-        0.3284899999925983,
-        0.34321999998307245
-      ],
-      "extra_peak_bytes": 74809344
-    }
-  ]
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.json
deleted file mode 100644
index 468ce07af..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.json
+++ /dev/null
@@ -1,37 +0,0 @@
-{
-  "package": "D:\\dandan123\\dandan\\issue\\investigation-rl-kernel\\RL-Kernel\\.venv\\Lib\\site-packages\\rl_engine",
-  "companion_library": "D:\\dandan123\\dandan\\issue\\investigation-rl-kernel\\RL-Kernel\\.venv\\Lib\\site-packages\\rl_engine\\_rope_cuda.cp313-win_amd64.pyd",
-  "native_dispatcher_registered": true,
-  "torch": "2.9.1+cu128",
-  "gpu": "NVIDIA GeForce RTX 5080",
-  "license_present": true,
-  "packed_qkv_stride": [
-    17280,
-    1152,
-    128,
-    1
-  ],
-  "checks": {
-    "cuda_eager": "passed",
-    "cuda_compiled": "passed",
-    "cuda_trace": {
-      "contract": "qwenimage-interleaved-rope-fp32-v1",
-      "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
-      "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
-      "layout": "BSHD, adjacent channel pairs",
-      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-    },
-    "triton_eager": "passed",
-    "triton_compiled": "passed",
-    "triton_trace": {
-      "contract": "qwenimage-interleaved-rope-fp32-v1",
-      "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
-      "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
-      "layout": "BSHD, adjacent channel pairs",
-      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-    }
-  },
-  "passed": true
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.txt
deleted file mode 100644
index 01c1dd377..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/installed-wheel.txt
+++ /dev/null
@@ -1,38 +0,0 @@
-WARNING 10-28 17:08:06 [RL-Kernel]: Core binary extension (_C) unavailable: cannot import name '_C' from 'rl_engine' (D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\.venv\Lib\site-packages\rl_engine\__init__.py). Falling back to native code.
-{
-  "package": "D:\\dandan123\\dandan\\issue\\investigation-rl-kernel\\RL-Kernel\\.venv\\Lib\\site-packages\\rl_engine",
-  "companion_library": "D:\\dandan123\\dandan\\issue\\investigation-rl-kernel\\RL-Kernel\\.venv\\Lib\\site-packages\\rl_engine\\_rope_cuda.cp313-win_amd64.pyd",
-  "native_dispatcher_registered": true,
-  "torch": "2.9.1+cu128",
-  "gpu": "NVIDIA GeForce RTX 5080",
-  "license_present": true,
-  "packed_qkv_stride": [
-    17280,
-    1152,
-    128,
-    1
-  ],
-  "checks": {
-    "cuda_eager": "passed",
-    "cuda_compiled": "passed",
-    "cuda_trace": {
-      "contract": "qwenimage-interleaved-rope-fp32-v1",
-      "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
-      "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
-      "layout": "BSHD, adjacent channel pairs",
-      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-    },
-    "triton_eager": "passed",
-    "triton_compiled": "passed",
-    "triton_trace": {
-      "contract": "qwenimage-interleaved-rope-fp32-v1",
-      "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
-      "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
-      "layout": "BSHD, adjacent channel pairs",
-      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-    }
-  },
-  "passed": true
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/layout.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/layout.txt
deleted file mode 100644
index 0086d019b..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/layout.txt
+++ /dev/null
@@ -1,138 +0,0 @@
-.FF.............s............F.FF....                                    [100%]
-================================== FAILURES ===================================
-________________ test_cuda_build_keeps_existing_ipc_collective ________________
-
-monkeypatch = <_pytest.monkeypatch.MonkeyPatch object at 0x000001D4B0E668B0>
-
-    def test_cuda_build_keeps_existing_ipc_collective(monkeypatch) -> None:
-        extension = _load_extension_config(monkeypatch, hip=None)
-
-        assert "csrc/cuda/collectives/deterministic_collective.cu" in extension["sources"]
-        assert "-DKERNEL_ALIGN_WITH_CUDA" in extension["extra_compile_args"]["cxx"]
-        assert "-DKERNEL_ALIGN_WITH_ROCM" not in extension["extra_compile_args"]["cxx"]
->       assert "-lcuda" in extension["extra_link_args"]
-E       AssertionError: assert '-lcuda' in ['-Wl,-rpath,$ORIGIN/../torch/lib']
-
-tests\build\test_build_platform_collectives.py:61: AssertionError
-____________ test_package_data_json_links_resolve_to_real_content _____________
-
-tmp_path = WindowsPath('C:/Users/123/AppData/Local/Temp/pytest-of-123/pytest-3/test_package_data_json_links_r0')
-
-    def test_package_data_json_links_resolve_to_real_content(tmp_path) -> None:
-        target = tmp_path / "configs" / "profile.json"
-        target.parent.mkdir()
-        target.write_text('{"name": "profile"}\n', encoding="utf-8")
-        source = tmp_path / "rl_engine" / "profile.json"
-        source.parent.mkdir()
->       source.symlink_to("../configs/profile.json")
-
-tests\build\test_build_platform_collectives.py:70:
-_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
-
-self = WindowsPath('C:/Users/123/AppData/Local/Temp/pytest-of-123/pytest-3/test_package_data_json_links_r0/rl_engine/profile.json')
-target = '../configs/profile.json', target_is_directory = False
-
-    def symlink_to(self, target, target_is_directory=False):
-        """
-        Make this path a symlink pointing to the target path.
-        Note the order of arguments (link, target) is the reverse of os.symlink.
-        """
->       os.symlink(target, self, target_is_directory)
-E       OSError: [WinError 1314] �ͻ���û���������Ȩ��: '../configs/profile.json' -> 'C:\\Users\\123\\AppData\\Local\\Temp\\pytest-of-123\\pytest-3\\test_package_data_json_links_r0\\rl_engine\\profile.json'
-
-D:\anaconda\Lib\pathlib\_local.py:789: OSError
-_______ test_recorded_algorithm_sources_resolve_to_real_implementations _______
-
-    def test_recorded_algorithm_sources_resolve_to_real_implementations():
-        from rl_engine.config.workload import load_manifest
-
-        moves = json.loads((ROOT / "tools/migration/layout.json").read_text())["moves"]
-        for case in load_manifest().representative_cases:
-            source, symbol = case["provenance_evidence"]["algorithm_source"].rsplit(":", 1)
-            implementation = ROOT / moves.get(source, source)
->           text = implementation.read_text()
-                   ^^^^^^^^^^^^^^^^^^^^^^^^^^
-
-tests\runtime\test_layout_compatibility.py:87:
-_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
-D:\anaconda\Lib\pathlib\_local.py:546: in read_text
-    return PathBase.read_text(self, encoding, errors, newline)
-           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
-_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
-
-self = WindowsPath('D:/dandan123/dandan/issue/investigation-rl-kernel/RL-Kernel/csrc/cuda/attention/deterministic_attention.cu')
-encoding = 'locale', errors = None, newline = None
-
-    def read_text(self, encoding=None, errors=None, newline=None):
-        """
-        Open the file in text mode, read it, and close the file.
-        """
-        with self.open(mode='r', encoding=encoding, errors=errors, newline=newline) as f:
->           return f.read()
-                   ^^^^^^^^
-E           UnicodeDecodeError: 'gbk' codec can't decode byte 0x92 in position 173: illegal multibyte sequence
-
-D:\anaconda\Lib\pathlib\_abc.py:633: UnicodeDecodeError
-_ test_packaged_contract_bytes_match_main[rl_engine.config-workloads/qwen3_8b.json-80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6] _
-
-package = 'rl_engine.config', resource = 'workloads/qwen3_8b.json'
-digest = '80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6'
-
-    @pytest.mark.parametrize(
-        ("package", "resource", "digest"),
-        [
-            (
-                "rl_engine.config",
-                "workloads/qwen3_8b.json",
-                "80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6",
-            ),
-            (
-                "rl_engine.contracts",
-                "profiles/precision/ws1.json",
-                "044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532",
-            ),
-        ],
-    )
-    def test_packaged_contract_bytes_match_main(package, resource, digest):
->       assert hashlib.sha256(files(package).joinpath(resource).read_bytes()).hexdigest() == digest
-E       AssertionError: assert 'c97e966fb075...75e111e052d7d' == '80395644b5e8...f19635d3df0f6'
-E
-E         - 80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6
-E         + c97e966fb075dad8ce7dcd7a23a58ffb856e7474dfb3c50510675e111e052d7d
-
-tests\runtime\test_layout_compatibility.py:118: AssertionError
-_ test_packaged_contract_bytes_match_main[rl_engine.contracts-profiles/precision/ws1.json-044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532] _
-
-package = 'rl_engine.contracts', resource = 'profiles/precision/ws1.json'
-digest = '044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532'
-
-    @pytest.mark.parametrize(
-        ("package", "resource", "digest"),
-        [
-            (
-                "rl_engine.config",
-                "workloads/qwen3_8b.json",
-                "80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6",
-            ),
-            (
-                "rl_engine.contracts",
-                "profiles/precision/ws1.json",
-                "044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532",
-            ),
-        ],
-    )
-    def test_packaged_contract_bytes_match_main(package, resource, digest):
->       assert hashlib.sha256(files(package).joinpath(resource).read_bytes()).hexdigest() == digest
-E       AssertionError: assert '2020c02e7af4...313d6752cd34e' == '044167ec21e3...7628cbd101532'
-E
-E         - 044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532
-E         + 2020c02e7af4e2ca4163e1ecfb9c491baef07b73deba39b0f44313d6752cd34e
-
-tests\runtime\test_layout_compatibility.py:118: AssertionError
-=========================== short test summary info ===========================
-FAILED tests/build/test_build_platform_collectives.py::test_cuda_build_keeps_existing_ipc_collective
-FAILED tests/build/test_build_platform_collectives.py::test_package_data_json_links_resolve_to_real_content
-FAILED tests/runtime/test_layout_compatibility.py::test_recorded_algorithm_sources_resolve_to_real_implementations
-FAILED tests/runtime/test_layout_compatibility.py::test_packaged_contract_bytes_match_main[rl_engine.config-workloads/qwen3_8b.json-80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6]
-FAILED tests/runtime/test_layout_compatibility.py::test_packaged_contract_bytes_match_main[rl_engine.contracts-profiles/precision/ws1.json-044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532]
-5 failed, 31 passed, 1 skipped in 5.67s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/manifest.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/manifest.json
deleted file mode 100644
index 190cf7834..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/manifest.json
+++ /dev/null
@@ -1,65 +0,0 @@
-{
-  "recorded_at_utc": "2026-10-28T09:12:00.361759+00:00",
-  "base_commit": "16b81175874ecf11d5ff05c00de3b69e9457591b",
-  "source_sha256": {
-    "MANIFEST.in": "706ec113547f7012cfa6d15d062817d5d2d9e0c0e18e0fb706eca5ba2335bcc5",
-    "benchmarks/operators/rope/benchmark_multi_axis_rope.py": "6961a55803fbd4e566bd060558fb59e658e7f1f704e61ca97f912e5d16b55884",
-    "build_tools/extensions.py": "cf07cd49cba0ba7b4da459c26546380b7b288232fcc032e7641863e09ae3014a",
-    "csrc/bindings/rope_module.cpp": "c69eb3f200de306ca5885fd7c4c7ccf49c822ed1f2f1f639223ec36f0c51b524",
-    "csrc/cuda/rope/multi_axis_rope.cu": "f5db5cf5d4962053194455d9cde8cf91cf4e824bad7abf72e8410a404bcb0e44",
-    "docs/.nav.yml": "814a6d60839465e8f9405a9a26306dbc203b0fcb6aa36db264fd7651624d7ef3",
-    "docs/operators/README.md": "ab926bf5a3fb719e4a57cb65f0784a52019f6ed4bd49c0b5f1b15a1765fd6d7a",
-    "docs/operators/multi-axis-rope.md": "79f94810ee15ead5f1ce24715456f65f12b383728f2a3571104614b1a33b47d8",
-    "pyproject.toml": "a132159ae2a2c077d9fe001194fa2b9ab382acbcd58a9eccdea9387286bfb258",
-    "rl_engine/backends/cuda/rope/multi_axis_rope.py": "d779d9b60ce5d1d6651925f723049930d898024fdf0548ce29a2019b84753d65",
-    "rl_engine/backends/shared/triton/rope/FLASH_ATTENTION_LICENSE": "8c9ccb96c065e706135b6cbad279b721da6156e51f3a5f27c6b3329af9416d73",
-    "rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py": "5c6984da351e5228f7549629d4ff53e5772b9a6ef2a01ba2a38493e769b15ab5",
-    "rl_engine/backends/shared/triton/rope/multi_axis_rope.py": "5b989bd63efd77913267b48654d10204bc563768c6c5d4e1cde9fb55e1354547",
-    "rl_engine/ops/rope/_registration.py": "add0375d1639444908091d01e8d29781f7a975beffec64b77bfaa81a5b9a0ae3",
-    "rl_engine/ops/rope/multi_axis_rope.py": "07bcff1160f22c1ff6730be731206b2f4a153e2fada26afc93fc884e780568cb",
-    "rl_engine/reference/rope/multi_axis_rope.py": "c9a94730047f51659f7aa0a443e88d6ce7d03d00d6813a9aed548cafbf6347bf",
-    "rl_engine/runtime/registry.py": "9f9c69bb6c2e941b057a66ac2d388be021cbb4b55cb3c602976844aabc6106e1",
-    "rl_engine/validation/operators/gradient_adapters.py": "5ee2e9ad619a4cad6e81d8f4a6ccf81ecf45e63a7764d0fcb4c204206b37c16b",
-    "rl_engine/validation/operators/operator_inputs.py": "55c8d9e7cbde2070662dd1d0ebe9888aef106b45fb05902057aec032d29a9836",
-    "rl_engine/validation/operators/operator_specs.py": "cef2e5bed8075509d5d5a538b378477b09f1adfe5fd01829156357cefa99f983",
-    "rl_engine/validation/operators/qwen_image_rope.py": "c3ddfc732f32de3c633de16064810d67202254ea78fd8107d558b471bb15a37f",
-    "tests/build/test_multi_axis_rope_build.py": "90c3740b2d4946acc6206d952622c19afaa0089cca904919088563427555b956",
-    "tests/ops/rope/test_multi_axis_rope.py": "1a0950b09eec407e8ffd490b40dda4118f8d9b876f4db4af8aac648af1a76687",
-    "tests/validation/operators/test_operator_inputs.py": "36365f662194f158276b3c07f6447471a0cf17253ec420ab63c8b9149c511b49",
-    "tools/validation/operators/check_multi_axis_rope_install.py": "ff8198ff55779ea2ede88d4ecfe07caa627631b4b08bce38139ccfa147abf538",
-    "tools/validation/operators/check_qwenimage_rope_upstream.py": "cc368d5bebf0f0acb21b7455348592d30442b6a8d349de0f49d5a09c5f20ce4e",
-    "tools/validation/operators/validate_multi_axis_rope.py": "def4747da9ea9194c761bef24479017ac74ac10de46a3fbe1a9528ec6ee07731"
-  },
-  "evidence_sha256": {
-    "baseline-layout.txt": "b050d798a5e3a6c2d43c7c22e4e1140529ebf05ebab982df17f501f5b69cff12",
-    "benchmark-bfloat16.json": "0679a422e8129a8a36ada938b024794334f784cdea28013574f99c70673e13a0",
-    "installed-wheel.json": "60f6a1a48d35fa25a639a1ff168ce0c37659e0258737422cafda3741d8ef0200",
-    "installed-wheel.txt": "2be6479a8e30938410095dbcc7724ad114cb61cb35ef86c7f86b8d507a3193f6",
-    "layout.txt": "83169d7995bdadc14744a0ed0fff69b334f2ed6ff0b92cda005a9cf0f733116d",
-    "memcheck.txt": "f6f05d4e5b772909f29baf2a1d6be04b4456660ca024e5067c3f64c7efa4b81f",
-    "operator-cuda.json": "1a231d26f2411a38cffe70fdb59b388227fc0e7da9625370b165d7d590e1e011",
-    "operator-triton.json": "af2be1f4df1bf9d7afb4802ca02c81826e64acacae891436f83ad6b3342109f5",
-    "package-build.txt": "34b11d626dc4988f3d862e4122599a0b1a949b1d799436981dd74f3ee5c1b07e",
-    "package.json": "e267d82707cdeb0c53d120928cd32a544a44190e8e41c43f3b6c1d4668275a7e",
-    "precommit.txt": "04f9db16780e0ad0b41e6450d362e6f0a7bdb07128596c8ea13f8ea8bb2cc674",
-    "quality.txt": "d4728355fe488102ca644f510795c474b54f8d42bb92ddfda34557ba521c6afd",
-    "README.md": "aa2cabc1e8423f4d737d0db90786142c4a919041dbf97335b17991475a4cbaca",
-    "regression.txt": "5d058ac93516d9ce1f486fbd4f35b84a254a32d8ddf26eedf41cf2388153e004",
-    "tests.txt": "f6af2230f6a0b5d993cb6bca2724a82f772c6b33df456edce6e728996f2a76d5",
-    "upstream.json": "8c087d60013f0266303c6d60237d87e5f1ff67b03e8e8c1900de4b20349e2e93",
-    "ws1-cuda.json": "531124223ae5fe43eca890fb1ccf80e09f33affe34bab132c7c17742081d85a3",
-    "ws1-cuda.txt": "b4cf09ef5165e535dba9ed19eeec08f15eccaf86b549bb1a253fd08e87be35ab",
-    "ws1-triton.json": "29eedee8e50b8750d0ce9be95deb39fc6aa9ab47b0867fe20ce2e7ad5cb47548",
-    "ws1-triton.txt": "a1daefba35f7c1a06d222750ad1ce9c4bdc97cc03410969e7b4f4ce011a6770a"
-  },
-  "layout_failures_equal_clean_base": true,
-  "layout_failure_ids": [
-    "tests/build/test_build_platform_collectives.py::test_cuda_build_keeps_existing_ipc_collective",
-    "tests/build/test_build_platform_collectives.py::test_package_data_json_links_resolve_to_real_content",
-    "tests/runtime/test_layout_compatibility.py::test_recorded_algorithm_sources_resolve_to_real_implementations",
-    "tests/runtime/test_layout_compatibility.py::test_packaged_contract_bytes_match_main[rl_engine.config-workloads/qwen3_8b.json-80395644b5e84ed1a76729ffc6f78eb11f41b9ab42fb36e2d00f19635d3df0f6]",
-    "tests/runtime/test_layout_compatibility.py::test_packaged_contract_bytes_match_main[rl_engine.contracts-profiles/precision/ws1.json-044167ec21e372017a2758dc09f805b368283c06b83f64061877628cbd101532]"
-  ],
-  "package_source_and_license_checks": "passed",
-  "hash_mode": "SHA256 of recorded working-tree bytes; upstream baselines separately normalize newlines"
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/memcheck.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/memcheck.txt
deleted file mode 100644
index dd59e40b0..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/memcheck.txt
+++ /dev/null
@@ -1,8 +0,0 @@
-========= COMPUTE-SANITIZER
-WARNING 10-28 16:46:09 [RL-Kernel]: Core binary extension (_C) unavailable: cannot import name '_C' from 'rl_engine' (D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\rl_engine\__init__.py). Falling back to native code.
-W1028 16:46:10.075000 16608 site-packages\torch\utils\cpp_extension.py:480] Error checking compiler version for cl: 'cp1' codec can't decode bytes in position 0--1: No mapping for the Unicode character exists in the target code page.
-ninja: no work to do.
-Focused CUDA source build; torch=2.9.1+cu128, GPU=NVIDIA GeForce RTX 5080
-..................................                                       [100%]
-34 passed, 48 deselected in 3.96s
-========= ERROR SUMMARY: 0 errors
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-cuda.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-cuda.json
deleted file mode 100644
index 03d59bb4b..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-cuda.json
+++ /dev/null
@@ -1,70 +0,0 @@
-{
-  "suite_name": "multi_axis_rope",
-  "total_candidates": 1,
-  "passed_candidates": 1,
-  "pass_rate": 1.0,
-  "passed": true,
-  "candidates": [
-    {
-      "candidate_name": "cuda-multi_axis_rope",
-      "backend": "cuda",
-      "total_outputs": 2,
-      "passed_outputs": 2,
-      "pass_rate": 1.0,
-      "passed": true,
-      "cases": [
-        {
-          "case_name": "multi_axis_rope-torch.bfloat16-2x7x3x128",
-          "dtype": "torch.bfloat16",
-          "op_class": "elementwise",
-          "passed": true,
-          "outputs": [
-            {
-              "output_index": 0,
-              "shape": [
-                2,
-                7,
-                3,
-                128
-              ],
-              "candidate_dtype": "torch.bfloat16",
-              "gold_dtype": "torch.float32",
-              "atol": 0.02,
-              "rtol": 0.016,
-              "max_abs_error": 0.007704973220825195,
-              "mean_abs_error": 0.0006018931162543595,
-              "max_rel_error": 0.003828622167930007,
-              "passed": true,
-              "judgment": "forward_accuracy",
-              "comparison_lhs_role": "bf16_candidate",
-              "comparison_rhs_role": "fp32_reference",
-              "message": ""
-            },
-            {
-              "output_index": 1,
-              "shape": [
-                2,
-                7,
-                3,
-                128
-              ],
-              "candidate_dtype": "torch.bfloat16",
-              "gold_dtype": "torch.bfloat16",
-              "atol": 0.02,
-              "rtol": 0.016,
-              "max_abs_error": 0.0,
-              "mean_abs_error": 0.0,
-              "max_rel_error": 0.0,
-              "passed": true,
-              "judgment": "gradient_accuracy",
-              "comparison_lhs_role": "bf16_candidate",
-              "comparison_rhs_role": "fp32_reference",
-              "message": "gradient:x"
-            }
-          ]
-        }
-      ],
-      "backend_provenance": null
-    }
-  ]
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-triton.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-triton.json
deleted file mode 100644
index 6f086edac..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/operator-triton.json
+++ /dev/null
@@ -1,70 +0,0 @@
-{
-  "suite_name": "multi_axis_rope",
-  "total_candidates": 1,
-  "passed_candidates": 1,
-  "pass_rate": 1.0,
-  "passed": true,
-  "candidates": [
-    {
-      "candidate_name": "triton-multi_axis_rope",
-      "backend": "triton",
-      "total_outputs": 2,
-      "passed_outputs": 2,
-      "pass_rate": 1.0,
-      "passed": true,
-      "cases": [
-        {
-          "case_name": "multi_axis_rope-torch.bfloat16-2x7x3x128",
-          "dtype": "torch.bfloat16",
-          "op_class": "elementwise",
-          "passed": true,
-          "outputs": [
-            {
-              "output_index": 0,
-              "shape": [
-                2,
-                7,
-                3,
-                128
-              ],
-              "candidate_dtype": "torch.bfloat16",
-              "gold_dtype": "torch.float32",
-              "atol": 0.02,
-              "rtol": 0.016,
-              "max_abs_error": 0.007704973220825195,
-              "mean_abs_error": 0.0006018931162543595,
-              "max_rel_error": 0.003828622167930007,
-              "passed": true,
-              "judgment": "forward_accuracy",
-              "comparison_lhs_role": "bf16_candidate",
-              "comparison_rhs_role": "fp32_reference",
-              "message": ""
-            },
-            {
-              "output_index": 1,
-              "shape": [
-                2,
-                7,
-                3,
-                128
-              ],
-              "candidate_dtype": "torch.bfloat16",
-              "gold_dtype": "torch.bfloat16",
-              "atol": 0.02,
-              "rtol": 0.016,
-              "max_abs_error": 0.0,
-              "mean_abs_error": 0.0,
-              "max_rel_error": 0.0,
-              "passed": true,
-              "judgment": "gradient_accuracy",
-              "comparison_lhs_role": "bf16_candidate",
-              "comparison_rhs_role": "fp32_reference",
-              "message": "gradient:x"
-            }
-          ]
-        }
-      ],
-      "backend_provenance": null
-    }
-  ]
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package-build.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package-build.txt
deleted file mode 100644
index d63d79d52..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package-build.txt
+++ /dev/null
@@ -1,40 +0,0 @@
-Selected lines from the completed sdist/wheel build log.
-Full local log SHA256: bfbcd342efaf8caa36adcd208acee385db1ba6fbc4326fe8e0325f9523f199b0
-running sdist
-copying rl_engine\backends\shared\triton\rope\FLASH_ATTENTION_LICENSE -> rl_kernel-0.1.0\rl_engine\backends\shared\triton\rope
-running bdist_wheel
-copying rl_engine\backends\shared\triton\rope\FLASH_ATTENTION_LICENSE -> build\lib.win-amd64-cpython-313\rl_engine\backends\shared\triton\rope
-copying rl_engine\backends\shared\triton\rope\FLASH_ATTENTION_LICENSE -> build\lib.win-amd64-cpython-313\rl_engine\backends\shared\triton\rope
-running build_ext
-ninja: no work to do.
-"C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.44.35207\bin\HostX64\x64\link.exe" /nologo /INCREMENTAL:NO /LTCG /DLL /MANIFEST:EMBED,ID=2 /MANIFESTUAC:NO /LIBPATH:D:\dandan123\dandan\myvllm\.cuda-12.8\Library\lib /LIBPATH:D:\anaconda\Lib\site-packages\torch\lib /LIBPATH:D:\dandan123\dandan\myvllm\.cuda-12.8\Library\lib\x64 /LIBPATH:D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\.venv\libs /LIBPATH:D:\anaconda\libs /LIBPATH:D:\anaconda /LIBPATH:D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\.venv\PCbuild\amd64 "/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.44.35207\lib\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\10\lib\10.0.26100.0\ucrt\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\10\\lib\10.0.26100.0\\um\x64" c10.lib torch.lib torch_cpu.lib torch_python.lib cudart.lib c10_cuda.lib torch_cuda.lib /EXPORT:PyInit__rope_cuda D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\bindings\rope_module.obj D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\cuda\rope\multi_axis_rope.obj /OUT:build\lib.win-amd64-cpython-313\rl_engine\_rope_cuda.cp313-win_amd64.pyd /IMPLIB:D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\bindings\_rope_cuda.cp313-win_amd64.lib
-copying build\lib.win-amd64-cpython-313\rl_engine\backends\shared\triton\rope\FLASH_ATTENTION_LICENSE -> build\bdist.win-amd64\wheel\.\rl_engine\backends\shared\triton\rope
-copying build\lib.win-amd64-cpython-313\rl_engine\_rope_cuda.cp313-win_amd64.pyd -> build\bdist.win-amd64\wheel\.\rl_engine
-adding 'rl_engine/_rope_cuda.cp313-win_amd64.pyd'
-adding 'rl_engine/backends/shared/triton/rope/FLASH_ATTENTION_LICENSE'
-
-In-place build log:
-D:\anaconda\Lib\site-packages\setuptools\config\_apply_pyprojecttoml.py:82: SetuptoolsDeprecationWarning: `project.license` as a TOML table is deprecated
-!!
-
-        ********************************************************************************
-        Please use a simple string containing a SPDX expression for `project.license`. You can also use `project.license-files`. (Both options available on setuptools>=77.0.0).
-
-        By 2027-Feb-18, you need to update your project and remove deprecated calls
-        or your builds will no longer be supported.
-
-        See https://packaging.python.org/en/latest/guides/writing-pyproject-toml/#license for details.
-        ********************************************************************************
-
-!!
-  corresp(dist, value, root_dir)
-running build_ext
-W1028 16:52:54.107000 9384 site-packages\torch\utils\cpp_extension.py:480] Error checking compiler version for cl: 'cp1' codec can't decode bytes in position 0--1: No mapping for the Unicode character exists in the target code page.
-building 'rl_engine._rope_cuda' extension
-W1028 16:52:54.150000 9384 site-packages\torch\utils\cpp_extension.py:480] Error checking compiler version for cl: 'cp1' codec can't decode bytes in position 0--1: No mapping for the Unicode character exists in the target code page.
-ninja: no work to do.
-"C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.44.35207\bin\HostX64\x64\link.exe" /nologo /INCREMENTAL:NO /LTCG /DLL /MANIFEST:EMBED,ID=2 /MANIFESTUAC:NO /LIBPATH:D:\dandan123\dandan\myvllm\.cuda-12.8\Library\lib /LIBPATH:D:\anaconda\Lib\site-packages\torch\lib /LIBPATH:D:\dandan123\dandan\myvllm\.cuda-12.8\Library\lib\x64 /LIBPATH:D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\.venv\libs /LIBPATH:D:\anaconda\libs /LIBPATH:D:\anaconda /LIBPATH:D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\.venv\PCbuild\amd64 "/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.44.35207\lib\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\10\lib\10.0.26100.0\ucrt\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\10\\lib\10.0.26100.0\\um\x64" c10.lib torch.lib torch_cpu.lib torch_python.lib cudart.lib c10_cuda.lib torch_cuda.lib /EXPORT:PyInit__rope_cuda D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\bindings\rope_module.obj D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\cuda\rope\multi_axis_rope.obj /OUT:build\lib.win-amd64-cpython-313\rl_engine\_rope_cuda.cp313-win_amd64.pyd /IMPLIB:D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\bindings\_rope_cuda.cp313-win_amd64.lib
-  正在创建库 D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\bindings\_rope_cuda.cp313-win_amd64.lib 和对象 D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\build\temp.win-amd64-cpython-313\Release\csrc\bindings\_rope_cuda.cp313-win_amd64.exp
-正在生成代码
-已完成代码的生成
-copying build\lib.win-amd64-cpython-313\rl_engine\_rope_cuda.cp313-win_amd64.pyd -> rl_engine
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package.json
deleted file mode 100644
index 51e102428..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/package.json
+++ /dev/null
@@ -1,31 +0,0 @@
-{
-  "wheel": "rl_kernel-0.1.0-cp313-cp313-win_amd64.whl",
-  "wheel_sha256": "bc5418bbc4be779ec6d5b4fa99d54b3244f2570d60a29adc3e6a1c38be25a864",
-  "sdist": "rl_kernel-0.1.0.tar.gz",
-  "sdist_sha256": "56026574fa3307e85e6e278298a09deee13b43d4ec5c61899698ac5e55163a12",
-  "native_library_member": "rl_engine/_rope_cuda.cp313-win_amd64.pyd",
-  "native_library_sha256": "c0878eeac05b6b400ed7c3935561de525ce3af8b2a050b5d8bfe5cdd66650940",
-  "inplace_library_sha256": "aab5bf0777da43609e067e6a41bc37c46fdf6b1dd0c60d1071c15bbc4f707242",
-  "binary_hash_scope": "Wheel and in-place libraries were linked separately; record both hashes",
-  "license_in_both_archives": true,
-  "license_sha256": "8c9ccb96c065e706135b6cbad279b721da6156e51f3a5f27c6b3329af9416d73",
-  "wheel_sources_match_checkout": [
-    "rl_engine/backends/cuda/rope/multi_axis_rope.py",
-    "rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py",
-    "rl_engine/backends/shared/triton/rope/multi_axis_rope.py",
-    "rl_engine/ops/rope/_registration.py",
-    "rl_engine/ops/rope/multi_axis_rope.py",
-    "rl_engine/reference/rope/multi_axis_rope.py",
-    "rl_engine/runtime/registry.py",
-    "rl_engine/validation/operators/gradient_adapters.py",
-    "rl_engine/validation/operators/operator_inputs.py",
-    "rl_engine/validation/operators/operator_specs.py",
-    "rl_engine/validation/operators/qwen_image_rope.py"
-  ],
-  "sdist_required_sources_match_checkout": [
-    "csrc/bindings/rope_module.cpp",
-    "csrc/cuda/rope/multi_axis_rope.cu",
-    "rl_engine/backends/shared/triton/rope/FLASH_ATTENTION_LICENSE"
-  ],
-  "profile": "RL_KERNEL_ROPE_ONLY=1; full monolithic _C build not qualified"
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/precommit.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/precommit.txt
deleted file mode 100644
index 475a8c0d8..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/precommit.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-trim trailing whitespace.................................................Passed
-fix end of files.........................................................Passed
-check yaml...............................................................Passed
-check for added large files..............................................Passed
-black....................................................................Passed
-isort....................................................................Passed
-flake8...................................................................Passed
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/quality.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/quality.txt
deleted file mode 100644
index 2756f4004..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/quality.txt
+++ /dev/null
@@ -1,58 +0,0 @@
-Changed-file checks; repository-pinned Black 24.4.2, isort 5.13.2 and Flake8 7.0.0.
-Test commands are separate from the static checks below.
-Regression command: python -m pytest tests/build/test_multi_axis_rope_build.py tests/validation/operators/test_operator_inputs.py tests/validation/operators/test_gradient_invariance.py tests/validation/operators/test_forward_invariance.py tests/validation/operators/test_elementwise_inventory.py tests/validation/operators/test_four_judgment_matrix.py tests/validation/operators/test_ws1_workload.py tests/runtime/test_kernel_registry.py tests/runtime/test_dispatch.py tests/ops/rope/test_rope.py -q -rs
-Layout output copies have only trailing whitespace normalized by pre-commit.
-
-Command: python -m isort --check-only --profile black --line-length 100 benchmarks/operators/rope/benchmark_multi_axis_rope.py build_tools/extensions.py rl_engine/backends/cuda/rope/multi_axis_rope.py rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py rl_engine/backends/shared/triton/rope/multi_axis_rope.py rl_engine/ops/rope/_registration.py rl_engine/ops/rope/multi_axis_rope.py rl_engine/reference/rope/multi_axis_rope.py rl_engine/runtime/registry.py rl_engine/validation/operators/gradient_adapters.py rl_engine/validation/operators/operator_inputs.py rl_engine/validation/operators/operator_specs.py rl_engine/validation/operators/qwen_image_rope.py tests/build/test_multi_axis_rope_build.py tests/ops/rope/test_multi_axis_rope.py tests/validation/operators/test_operator_inputs.py tools/validation/operators/check_multi_axis_rope_install.py tools/validation/operators/check_qwenimage_rope_upstream.py tools/validation/operators/validate_multi_axis_rope.py
-Exit code: 0
-
-
-Command: python -m black --check --line-length 100 benchmarks/operators/rope/benchmark_multi_axis_rope.py build_tools/extensions.py rl_engine/backends/cuda/rope/multi_axis_rope.py rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py rl_engine/backends/shared/triton/rope/multi_axis_rope.py rl_engine/ops/rope/_registration.py rl_engine/ops/rope/multi_axis_rope.py rl_engine/reference/rope/multi_axis_rope.py rl_engine/runtime/registry.py rl_engine/validation/operators/gradient_adapters.py rl_engine/validation/operators/operator_inputs.py rl_engine/validation/operators/operator_specs.py rl_engine/validation/operators/qwen_image_rope.py tests/build/test_multi_axis_rope_build.py tests/ops/rope/test_multi_axis_rope.py tests/validation/operators/test_operator_inputs.py tools/validation/operators/check_multi_axis_rope_install.py tools/validation/operators/check_qwenimage_rope_upstream.py tools/validation/operators/validate_multi_axis_rope.py
-Exit code: 0
-All done! ✨ 🍰 ✨
-
-19 files would be left unchanged.
-
-Command: python -m flake8 --max-line-length=100 --extend-ignore=E203,E704 benchmarks/operators/rope/benchmark_multi_axis_rope.py build_tools/extensions.py rl_engine/backends/cuda/rope/multi_axis_rope.py rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py rl_engine/backends/shared/triton/rope/multi_axis_rope.py rl_engine/ops/rope/_registration.py rl_engine/ops/rope/multi_axis_rope.py rl_engine/reference/rope/multi_axis_rope.py rl_engine/runtime/registry.py rl_engine/validation/operators/gradient_adapters.py rl_engine/validation/operators/operator_inputs.py rl_engine/validation/operators/operator_specs.py rl_engine/validation/operators/qwen_image_rope.py tests/build/test_multi_axis_rope_build.py tests/ops/rope/test_multi_axis_rope.py tests/validation/operators/test_operator_inputs.py tools/validation/operators/check_multi_axis_rope_install.py tools/validation/operators/check_qwenimage_rope_upstream.py tools/validation/operators/validate_multi_axis_rope.py
-Exit code: 0
-
-
-Command: python -m ruff check benchmarks/operators/rope/benchmark_multi_axis_rope.py build_tools/extensions.py rl_engine/backends/cuda/rope/multi_axis_rope.py rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py rl_engine/backends/shared/triton/rope/multi_axis_rope.py rl_engine/ops/rope/_registration.py rl_engine/ops/rope/multi_axis_rope.py rl_engine/reference/rope/multi_axis_rope.py rl_engine/runtime/registry.py rl_engine/validation/operators/gradient_adapters.py rl_engine/validation/operators/operator_inputs.py rl_engine/validation/operators/operator_specs.py rl_engine/validation/operators/qwen_image_rope.py tests/build/test_multi_axis_rope_build.py tests/ops/rope/test_multi_axis_rope.py tests/validation/operators/test_operator_inputs.py tools/validation/operators/check_multi_axis_rope_install.py tools/validation/operators/check_qwenimage_rope_upstream.py tools/validation/operators/validate_multi_axis_rope.py
-Exit code: 0
-All checks passed!
-
-Command: python -m mkdocs build --strict -f mkdocs.yaml
-Exit code: 0
-INFO    -  Cleaning site directory
-
-INFO    -  Building documentation to directory: D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\site
-
-mkdocs_git_revision_date_localized_plugin: 'D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\docs\operators\multi-axis-rope.md' has no git logs, using current timestamp
-
-INFO    -  The following pages exist in the docs directory, but are not included in the "nav" configuration:
-
-  - blog\2026-07-08-announcing-rl-kernel-linear-logp-for-vime-zh.md
-
-  - blog\2026-08-09-dsv4-flash-moe-consistency-roadmap.md
-
-  - operators\matmul.md
-
-  - operators\rope.md
-
-  - usage\cross-config-drift-report.md
-
-  - usage\cuda-rocm-consistency-audit.md
-
-  - usage\h100-configurable.md
-
-  - usage\h100-cp-gradient-validation.md
-
-  - usage\h100-matrix-validation.md
-
-  - usage\h100-pcp-validation.md
-
-  - usage\rocm-sparse-performance.md
-
-INFO    -  Documentation built in 4.01 seconds
-
-Repository pre-commit: passed; see precommit.txt.
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/regression.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/regression.txt
deleted file mode 100644
index 066e09cec..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/regression.txt
+++ /dev/null
@@ -1,10 +0,0 @@
-........................................................................ [ 35%]
-........................................................................ [ 71%]
-...................................................sssssss               [100%]
-=========================== short test summary info ===========================
-SKIPPED [3] tests\ops\rope\test_rope.py:390: rope_apply_ascend is not compiled (requires an Ascend NPU build).
-SKIPPED [1] tests\ops\rope\test_rope.py:406: rope_apply_ascend is not compiled (requires an Ascend NPU build).
-SKIPPED [1] tests\ops\rope\test_rope.py:413: rope_apply_ascend is not compiled (requires an Ascend NPU build).
-SKIPPED [1] tests\ops\rope\test_rope.py:420: rope_apply_ascend is not compiled (requires an Ascend NPU build).
-SKIPPED [1] tests\ops\rope\test_rope.py:433: rope_apply_ascend is not compiled (requires an Ascend NPU build).
-195 passed, 7 skipped in 4.31s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/tests.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/tests.txt
deleted file mode 100644
index 38e8e21e6..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/tests.txt
+++ /dev/null
@@ -1,5 +0,0 @@
-........................................................................ [ 87%]
-s.........                                                               [100%]
-=========================== short test summary info ===========================
-SKIPPED [1] tests\ops\rope\test_multi_axis_rope.py:336: requires two NVIDIA GPUs
-81 passed, 1 skipped in 5.55s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/upstream.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/upstream.json
deleted file mode 100644
index 1f89880c1..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/upstream.json
+++ /dev/null
@@ -1,335 +0,0 @@
-{
-  "source_sha256": "0bad11901deae69f05aec01fd464e51e38582d455068f2851ee665036179321f",
-  "cases": [
-    {
-      "grid": [
-        2,
-        3
-      ],
-      "stream": "image",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        2,
-        3
-      ],
-      "stream": "image",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        2,
-        3
-      ],
-      "stream": "image",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        2,
-        3
-      ],
-      "stream": "text",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        2,
-        3
-      ],
-      "stream": "text",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        2,
-        3
-      ],
-      "stream": "text",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        3,
-        5
-      ],
-      "stream": "image",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        3,
-        5
-      ],
-      "stream": "image",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        3,
-        5
-      ],
-      "stream": "image",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        3,
-        5
-      ],
-      "stream": "text",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        3,
-        5
-      ],
-      "stream": "text",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        3,
-        5
-      ],
-      "stream": "text",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        64,
-        64
-      ],
-      "stream": "image",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        64,
-        64
-      ],
-      "stream": "image",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        64,
-        64
-      ],
-      "stream": "image",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        64,
-        64
-      ],
-      "stream": "text",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        64,
-        64
-      ],
-      "stream": "text",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        64,
-        64
-      ],
-      "stream": "text",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        83,
-        83
-      ],
-      "stream": "image",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        83,
-        83
-      ],
-      "stream": "image",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        83,
-        83
-      ],
-      "stream": "image",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        83,
-        83
-      ],
-      "stream": "text",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        83,
-        83
-      ],
-      "stream": "text",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        83,
-        83
-      ],
-      "stream": "text",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        104,
-        58
-      ],
-      "stream": "image",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        104,
-        58
-      ],
-      "stream": "image",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        104,
-        58
-      ],
-      "stream": "image",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        104,
-        58
-      ],
-      "stream": "text",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        104,
-        58
-      ],
-      "stream": "text",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        104,
-        58
-      ],
-      "stream": "text",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    }
-  ]
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.json
deleted file mode 100644
index 20ad4dd78..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.json
+++ /dev/null
@@ -1,1962 +0,0 @@
-{
-  "scope": "Qwen-Image multi_axis_rope operator; C1 elementwise and C2 config transforms",
-  "trace": {
-    "contract": "qwenimage-interleaved-rope-fp32-v1",
-    "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
-    "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
-    "layout": "BSHD, adjacent channel pairs",
-    "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-    "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-  },
-  "passed": true,
-  "forward": {
-    "op_name": "multi_axis_rope",
-    "backend_profile": "cuda_bf16",
-    "accuracy_reports": [
-      {
-        "config_id": "BN/full",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/chunked",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/chunked",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004673302173614502,
-            "mean_abs_error": 6.917401333339512e-05,
-            "max_rel_error": 0.003753266530111432,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.00048516690731048584,
-            "mean_abs_error": 7.802647451171651e-05,
-            "max_rel_error": 0.0037793784867972136,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004864335060119629,
-            "mean_abs_error": 7.697804539930075e-05,
-            "max_rel_error": 0.003821910358965397,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.710371912457049e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004673302173614502,
-            "mean_abs_error": 6.917401333339512e-05,
-            "max_rel_error": 0.003753266530111432,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.00048516690731048584,
-            "mean_abs_error": 7.802647451171651e-05,
-            "max_rel_error": 0.0037793784867972136,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004864335060119629,
-            "mean_abs_error": 7.697804539930075e-05,
-            "max_rel_error": 0.003821910358965397,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.710371912457049e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/permuted",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/permuted",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_right",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/padded_right",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_left",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/padded_left",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      }
-    ],
-    "invariance_reports": [
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/chunked",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/chunked"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s0",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s1",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s2",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s3",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/permuted",
-        "transform_kind": "permutation",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/permuted"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_right",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_right"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_left",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_left"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      }
-    ],
-    "logprob_smoke": null,
-    "backend_provenance": {
-      "backend_profile": "cuda_bf16",
-      "requested_backend": "cuda",
-      "actual_backend": "cuda",
-      "execution_dtype": "bfloat16",
-      "accumulation_dtype": "float32",
-      "output_dtype": "bfloat16",
-      "reference_dtype": "float32",
-      "candidate_tf32_enabled": false,
-      "reference_tf32_enabled": false
-    },
-    "candidate_id": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
-    "device": "cuda",
-    "compute_capability": "sm120",
-    "seed": 20260812,
-    "fallback_reason": null,
-    "passed": true,
-    "provenance_valid": true,
-    "metadata_valid": true,
-    "observed_kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope"
-  },
-  "gradient": {
-    "op_name": "multi_axis_rope",
-    "backend_profile": "cuda_bf16",
-    "accuracy_reports": [
-      {
-        "config_id": "BN/full",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/chunked",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/chunked",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.7887988835573196e-05,
-            "mean_abs_error": 3.6821968478761846e-06,
-            "max_rel_error": 1.8164952993392944,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.7113780965446495e-06,
-            "max_rel_error": 0.4316565990447998,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.995925024151802e-05,
-            "mean_abs_error": 4.7168209675874095e-06,
-            "max_rel_error": 0.7439075112342834,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.0211638659238815e-05,
-            "mean_abs_error": 4.9605891945248e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.7887988835573196e-05,
-            "mean_abs_error": 3.6821968478761846e-06,
-            "max_rel_error": 1.8164952993392944,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.7113780965446495e-06,
-            "max_rel_error": 0.4316565990447998,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.995925024151802e-05,
-            "mean_abs_error": 4.7168209675874095e-06,
-            "max_rel_error": 0.7439075112342834,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.0211638659238815e-05,
-            "mean_abs_error": 4.9605891945248e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/permuted",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/permuted",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_right",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/padded_right",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_left",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/padded_left",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      }
-    ],
-    "invariance_reports": [
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/chunked",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/chunked"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s0",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s1",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s2",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s3",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/permuted",
-        "transform_kind": "permutation",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/permuted"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_right",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_right"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_left",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_left"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      }
-    ],
-    "singleton_aggregate_reports": [],
-    "backend_provenance": {
-      "backend_profile": "cuda_bf16",
-      "requested_backend": "cuda",
-      "actual_backend": "cuda",
-      "execution_dtype": "bfloat16",
-      "accumulation_dtype": "float32",
-      "output_dtype": "bfloat16",
-      "reference_dtype": "float32",
-      "candidate_tf32_enabled": false,
-      "reference_tf32_enabled": false
-    },
-    "candidate_id": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
-    "device": "cuda",
-    "compute_capability": "sm120",
-    "seed": 20260812,
-    "fallback_reason": null,
-    "passed": true,
-    "provenance_valid": true,
-    "metadata_valid": true,
-    "loss_reduction": "sum_over_active_tokens_then_optional_mean_by_active_count",
-    "active_token_denominator": 27,
-    "grad_tensor_names": [
-      "dx"
-    ],
-    "first_failing_op": null,
-    "first_failing_tensor": null,
-    "first_failing_config_pair": null,
-    "observed_kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope"
-  }
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.txt
deleted file mode 100644
index e69d60986..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-cuda.txt
+++ /dev/null
@@ -1,3 +0,0 @@
-WARNING 10-28 16:54:40 [RL-Kernel]: Core binary extension (_C) unavailable: cannot import name '_C' from 'rl_engine' (D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\rl_engine\__init__.py). Falling back to native code.
-Using the installed production RoPE dispatcher library
-WS1 four judgments backend=cuda: passed=True
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.json b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.json
deleted file mode 100644
index d81adfea1..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.json
+++ /dev/null
@@ -1,1962 +0,0 @@
-{
-  "scope": "Qwen-Image multi_axis_rope operator; C1 elementwise and C2 config transforms",
-  "trace": {
-    "contract": "qwenimage-interleaved-rope-fp32-v1",
-    "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
-    "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
-    "layout": "BSHD, adjacent channel pairs",
-    "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-    "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-  },
-  "passed": true,
-  "forward": {
-    "op_name": "multi_axis_rope",
-    "backend_profile": "triton_cuda_bf16",
-    "accuracy_reports": [
-      {
-        "config_id": "BN/full",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/chunked",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/chunked",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004673302173614502,
-            "mean_abs_error": 6.917401333339512e-05,
-            "max_rel_error": 0.003753266530111432,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.00048516690731048584,
-            "mean_abs_error": 7.802647451171651e-05,
-            "max_rel_error": 0.0037793784867972136,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004864335060119629,
-            "mean_abs_error": 7.697804539930075e-05,
-            "max_rel_error": 0.003821910358965397,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.710371912457049e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004673302173614502,
-            "mean_abs_error": 6.917401333339512e-05,
-            "max_rel_error": 0.003753266530111432,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.00048516690731048584,
-            "mean_abs_error": 7.802647451171651e-05,
-            "max_rel_error": 0.0037793784867972136,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004864335060119629,
-            "mean_abs_error": 7.697804539930075e-05,
-            "max_rel_error": 0.003821910358965397,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.710371912457049e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/permuted",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/permuted",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_right",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/padded_right",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_left",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/padded_left",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      }
-    ],
-    "invariance_reports": [
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/chunked",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/chunked"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s0",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s1",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s2",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s3",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/permuted",
-        "transform_kind": "permutation",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/permuted"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_right",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_right"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_left",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_left"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      }
-    ],
-    "logprob_smoke": null,
-    "backend_provenance": {
-      "backend_profile": "triton_cuda_bf16",
-      "requested_backend": "triton",
-      "actual_backend": "triton",
-      "execution_dtype": "bfloat16",
-      "accumulation_dtype": "float32",
-      "output_dtype": "bfloat16",
-      "reference_dtype": "float32",
-      "candidate_tf32_enabled": false,
-      "reference_tf32_enabled": false
-    },
-    "candidate_id": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
-    "device": "cuda",
-    "compute_capability": "sm120",
-    "seed": 20260812,
-    "fallback_reason": null,
-    "passed": true,
-    "provenance_valid": true,
-    "metadata_valid": true,
-    "observed_kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel"
-  },
-  "gradient": {
-    "op_name": "multi_axis_rope",
-    "backend_profile": "triton_cuda_bf16",
-    "accuracy_reports": [
-      {
-        "config_id": "BN/full",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/chunked",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/chunked",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.7887988835573196e-05,
-            "mean_abs_error": 3.6821968478761846e-06,
-            "max_rel_error": 1.8164952993392944,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.7113780965446495e-06,
-            "max_rel_error": 0.4316565990447998,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.995925024151802e-05,
-            "mean_abs_error": 4.7168209675874095e-06,
-            "max_rel_error": 0.7439075112342834,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.0211638659238815e-05,
-            "mean_abs_error": 4.9605891945248e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.7887988835573196e-05,
-            "mean_abs_error": 3.6821968478761846e-06,
-            "max_rel_error": 1.8164952993392944,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.7113780965446495e-06,
-            "max_rel_error": 0.4316565990447998,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.995925024151802e-05,
-            "mean_abs_error": 4.7168209675874095e-06,
-            "max_rel_error": 0.7439075112342834,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.0211638659238815e-05,
-            "mean_abs_error": 4.9605891945248e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/permuted",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/permuted",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_right",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/padded_right",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_left",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/padded_left",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      }
-    ],
-    "invariance_reports": [
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/chunked",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/chunked"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s0",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s1",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s2",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s3",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/permuted",
-        "transform_kind": "permutation",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/permuted"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_right",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_right"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_left",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_left"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      }
-    ],
-    "singleton_aggregate_reports": [],
-    "backend_provenance": {
-      "backend_profile": "triton_cuda_bf16",
-      "requested_backend": "triton",
-      "actual_backend": "triton",
-      "execution_dtype": "bfloat16",
-      "accumulation_dtype": "float32",
-      "output_dtype": "bfloat16",
-      "reference_dtype": "float32",
-      "candidate_tf32_enabled": false,
-      "reference_tf32_enabled": false
-    },
-    "candidate_id": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
-    "device": "cuda",
-    "compute_capability": "sm120",
-    "seed": 20260812,
-    "fallback_reason": null,
-    "passed": true,
-    "provenance_valid": true,
-    "metadata_valid": true,
-    "loss_reduction": "sum_over_active_tokens_then_optional_mean_by_active_count",
-    "active_token_denominator": 27,
-    "grad_tensor_names": [
-      "dx"
-    ],
-    "first_failing_op": null,
-    "first_failing_tensor": null,
-    "first_failing_config_pair": null,
-    "observed_kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel"
-  }
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.txt b/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.txt
deleted file mode 100644
index 283d1933a..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_v2_sm120/ws1-triton.txt
+++ /dev/null
@@ -1,3 +0,0 @@
-WARNING 10-28 16:54:47 [RL-Kernel]: Core binary extension (_C) unavailable: cannot import name '_C' from 'rl_engine' (D:\dandan123\dandan\issue\investigation-rl-kernel\RL-Kernel\rl_engine\__init__.py). Falling back to native code.
-Using the installed production RoPE dispatcher library
-WS1 four judgments backend=triton: passed=True
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/README.md b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/README.md
deleted file mode 100644
index 9ef00fdb1..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/README.md
+++ /dev/null
@@ -1,136 +0,0 @@
-# Qwen-Image RoPE: full native package qualification on WSL2
-
-Recorded 2026-10-28 for [PR #623](https://github.com/RL-Align/RL-Kernel/pull/623).
-Numerical/native sources are based on `ac17ab87e8d93dabe93dfdc134e9853d47de7af1`;
-this follow-up adds the reproducible gate, CPU CI routing and package provenance.
-Source fingerprints are in `manifest.json`. Historical Windows measurements stay
-in the separate `qwenimage_multi_axis_rope_v2_sm120` directory.
-
-## Environment and build
-
-| Item | Recorded value |
-| --- | --- |
-| Linux distribution | Ubuntu 24.04.5 LTS, x86_64, glibc 2.39 |
-| Virtualization | WSL 3.0.3 on Windows 11 |
-| Linux kernel | 6.18.54.1-microsoft-standard-WSL2 |
-| GPU / driver | RTX 5080, SM120, 16 GB / Windows driver 617.42 |
-| Python / PyTorch / Triton | 3.12.3 / 2.9.1+cu128 / 3.5.1 |
-| CUDA compiler / host compiler | CUDA 12.8.93 / GCC 13.3.0 |
-| Compute Sanitizer | 2025.1.0.0, CUDA 12.8 |
-| Build profile | Ordinary complete `rl_engine._C`; 10 source translation units |
-| CUDA architecture / parallelism | `TORCH_CUDA_ARCH_LIST=12.0`, `MAX_JOBS=2` |
-
-CUDA compiler, development libraries and command-line tools were installed from
-NVIDIA's WSL-Ubuntu repository; the existing Windows GPU driver provides CUDA
-driver access. The WSL driver directory `/usr/lib/wsl/lib` was added to
-`LIBRARY_PATH` for linking the package's existing CUDA driver API calls.
-`RL_KERNEL_ROPE_ONLY`, forced SM90 and fast-math flags were unset.
-
-The normal PEP 517 wheel build compiled all ten configured translation units,
-including the existing logprob/GEMM/norm/activation/attention/collective code and
-the new RoPE source. The complete `_C` loaded and registered the native RoPE
-dispatcher. `full-extension.json` records sources, flags, binary path and hash;
-`package.json` records the built wheel and installed library fingerprints.
-
-The checkout, virtual environments, build outputs and caches are inside the
-Linux filesystem, backed by the D: virtual disk. A fresh Git clone preserved all
-27 tracked symlinks. This is a WSL2 Linux userspace/compiler qualification on
-SM120; bare-metal Linux and A100/H100 remain separate hardware profiles.
-
-## Results
-
-| Check | Result | Evidence |
-| --- | --- | --- |
-| Ordinary complete `_C` wheel and in-place build/import | Passed | `full-extension.json`, `full-wheel-build.txt`, `package.json` |
-| Operator accuracy/invariance/autograd/graphs/opcheck/fullgraph | 81 passed, 1 two-GPU guard skip | `tests-and-ws1.txt`, `operator-tests.xml` |
-| CUDA and Triton WS1 four judgments | Both passed with explicit provenance | `ws1-cuda.json`, `ws1-triton.json` |
-| Standard operator CLI forward/input gradient | Both passed | `operator-cuda.json`, `operator-triton.json` |
-| Related regressions | 195 passed, 7 Ascend skips | `regression.txt` |
-| Repository layout with extension enforcement enabled | 37 passed, zero skips/failures | `layout.txt` |
-| Fresh installed full wheel, outside source checkout | Eager/compiled forward and input-gradient bytes exact for both backends | `installed-wheel.json` |
-| Actual CPU-only PyTorch wheel/reference installation, Python 3.10.22 | 57 passed, 52 expected GPU skips; reference forward/backward passed | `cpu310-contracts.txt`, `cpu310-environment.json` |
-| Actual CPU-only PyTorch wheel/reference installation, Python 3.12.3 | 57 passed, 52 expected GPU skips; reference forward/backward passed | `cpu-contracts.txt`, `cpu-environment.json` |
-| Compute Sanitizer accuracy/stride/empty/stream subset | 34 passed; zero errors | `memcheck.txt` |
-| Pinned actual Diffusers symbols | 30 forward/backward cases passed; image/text tables exact | `upstream.json` |
-| Reference-grid benchmark correctness | 40 comparisons passed | `benchmark-bfloat16.json` |
-| Shell/workflow syntax, changed-file checks, strict docs | Passed locally | `quality.txt` |
-
-The full-extension GPU gate rejects missing native/Triton execution. Its only
-allowed operator-test skip is the two-device guard on a one-GPU runner.
-The CPU-only results are separate: both Python versions use `torch==2.9.1+cpu`,
-and their installed packages contain no native CUDA binary. CPU skips are not
-GPU acceptance evidence. Both runs install the normal CPU package and exercise
-the workflow's three test files locally; these are not hosted CI results.
-The 37 passing Linux layout tests contrast with the five Windows baseline
-failures recorded earlier; those historical logs were not rewritten.
-
-The installed-wheel probe starts in a fresh process outside the checkout,
-loads `_C` from site-packages and reports that exact path. It tests packed-QKV
-forward/input-gradient equality in eager and `torch.compile(fullgraph=True)`
-for both CUDA and Triton. The FlashAttention BSD license is present in the wheel.
-
-## Standalone benchmark
-
-BF16, batch 1, 24 heads, head dimension 128. CUDA Graph medians below are
-**forward / forward+backward milliseconds per call**:
-
-| Grid / layout | Native CUDA | Adapted Triton | FlashAttention, no FMA | Diffusers complex |
-| --- | --- | --- | --- | --- |
-| 64x64 contiguous | 0.05780 / 0.12967 | 0.04929 / 0.12414 | 0.04898 / 0.12407 | 0.24446 / 0.51281 |
-| 64x64 packed QKV | 0.05868 / 0.13079 | 0.04731 / 0.11841 | 0.04722 / 0.11847 | 0.24754 / 0.51583 |
-| 83x83 contiguous | 0.11198 / 0.22392 | 0.10720 / 0.22538 | 0.10739 / 0.21420 | 0.52579 / 1.06721 |
-| 83x83 packed QKV | 0.11233 / 0.22486 | 0.10745 / 0.21478 | 0.10744 / 0.21469 | 0.52820 / 1.06947 |
-| 104x58 contiguous | 0.09797 / 0.19708 | 0.09366 / 0.18497 | 0.09362 / 0.18458 | 0.44992 / 0.90110 |
-| 104x58 packed QKV | 0.09610 / 0.19612 | 0.09310 / 0.18828 | 0.09323 / 0.18817 | 0.44979 / 0.90276 |
-
-These are 20 calls per graph, five warm calls/replays and 30 replay measurements.
-Three synchronized eager-wall trials of 30 calls are also in the JSON. Tables,
-compilation and input/backing allocations are outside timing. Extra peak includes
-the returned output or gradient. The short synthetic case is included in raw data.
-At 4096 tokens, both GPU paths' forward extra peak is 24 MiB for either layout,
-the returned BF16 output size. Packed inputs are read without materialization.
-
-Original FlashAttention and its autograd wrapper are checksum-pinned to
-`98c1cee0ed59307833841cdbf567fe336ceec13a`. The no-FMA baseline changes only the
-launch contraction option. CUDA, adapted Triton and that baseline have zero
-differing forward/input-gradient bytes against the unfused FP32 gold in every
-benchmark case. Default FlashAttention and complex arithmetic use declared tolerances.
-The adapted Triton path is close to the mature baseline; no performance advantage
-over FlashAttention is established. Timings are one-process measurements on a
-shared Windows/WSL GPU without clock pinning, not a model-throughput result or
-a controlled comparison against the earlier Windows process.
-
-## Reproduction and CI
-
-Use the matching CUDA 12.8 toolkit and Linux PyTorch 2.9.1+cu128 in a fresh
-Linux virtual environment. Other Python package versions are recorded in
-`python-packages.json`. From the Linux checkout:
-
-```bash
-export CUDA_HOME=/usr/local/cuda-12.8
-export PATH="$CUDA_HOME/bin:$PATH"
-export LIBRARY_PATH="/usr/lib/wsl/lib${LIBRARY_PATH:+:$LIBRARY_PATH}"
-export MAX_JOBS=2 TORCH_CUDA_ARCH_LIST=12.0 RL_KERNEL_REQUIRE_EXT=1
-unset RL_KERNEL_ROPE_ONLY KERNEL_ALIGN_FORCE_SM90 KERNEL_ALIGN_USE_FAST_MATH
-python -m pip wheel --no-build-isolation --no-deps . -w /tmp/rope-full-wheel
-python -m pip install --force-reinstall --no-deps /tmp/rope-full-wheel/rl_kernel-*.whl
-python setup.py build_ext --inplace
-ROPE_PYTHON=python ROPE_WS1_OUTPUT=/tmp/rope-ws1 bash ci/scripts/run_qwenimage_rope_ws1.sh
-python ci/run.py layout -q -rs
-compute-sanitizer --tool memcheck --error-exitcode 1 python -m pytest \
-  tests/ops/rope/test_multi_axis_rope.py -q \
-  -k "forward_backward_accuracy or noncontiguous_input_tables_and_gradient or empty_forward_backward or packed_qkv_and_independent_table_strides or native_entrypoint_checks"
-```
-
-The new [operator gate](../../../ci/scripts/run_qwenimage_rope_ws1.sh) requires
-the complete `_C` registration, records actual library/backend identity and
-enforces the GPU skip policy. The [CPU workflow](../../../.github/workflows/qwenimage-rope.yml)
-targets `test-qwenimage`, runs Python 3.10/3.12 CPU contracts, checks changed files
-and builds documentation. This report includes local CPU-only runs on both
-Python 3.10 and 3.12; hosted CI status is tracked on the PR. GPU runners use the
-explicit gate.
-
-Bare-metal Linux, A100/H100, two-GPU guards, QK-norm fusion, model/block assembly
-and trajectories remain unqualified. Building the entire library does not
-certify every native operator. No Dense Qwen3/distributed certification or
-cross-device/version byte-equality claim is made.
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/benchmark-bfloat16.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/benchmark-bfloat16.json
deleted file mode 100644
index ca80105b0..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/benchmark-bfloat16.json
+++ /dev/null
@@ -1,2781 +0,0 @@
-{
-  "gpu": "NVIDIA GeForce RTX 5080",
-  "capability": [
-    12,
-    0
-  ],
-  "os": "Linux-6.18.54.1-microsoft-standard-WSL2-x86_64-with-glibc2.39",
-  "python": "3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0]",
-  "torch": "2.9.1+cu128",
-  "cuda_runtime": "12.8",
-  "triton": "3.5.1",
-  "dtype": "bfloat16",
-  "base_commit": "ac17ab87e8d93dabe93dfdc134e9853d47de7af1",
-  "source_sha256": {
-    "csrc/cuda/rope/multi_axis_rope.cu": "f5db5cf5d4962053194455d9cde8cf91cf4e824bad7abf72e8410a404bcb0e44",
-    "rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py": "cf0eebf315830a4178053a71a77e650b58aa85bba1d1a08aff7677350af755d9",
-    "rl_engine/backends/shared/triton/rope/multi_axis_rope.py": "f08ff90627c66e15e9c558f8a68f132434a741a76997b3800bfbf73dc8940e9e",
-    "rl_engine/backends/cuda/rope/multi_axis_rope.py": "d779d9b60ce5d1d6651925f723049930d898024fdf0548ce29a2019b84753d65",
-    "rl_engine/ops/rope/_registration.py": "add0375d1639444908091d01e8d29781f7a975beffec64b77bfaa81a5b9a0ae3",
-    "rl_engine/ops/rope/multi_axis_rope.py": "07bcff1160f22c1ff6730be731206b2f4a153e2fada26afc93fc884e780568cb",
-    "rl_engine/reference/rope/multi_axis_rope.py": "c9a94730047f51659f7aa0a443e88d6ce7d03d00d6813a9aed548cafbf6347bf",
-    "benchmarks/operators/rope/benchmark_multi_axis_rope.py": "6961a55803fbd4e566bd060558fb59e658e7f1f704e61ca97f912e5d16b55884"
-  },
-  "timing": "CUDA graph ms/call and synchronized eager wall ms/call; warm allocation; excludes tables, compilation, and initial allocation of inputs",
-  "memory": "extra peak allocated bytes excluding inputs/tables/grad; includes returned output or gradient; packed QKV backing storage allocated before timing",
-  "iterations": 30,
-  "warmup": 5,
-  "trials": 3,
-  "graph_calls": 20,
-  "flash_baseline": {
-    "included": true,
-    "revision": "98c1cee0ed59307833841cdbf567fe336ceec13a",
-    "kernel_sha256_normalized": "d2037e623854c702fb075784fe4292d4a89d6b9ac1ae33cab3d6788f8f94f240",
-    "layer_sha256_normalized": "a79cb42dfc33e126227bdba27ae7a024d7d5cc5c8fd7f7a5f26c68d501024e0a",
-    "mode": "upstream rotary_kernel, apply_rotary and ApplyRotaryEmb, interleaved=True",
-    "no_fma_change": "only add enable_fp_fusion=False at the original kernel launch"
-  },
-  "trace": {
-    "cuda": {
-      "contract": "qwenimage-interleaved-rope-fp32-v1",
-      "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
-      "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
-      "layout": "BSHD, adjacent channel pairs",
-      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-    },
-    "triton": {
-      "contract": "qwenimage-interleaved-rope-fp32-v1",
-      "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
-      "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
-      "layout": "BSHD, adjacent channel pairs",
-      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-    }
-  },
-  "qualification": "Exact bytes vs unfused FP32 gold for CUDA, Triton and flash_no_fma; declared tolerance for upstream complex and default FlashAttention",
-  "correctness": [
-    {
-      "backend": "upstream_complex",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        130,
-        139
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        130,
-        139
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        141,
-        139
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        141,
-        139
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        235,
-        224
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        235,
-        224
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        221,
-        226
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        221,
-        226
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        195,
-        199
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        195,
-        199
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "upstream_complex",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        230,
-        220
-      ]
-    },
-    {
-      "backend": "cuda",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "triton",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    },
-    {
-      "backend": "flash_default",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        230,
-        220
-      ]
-    },
-    {
-      "backend": "flash_no_fma",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "different_bytes_y_dx": [
-        0,
-        0
-      ]
-    }
-  ],
-  "results": [
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0034856000915169717,
-      "graph_min_ms": 0.0034543998539447785,
-      "graph_max_ms": 0.0035312000662088396,
-      "eager_wall_median_ms": 0.08385686669498682,
-      "eager_wall_trials_ms": [
-        0.09771363335554877,
-        0.08163450002030004,
-        0.08385686669498682
-      ],
-      "extra_peak_bytes": 184320
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.007651200145483017,
-      "graph_min_ms": 0.007601600140333176,
-      "graph_max_ms": 0.007692799717187881,
-      "eager_wall_median_ms": 0.19571573332844613,
-      "eager_wall_trials_ms": [
-        0.2535254000273805,
-        0.19571573332844613,
-        0.18590186670432254
-      ],
-      "extra_peak_bytes": 187392
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0012047999538481236,
-      "graph_min_ms": 0.0011695999652147292,
-      "graph_max_ms": 0.0012400000356137752,
-      "eager_wall_median_ms": 0.05762376664885475,
-      "eager_wall_trials_ms": [
-        0.07130193328824437,
-        0.05762376664885475,
-        0.057149166665719044
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.002329600043594837,
-      "graph_min_ms": 0.002223999984562397,
-      "graph_max_ms": 0.0026383999735116958,
-      "eager_wall_median_ms": 0.17063139997238372,
-      "eager_wall_trials_ms": [
-        0.17063139997238372,
-        0.16431806670880178,
-        0.24000986668397672
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0013432000298053027,
-      "graph_min_ms": 0.0010928000323474408,
-      "graph_max_ms": 0.006166400015354156,
-      "eager_wall_median_ms": 0.08490003328915918,
-      "eager_wall_trials_ms": [
-        0.08490003328915918,
-        0.0852208666401566,
-        0.08217023332690587
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.002125599980354309,
-      "graph_min_ms": 0.0020735999569296837,
-      "graph_max_ms": 0.0021679999306797983,
-      "eager_wall_median_ms": 0.3603733666144156,
-      "eager_wall_trials_ms": [
-        0.3862667332820517,
-        0.31255590001819655,
-        0.3603733666144156
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0011024000123143196,
-      "graph_min_ms": 0.0010544000193476676,
-      "graph_max_ms": 0.01971679925918579,
-      "eager_wall_median_ms": 0.07505266664035541,
-      "eager_wall_trials_ms": [
-        0.09395336667997374,
-        0.07483469998987857,
-        0.07505266664035541
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0020584000274538996,
-      "graph_min_ms": 0.002022399939596653,
-      "graph_max_ms": 0.002151999995112419,
-      "eager_wall_median_ms": 0.3825640333161573,
-      "eager_wall_trials_ms": [
-        0.3916477333405055,
-        0.38042896667320747,
-        0.3825640333161573
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0011288000270724297,
-      "graph_min_ms": 0.0010703999549150466,
-      "graph_max_ms": 0.0025040000677108766,
-      "eager_wall_median_ms": 0.07896646663236122,
-      "eager_wall_trials_ms": [
-        0.07896646663236122,
-        0.08368196668016026,
-        0.07650009999148703
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        18432,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.002057600021362305,
-      "graph_min_ms": 0.002022399939596653,
-      "graph_max_ms": 0.0021487999707460403,
-      "eager_wall_median_ms": 0.3408336999806731,
-      "eager_wall_trials_ms": [
-        0.3408336999806731,
-        0.32808933331883355,
-        0.35737479999321903
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0037903999909758568,
-      "graph_min_ms": 0.003734400123357773,
-      "graph_max_ms": 0.0038384001702070234,
-      "eager_wall_median_ms": 0.026099166704322368,
-      "eager_wall_trials_ms": [
-        0.027518266688275617,
-        0.02535739998468974,
-        0.026099166704322368
-      ],
-      "extra_peak_bytes": 184320
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.008167200163006781,
-      "graph_min_ms": 0.00809599980711937,
-      "graph_max_ms": 0.008207999914884568,
-      "eager_wall_median_ms": 0.22357730000900725,
-      "eager_wall_trials_ms": [
-        0.255850233346185,
-        0.22357730000900725,
-        0.2018136666568656
-      ],
-      "extra_peak_bytes": 187392
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.001232000021263957,
-      "graph_min_ms": 0.0011552000418305398,
-      "graph_max_ms": 0.0026464000344276427,
-      "eager_wall_median_ms": 0.05505463335187718,
-      "eager_wall_trials_ms": [
-        0.05505463335187718,
-        0.052852566659566946,
-        0.05531913332864254
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.002279200032353401,
-      "graph_min_ms": 0.00220320001244545,
-      "graph_max_ms": 0.005976000055670738,
-      "eager_wall_median_ms": 0.2586098333267728,
-      "eager_wall_trials_ms": [
-        0.2587837333218583,
-        0.25146820001585485,
-        0.2586098333267728
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0011359999887645244,
-      "graph_min_ms": 0.001071999967098236,
-      "graph_max_ms": 0.002603200078010559,
-      "eager_wall_median_ms": 0.03804060003555302,
-      "eager_wall_trials_ms": [
-        0.03804060003555302,
-        0.04575726664673615,
-        0.03741419999035619
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0021271999925374986,
-      "graph_min_ms": 0.0020896000787615776,
-      "graph_max_ms": 0.016212800145149232,
-      "eager_wall_median_ms": 0.24443720000514682,
-      "eager_wall_trials_ms": [
-        0.281011633342132,
-        0.24443720000514682,
-        0.23313996668245332
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.001104000024497509,
-      "graph_min_ms": 0.001065600011497736,
-      "graph_max_ms": 0.0011983999982476234,
-      "eager_wall_median_ms": 0.078476833308135,
-      "eager_wall_trials_ms": [
-        0.07911106664929927,
-        0.07519836666081876,
-        0.078476833308135
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0020527999848127367,
-      "graph_min_ms": 0.0020207999274134635,
-      "graph_max_ms": 0.0021023999899625776,
-      "eager_wall_median_ms": 0.30830893338134047,
-      "eager_wall_trials_ms": [
-        0.3673079666744646,
-        0.29416493331761256,
-        0.30830893338134047
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0011272000148892403,
-      "graph_min_ms": 0.0010511999949812888,
-      "graph_max_ms": 0.002239999920129776,
-      "eager_wall_median_ms": 0.07592803334167304,
-      "eager_wall_trials_ms": [
-        0.0775905666766145,
-        0.07592803334167304,
-        0.0742380666755101
-      ],
-      "extra_peak_bytes": 36864
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6,
-        24,
-        128
-      ],
-      "input_stride": [
-        55296,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0020535999909043314,
-      "graph_min_ms": 0.002014400064945221,
-      "graph_max_ms": 0.0020975999534130095,
-      "eager_wall_median_ms": 0.31871896662778454,
-      "eager_wall_trials_ms": [
-        0.31871896662778454,
-        0.33167873334605247,
-        0.2915063333299865
-      ],
-      "extra_peak_bytes": 73728
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.24445840120315554,
-      "graph_min_ms": 0.24224960803985596,
-      "graph_max_ms": 0.29114880561828616,
-      "eager_wall_median_ms": 0.247559633377629,
-      "eager_wall_trials_ms": [
-        0.25793806665509084,
-        0.2471523333345734,
-        0.247559633377629
-      ],
-      "extra_peak_bytes": 125829120
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.5128072023391723,
-      "graph_min_ms": 0.49546399116516116,
-      "graph_max_ms": 0.5571055889129639,
-      "eager_wall_median_ms": 0.5245368333514003,
-      "eager_wall_trials_ms": [
-        0.5245368333514003,
-        0.518849266639639,
-        0.5350402333230401
-      ],
-      "extra_peak_bytes": 127926272
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.05780480206012726,
-      "graph_min_ms": 0.057633602619171144,
-      "graph_max_ms": 0.07439360022544861,
-      "eager_wall_median_ms": 0.06806933330760027,
-      "eager_wall_trials_ms": [
-        0.07017140002668991,
-        0.06806933330760027,
-        0.06168783335548748
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.1296720027923584,
-      "graph_min_ms": 0.12902239561080933,
-      "graph_max_ms": 0.15688639879226685,
-      "eager_wall_median_ms": 0.2333034999537631,
-      "eager_wall_trials_ms": [
-        0.23402639999403618,
-        0.19631726663646987,
-        0.2333034999537631
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.049288800358772276,
-      "graph_min_ms": 0.04815520048141479,
-      "graph_max_ms": 0.06759200096130372,
-      "eager_wall_median_ms": 0.08061226665934858,
-      "eager_wall_trials_ms": [
-        0.07100329997532147,
-        0.0930295666573026,
-        0.08061226665934858
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.12414479851722718,
-      "graph_min_ms": 0.12152320146560669,
-      "graph_max_ms": 0.16048799753189086,
-      "eager_wall_median_ms": 0.35214599999259616,
-      "eager_wall_trials_ms": [
-        0.35569400000288925,
-        0.35214599999259616,
-        0.32247826663175755
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.04924079924821853,
-      "graph_min_ms": 0.04797120094299316,
-      "graph_max_ms": 0.068995201587677,
-      "eager_wall_median_ms": 0.13288023334704727,
-      "eager_wall_trials_ms": [
-        0.14359453337723002,
-        0.12000013339275029,
-        0.13288023334704727
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.12360240221023559,
-      "graph_min_ms": 0.1215999960899353,
-      "graph_max_ms": 0.16038399934768677,
-      "eager_wall_median_ms": 0.35429976663484314,
-      "eager_wall_trials_ms": [
-        0.34954116666388774,
-        0.3772011999899405,
-        0.35429976663484314
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0489767998456955,
-      "graph_min_ms": 0.0479856014251709,
-      "graph_max_ms": 0.06645600199699402,
-      "eager_wall_median_ms": 0.1330274333061728,
-      "eager_wall_trials_ms": [
-        0.1330274333061728,
-        0.1200082000044252,
-        0.13366413331823424
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        12582912,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.12407039999961852,
-      "graph_min_ms": 0.12193280458450317,
-      "graph_max_ms": 0.1605712056159973,
-      "eager_wall_median_ms": 0.36529683329717955,
-      "eager_wall_trials_ms": [
-        0.3895335000076254,
-        0.36529683329717955,
-        0.3503133667133322
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.2475391983985901,
-      "graph_min_ms": 0.24529919624328614,
-      "graph_max_ms": 0.27714240550994873,
-      "eager_wall_median_ms": 0.25804310001452296,
-      "eager_wall_trials_ms": [
-        0.25804310001452296,
-        0.26510979999632883,
-        0.2503190333300154
-      ],
-      "extra_peak_bytes": 125829120
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.5158272027969361,
-      "graph_min_ms": 0.49741439819335936,
-      "graph_max_ms": 0.5595920085906982,
-      "eager_wall_median_ms": 0.5221092666639985,
-      "eager_wall_trials_ms": [
-        0.5362037999778598,
-        0.5221092666639985,
-        0.5111432000073061
-      ],
-      "extra_peak_bytes": 127926272
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.05868320167064667,
-      "graph_min_ms": 0.058520001173019406,
-      "graph_max_ms": 0.0754975974559784,
-      "eager_wall_median_ms": 0.06486223331497361,
-      "eager_wall_trials_ms": [
-        0.06433113333817649,
-        0.08489736665069358,
-        0.06486223331497361
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.13079119920730592,
-      "graph_min_ms": 0.13011519908905028,
-      "graph_max_ms": 0.1711359977722168,
-      "eager_wall_median_ms": 0.21280260001124893,
-      "eager_wall_trials_ms": [
-        0.2440493333779159,
-        0.20012976668416135,
-        0.21280260001124893
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.04731200039386749,
-      "graph_min_ms": 0.04663040041923523,
-      "graph_max_ms": 0.06482719779014587,
-      "eager_wall_median_ms": 0.08378989999376547,
-      "eager_wall_trials_ms": [
-        0.08378989999376547,
-        0.07920856666411662,
-        0.09377416663483018
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.11841280460357666,
-      "graph_min_ms": 0.11643359661102295,
-      "graph_max_ms": 0.14446239471435546,
-      "eager_wall_median_ms": 0.3498463667104564,
-      "eager_wall_trials_ms": [
-        0.3498463667104564,
-        0.35651856663510745,
-        0.3451830999741408
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.047358399629592894,
-      "graph_min_ms": 0.04637599885463715,
-      "graph_max_ms": 0.06467199921607972,
-      "eager_wall_median_ms": 0.12952533331069085,
-      "eager_wall_trials_ms": [
-        0.12952533331069085,
-        0.1346659000167468,
-        0.12932516668418734
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.11804000139236451,
-      "graph_min_ms": 0.11632959842681885,
-      "graph_max_ms": 0.13619199991226197,
-      "eager_wall_median_ms": 0.3809861666619933,
-      "eager_wall_trials_ms": [
-        0.3809861666619933,
-        0.39690620002753957,
-        0.3660936666468236
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.04721840023994446,
-      "graph_min_ms": 0.04616160094738007,
-      "graph_max_ms": 0.06441760063171387,
-      "eager_wall_median_ms": 0.1290158666355031,
-      "eager_wall_trials_ms": [
-        0.13406613331123177,
-        0.1290158666355031,
-        0.12222770001244498
-      ],
-      "extra_peak_bytes": 25165824
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        4096,
-        24,
-        128
-      ],
-      "input_stride": [
-        37748736,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.11847440004348755,
-      "graph_min_ms": 0.11651359796524048,
-      "graph_max_ms": 0.1468176007270813,
-      "eager_wall_median_ms": 0.4179154666417162,
-      "eager_wall_trials_ms": [
-        0.4434371333142432,
-        0.4179154666417162,
-        0.4101795332947707
-      ],
-      "extra_peak_bytes": 50331648
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.5257872104644776,
-      "graph_min_ms": 0.5080063819885254,
-      "graph_max_ms": 0.5525023937225342,
-      "eager_wall_median_ms": 0.5246295667044857,
-      "eager_wall_trials_ms": [
-        0.5097412666752158,
-        0.528316466625256,
-        0.5246295667044857
-      ],
-      "extra_peak_bytes": 211630080
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 1.067212772369385,
-      "graph_min_ms": 1.0389776229858398,
-      "graph_max_ms": 1.0976528167724608,
-      "eager_wall_median_ms": 1.0610819333426964,
-      "eager_wall_trials_ms": [
-        1.0702722999970622,
-        1.0479776333644015,
-        1.0610819333426964
-      ],
-      "extra_peak_bytes": 215157248
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.11197760105133056,
-      "graph_min_ms": 0.11102559566497802,
-      "graph_max_ms": 0.13988800048828126,
-      "eager_wall_median_ms": 0.12250546666715915,
-      "eager_wall_trials_ms": [
-        0.11638536664880424,
-        0.13100786666958206,
-        0.12250546666715915
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.22392239570617678,
-      "graph_min_ms": 0.22292959690093994,
-      "graph_max_ms": 0.26030399799346926,
-      "eager_wall_median_ms": 0.24186149997452353,
-      "eager_wall_trials_ms": [
-        0.24518829995940905,
-        0.23021603331775017,
-        0.24186149997452353
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.10719599723815917,
-      "graph_min_ms": 0.10527039766311645,
-      "graph_max_ms": 0.123745596408844,
-      "eager_wall_median_ms": 0.10772686667526916,
-      "eager_wall_trials_ms": [
-        0.10711883332987782,
-        0.10772686667526916,
-        0.10832119999880281
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.2253783941268921,
-      "graph_min_ms": 0.21142399311065674,
-      "graph_max_ms": 0.25633440017700193,
-      "eager_wall_median_ms": 0.29183916667534504,
-      "eager_wall_trials_ms": [
-        0.29183916667534504,
-        0.31016709999676095,
-        0.27477109997562366
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.10732319951057434,
-      "graph_min_ms": 0.10543199777603149,
-      "graph_max_ms": 0.12232480049133301,
-      "eager_wall_median_ms": 0.11885433329249888,
-      "eager_wall_trials_ms": [
-        0.11160320003303544,
-        0.11885433329249888,
-        0.148421633275575
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.22528079748153684,
-      "graph_min_ms": 0.21176319122314452,
-      "graph_max_ms": 0.24576001167297362,
-      "eager_wall_median_ms": 0.41978080001475365,
-      "eager_wall_trials_ms": [
-        0.4051902666712219,
-        0.42735956667456776,
-        0.41978080001475365
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.10738959908485413,
-      "graph_min_ms": 0.10557440519332886,
-      "graph_max_ms": 0.1310464024543762,
-      "eager_wall_median_ms": 0.1100779333379857,
-      "eager_wall_trials_ms": [
-        0.11928496666466042,
-        0.10885563333431492,
-        0.1100779333379857
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        21163008,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.21419600248336793,
-      "graph_min_ms": 0.2123471975326538,
-      "graph_max_ms": 0.24199841022491456,
-      "eager_wall_median_ms": 0.42007340004298993,
-      "eager_wall_trials_ms": [
-        0.4409272999813159,
-        0.42007340004298993,
-        0.40205200002674246
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.5282040119171143,
-      "graph_min_ms": 0.5110288143157959,
-      "graph_max_ms": 0.5885712146759033,
-      "eager_wall_median_ms": 0.547415499992591,
-      "eager_wall_trials_ms": [
-        0.547415499992591,
-        0.5160773666527044,
-        0.5478402666388623
-      ],
-      "extra_peak_bytes": 211630080
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 1.0694655895233156,
-      "graph_min_ms": 1.0273263931274415,
-      "graph_max_ms": 1.0971407890319824,
-      "eager_wall_median_ms": 1.0825770000034634,
-      "eager_wall_trials_ms": [
-        1.0599153666286536,
-        1.0825770000034634,
-        1.0843000333503976
-      ],
-      "extra_peak_bytes": 215157248
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.11232720017433166,
-      "graph_min_ms": 0.11165920495986939,
-      "graph_max_ms": 0.13819040060043336,
-      "eager_wall_median_ms": 0.12610396667393312,
-      "eager_wall_trials_ms": [
-        0.12610396667393312,
-        0.11783170005704353,
-        0.1361641000282058
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.22485519647598268,
-      "graph_min_ms": 0.2238991975784302,
-      "graph_max_ms": 0.2728208065032959,
-      "eager_wall_median_ms": 0.2468356666213367,
-      "eager_wall_trials_ms": [
-        0.2468356666213367,
-        0.25081796666199807,
-        0.23804076669572774
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.10744640231132507,
-      "graph_min_ms": 0.10558079481124878,
-      "graph_max_ms": 0.12357280254364014,
-      "eager_wall_median_ms": 0.10917900005248764,
-      "eager_wall_trials_ms": [
-        0.108484466666899,
-        0.11029286670236615,
-        0.10917900005248764
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.21478239297866822,
-      "graph_min_ms": 0.21232800483703612,
-      "graph_max_ms": 0.2458656072616577,
-      "eager_wall_median_ms": 0.3064794333719571,
-      "eager_wall_trials_ms": [
-        0.3064794333719571,
-        0.29039823330094805,
-        0.3166810000038822
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.10760959982872009,
-      "graph_min_ms": 0.10598560571670532,
-      "graph_max_ms": 0.13379520177841187,
-      "eager_wall_median_ms": 0.11580663334219328,
-      "eager_wall_trials_ms": [
-        0.11580663334219328,
-        0.12037279999882837,
-        0.10920460002428929
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.2147576093673706,
-      "graph_min_ms": 0.21242880821228027,
-      "graph_max_ms": 0.2415616035461426,
-      "eager_wall_median_ms": 0.42939333337320323,
-      "eager_wall_trials_ms": [
-        0.4258981332895928,
-        0.45155456670424127,
-        0.42939333337320323
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.1074351966381073,
-      "graph_min_ms": 0.10590399503707885,
-      "graph_max_ms": 0.13916159868240358,
-      "eager_wall_median_ms": 0.11420323335187277,
-      "eager_wall_trials_ms": [
-        0.11669516667704254,
-        0.11420323335187277,
-        0.11166136664542137
-      ],
-      "extra_peak_bytes": 42326016
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6889,
-        24,
-        128
-      ],
-      "input_stride": [
-        63489024,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.21468960046768187,
-      "graph_min_ms": 0.21285600662231446,
-      "graph_max_ms": 0.24125280380249023,
-      "eager_wall_median_ms": 0.4036813999846345,
-      "eager_wall_trials_ms": [
-        0.40494703334843507,
-        0.4027272333284297,
-        0.4036813999846345
-      ],
-      "extra_peak_bytes": 84652032
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.44991838932037354,
-      "graph_min_ms": 0.4312863826751709,
-      "graph_max_ms": 0.4923136234283447,
-      "eager_wall_median_ms": 0.4565672999888193,
-      "eager_wall_trials_ms": [
-        0.46419443333434174,
-        0.4565672999888193,
-        0.43665636664324364
-      ],
-      "extra_peak_bytes": 185303040
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.9010960102081299,
-      "graph_min_ms": 0.8724479675292969,
-      "graph_max_ms": 0.9627167701721191,
-      "eager_wall_median_ms": 0.9057802333093908,
-      "eager_wall_trials_ms": [
-        0.9057802333093908,
-        0.9399917666087276,
-        0.8861856666650662
-      ],
-      "extra_peak_bytes": 188391424
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09797199964523315,
-      "graph_min_ms": 0.0971664011478424,
-      "graph_max_ms": 0.12116639614105225,
-      "eager_wall_median_ms": 0.10149109997049284,
-      "eager_wall_trials_ms": [
-        0.10149109997049284,
-        0.10145610000715048,
-        0.10150603332779913
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.19708160161972046,
-      "graph_min_ms": 0.19586080312728882,
-      "graph_max_ms": 0.2340928077697754,
-      "eager_wall_median_ms": 0.23985833334639514,
-      "eager_wall_trials_ms": [
-        0.2599421666673152,
-        0.21093993333731,
-        0.23985833334639514
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09365599751472473,
-      "graph_min_ms": 0.09192960262298584,
-      "graph_max_ms": 0.1098512053489685,
-      "eager_wall_median_ms": 0.09574563337082509,
-      "eager_wall_trials_ms": [
-        0.10015153335795428,
-        0.09574563337082509,
-        0.09447829997952795
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.18496879935264587,
-      "graph_min_ms": 0.18224799633026123,
-      "graph_max_ms": 0.20439839363098145,
-      "eager_wall_median_ms": 0.27221376664480584,
-      "eager_wall_trials_ms": [
-        0.30083316663270426,
-        0.2585961666530541,
-        0.27221376664480584
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09344720244407653,
-      "graph_min_ms": 0.09223840236663819,
-      "graph_max_ms": 0.10959680080413818,
-      "eager_wall_median_ms": 0.10214689997762132,
-      "eager_wall_trials_ms": [
-        0.10214689997762132,
-        0.09693080003974804,
-        0.10462310001457809
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.18451839685440063,
-      "graph_min_ms": 0.1817072033882141,
-      "graph_max_ms": 0.2279695987701416,
-      "eager_wall_median_ms": 0.38619293330460397,
-      "eager_wall_trials_ms": [
-        0.4084413000479496,
-        0.38619293330460397,
-        0.3539249666573596
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09362159967422486,
-      "graph_min_ms": 0.09215999841690063,
-      "graph_max_ms": 0.1108240008354187,
-      "eager_wall_median_ms": 0.10443516669814319,
-      "eager_wall_trials_ms": [
-        0.10080796667656007,
-        0.10632786670612404,
-        0.10443516669814319
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "contiguous",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        18530304,
-        3072,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.18458240032196044,
-      "graph_min_ms": 0.18273439407348632,
-      "graph_max_ms": 0.21199359893798828,
-      "eager_wall_median_ms": 0.3521715333287527,
-      "eager_wall_trials_ms": [
-        0.3573932999643148,
-        0.32707516663019004,
-        0.3521715333287527
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.44979281425476075,
-      "graph_min_ms": 0.4314896106719971,
-      "graph_max_ms": 0.48181757926940916,
-      "eager_wall_median_ms": 0.4543782333106113,
-      "eager_wall_trials_ms": [
-        0.4543782333106113,
-        0.4393956000058097,
-        0.4658290666460137
-      ],
-      "extra_peak_bytes": 185303040
-    },
-    {
-      "backend": "upstream_complex",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.9027584075927735,
-      "graph_min_ms": 0.8707056045532227,
-      "graph_max_ms": 0.9478416442871094,
-      "eager_wall_median_ms": 0.9098065333091654,
-      "eager_wall_trials_ms": [
-        0.92959719998665,
-        0.9058386999943954,
-        0.9098065333091654
-      ],
-      "extra_peak_bytes": 188391424
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.0960999995470047,
-      "graph_min_ms": 0.09566079974174499,
-      "graph_max_ms": 0.1124384045600891,
-      "eager_wall_median_ms": 0.10085823329670045,
-      "eager_wall_trials_ms": [
-        0.10085823329670045,
-        0.1027275666880693,
-        0.10056819998377857
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "cuda",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.19612080454826356,
-      "graph_min_ms": 0.19543039798736572,
-      "graph_max_ms": 0.22353920936584473,
-      "eager_wall_median_ms": 0.21260113335301867,
-      "eager_wall_trials_ms": [
-        0.21260113335301867,
-        0.23863556671130937,
-        0.20749836670195995
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "triton",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09310480058193207,
-      "graph_min_ms": 0.09154719710350037,
-      "graph_max_ms": 0.1167631983757019,
-      "eager_wall_median_ms": 0.10304526670855314,
-      "eager_wall_trials_ms": [
-        0.09351696662633913,
-        0.10913006669094709,
-        0.10304526670855314
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "triton",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.18827759623527526,
-      "graph_min_ms": 0.18518240451812745,
-      "graph_max_ms": 0.21609120368957518,
-      "eager_wall_median_ms": 0.33319789999950444,
-      "eager_wall_trials_ms": [
-        0.33319789999950444,
-        0.3456709999833644,
-        0.3028138333320385
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09317199885845184,
-      "graph_min_ms": 0.09170079827308655,
-      "graph_max_ms": 0.11182240247726441,
-      "eager_wall_median_ms": 0.13384176666780453,
-      "eager_wall_trials_ms": [
-        0.12049876671274735,
-        0.15613870000379393,
-        0.13384176666780453
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "flash_default",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.18772000074386597,
-      "graph_min_ms": 0.18648799657821655,
-      "graph_max_ms": 0.2264847993850708,
-      "eager_wall_median_ms": 0.3818082333358082,
-      "eager_wall_trials_ms": [
-        0.44640976666414645,
-        0.3674994000296768,
-        0.3818082333358082
-      ],
-      "extra_peak_bytes": 74809344
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.09322640001773834,
-      "graph_min_ms": 0.09153280258178711,
-      "graph_max_ms": 0.11082240343093872,
-      "eager_wall_median_ms": 0.13263150000663396,
-      "eager_wall_trials_ms": [
-        0.13263150000663396,
-        0.13320910002221353,
-        0.12726086667195582
-      ],
-      "extra_peak_bytes": 37060608
-    },
-    {
-      "backend": "flash_no_fma",
-      "phase": "forward_backward",
-      "layout": "packed_qkv",
-      "shape": [
-        1,
-        6032,
-        24,
-        128
-      ],
-      "input_stride": [
-        55590912,
-        9216,
-        128,
-        1
-      ],
-      "graph_median_ms": 0.18817280530929564,
-      "graph_min_ms": 0.18564800024032593,
-      "graph_max_ms": 0.21453280448913575,
-      "eager_wall_median_ms": 0.3810788333794335,
-      "eager_wall_trials_ms": [
-        0.3668230666638313,
-        0.3885829333739821,
-        0.3810788333794335
-      ],
-      "extra_peak_bytes": 74809344
-    }
-  ]
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-contracts.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-contracts.txt
deleted file mode 100644
index de9819e68..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-contracts.txt
+++ /dev/null
@@ -1,20 +0,0 @@
-...........ssssssssssss.ss.ss.ss...ssssss.ss......ssssssssssss.ss...ssss [ 66%]
-s.s.ssssss...........................                                    [100%]
-=========================== short test summary info ============================
-SKIPPED [12] tests/ops/rope/test_multi_axis_rope.py:79: NVIDIA CUDA required
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:100: NVIDIA CUDA required
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:110: NVIDIA CUDA required
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:133: NVIDIA CUDA required
-SKIPPED [6] tests/ops/rope/test_multi_axis_rope.py:148: NVIDIA CUDA required
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:158: NVIDIA CUDA required
-SKIPPED [12] tests/ops/rope/test_multi_axis_rope.py:174: NVIDIA CUDA required
-SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:217: native CUDA multi_axis_rope extension required
-SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:230: NVIDIA CUDA required
-SKIPPED [3] tests/ops/rope/test_multi_axis_rope.py:255: NVIDIA CUDA required
-SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:315: NVIDIA CUDA required
-SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:336: requires two NVIDIA GPUs
-SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:355: CUDA context requires CUDA
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:368: NVIDIA CUDA required
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:399: NVIDIA CUDA required
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:427: NVIDIA CUDA required
-57 passed, 52 skipped in 1.70s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-environment.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-environment.json
deleted file mode 100644
index e41fd057f..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu-environment.json
+++ /dev/null
@@ -1,7 +0,0 @@
-{
-  "torch": "2.9.1+cpu",
-  "cuda": null,
-  "package": "/home/dandan/venvs/rlk-rope-cpu/lib/python3.12/site-packages/rl_engine/__init__.py",
-  "reference_forward_backward": true,
-  "passed": true
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-contracts.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-contracts.txt
deleted file mode 100644
index 6f0703516..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-contracts.txt
+++ /dev/null
@@ -1,20 +0,0 @@
-...........ssssssssssss.ss.ss.ss...ssssss.ss......ssssssssssss.ss...ssss [ 66%]
-s.s.ssssss...........................                                    [100%]
-=========================== short test summary info ============================
-SKIPPED [12] tests/ops/rope/test_multi_axis_rope.py:79: NVIDIA CUDA required
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:100: NVIDIA CUDA required
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:110: NVIDIA CUDA required
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:133: NVIDIA CUDA required
-SKIPPED [6] tests/ops/rope/test_multi_axis_rope.py:148: NVIDIA CUDA required
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:158: NVIDIA CUDA required
-SKIPPED [12] tests/ops/rope/test_multi_axis_rope.py:174: NVIDIA CUDA required
-SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:217: native CUDA multi_axis_rope extension required
-SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:230: NVIDIA CUDA required
-SKIPPED [3] tests/ops/rope/test_multi_axis_rope.py:255: NVIDIA CUDA required
-SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:315: NVIDIA CUDA required
-SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:336: requires two NVIDIA GPUs
-SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:355: CUDA context requires CUDA
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:368: NVIDIA CUDA required
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:399: NVIDIA CUDA required
-SKIPPED [2] tests/ops/rope/test_multi_axis_rope.py:427: NVIDIA CUDA required
-57 passed, 52 skipped in 1.72s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-environment.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-environment.json
deleted file mode 100644
index 833e69038..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/cpu310-environment.json
+++ /dev/null
@@ -1,9 +0,0 @@
-{
-  "python": "3.10.22",
-  "torch": "2.9.1+cpu",
-  "cuda": null,
-  "package": "/home/dandan/venvs/rlk-rope-cpu310/lib/python3.10/site-packages/rl_engine",
-  "reference_forward_backward": true,
-  "cpu_wheel_contains_no_cuda_binary": true,
-  "passed": true
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-extension.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-extension.json
deleted file mode 100644
index 420aa5da8..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-extension.json
+++ /dev/null
@@ -1,22 +0,0 @@
-{
-  "profile": "ordinary complete _C; no RL_KERNEL_ROPE_ONLY",
-  "module": "/home/dandan/src/RL-Kernel/rl_engine/_C.cpython-312-x86_64-linux-gnu.so",
-  "binary_sha256": "2661e7188984c7d78acae1ef5186a3e90ec4aa8bb6f457522b23435a1f24f94a",
-  "sources": [
-    "csrc/bindings/ops.cpp",
-    "csrc/cuda/logprob/fused_logp_kernel.cu",
-    "csrc/cuda/logprob/deterministic_logp_kernel.cu",
-    "csrc/cuda/gemm/det_gemm_kernel.cu",
-    "csrc/cuda/norm/rmsnorm.cu",
-    "csrc/cuda/activation/activation.cu",
-    "csrc/cuda/attention/deterministic_attention.cu",
-    "csrc/cuda/rope/multi_axis_rope.cu",
-    "csrc/cuda/collectives/deterministic_collective.cu",
-    "csrc/cuda/attention/prefix_shared_attention.cu"
-  ],
-  "cuda_home": "/usr/local/cuda-12.8",
-  "driver_link_search": "/usr/lib/wsl/lib",
-  "torch": "2.9.1+cu128",
-  "architecture": "12.0",
-  "max_jobs": "2"
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-inplace-build.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-inplace-build.txt
deleted file mode 100644
index d61fb7517..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-inplace-build.txt
+++ /dev/null
@@ -1,20 +0,0 @@
-/home/dandan/src/RL-Kernel/.venv/lib/python3.12/site-packages/setuptools/config/_apply_pyprojecttoml.py:82: SetuptoolsDeprecationWarning: `project.license` as a TOML table is deprecated
-!!
-
-        ********************************************************************************
-        Please use a simple string containing a SPDX expression for `project.license`. You can also use `project.license-files`. (Both options available on setuptools>=77.0.0).
-
-        This deprecation is overdue, please update your project and remove deprecated
-        calls to avoid build errors in the future.
-
-        See https://packaging.python.org/en/latest/guides/writing-pyproject-toml/#license for details.
-        ********************************************************************************
-
-!!
-  corresp(dist, value, root_dir)
-running build_ext
-W1028 18:37:59.459000 9791 .venv/lib/python3.12/site-packages/torch/utils/cpp_extension.py:531] There are no x86_64-linux-gnu-g++ version bounds defined for CUDA version 12.8
-building 'rl_engine._C' extension
-ninja: no work to do.
-x86_64-linux-gnu-g++ -fno-strict-overflow -Wsign-compare -DNDEBUG -g -O2 -Wall -shared -Wl,-O1 -Wl,-Bsymbolic-functions /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/bindings/ops.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/activation/activation.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/attention/deterministic_attention.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/attention/prefix_shared_attention.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/collectives/deterministic_collective.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/gemm/det_gemm_kernel.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/logprob/deterministic_logp_kernel.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/logprob/fused_logp_kernel.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/norm/rmsnorm.o /home/dandan/src/RL-Kernel/build/temp.linux-x86_64-cpython-312/csrc/cuda/rope/multi_axis_rope.o -L/home/dandan/src/RL-Kernel/.venv/lib/python3.12/site-packages/torch/lib -L/usr/local/cuda-12.8/lib64 -L/usr/lib/x86_64-linux-gnu -lc10 -ltorch -ltorch_cpu -ltorch_python -lcudart -lc10_cuda -ltorch_cuda -o build/lib.linux-x86_64-cpython-312/rl_engine/_C.cpython-312-x86_64-linux-gnu.so -Wl,-rpath,$ORIGIN/../torch/lib -lcuda
-copying build/lib.linux-x86_64-cpython-312/rl_engine/_C.cpython-312-x86_64-linux-gnu.so -> rl_engine
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-wheel-build.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-wheel-build.txt
deleted file mode 100644
index 592c254f2..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/full-wheel-build.txt
+++ /dev/null
@@ -1,10 +0,0 @@
-Processing /home/dandan/src/RL-Kernel
-  Preparing metadata (pyproject.toml): started
-  Preparing metadata (pyproject.toml): finished with status 'done'
-Building wheels for collected packages: RL-Kernel
-  Building wheel for RL-Kernel (pyproject.toml): started
-  Building wheel for RL-Kernel (pyproject.toml): still running...
-  Building wheel for RL-Kernel (pyproject.toml): finished with status 'done'
-  Created wheel for RL-Kernel: filename=rl_kernel-0.1.0-cp312-cp312-linux_x86_64.whl size=5020592 sha256=32a39cd9cbb30b87816f6584096238fc2e4b0803eb387a2d3e1e80f3d1b40d8a
-  Stored in directory: /tmp/pip-ephem-wheel-cache-8gtts1_e/wheels/de/2c/1f/c224e9a9cc43d1b39ee96ba5703938f5f1d66da44d8695a43a
-Successfully built RL-Kernel
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/gpu-environment.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/gpu-environment.json
deleted file mode 100644
index e40096583..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/gpu-environment.json
+++ /dev/null
@@ -1,15 +0,0 @@
-{
-  "scope": "Qwen-Image multi_axis_rope operator; ordinary full _C build",
-  "python": "3.12.3",
-  "platform": "Linux-6.18.54.1-microsoft-standard-WSL2-x86_64-with-glibc2.39",
-  "torch": "2.9.1+cu128",
-  "cuda": "12.8",
-  "triton": "3.5.1",
-  "gpu": "NVIDIA GeForce RTX 5080",
-  "capability": [
-    12,
-    0
-  ],
-  "native_library": "/home/dandan/src/RL-Kernel/rl_engine/_C.cpython-312-x86_64-linux-gnu.so",
-  "native_library_sha256": "2661e7188984c7d78acae1ef5186a3e90ec4aa8bb6f457522b23435a1f24f94a"
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/installed-wheel.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/installed-wheel.json
deleted file mode 100644
index 28d3066fc..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/installed-wheel.json
+++ /dev/null
@@ -1,38 +0,0 @@
-{
-  "package": "/home/dandan/src/RL-Kernel/.venv/lib/python3.12/site-packages/rl_engine",
-  "companion_library": null,
-  "full_extension_library": "/home/dandan/src/RL-Kernel/.venv/lib/python3.12/site-packages/rl_engine/_C.cpython-312-x86_64-linux-gnu.so",
-  "native_dispatcher_registered": true,
-  "torch": "2.9.1+cu128",
-  "gpu": "NVIDIA GeForce RTX 5080",
-  "license_present": true,
-  "packed_qkv_stride": [
-    17280,
-    1152,
-    128,
-    1
-  ],
-  "checks": {
-    "cuda_eager": "passed",
-    "cuda_compiled": "passed",
-    "cuda_trace": {
-      "contract": "qwenimage-interleaved-rope-fp32-v1",
-      "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
-      "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
-      "layout": "BSHD, adjacent channel pairs",
-      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-    },
-    "triton_eager": "passed",
-    "triton_compiled": "passed",
-    "triton_trace": {
-      "contract": "qwenimage-interleaved-rope-fp32-v1",
-      "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
-      "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
-      "layout": "BSHD, adjacent channel pairs",
-      "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-      "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-    }
-  },
-  "passed": true
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/layout.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/layout.txt
deleted file mode 100644
index 8c81ecbf5..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/layout.txt
+++ /dev/null
@@ -1,2 +0,0 @@
-.....................................                                    [100%]
-37 passed in 3.81s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/manifest.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/manifest.json
deleted file mode 100644
index 9db29dacc..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/manifest.json
+++ /dev/null
@@ -1,60 +0,0 @@
-{
-  "profile": "Ubuntu 24.04.5 / WSL2 / RTX 5080 SM120 / ordinary complete _C",
-  "numerical_source_commit": "ac17ab87e8d93dabe93dfdc134e9853d47de7af1",
-  "source_sha256": {
-    ".github/workflows/qwenimage-rope.yml": "39e7e02a1f1f75603482e6acce9eb99ba8b5c5260ec6043271714061843651c3",
-    "MANIFEST.in": "87e59e6c40c67e54ae9bcd057957ac7d7823067cb455cbc4dff45b1ec88bdb3c",
-    "benchmarks/operators/rope/benchmark_multi_axis_rope.py": "6961a55803fbd4e566bd060558fb59e658e7f1f704e61ca97f912e5d16b55884",
-    "build_tools/extensions.py": "3723f14a3fb9307889790af15b5e3f360c4c46e12b8fef22423070e9660f70c5",
-    "ci/scripts/run_qwenimage_rope_ws1.sh": "8e1e12c1aad9231f215505a172b3a5416c54e7cbfb8856f5ff00d225be7bfa95",
-    "csrc/bindings/ops.cpp": "10b76a5e236c0dc3715f9703e28c057f404f21f2a9802245d5221c4997b2c3f4",
-    "csrc/cuda/activation/activation.cu": "35ef0cefcaa6167aec7a901fd19a2f6c90a8b9f01c2e786187aacdeee10c4b63",
-    "csrc/cuda/attention/deterministic_attention.cu": "99b44addd61b999a5fb6a04ca227ed8814d454368f361a8fd032e83d4d83a561",
-    "csrc/cuda/attention/prefix_shared_attention.cu": "218ab9647cd274f7a4d8246f8df8e875c5cf431a5a73f252e946b10c74a2d63f",
-    "csrc/cuda/collectives/deterministic_collective.cu": "9e114d5aae7a73a912e4e4afaa23a74126bef5557ba3b54832da0c0cdf9d97a2",
-    "csrc/cuda/gemm/det_gemm_kernel.cu": "08d08b0a32900fd9590d47f04819b34a2d37cbf53996aa5ed1373d88eae29f33",
-    "csrc/cuda/logprob/deterministic_logp_kernel.cu": "271399040a0489a3cabd9e0c06b64fa8b1bac92ab08710fce8922bcb1ff2b04f",
-    "csrc/cuda/logprob/fused_logp_kernel.cu": "ea67bcd307f0c63753397225543d333f58cf3d66a0ae69b643499d1e6729d8e7",
-    "csrc/cuda/norm/rmsnorm.cu": "fbb765563dbe9c063b35f38354321eaf522cf62ba03addf327154395db9683b9",
-    "csrc/cuda/rope/multi_axis_rope.cu": "f5db5cf5d4962053194455d9cde8cf91cf4e824bad7abf72e8410a404bcb0e44",
-    "docs/operators/multi-axis-rope.md": "1fb3859a9a2639aae809ece52a785737e9966098c24a08626b9240dad0d6cbf6",
-    "pyproject.toml": "c2d948784b379a0a828bbf3d8cfd9e0e7398a1ee5694fe8298509514684b2709",
-    "rl_engine/backends/cuda/rope/multi_axis_rope.py": "d779d9b60ce5d1d6651925f723049930d898024fdf0548ce29a2019b84753d65",
-    "rl_engine/backends/shared/triton/rope/_flash_attn_rotary.py": "cf0eebf315830a4178053a71a77e650b58aa85bba1d1a08aff7677350af755d9",
-    "rl_engine/backends/shared/triton/rope/multi_axis_rope.py": "f08ff90627c66e15e9c558f8a68f132434a741a76997b3800bfbf73dc8940e9e",
-    "rl_engine/ops/rope/_registration.py": "add0375d1639444908091d01e8d29781f7a975beffec64b77bfaa81a5b9a0ae3",
-    "rl_engine/ops/rope/multi_axis_rope.py": "07bcff1160f22c1ff6730be731206b2f4a153e2fada26afc93fc884e780568cb",
-    "rl_engine/reference/rope/multi_axis_rope.py": "c9a94730047f51659f7aa0a443e88d6ce7d03d00d6813a9aed548cafbf6347bf",
-    "rl_engine/validation/operators/qwen_image_rope.py": "c3ddfc732f32de3c633de16064810d67202254ea78fd8107d558b471bb15a37f",
-    "tests/ops/rope/test_multi_axis_rope.py": "c9ff3eb6e6e013243fa741f853debb256fdac6c1d3a3820eb7de764a830e240c",
-    "tools/validation/operators/check_multi_axis_rope_install.py": "f295fc8ac5878ac13b6255c2f0be83b66d34ff2ad2f75111bb8b633508a865a7"
-  },
-  "evidence_sha256": {
-    "README.md": "3ad6f8ef514b4fc61d51fb0ce8cb6d420d2180251c9ee4f84b8c931aa3a75094",
-    "benchmark-bfloat16.json": "e9c1e4b1085623d84800656c179e4a0e86d70a0ef7fc9ed92f9e6ccc50f0f491",
-    "cpu-contracts.txt": "9518f53a6efc603bca1c52f160e5430f1006bb2abc9a7ef57694dae6feac5d78",
-    "cpu-environment.json": "4dfc68f3d76af05a050d830840b4a6b8c8fa0a4c12c29d11d632a56a57a2d4c6",
-    "cpu310-contracts.txt": "02f57808b9553cefe6f4dcec06d46a5163157b4290cc0c590b344a1616057852",
-    "cpu310-environment.json": "7130d4c164c98804284cc0f1e07ae6265c35e7498c2fb8285a5e6df87cb010b2",
-    "full-extension.json": "8fae782e3c5e6aae59de6fe6b4e7ae185729e49d2282dbd0cb3b9ca08c0274cd",
-    "full-inplace-build.txt": "04d5ef122650ab305847d58850a6c3b44438eacf97920492e1eb1c5580e9a21e",
-    "full-wheel-build.txt": "a7883dfb1211798c869ede1ee0b6f1dfb32edd7b27b359ee18c6e5ab9fe449d5",
-    "gpu-environment.json": "bf808ab535d90b9b52dfcfb86543809e7d412cb48418677923cad22783bb9b8e",
-    "installed-wheel.json": "c726717106f30336e0a541e4a38d1f1b3858ed5dd262516c86618d4511a43554",
-    "layout.txt": "daef0733010cdad31123cb6b8cb041fe6c626319d527a5d97c88d1b958479039",
-    "memcheck.txt": "1a0c572f8f0d18e3699c847dbb69b345e4fd19200cabda3a39385d588cff0f10",
-    "operator-cuda.json": "fd95922bac2c5921744a8e601e9a267e51b7b7d7d44d94745f53964c2da2feb1",
-    "operator-tests.xml": "0ed98ddf85e34ba87f41465be1db4f417997e82b50cc5f8f5a83bd00f1f1543e",
-    "operator-triton.json": "7791aa1dfa68652dd66e3c332d8f036e9e21851e1df7fd41a090d594e22b06ff",
-    "package.json": "3f75c06fdee58afb1cdd9a2a611b432c8be1960fa640d96263c86b87423fe7bd",
-    "precommit.txt": "7cf90e06cf73194271c176ebc2d147e64cf8f5ed522d8e0e03093a4647f63759",
-    "python-packages.json": "617f8cef5a24f3b2a683e70e0b491e47db6d24a7ca848e5b3e461e87993a7201",
-    "quality.txt": "8826f0e406946a923400d792dd60d0f545b32f9fcffa3a148268c969e30e14c0",
-    "regression.txt": "039cd9813a2ea54bfc5b7750fc0555b54d2bf30b09dbca2eba4a0403578ab9b0",
-    "tests-and-ws1.txt": "68397e27144c9244743dbcff39ffdca8a77a30eaf1e8180b2ca84dfc1216d633",
-    "upstream.json": "04e5472a82477a305b996ccbb420d0f3cc7b39af4053c867d37916a9bf5de51a",
-    "ws1-cuda.json": "7302e11f798816178622393095858871e2dc53ab925509eb130196a40b3c18ad",
-    "ws1-triton.json": "b21f985fbdff7e9b14f689d5eac25c9df5d6d3b0189900085837d5fc400f9d38"
-  },
-  "hash_mode": "SHA256 of recorded Linux file bytes; text is LF"
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/memcheck.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/memcheck.txt
deleted file mode 100644
index 915746141..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/memcheck.txt
+++ /dev/null
@@ -1,4 +0,0 @@
-========= COMPUTE-SANITIZER
-..................................                                       [100%]
-34 passed, 48 deselected in 3.49s
-========= ERROR SUMMARY: 0 errors
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-cuda.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-cuda.json
deleted file mode 100644
index 801312e4a..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-cuda.json
+++ /dev/null
@@ -1,70 +0,0 @@
-{
-  "suite_name": "multi_axis_rope",
-  "total_candidates": 1,
-  "passed_candidates": 1,
-  "pass_rate": 1.0,
-  "passed": true,
-  "candidates": [
-    {
-      "candidate_name": "cuda-multi_axis_rope",
-      "backend": "cuda",
-      "total_outputs": 2,
-      "passed_outputs": 2,
-      "pass_rate": 1.0,
-      "passed": true,
-      "cases": [
-        {
-          "case_name": "multi_axis_rope-torch.bfloat16-2x7x3x128",
-          "dtype": "torch.bfloat16",
-          "op_class": "elementwise",
-          "passed": true,
-          "outputs": [
-            {
-              "output_index": 0,
-              "shape": [
-                2,
-                7,
-                3,
-                128
-              ],
-              "candidate_dtype": "torch.bfloat16",
-              "gold_dtype": "torch.float32",
-              "atol": 0.02,
-              "rtol": 0.016,
-              "max_abs_error": 0.007704973220825195,
-              "mean_abs_error": 0.0006018930580466986,
-              "max_rel_error": 0.003828622167930007,
-              "passed": true,
-              "judgment": "forward_accuracy",
-              "comparison_lhs_role": "bf16_candidate",
-              "comparison_rhs_role": "fp32_reference",
-              "message": ""
-            },
-            {
-              "output_index": 1,
-              "shape": [
-                2,
-                7,
-                3,
-                128
-              ],
-              "candidate_dtype": "torch.bfloat16",
-              "gold_dtype": "torch.bfloat16",
-              "atol": 0.02,
-              "rtol": 0.016,
-              "max_abs_error": 0.0,
-              "mean_abs_error": 0.0,
-              "max_rel_error": 0.0,
-              "passed": true,
-              "judgment": "gradient_accuracy",
-              "comparison_lhs_role": "bf16_candidate",
-              "comparison_rhs_role": "fp32_reference",
-              "message": "gradient:x"
-            }
-          ]
-        }
-      ],
-      "backend_provenance": null
-    }
-  ]
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-tests.xml b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-tests.xml
deleted file mode 100644
index 9e26942b9..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-tests.xml
+++ /dev/null
@@ -1 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?><testsuites name="pytest tests"><testsuite name="pytest" errors="0" failures="0" skipped="1" tests="82" time="17.488" timestamp="2026-10-28T18:40:27.472020+08:00" hostname="DESKTOP-J18U1QU"><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_positions_and_axis_frequencies[2-4]" time="0.009" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_positions_and_axis_frequencies[3-5]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_positions_and_axis_frequencies[64-64]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_positions_and_axis_frequencies[83-83]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_positions_and_axis_frequencies[104-58]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[reference-False-dtype0]" time="1.137" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[reference-False-dtype1]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[reference-False-dtype2]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[reference-True-dtype0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[reference-True-dtype1]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[reference-True-dtype2]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[cuda-False-dtype0]" time="3.286" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[cuda-False-dtype1]" time="0.003" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[cuda-False-dtype2]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[cuda-True-dtype0]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[cuda-True-dtype1]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[cuda-True-dtype2]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[triton-False-dtype0]" time="0.900" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[triton-False-dtype1]" time="0.137" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[triton-False-dtype2]" time="0.144" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[triton-True-dtype0]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[triton-True-dtype1]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_forward_backward_accuracy[triton-True-dtype2]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_adjacent_pair_layout_not_rotate_half[reference]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_adjacent_pair_layout_not_rotate_half[cuda]" time="0.767" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_adjacent_pair_layout_not_rotate_half[triton]" time="0.266" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_batch_position_padding_and_chunk_invariance[reference]" time="0.026" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_batch_position_padding_and_chunk_invariance[cuda]" time="0.097" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_batch_position_padding_and_chunk_invariance[triton]" time="0.352" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_noncontiguous_input_tables_and_gradient[reference]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_noncontiguous_input_tables_and_gradient[cuda]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_noncontiguous_input_tables_and_gradient[triton]" time="0.002" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[reference-0-7]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[reference-2-0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[reference-0-0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[cuda-0-7]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[cuda-2-0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[cuda-0-0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[triton-0-7]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[triton-2-0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_empty_forward_backward[triton-0-0]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_second_derivative[reference]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_second_derivative[cuda]" time="0.931" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_second_derivative[triton]" time="0.292" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[reference-shape]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[reference-dtype]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[reference-table_shape]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[reference-table_dtype]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[reference-table_grad]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[reference-device]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[cuda-shape]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[cuda-dtype]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[cuda-table_shape]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[cuda-table_dtype]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[cuda-table_grad]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[cuda-device]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[triton-shape]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[triton-dtype]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[triton-table_shape]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[triton-table_dtype]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[triton-table_grad]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_inputs[triton-device]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_explicit_backend_and_reference_dispatch" time="0.022" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_cuda_registry_uses_native_when_present" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_cuda_registry_falls_back_to_triton" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_positions[bad0]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_positions[bad1]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_reject_invalid_positions[bad2]" time="0.000" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_cuda_triton_byte_parity_and_graph[dtype0]" time="0.104" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_cuda_triton_byte_parity_and_graph[dtype1]" time="0.107" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_cuda_triton_byte_parity_and_graph[dtype2]" time="0.099" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_native_entrypoint_checks_and_nondefault_stream" time="0.014" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_device_guard" time="0.000"><skipped type="pytest.skip" message="requires two NVIDIA GPUs">/home/dandan/src/RL-Kernel/tests/ops/rope/test_multi_axis_rope.py:336: requires two NVIDIA GPUs</skipped></testcase><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_canonical_tables_ignore_default_device[meta]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_canonical_tables_ignore_default_device[cuda]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_packed_qkv_and_independent_table_strides[reference]" time="0.001" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_packed_qkv_and_independent_table_strides[cuda]" time="0.013" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_packed_qkv_and_independent_table_strides[triton]" time="0.596" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_dispatcher_fake_autograd_and_fullgraph_compile[cuda]" time="2.314" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_dispatcher_fake_autograd_and_fullgraph_compile[triton]" time="0.265" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_standard_ws1_four_judgments[cuda]" time="4.259" /><testcase classname="tests.ops.rope.test_multi_axis_rope" name="test_standard_ws1_four_judgments[triton]" time="0.596" /></testsuite></testsuites>
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-triton.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-triton.json
deleted file mode 100644
index e105fb723..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/operator-triton.json
+++ /dev/null
@@ -1,70 +0,0 @@
-{
-  "suite_name": "multi_axis_rope",
-  "total_candidates": 1,
-  "passed_candidates": 1,
-  "pass_rate": 1.0,
-  "passed": true,
-  "candidates": [
-    {
-      "candidate_name": "triton-multi_axis_rope",
-      "backend": "triton",
-      "total_outputs": 2,
-      "passed_outputs": 2,
-      "pass_rate": 1.0,
-      "passed": true,
-      "cases": [
-        {
-          "case_name": "multi_axis_rope-torch.bfloat16-2x7x3x128",
-          "dtype": "torch.bfloat16",
-          "op_class": "elementwise",
-          "passed": true,
-          "outputs": [
-            {
-              "output_index": 0,
-              "shape": [
-                2,
-                7,
-                3,
-                128
-              ],
-              "candidate_dtype": "torch.bfloat16",
-              "gold_dtype": "torch.float32",
-              "atol": 0.02,
-              "rtol": 0.016,
-              "max_abs_error": 0.007704973220825195,
-              "mean_abs_error": 0.0006018930580466986,
-              "max_rel_error": 0.003828622167930007,
-              "passed": true,
-              "judgment": "forward_accuracy",
-              "comparison_lhs_role": "bf16_candidate",
-              "comparison_rhs_role": "fp32_reference",
-              "message": ""
-            },
-            {
-              "output_index": 1,
-              "shape": [
-                2,
-                7,
-                3,
-                128
-              ],
-              "candidate_dtype": "torch.bfloat16",
-              "gold_dtype": "torch.bfloat16",
-              "atol": 0.02,
-              "rtol": 0.016,
-              "max_abs_error": 0.0,
-              "mean_abs_error": 0.0,
-              "max_rel_error": 0.0,
-              "passed": true,
-              "judgment": "gradient_accuracy",
-              "comparison_lhs_role": "bf16_candidate",
-              "comparison_rhs_role": "fp32_reference",
-              "message": "gradient:x"
-            }
-          ]
-        }
-      ],
-      "backend_provenance": null
-    }
-  ]
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/package.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/package.json
deleted file mode 100644
index 1dce0c941..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/package.json
+++ /dev/null
@@ -1,11 +0,0 @@
-{
-  "profile": "ordinary complete _C wheel; RL_KERNEL_ROPE_ONLY unset",
-  "wheel": "rl_kernel-0.1.0-cp312-cp312-linux_x86_64.whl",
-  "wheel_sha256": "32a39cd9cbb30b87816f6584096238fc2e4b0803eb387a2d3e1e80f3d1b40d8a",
-  "native_member": "rl_engine/_C.cpython-312-x86_64-linux-gnu.so",
-  "native_sha256": "2661e7188984c7d78acae1ef5186a3e90ec4aa8bb6f457522b23435a1f24f94a",
-  "license_present": true,
-  "installed_library": "/home/dandan/src/RL-Kernel/.venv/lib/python3.12/site-packages/rl_engine/_C.cpython-312-x86_64-linux-gnu.so",
-  "installed_library_sha256": "2661e7188984c7d78acae1ef5186a3e90ec4aa8bb6f457522b23435a1f24f94a",
-  "cpu_wheel_contains_no_cuda_binary": true
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/precommit.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/precommit.txt
deleted file mode 100644
index 475a8c0d8..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/precommit.txt
+++ /dev/null
@@ -1,7 +0,0 @@
-trim trailing whitespace.................................................Passed
-fix end of files.........................................................Passed
-check yaml...............................................................Passed
-check for added large files..............................................Passed
-black....................................................................Passed
-isort....................................................................Passed
-flake8...................................................................Passed
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/python-packages.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/python-packages.json
deleted file mode 100644
index 41997f91d..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/python-packages.json
+++ /dev/null
@@ -1,114 +0,0 @@
-{
-  "GitPython": "3.2.1",
-  "Jinja2": "3.1.6",
-  "Markdown": "3.11.2",
-  "MarkupSafe": "3.0.3",
-  "PyYAML": "6.0.3",
-  "Pygments": "2.21.0",
-  "RL-Kernel": "0.1.0",
-  "accelerate": "1.15.0",
-  "annotated-doc": "0.0.5",
-  "annotated-types": "0.8.0",
-  "anyio": "4.15.1",
-  "ast_serialize": "0.13.1",
-  "babel": "2.18.0",
-  "backrefs": "8.1",
-  "black": "24.4.2",
-  "bracex": "3.2",
-  "certifi": "2026.10.23",
-  "cfgv": "3.5.0",
-  "charset-normalizer": "3.5.2",
-  "click": "8.5.0",
-  "colorama": "0.4.6",
-  "distlib": "0.4.4",
-  "filelock": "3.32.3",
-  "flake8": "7.0.0",
-  "fsspec": "2026.7.0",
-  "ghp-import": "2.1.0",
-  "gitdb": "4.0.12",
-  "h11": "0.16.0",
-  "hf-xet": "1.7.0",
-  "httpcore": "1.0.9",
-  "httpx": "0.28.1",
-  "huggingface_hub": "1.33.0",
-  "identify": "2.6.20",
-  "idna": "3.20",
-  "iniconfig": "2.3.1",
-  "isort": "5.13.2",
-  "librt": "0.16.0",
-  "markdown-it-py": "4.2.0",
-  "mccabe": "0.7.0",
-  "mdurl": "0.1.2",
-  "mergedeep": "1.3.4",
-  "mkdocs": "1.6.1",
-  "mkdocs-awesome-nav": "3.3.0",
-  "mkdocs-get-deps": "0.2.2",
-  "mkdocs-git-revision-date-localized-plugin": "1.6.2",
-  "mkdocs-material": "9.8.0",
-  "mkdocs-material-extensions": "1.3.1",
-  "mpmath": "1.3.0",
-  "mypy": "2.4.0",
-  "mypy_extensions": "1.1.0",
-  "natsort": "8.4.0",
-  "networkx": "3.6.1",
-  "ninja": "1.13.2",
-  "nodeenv": "1.11.0",
-  "numpy": "2.5.4",
-  "nvidia-cublas-cu12": "12.8.4.1",
-  "nvidia-cuda-cupti-cu12": "12.8.90",
-  "nvidia-cuda-nvrtc-cu12": "12.8.93",
-  "nvidia-cuda-runtime-cu12": "12.8.90",
-  "nvidia-cudnn-cu12": "9.10.2.21",
-  "nvidia-cufft-cu12": "11.3.3.83",
-  "nvidia-cufile-cu12": "1.13.1.3",
-  "nvidia-curand-cu12": "10.3.9.90",
-  "nvidia-cusolver-cu12": "11.7.3.90",
-  "nvidia-cusparse-cu12": "12.5.8.93",
-  "nvidia-cusparselt-cu12": "0.7.1",
-  "nvidia-nccl-cu12": "2.27.5",
-  "nvidia-nvjitlink-cu12": "12.8.93",
-  "nvidia-nvshmem-cu12": "3.3.20",
-  "nvidia-nvtx-cu12": "12.8.90",
-  "packaging": "26.3",
-  "paginate": "0.5.7",
-  "pathspec": "1.1.1",
-  "pip": "25.3",
-  "platformdirs": "4.13.7",
-  "pluggy": "1.7.0",
-  "pre_commit": "4.7.0",
-  "psutil": "7.2.2",
-  "pycodestyle": "2.11.1",
-  "pydantic": "2.14.1",
-  "pydantic_core": "2.50.1",
-  "pyflakes": "3.2.0",
-  "pymdown-extensions": "12.3",
-  "pytest": "9.2.0",
-  "python-dateutil": "3.0.0",
-  "python-discovery": "1.6.7",
-  "pyyaml_env_tag": "1.1",
-  "regex": "2026.9.29",
-  "requests": "2.34.2",
-  "rich": "15.0.0",
-  "ruff": "0.17.3",
-  "safetensors": "0.8.0",
-  "setuptools": "80.9.0",
-  "shellingham": "1.5.4",
-  "six": "1.17.0",
-  "smmap": "5.0.3",
-  "sympy": "1.14.0",
-  "tabulate": "0.10.0",
-  "tokenizers": "0.22.2",
-  "torch": "2.9.1+cu128",
-  "tqdm": "4.70.1",
-  "transformers": "5.13.1",
-  "triton": "3.5.1",
-  "typer": "0.27.3",
-  "typing-inspection": "0.4.4",
-  "typing_extensions": "4.16.0",
-  "tzdata": "2026.6",
-  "urllib3": "2.8.0",
-  "virtualenv": "21.15.1",
-  "watchdog": "6.0.0",
-  "wcmatch": "11.2",
-  "wheel": "0.45.1"
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/quality.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/quality.txt
deleted file mode 100644
index 8e84b2473..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/quality.txt
+++ /dev/null
@@ -1,34 +0,0 @@
-Local verification; no CUDA numerical/kernel sources changed for this follow-up.
-
-workflow-validation.log
-Workflow YAML, branch coverage and embedded Python syntax validated
-
-isort.log
-
-
-black.log
-All done! ✨ 🍰 ✨
-19 files would be left unchanged.
-
-flake8.log
-
-
-ruff.log
-All checks passed!
-
-docs.log
-INFO    -  Cleaning site directory
-INFO    -  Building documentation to directory: /home/dandan/src/RL-Kernel/site
-INFO    -  The following pages exist in the docs directory, but are not included in the "nav" configuration:
-  - blog/2026-07-08-announcing-rl-kernel-linear-logp-for-vime-zh.md
-  - blog/2026-08-09-dsv4-flash-moe-consistency-roadmap.md
-  - operators/matmul.md
-  - operators/rope.md
-  - usage/cross-config-drift-report.md
-  - usage/cuda-rocm-consistency-audit.md
-  - usage/h100-configurable.md
-  - usage/h100-cp-gradient-validation.md
-  - usage/h100-matrix-validation.md
-  - usage/h100-pcp-validation.md
-  - usage/rocm-sparse-performance.md
-INFO    -  Documentation built in 1.15 seconds
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/regression.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/regression.txt
deleted file mode 100644
index 95d27f812..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/regression.txt
+++ /dev/null
@@ -1,10 +0,0 @@
-........................................................................ [ 35%]
-........................................................................ [ 71%]
-...................................................sssssss               [100%]
-=========================== short test summary info ============================
-SKIPPED [3] tests/ops/rope/test_rope.py:390: rope_apply_ascend is not compiled (requires an Ascend NPU build).
-SKIPPED [1] tests/ops/rope/test_rope.py:406: rope_apply_ascend is not compiled (requires an Ascend NPU build).
-SKIPPED [1] tests/ops/rope/test_rope.py:413: rope_apply_ascend is not compiled (requires an Ascend NPU build).
-SKIPPED [1] tests/ops/rope/test_rope.py:420: rope_apply_ascend is not compiled (requires an Ascend NPU build).
-SKIPPED [1] tests/ops/rope/test_rope.py:433: rope_apply_ascend is not compiled (requires an Ascend NPU build).
-195 passed, 7 skipped in 4.56s
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/tests-and-ws1.txt b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/tests-and-ws1.txt
deleted file mode 100644
index 0b8ce46dd..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/tests-and-ws1.txt
+++ /dev/null
@@ -1,11 +0,0 @@
-{"scope": "Qwen-Image multi_axis_rope operator; ordinary full _C build", "python": "3.12.3", "platform": "Linux-6.18.54.1-microsoft-standard-WSL2-x86_64-with-glibc2.39", "torch": "2.9.1+cu128", "cuda": "12.8", "triton": "3.5.1", "gpu": "NVIDIA GeForce RTX 5080", "capability": [12, 0], "native_library": "/home/dandan/src/RL-Kernel/rl_engine/_C.cpython-312-x86_64-linux-gnu.so", "native_library_sha256": "2661e7188984c7d78acae1ef5186a3e90ec4aa8bb6f457522b23435a1f24f94a"}
-........................................................................ [ 87%]
-s.........                                                               [100%]
-=========================== short test summary info ============================
-SKIPPED [1] tests/ops/rope/test_multi_axis_rope.py:336: requires two NVIDIA GPUs
-81 passed, 1 skipped in 17.49s
-Using the installed production RoPE dispatcher library
-WS1 four judgments backend=cuda: passed=True
-Using the installed production RoPE dispatcher library
-WS1 four judgments backend=triton: passed=True
-Qwen-Image RoPE full-extension operator gate passed
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/upstream.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/upstream.json
deleted file mode 100644
index 032e23098..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/upstream.json
+++ /dev/null
@@ -1,335 +0,0 @@
-{
-  "source_sha256": "0bad11901deae69f05aec01fd464e51e38582d455068f2851ee665036179321f",
-  "cases": [
-    {
-      "grid": [
-        2,
-        3
-      ],
-      "stream": "image",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        2,
-        3
-      ],
-      "stream": "image",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        2,
-        3
-      ],
-      "stream": "image",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        2,
-        3
-      ],
-      "stream": "text",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        2,
-        3
-      ],
-      "stream": "text",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        2,
-        3
-      ],
-      "stream": "text",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        3,
-        5
-      ],
-      "stream": "image",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        3,
-        5
-      ],
-      "stream": "image",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        3,
-        5
-      ],
-      "stream": "image",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        3,
-        5
-      ],
-      "stream": "text",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        3,
-        5
-      ],
-      "stream": "text",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        3,
-        5
-      ],
-      "stream": "text",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        64,
-        64
-      ],
-      "stream": "image",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        64,
-        64
-      ],
-      "stream": "image",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        64,
-        64
-      ],
-      "stream": "image",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        64,
-        64
-      ],
-      "stream": "text",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        64,
-        64
-      ],
-      "stream": "text",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        64,
-        64
-      ],
-      "stream": "text",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        83,
-        83
-      ],
-      "stream": "image",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 2.384185791015625e-07,
-      "dx_max_abs": 2.384185791015625e-07
-    },
-    {
-      "grid": [
-        83,
-        83
-      ],
-      "stream": "image",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        83,
-        83
-      ],
-      "stream": "image",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        83,
-        83
-      ],
-      "stream": "text",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        83,
-        83
-      ],
-      "stream": "text",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        83,
-        83
-      ],
-      "stream": "text",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        104,
-        58
-      ],
-      "stream": "image",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 2.384185791015625e-07,
-      "dx_max_abs": 2.384185791015625e-07
-    },
-    {
-      "grid": [
-        104,
-        58
-      ],
-      "stream": "image",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        104,
-        58
-      ],
-      "stream": "image",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        104,
-        58
-      ],
-      "stream": "text",
-      "dtype": "torch.float32",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        104,
-        58
-      ],
-      "stream": "text",
-      "dtype": "torch.float16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    },
-    {
-      "grid": [
-        104,
-        58
-      ],
-      "stream": "text",
-      "dtype": "torch.bfloat16",
-      "table_bitwise": true,
-      "forward_max_abs": 0.0,
-      "dx_max_abs": 0.0
-    }
-  ]
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-cuda.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-cuda.json
deleted file mode 100644
index 3a3e06daf..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-cuda.json
+++ /dev/null
@@ -1,1962 +0,0 @@
-{
-  "scope": "Qwen-Image multi_axis_rope operator; C1 elementwise and C2 config transforms",
-  "trace": {
-    "contract": "qwenimage-interleaved-rope-fp32-v1",
-    "actual_backend": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
-    "kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope",
-    "layout": "BSHD, adjacent channel pairs",
-    "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-    "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-  },
-  "passed": true,
-  "forward": {
-    "op_name": "multi_axis_rope",
-    "backend_profile": "cuda_bf16",
-    "accuracy_reports": [
-      {
-        "config_id": "BN/full",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/chunked",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/chunked",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004673302173614502,
-            "mean_abs_error": 6.917398422956467e-05,
-            "max_rel_error": 0.003753266530111432,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.00048516690731048584,
-            "mean_abs_error": 7.802648178767413e-05,
-            "max_rel_error": 0.0037793784867972136,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004864335060119629,
-            "mean_abs_error": 7.697803084738553e-05,
-            "max_rel_error": 0.003821910358965397,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.710371912457049e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004673302173614502,
-            "mean_abs_error": 6.917398422956467e-05,
-            "max_rel_error": 0.003753266530111432,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.00048516690731048584,
-            "mean_abs_error": 7.802648178767413e-05,
-            "max_rel_error": 0.0037793784867972136,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004864335060119629,
-            "mean_abs_error": 7.697803084738553e-05,
-            "max_rel_error": 0.003821910358965397,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.710371912457049e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/permuted",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/permuted",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_right",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/padded_right",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_left",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/padded_left",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      }
-    ],
-    "invariance_reports": [
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/chunked",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/chunked"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s0",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s1",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s2",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s3",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/permuted",
-        "transform_kind": "permutation",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/permuted"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_right",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_right"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_left",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_left"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      }
-    ],
-    "logprob_smoke": null,
-    "backend_provenance": {
-      "backend_profile": "cuda_bf16",
-      "requested_backend": "cuda",
-      "actual_backend": "cuda",
-      "execution_dtype": "bfloat16",
-      "accumulation_dtype": "float32",
-      "output_dtype": "bfloat16",
-      "reference_dtype": "float32",
-      "candidate_tf32_enabled": false,
-      "reference_tf32_enabled": false
-    },
-    "candidate_id": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
-    "device": "cuda",
-    "compute_capability": "sm120",
-    "seed": 20260812,
-    "fallback_reason": null,
-    "passed": true,
-    "provenance_valid": true,
-    "metadata_valid": true,
-    "observed_kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope"
-  },
-  "gradient": {
-    "op_name": "multi_axis_rope",
-    "backend_profile": "cuda_bf16",
-    "accuracy_reports": [
-      {
-        "config_id": "BN/full",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/chunked",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/chunked",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.7887988835573196e-05,
-            "mean_abs_error": 3.6821973026235355e-06,
-            "max_rel_error": 1.8164952993392944,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.7113780965446495e-06,
-            "max_rel_error": 0.4316565990447998,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.995925024151802e-05,
-            "mean_abs_error": 4.716822331829462e-06,
-            "max_rel_error": 0.7439075112342834,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.0211638659238815e-05,
-            "mean_abs_error": 4.9605891945248e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.7887988835573196e-05,
-            "mean_abs_error": 3.6821973026235355e-06,
-            "max_rel_error": 1.8164952993392944,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.7113780965446495e-06,
-            "max_rel_error": 0.4316565990447998,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.995925024151802e-05,
-            "mean_abs_error": 4.716822331829462e-06,
-            "max_rel_error": 0.7439075112342834,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.0211638659238815e-05,
-            "mean_abs_error": 4.9605891945248e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/permuted",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/permuted",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_right",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/padded_right",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_left",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/padded_left",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "cuda_bf16",
-          "requested_backend": "cuda",
-          "actual_backend": "cuda",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      }
-    ],
-    "invariance_reports": [
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/chunked",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/chunked"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s0",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s1",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s2",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s3",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/permuted",
-        "transform_kind": "permutation",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/permuted"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_right",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_right"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_left",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_left"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      }
-    ],
-    "singleton_aggregate_reports": [],
-    "backend_provenance": {
-      "backend_profile": "cuda_bf16",
-      "requested_backend": "cuda",
-      "actual_backend": "cuda",
-      "execution_dtype": "bfloat16",
-      "accumulation_dtype": "float32",
-      "output_dtype": "bfloat16",
-      "reference_dtype": "float32",
-      "candidate_tf32_enabled": false,
-      "reference_tf32_enabled": false
-    },
-    "candidate_id": "rlkernel.cuda.qwenimage_multi_axis_rope.v1",
-    "device": "cuda",
-    "compute_capability": "sm120",
-    "seed": 20260812,
-    "fallback_reason": null,
-    "passed": true,
-    "provenance_valid": true,
-    "metadata_valid": true,
-    "loss_reduction": "sum_over_active_tokens_then_optional_mean_by_active_count",
-    "active_token_denominator": 27,
-    "grad_tensor_names": [
-      "dx"
-    ],
-    "first_failing_op": null,
-    "first_failing_tensor": null,
-    "first_failing_config_pair": null,
-    "observed_kernel_id": "torch.ops.rl_kernel_native.multi_axis_rope"
-  }
-}
diff --git a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-triton.json b/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-triton.json
deleted file mode 100644
index 628fcbf70..000000000
--- a/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120/ws1-triton.json
+++ /dev/null
@@ -1,1962 +0,0 @@
-{
-  "scope": "Qwen-Image multi_axis_rope operator; C1 elementwise and C2 config transforms",
-  "trace": {
-    "contract": "qwenimage-interleaved-rope-fp32-v1",
-    "actual_backend": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
-    "kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel",
-    "layout": "BSHD, adjacent channel pairs",
-    "arithmetic": "fp32 separate multiply/add, one output cast, no reduction",
-    "table_revision": "40f63690da440e8c382ab53e67508835deb1ed2f"
-  },
-  "passed": true,
-  "forward": {
-    "op_name": "multi_axis_rope",
-    "backend_profile": "triton_cuda_bf16",
-    "accuracy_reports": [
-      {
-        "config_id": "BN/full",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/chunked",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/chunked",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004673302173614502,
-            "mean_abs_error": 6.917398422956467e-05,
-            "max_rel_error": 0.003753266530111432,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.00048516690731048584,
-            "mean_abs_error": 7.802648178767413e-05,
-            "max_rel_error": 0.0037793784867972136,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004864335060119629,
-            "mean_abs_error": 7.697803084738553e-05,
-            "max_rel_error": 0.003821910358965397,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.710371912457049e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004673302173614502,
-            "mean_abs_error": 6.917398422956467e-05,
-            "max_rel_error": 0.003753266530111432,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.00048516690731048584,
-            "mean_abs_error": 7.802648178767413e-05,
-            "max_rel_error": 0.0037793784867972136,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004864335060119629,
-            "mean_abs_error": 7.697803084738553e-05,
-            "max_rel_error": 0.003821910358965397,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.710371912457049e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/permuted",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/permuted",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_right",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/padded_right",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_left",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/padded_left",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 0.0004878789186477661,
-            "mean_abs_error": 8.054684440139681e-05,
-            "max_rel_error": 0.0038475203327834606,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "forward_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      }
-    ],
-    "invariance_reports": [
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/chunked",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/chunked"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s0",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s1",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s2",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s3",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/permuted",
-        "transform_kind": "permutation",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/permuted"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_right",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_right"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_left",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "output",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_left"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "forward_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      }
-    ],
-    "logprob_smoke": null,
-    "backend_provenance": {
-      "backend_profile": "triton_cuda_bf16",
-      "requested_backend": "triton",
-      "actual_backend": "triton",
-      "execution_dtype": "bfloat16",
-      "accumulation_dtype": "float32",
-      "output_dtype": "bfloat16",
-      "reference_dtype": "float32",
-      "candidate_tf32_enabled": false,
-      "reference_tf32_enabled": false
-    },
-    "candidate_id": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
-    "device": "cuda",
-    "compute_capability": "sm120",
-    "seed": 20260812,
-    "fallback_reason": null,
-    "passed": true,
-    "provenance_valid": true,
-    "metadata_valid": true,
-    "observed_kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel"
-  },
-  "gradient": {
-    "op_name": "multi_axis_rope",
-    "backend_profile": "triton_cuda_bf16",
-    "accuracy_reports": [
-      {
-        "config_id": "BN/full",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/chunked",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/chunked",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.7887988835573196e-05,
-            "mean_abs_error": 3.6821973026235355e-06,
-            "max_rel_error": 1.8164952993392944,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.7113780965446495e-06,
-            "max_rel_error": 0.4316565990447998,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.995925024151802e-05,
-            "mean_abs_error": 4.716822331829462e-06,
-            "max_rel_error": 0.7439075112342834,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/full/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/full/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.0211638659238815e-05,
-            "mean_abs_error": 4.9605891945248e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s0",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s0",
-              "fp32_reference"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.7887988835573196e-05,
-            "mean_abs_error": 3.6821973026235355e-06,
-            "max_rel_error": 1.8164952993392944,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s1",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s1",
-              "fp32_reference"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.7113780965446495e-06,
-            "max_rel_error": 0.4316565990447998,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s2",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s2",
-              "fp32_reference"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 2.995925024151802e-05,
-            "mean_abs_error": 4.716822331829462e-06,
-            "max_rel_error": 0.7439075112342834,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "B1-singleton_aggregate/chunked/s3",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "B1-singleton_aggregate/chunked/s3",
-              "fp32_reference"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.0211638659238815e-05,
-            "mean_abs_error": 4.9605891945248e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/permuted",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/permuted",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_right",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/padded_right",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      },
-      {
-        "config_id": "BN/padded_left",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/padded_left",
-              "fp32_reference"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "float32",
-            "max_abs_error": 3.213249146938324e-05,
-            "mean_abs_error": 4.699562850873917e-06,
-            "max_rel_error": 5.814913272857666,
-            "atol": 0.02,
-            "rtol": 0.016,
-            "passed": true,
-            "judgment": "gradient_accuracy",
-            "comparison_lhs_role": "bf16_candidate",
-            "comparison_rhs_role": "fp32_reference"
-          }
-        ],
-        "passed": true,
-        "backend_provenance": {
-          "backend_profile": "triton_cuda_bf16",
-          "requested_backend": "triton",
-          "actual_backend": "triton",
-          "execution_dtype": "bfloat16",
-          "accumulation_dtype": "float32",
-          "output_dtype": "bfloat16",
-          "reference_dtype": "float32",
-          "candidate_tf32_enabled": false,
-          "reference_tf32_enabled": false
-        }
-      }
-    ],
-    "invariance_reports": [
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/chunked",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/chunked"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s0",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s1",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s2",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/full/s3",
-        "transform_kind": "batch_size",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/full/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s0",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s0"
-            ],
-            "shape": [
-              3,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s1",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s1"
-            ],
-            "shape": [
-              8,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s2",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s2"
-            ],
-            "shape": [
-              5,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "B1-singleton_aggregate/chunked/s3",
-        "transform_kind": "chunk",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "B1-singleton_aggregate/chunked/s3"
-            ],
-            "shape": [
-              11,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/permuted",
-        "transform_kind": "permutation",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/permuted"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_right",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_right"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      },
-      {
-        "canonical_config_id": "BN/full",
-        "transformed_config_id": "BN/padded_left",
-        "transform_kind": "padding",
-        "op_class": "elementwise",
-        "dtype": "bfloat16",
-        "backend_profile": "triton_cuda_bf16",
-        "details": [
-          {
-            "tensor_name": "dx",
-            "config_pair": [
-              "BN/full",
-              "BN/padded_left"
-            ],
-            "shape": [
-              27,
-              3,
-              128
-            ],
-            "dtype": "bfloat16",
-            "max_abs_error": 0.0,
-            "mean_abs_error": 0.0,
-            "max_rel_error": 0.0,
-            "atol": 0.0,
-            "rtol": 0.0,
-            "passed": true,
-            "judgment": "gradient_invariance",
-            "comparison_lhs_role": "transformed_config",
-            "comparison_rhs_role": "canonical_config"
-          }
-        ],
-        "passed": true
-      }
-    ],
-    "singleton_aggregate_reports": [],
-    "backend_provenance": {
-      "backend_profile": "triton_cuda_bf16",
-      "requested_backend": "triton",
-      "actual_backend": "triton",
-      "execution_dtype": "bfloat16",
-      "accumulation_dtype": "float32",
-      "output_dtype": "bfloat16",
-      "reference_dtype": "float32",
-      "candidate_tf32_enabled": false,
-      "reference_tf32_enabled": false
-    },
-    "candidate_id": "rlkernel.triton.qwenimage_multi_axis_rope.v1",
-    "device": "cuda",
-    "compute_capability": "sm120",
-    "seed": 20260812,
-    "fallback_reason": null,
-    "passed": true,
-    "provenance_valid": true,
-    "metadata_valid": true,
-    "loss_reduction": "sum_over_active_tokens_then_optional_mean_by_active_count",
-    "active_token_denominator": 27,
-    "grad_tensor_names": [
-      "dx"
-    ],
-    "first_failing_op": null,
-    "first_failing_tensor": null,
-    "first_failing_config_pair": null,
-    "observed_kernel_id": "rl_engine.backends.shared.triton.rope._flash_attn_rotary.rotary_kernel"
-  }
-}

From 9e2e50fcd58e426d5fb67f1e1a2b79ef2ceaa5ef Mon Sep 17 00:00:00 2001
From: dandan20031031 <2425926169@qq.com>
Date: Wed, 28 Oct 2026 23:35:02 +0800
Subject: [PATCH 5/5] test(rope): align validation fixtures and cover review
 findings

Signed-off-by: dandan20031031 <2425926169@qq.com>
---
 csrc/cuda/rope/multi_axis_rope.cu             | 13 ++--
 docs/operators/multi-axis-rope.md             | 11 +++
 .../qwenimage_multi_axis_rope_sm120.md        | 39 +++++++++-
 rl_engine/ops/rope/_registration.py           |  3 +
 .../validation/operators/gradient_adapters.py | 16 +++-
 .../validation/operators/operator_inputs.py   |  5 +-
 rl_engine/validation/operators/rope_inputs.py | 21 +++++
 tests/ops/rope/test_multi_axis_rope.py        | 78 ++++++++++++++++++-
 .../operators/test_operator_inputs.py         | 65 ++++++++++++++++
 9 files changed, 236 insertions(+), 15 deletions(-)
 create mode 100644 rl_engine/validation/operators/rope_inputs.py

diff --git a/csrc/cuda/rope/multi_axis_rope.cu b/csrc/cuda/rope/multi_axis_rope.cu
index 3de49caba..27459c5fb 100644
--- a/csrc/cuda/rope/multi_axis_rope.cu
+++ b/csrc/cuda/rope/multi_axis_rope.cu
@@ -9,6 +9,9 @@
 #include <algorithm>
 
 namespace {
+constexpr int64_t kHeadDim = 128;
+constexpr int64_t kPairsPerHead = kHeadDim / 2;
+
 struct Strides {
   int64_t xb, xs, xh, xd;
   int64_t cb, cs, cd, sb, ss, sd;
@@ -21,8 +24,8 @@ __global__ void rotate_pairs(const scalar_t* x, const float* cos, const float* s
   const int64_t step = static_cast<int64_t>(blockDim.x) * gridDim.x;
   for (int64_t i = static_cast<int64_t>(blockIdx.x) * blockDim.x + threadIdx.x;
        i < n; i += step) {
-    const int64_t pair = i % 64;
-    const int64_t row = i / 64;
+    const int64_t pair = i % kPairsPerHead;
+    const int64_t row = i / kPairsPerHead;
     const int64_t head = row % heads;
     const int64_t token = row / heads;
     const int64_t batch = token / seq;
@@ -43,15 +46,15 @@ at::Tensor multi_axis_rope_apply(const at::Tensor& x, const at::Tensor& cos,
                                 const at::Tensor& sin, bool transpose) {
   TORCH_CHECK(x.is_cuda() && cos.is_cuda() && sin.is_cuda(), "multi_axis_rope: CUDA tensors required");
   TORCH_CHECK(x.device() == cos.device() && x.device() == sin.device(), "multi_axis_rope: device mismatch");
-  TORCH_CHECK(x.dim() == 4 && x.size(3) == 128 && x.size(2) > 0,
+  TORCH_CHECK(x.dim() == 4 && x.size(3) == kHeadDim && x.size(2) > 0,
               "multi_axis_rope: x must be [B, S, H, 128], H > 0");
   TORCH_CHECK(x.scalar_type() == at::kFloat || x.scalar_type() == at::kHalf ||
               x.scalar_type() == at::kBFloat16, "multi_axis_rope: unsupported input dtype");
   TORCH_CHECK(cos.scalar_type() == at::kFloat && sin.scalar_type() == at::kFloat,
               "multi_axis_rope: tables must be fp32");
   TORCH_CHECK(cos.sizes() == sin.sizes(), "multi_axis_rope: table shape mismatch");
-  TORCH_CHECK((cos.dim() == 2 && cos.size(0) == x.size(1) && cos.size(1) == 64) ||
-              (cos.dim() == 3 && cos.size(0) == x.size(0) && cos.size(1) == x.size(1) && cos.size(2) == 64),
+  TORCH_CHECK((cos.dim() == 2 && cos.size(0) == x.size(1) && cos.size(1) == kPairsPerHead) ||
+              (cos.dim() == 3 && cos.size(0) == x.size(0) && cos.size(1) == x.size(1) && cos.size(2) == kPairsPerHead),
               "multi_axis_rope: tables must be [S, 64] or [B, S, 64]");
   TORCH_CHECK(!cos.requires_grad() && !sin.requires_grad(), "multi_axis_rope: table gradients unsupported");
   const c10::cuda::CUDAGuard guard(x.device());
diff --git a/docs/operators/multi-axis-rope.md b/docs/operators/multi-axis-rope.md
index 16899a755..a7c8f6495 100644
--- a/docs/operators/multi-axis-rope.md
+++ b/docs/operators/multi-axis-rope.md
@@ -172,6 +172,17 @@ with C1's elementwise tolerance profile and C2's logical configuration transform
 This gate qualifies the Qwen-Image operator and does not certify the Dense Qwen3
 chain, distributed execution, or model assembly.
 
+The standard CLI and WS1 forward/gradient adapters share a centered image-grid
+prefix from `qwen_image_positions`. The square grid contains the full logical
+sample; a nonsquare sequence length uses its row-major prefix. WS1 selects rows
+by `(sample_id, token_position)` from that sample's grid, so physical chunking,
+reordering and padding cannot change active-token coordinates. Padding uses
+zero coordinates. These are operator fixtures, not a model's image/text assembly.
+
+Backward provenance is recorded inside the opaque custom-op implementation,
+outside fake/AOT tracing. Its Python `execution_count` is not a CUDA Graph replay
+counter: replay executes the captured GPU kernels without re-entering Python.
+
 For a complete native package installation on Linux, run the dedicated gate:
 
 ```bash
diff --git a/reports/experiments/qwenimage_multi_axis_rope_sm120.md b/reports/experiments/qwenimage_multi_axis_rope_sm120.md
index a7d82f65b..a827bb42e 100644
--- a/reports/experiments/qwenimage_multi_axis_rope_sm120.md
+++ b/reports/experiments/qwenimage_multi_axis_rope_sm120.md
@@ -1,7 +1,7 @@
 # Qwen-Image multi-axis RoPE: SM120 validation
 
 Recorded 2026-10-28 for [PR #623](https://github.com/RL-Align/RL-Kernel/pull/623).
-The numerical/native implementation is based on `ac17ab87e8d93dabe93dfdc134e9853d47de7af1`.
+The rotation arithmetic is based on `ac17ab87e8d93dabe93dfdc134e9853d47de7af1`.
 The [operator documentation](../../docs/operators/multi-axis-rope.md) defines the
 contract, upstream reuse, API and detailed reproduction commands.
 
@@ -25,7 +25,7 @@ The normal full wheel was built with `RL_KERNEL_ROPE_ONLY`, forced SM90 and
 fast-math flags unset. Linking existing CUDA driver calls used
 `LIBRARY_PATH=/usr/lib/wsl/lib`; the Windows driver supplies CUDA driver access.
 
-## Recorded results
+## Initial qualification
 
 | Check | WSL2 Linux | Windows |
 | --- | --- | --- |
@@ -51,6 +51,37 @@ The five historical Windows layout failures also reproduce on its clean base
 placeholders. The later Linux run preserved all 27 tracked symlinks and passed.
 The earlier Windows logs remain unchanged in the archive.
 
+## Automated-review follow-up
+
+The standard CLI and WS1 adapters now share a centered image-grid prefix,
+selected by full logical sample length and token identity. Eleven CPU regression
+cases cover empty/square/nonsquare lengths and packed/permuted/chunked/padded
+layouts. Before the change, ten of those cases failed; all eleven now pass.
+Additional tests pin non-CUDA reference routing, library-load error propagation,
+shared-table broadcasting, and head-dimension stride 2 or outermost storage.
+Native shape checks and indexing share named dimension constants; the rotation
+arithmetic is unchanged. Backward telemetry documents its graph-replay limits.
+
+The following are new local WSL2 runs of the updated implementation and fixtures:
+
+| Check | Result |
+| --- | --- |
+| Rebuilt ordinary complete CUDA wheel; fresh installed-package eager/fullgraph forward/input gradient | Passed for both GPU backends |
+| Public operator suite | 105 passed, 1 two-GPU guard skip |
+| CUDA/Triton WS1 accuracy/invariance and ordinary CLI checks | Both passed with explicit backend provenance |
+| Related regressions, including the eleven new input cases | 206 passed, 7 Ascend skips |
+| Repository layout with native extension enforcement | 37 passed |
+| CPU-only PyTorch 2.9.1+cpu, Python 3.10.22 and 3.12.3 | Each: 80 passed, 64 expected GPU skips; installed reference forward/backward passed |
+| Expanded Compute Sanitizer subset, including the new stride cases | 52 passed, zero errors |
+| Changed-file checks and strict docs | Passed locally |
+
+The full wheel contains the shared validation helper and the FlashAttention BSD
+license. Wheel SHA256:
+`9feb8200b948021c268c933f93efcd424b6b68d6c55bc24c3004d6336deb3a13`.
+Non-CUDA routing tests exercise candidate selection with CPU tensors; they do
+not certify NPU/ROCm/MUSA hardware. The benchmark below retains its initial-run
+measurements; no new timing comparison was performed for this follow-up.
+
 ## Standalone benchmark
 
 WSL2 BF16, batch 1, 24 heads, D=128. CUDA Graph medians in ms/call;
@@ -115,8 +146,8 @@ source/binary SHA256 manifests remain available at immutable commit
 - [WSL2 complete-extension evidence](https://github.com/dandan20031031/RL-Kernel/tree/84751473006401c19c1c8c17844e8ef8b1c44375/reports/experiments/qwenimage_multi_axis_rope_wsl2_sm120)
 - [Historical Windows operator-profile evidence](https://github.com/dandan20031031/RL-Kernel/tree/84751473006401c19c1c8c17844e8ef8b1c44375/reports/experiments/qwenimage_multi_axis_rope_v2_sm120)
 
-This summary preserves those recorded results; consolidating reports does not
-rerun the numerical suites. Bare-metal Linux, A100/H100, two-GPU guards,
+Those archives identify the initial runs; the follow-up table records subsequent
+reruns. Bare-metal Linux, A100/H100, two-GPU guards,
 QK-norm fusion, model/block assembly and trajectories remain unqualified.
 Building the full library does not certify every native operator or Dense Qwen3.
 No distributed or cross-version/device byte-equality claim is made.
diff --git a/rl_engine/ops/rope/_registration.py b/rl_engine/ops/rope/_registration.py
index 343d41a9c..33798a03e 100644
--- a/rl_engine/ops/rope/_registration.py
+++ b/rl_engine/ops/rope/_registration.py
@@ -15,6 +15,9 @@ def rotation(
         validate_inputs(x, cos, sin)
         result = implementation(x, cos, sin, transpose)
         if transpose:
+            # Keep telemetry inside the opaque custom op; fake/AOT tracing
+            # must not execute this Python lock. CUDA Graph replay runs only
+            # captured kernels, so execution_count is not a replay counter.
             record_backward(
                 "multi_axis_rope",
                 kernel_id=kernel_id,
diff --git a/rl_engine/validation/operators/gradient_adapters.py b/rl_engine/validation/operators/gradient_adapters.py
index 6bc6e8104..48da77045 100644
--- a/rl_engine/validation/operators/gradient_adapters.py
+++ b/rl_engine/validation/operators/gradient_adapters.py
@@ -32,6 +32,7 @@
     MissingBackwardError,
 )
 from rl_engine.validation.operators.operator_specs import OP_SPECS, _load_object
+from rl_engine.validation.operators.rope_inputs import qwen_image_grid_prefix
 
 REPO_ROOT = Path(__file__).resolve().parents[3]
 
@@ -733,8 +734,19 @@ def _row_inputs(
         if head_dim != 128:
             raise ValueError("multi_axis_rope gradient adapter requires head_dim=128")
         rows = _stack_rows(keys, leading, (n_heads, head_dim), device=device, dtype=dtype)
-        positions = _row_positions(keys, device=torch.device("cpu"))
-        coords = torch.stack((positions % 7, positions % 127 - 63, positions % 61 - 30), dim=-1)
+        # Use the complete logical sample, not this physical call's rows, to
+        # keep grid centers stable under chunks, permutations and padding.
+        lengths: dict[str, int] = {}
+        for sample_id, position in tokens:
+            lengths[sample_id] = max(lengths.get(sample_id, 0), position + 1)
+        grids: dict[str, torch.Tensor] = {}
+        coords = torch.zeros((len(keys), 3), device="cpu", dtype=torch.int64)
+        for row, key in enumerate(keys):
+            if key is not None:
+                sample_id, position = key
+                if sample_id not in grids:
+                    grids[sample_id] = qwen_image_grid_prefix(lengths[sample_id])
+                coords[row] = grids[sample_id][position]
         cos, sin = build_multi_axis_cos_sin(coords, device=device)
         return {"x": rows.unsqueeze(0), "cos": cos, "sin": sin}
     if op_name == "rope":
diff --git a/rl_engine/validation/operators/operator_inputs.py b/rl_engine/validation/operators/operator_inputs.py
index 137d8d69b..0306a683d 100644
--- a/rl_engine/validation/operators/operator_inputs.py
+++ b/rl_engine/validation/operators/operator_inputs.py
@@ -8,6 +8,8 @@
 
 import torch
 
+from rl_engine.validation.operators.rope_inputs import qwen_image_grid_prefix
+
 DEFAULT_HIDDEN = 4096
 DEFAULT_N_HEADS = 32
 DEFAULT_N_KV_HEADS = 8
@@ -259,8 +261,7 @@ def _make_multi_axis_rope_inputs(args, dtype, device):
     if _arg_int(args, "head_dim", 128) != 128:
         raise ValueError("multi_axis_rope requires head_dim=128")
     heads = _arg_int(args, "n_heads", 24)
-    row = torch.arange(seq, device="cpu", dtype=torch.int64)
-    positions = torch.stack((row % 7, (row // 127) % 4096, row % 127 - 63), dim=-1)
+    positions = qwen_image_grid_prefix(seq)
     cos, sin = build_multi_axis_cos_sin(positions, device=device)
     return {
         "x": _floating_tensor((batch, seq, heads, 128), args, dtype, device, offset=0),
diff --git a/rl_engine/validation/operators/rope_inputs.py b/rl_engine/validation/operators/rope_inputs.py
new file mode 100644
index 000000000..5e22f977e
--- /dev/null
+++ b/rl_engine/validation/operators/rope_inputs.py
@@ -0,0 +1,21 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Shared Qwen-Image coordinates for operator and logical-layout validation."""
+
+import math
+
+import torch
+
+from rl_engine.ops.rope.multi_axis_rope import qwen_image_positions
+
+
+def qwen_image_grid_prefix(sequence_length: int) -> torch.Tensor:
+    """Use a centered square-grid prefix for a complete logical sequence.
+
+    CLI sequence lengths need not be square. Grid size depends on the full
+    sample length, never the physical batch, padding or current chunk length.
+    """
+    if sequence_length < 0:
+        raise ValueError("sequence_length must be nonnegative")
+    side = math.isqrt(max(0, sequence_length - 1)) + 1
+    image, _ = qwen_image_positions(side, side, 0)
+    return image[:sequence_length]
diff --git a/tests/ops/rope/test_multi_axis_rope.py b/tests/ops/rope/test_multi_axis_rope.py
index 6cbcb2b85..9f30c35b0 100644
--- a/tests/ops/rope/test_multi_axis_rope.py
+++ b/tests/ops/rope/test_multi_axis_rope.py
@@ -3,6 +3,7 @@
 
 import gc
 import math
+from types import SimpleNamespace
 
 import pytest
 import torch
@@ -86,12 +87,15 @@ def test_forward_backward_accuracy(implementation, dtype, batched):
     result = op(x, cos, sin)
     dx = torch.autograd.grad(result, x, grad)[0]
     tolerance = {torch.float32: 8e-7, torch.float16: 0.002, torch.bfloat16: 0.02}[dtype]
+    # Make the oracle's batch axis explicit when the candidate reads shared tables.
+    c = cos if batched else cos.unsqueeze(0).expand(x.shape[0], -1, -1)
+    s = sin if batched else sin.unsqueeze(0).expand(x.shape[0], -1, -1)
     torch.testing.assert_close(
-        result.double(), complex_reference(x, cos, sin), atol=tolerance, rtol=tolerance
+        result.double(), complex_reference(x, c, s), atol=tolerance, rtol=tolerance
     )
     torch.testing.assert_close(
         dx.double(),
-        complex_reference(grad, cos, sin, transpose=True),
+        complex_reference(grad, c, s, transpose=True),
         atol=tolerance,
         rtol=tolerance,
     )
@@ -145,6 +149,33 @@ def test_noncontiguous_input_tables_and_gradient(implementation):
     assert_bits(dx, dd)
 
 
+@pytest.mark.parametrize("dtype", [torch.float32, torch.float16, torch.bfloat16])
+@pytest.mark.parametrize("layout", ["head_dim_stride_2", "head_dim_outermost"])
+def test_arbitrary_head_dimension_stride_and_gradient(implementation, dtype, layout):
+    op, device = implementation
+    _, cos, sin = inputs(device, dtype)
+    if layout == "head_dim_stride_2":
+        x = torch.randn(2, 7, 3, 256, device=device, dtype=dtype)[..., ::2]
+        assert x.stride(-1) == 2
+    else:
+        x = torch.randn(128, 2, 7, 3, device=device, dtype=dtype).permute(1, 2, 3, 0)
+        assert x.stride(-1) > max(x.stride()[:-1])
+    x.requires_grad_()
+    grad = torch.randn(2, 7, 3, 256, device=device, dtype=dtype)[..., ::2]
+    result = op(x, cos, sin)
+    dx = torch.autograd.grad(result, x, grad)[0]
+    tolerance = {torch.float32: 8e-7, torch.float16: 0.002, torch.bfloat16: 0.02}[dtype]
+    torch.testing.assert_close(
+        result.double(), complex_reference(x, cos, sin), atol=tolerance, rtol=tolerance
+    )
+    torch.testing.assert_close(
+        dx.double(),
+        complex_reference(grad, cos, sin, transpose=True),
+        atol=tolerance,
+        rtol=tolerance,
+    )
+
+
 @pytest.mark.parametrize("batch,seq", [(0, 7), (2, 0), (0, 0)])
 def test_empty_forward_backward(implementation, batch, seq):
     op, device = implementation
@@ -209,6 +240,49 @@ def test_explicit_backend_and_reference_dispatch(monkeypatch):
     assert_bits(multi_axis_rope(x, cos, sin), op(x, cos, sin))
 
 
+@pytest.mark.parametrize("platform", ["cpu", "npu", "rocm", "musa"])
+def test_non_cuda_registry_keeps_multi_axis_reference(platform, monkeypatch):
+    """Exercise candidate selection on CPU; no accelerator execution is certified."""
+    from rl_engine.runtime.registry import KernelRegistry, OpBackend
+
+    registry = KernelRegistry()
+    assert registry._priority_map[platform]["multi_axis_rope"] == [
+        OpBackend.PYTORCH_MULTI_AXIS_ROPE
+    ]
+    monkeypatch.setattr(registry, "_platform", lambda: platform)
+    op = registry.get_op("multi_axis_rope")
+    assert op.trace()["actual_backend"] == "rlkernel.reference.qwenimage_multi_axis_rope.v1"
+    x, cos, sin = inputs("cpu")
+    assert_bits(op(x, cos, sin), get_multi_axis_rope_op("reference")(x, cos, sin))
+
+
+@pytest.mark.parametrize("exception_type", [OSError, RuntimeError])
+def test_operator_wheel_load_error_propagates(exception_type, monkeypatch):
+    from rl_engine.backends.cuda.rope import multi_axis_rope as cuda_module
+
+    error = exception_type("probe library load failure")
+
+    def fail_load(path):
+        raise error
+
+    monkeypatch.setattr(
+        cuda_module,
+        "torch",
+        SimpleNamespace(
+            version=SimpleNamespace(hip=None),
+            ops=SimpleNamespace(rl_kernel_native=SimpleNamespace(), load_library=fail_load),
+        ),
+    )
+    monkeypatch.setattr(
+        cuda_module.importlib.util,
+        "find_spec",
+        lambda name: SimpleNamespace(origin="probe.so"),
+    )
+    with pytest.raises(exception_type, match="probe library load failure") as caught:
+        cuda_module.native_available()
+    assert caught.value is error
+
+
 def test_cuda_registry_uses_native_when_present():
     from rl_engine.backends.cuda.rope.multi_axis_rope import native_available
     from rl_engine.runtime.registry import KernelRegistry
diff --git a/tests/validation/operators/test_operator_inputs.py b/tests/validation/operators/test_operator_inputs.py
index b2c054b18..0416d1540 100644
--- a/tests/validation/operators/test_operator_inputs.py
+++ b/tests/validation/operators/test_operator_inputs.py
@@ -8,6 +8,9 @@
 import pytest
 import torch
 
+from rl_engine.config.workload import LogicalSample
+from rl_engine.ops.rope.multi_axis_rope import build_multi_axis_cos_sin, qwen_image_positions
+from rl_engine.validation.operators.gradient_adapters import _row_inputs
 from rl_engine.validation.operators.operator_inputs import make_operator_inputs, operator_shape_name
 from rl_engine.validation.operators.operator_specs import (
     make_candidate,
@@ -121,6 +124,68 @@ def test_multi_axis_rope_operator_spec_and_seeded_inputs():
         )
 
 
+@pytest.mark.parametrize(
+    "seq,side", [(0, 1), (1, 1), (7, 3), (15, 4), (16, 4), (4096, 64), (6889, 83)]
+)
+def test_multi_axis_rope_validation_uses_centered_image_grid_prefix(seq, side):
+    inputs = make_operator_inputs(
+        "multi_axis_rope", _args(seq=seq, n_heads=1), torch.float32, torch.device("cpu")
+    )
+    image, _ = qwen_image_positions(side, side, 0)
+    cos, sin = build_multi_axis_cos_sin(image[:seq])
+    assert torch.equal(inputs["cos"], cos)
+    assert torch.equal(inputs["sin"], sin)
+
+
+@pytest.mark.parametrize("layout", ["packed", "permuted", "chunked", "padded"])
+def test_multi_axis_rope_validation_preserves_logical_tables_across_layouts(layout):
+    samples = (
+        LogicalSample("short", tuple(range(7)), prompt_len=1, seq_len=7),
+        LogicalSample("long", tuple(range(15)), prompt_len=2, seq_len=15),
+    )
+    tokens = {
+        (token.sample_id, token.token_position): token
+        for sample in samples
+        for token in sample.tokens()
+    }
+    expected = {
+        sample.sample_id: make_operator_inputs(
+            "multi_axis_rope",
+            _args(seq=sample.seq_len, n_heads=1),
+            torch.float32,
+            torch.device("cpu"),
+        )
+        for sample in samples
+    }
+    keys = list(tokens)
+    if layout == "permuted":
+        keys.reverse()
+    elif layout == "chunked":
+        keys = [("short", 5), ("short", 6), ("long", 12)]
+    elif layout == "padded":
+        keys = [None, ("long", 4), None, ("short", 6), ("short", 0), None]
+    inputs = _row_inputs(
+        "multi_axis_rope",
+        keys,
+        tokens,
+        {},
+        device=torch.device("cpu"),
+        dtype=torch.float32,
+        hidden=1,
+        vocab_size=17,
+        n_heads=1,
+        head_dim=128,
+    )
+    for row, key in enumerate(keys):
+        if key is None:
+            assert torch.equal(inputs["cos"][row], torch.ones(64))
+            assert torch.equal(inputs["sin"][row], torch.zeros(64))
+        else:
+            sample_id, position = key
+            for name in ("cos", "sin"):
+                assert torch.equal(inputs[name][row], expected[sample_id][name][position])
+
+
 def test_constant_linear_logp_inputs_match_operator_contract():
     args = _args(input_mode="constant", constant_value=0.5, token_value=3)
     inputs = make_operator_inputs("linear_logp", args, torch.float32, torch.device("cpu"))