[WS1][CUDA][MiniMax-H3] runtime modes, ablation overlay, Diffusers adapter, training provider, LoRA gate (RFC #420 stack 6/7) by fusheng-ji · Pull Request #574 · RL-Align/RL-Kernel

fusheng-ji · 2026-10-22T16:58:21Z

RFC #420 H3 stack 6/7 (base test-h3, merge in this order): #586 -> #587 -> #588 -> #567 -> #568 -> #569 -> #570 -> #571 -> this PR (#574) -> #575. This PR merges after #571; until then GitHub also shows the parent's commits. This PR's own diff.

What

h3_runtime_modes: RL_KERNEL_MODE strict / audit / auto / off decides whether the RL-Kernel graph is bound; strict raises instead of falling back, audit reports the fallback; the case id is R/R (strict, audit) or P/P (auto, off).
h3_ablation_overlay: AblationOverlay.parse("node:attention,boundary:token_refiner") keeps selected block nodes or boundary modules on provider arithmetic inside a strict run; the run is labelled ablation(...), never R/R; unknown names are rejected. full_model_execution gains native_boundaries for this.
diffusers_runtime_adapter_h3: H3DiffusersAdapter binds MiniMaxH3Transformer3DModel for the duration of a with adapter.active(): block, forward and backward (so checkpoint recomputation sees the same operators), and reports what was bound.
h3_training_provider: H3TrainingProvider gives the actor side a step() scope plus a per-sample transition.
h3_lora_gradient_gate: a LoRA over a strict projection trains bitwise repeatably, forward rows do not depend on batching, the base stays frozen, and the gradients are FP64-close.
transition.py (used by the training provider, not a separate RFC row): the cps, noise_level=0.8 step with layout-invariant per-sample noise.

Prior art & reuse decision

Existing	Used here
`rl_engine/runtime/policy.py` (`RL_KERNEL_MODE`, #361 pattern)	reused for the mode policy
`AblationOverlay` pattern of #230	extended with H3 node and boundary names
verl-omni `FlowMatchSDEDiscreteScheduler`	the transition is checked bitwise against its `sample_previous_step(sde_type="cps")`

Results (H100 / Hopper)

Measured on NVIDIA H100 (93 GiB, sm90), CUDA 13.0, torch 2.13.0+cu130, pinned MiniMax-H3@42ed227 weights, at head af3dcaef. Runner: tools/validation/models/h3_runtime_report.py. Data and figure on the fork's pr_assets branch (report.json, figure.png, environment.json).

LoRA gradient gate (3 shapes, N x K x T up to 5120 x 5120 x 33): bitwise repeatable, forward rows independent of batching, base frozen in all three. Relative error to FP64 autograd: dA 4.3e-7 .. 1.3e-6, dB 3.9e-7 .. 1.3e-6, dX 1.7e-3 (BF16 input rounding), forward 1.7e-3.

CPS transition (B x tokens x channels: 3x257x96, 4x4096x96, 2x8192x32): prev_sample and prev_sample_mean are bitwise equal to the scheduler formula and, with the verl-omni FlowMatchSDEDiscreteScheduler (cps, noise_level=0.8), bitwise equal in the mean. Log-prob max relative difference to the scheduler's torch.mean: 5.8e-6 / 1.5e-4 / 7.5e-5 (fixed-order GEMM reduction vs torch reduction, FP32 rounding). Each sample is independent of the batch (bitwise) and scoring a stored rollout reproduces the sampling log-prob (bitwise) in all three.

Runtime modes on the 33B model (forward + backward through H3TrainingProvider, request sizes 12, 925, 4253 and 8349 tokens; relative difference to a first off (provider) run):

variant	12 tok: video / audio / input grad	925 tok	4253 tok
off, 2nd provider run	0 / 0 / 0	0 / 0 / 1.3e-2	0 / 0 / 1.7e-2
strict (R/R)	2.07e-3 / 1.77e-2 / 0.226	5.52e-3 / 2.40e-2 / 0.129	5.30e-3 / 2.06e-2 / 0.175
strict, `node:attention` native	2.22e-3 / 1.59e-2 / 0.164	5.24e-3 / 2.21e-2 / 0.126	5.39e-3 / 1.98e-2 / 0.162
strict, `boundary:token_refiner` native	2.00e-3 / 2.05e-2 / 0.151	5.38e-3 / 2.06e-2 / 0.117	5.27e-3 / 1.91e-2 / 0.159

8349 tokens (video / audio / input grad; second provider run: 0 / 0 / 1.5e-2): strict 5.29e-3 / 1.95e-2 / 0.154 with h3_full_attention (#553) on #575 (4f7c5d11); node:attention native 5.31e-3 / 1.94e-2 / 0.158 (still on interim attention on this branch).

Strict is bitwise repeatable at all three sizes (video, audio, input gradient, transition log-prob), and audit gives the same bits as strict.
The provider is not: two provider runs differ in the input gradient by 1.3e-2 (925 tokens) and 1.7e-2 (4253 tokens); at 12 tokens they agree. That is the nondeterminism these rows remove.
The strict-vs-provider gap (input gradient 0.12 - 0.23, BF16 provider vs FP32-accumulated candidate, amplified by a 33B network) is the same effect as in [WS1][CUDA][MiniMax-H3] ws1_full_h3_transformer and ws1_dual_trajectory: 33B forward/backward and rollout-vs-train control (RFC #420 stack 5/5) #571. Switching one node or one boundary back to provider does not remove it at any size, so it is not one component's.
On this branch alone, 8349-token strict/audit runs hit CUDA out of memory in the interim DeterministicAttentionOp backward (FP32 attention matrix). The 8349-token strict numbers in the table were taken with h3_full_attention ([WS1][CUDA][MiniMax-H3] Add batch-invariant full attention #553) on the stacked branch [WS1][CUDA][MiniMax-H3] checkpoint mapper and the vLLM-Omni rollout adapter (RFC #420 stack 7/7) #575 at head 4f7c5d11 (video 5.29e-3 / audio 1.95e-2 / input grad 0.154; bitwise repeatable).
The adapter reports which boundary used which binding (boundary_bindings), including proj_in, proj_out, time_embedder and norm_out as provider.

No timing is claimed here; the speed of the underlying rows is in their own PRs.

Results (B200 / Blackwell)

Measured on NVIDIA B200 (179 GiB, sm100), driver 580.126, CUDA 13.0, torch 2.13.0+cu130, pinned MiniMax-H3@42ed227 weights, clean tree at af3dcaef. Optional dependencies for the two comparison tests are pinned to the revisions recorded for the H100 run (environment.json): vLLM-Omni 12e92800 and verl-omni efb9dc7f.

Tests: 19 passed (19 collected), none skipped, including the scheduler comparison against verl-omni (VERL_OMNI_PATH set); tests.json. Report: report.json.

Runtime modes strict, audit, auto, off bind and report as intended (R/R for strict and audit, P/P for auto and off).
LoRA gradient gate (3 shapes): repeatable, rows independent, base frozen; forward error against FP64 0.00166, gradients of A and B 3.9e-07 to 1.3e-06, of the input 0.0017 to 0.0017 (gate 5e-3).
CPS transition: prev_sample and mean bitwise equal, batch invariant, scoring reproduces the sampled log-prob in all 3 shapes; the mean is bitwise equal to verl-omni's and the largest log-prob relative difference to it is 5.8e-06, 1.5e-04, 7.5e-05 (shapes B3 x 257x96, B4 x 4096x96, B2 x 8192x32).
Full 33B model, strict (R/R) against the provider:

tokens	video rel. diff	audio rel. diff	input-grad rel. diff (strict)	input-grad rel. diff (provider vs provider, P/P)	strict bitwise repeatable	audit equals strict
12	2.29e-03	2.13e-02	0.200	0.000	True	True
925	5.40e-03	2.34e-02	0.134	0.014	True	True
4253	5.20e-03	2.04e-02	0.169	0.017	True	True

The input-gradient columns moved by up to about 8% between two B200 runs of this PR (strict at 925 tokens: 0.1339 in the first run, 0.1345 here; the P/P column, which compares the provider with itself, 0.0127 and 0.0136), so part of that column is run-to-run variation of the provider's backward; this was not investigated further. The log-prob difference to the provider is 0 in every row. Ablations that put attention or the token refiner back on the native path give differences of the same size as strict (in the report).

Not run on B200: the 8349-token strict / audit rows, which need the h3-flash-attention-v1 binding that is not in this PR (see the H100 section).

Tests

tests/models/minimax_h3/test_h3_runtime_integration.py (19 tests; the measurements above are from tools/validation/models/h3_runtime_report.py): runtime modes and the environment variable, overlay parsing and validation, adapter off / strict / audit, transition (scheduler comparison needs VERL_OMNI_PATH; skipped without it), LoRA gate and FP64 check. All 19 pass on H100 with a verl-omni checkout (18 pass and 1 skips without it). The #571 chain regression (refiner, projection rows, one block, full model) still passes.

B200: 19 passed, with VERL_OMNI_PATH set.

Scope & Known limitations

The noise law is the scheduler's current cps law (noise_level=0.8); if the recipe changes it, only transition.py and its test change. Maintainer: is this the law to pin?
Attention, rotary and AdaLN of other rows stay provider unless bound by their own rows.
Not in this PR: checkpoint mapper and the vLLM-Omni adapter (next PR in the stack).

coderabbitai · 2026-10-22T17:00:43Z

Important

Review skipped

Too many files!

This PR contains 160 files, which is 60 over the limit of 100.

To get a review, reduce the PR to 100 files or fewer by splitting it into smaller PRs or changing its base branch.

Upgrade to a paid plan to raise the limit.

This review couldn't start because sufficient usage credits or metered capacity aren't available. Add credits or update usage-based reviews in the billing tab, then retry.

Check out review usage here.

Run configuration

Source mode: Auto
Repository: RL-Align/RL-Kernel
Branch: test-h3
File: .coderabbit.yaml
Commit: d22d4b4
Layers: Repository YAML
Review profile: CHILL
Plan: Advanced
Run ID: e70533d9-d6b9-4505-b783-c03d642f3bd0

Commits

Reviewing files that changed from the base of the PR and between d22d4b4 and 7b284f6.

Files selected for processing: 160

.github/workflows/ci.yml
benchmarks/models/benchmark_h3_conditioning.py
build_tools/extensions.py
csrc/bindings/ops.cpp
csrc/cuda/h3/adaln_row_gather.cu
csrc/cuda/h3/det_linear.cu
csrc/cuda/h3/gate_residual.cu
csrc/cuda/h3/rmsnorm_modulate.cu
csrc/cuda/h3/timestep_sinusoid.cu
docs/.nav.yml
docs/operators/README.md
docs/operators/h3-adaln-gate-residual.md
docs/operators/h3-adaln-projection.md
docs/operators/h3-adaln-row-gather.md
docs/operators/h3-arch-fingerprint.md
docs/operators/h3-final-adaln-out.md
docs/operators/h3-full-transformer.md
docs/operators/h3-integration.md
docs/operators/h3-one-block.md
docs/operators/h3-operator-trace.md
docs/operators/h3-projection-rows.md
docs/operators/h3-rmsnorm.md
docs/operators/h3-sp-norm-adaln.md
docs/operators/h3-timestep-mlp.md
docs/operators/h3-timestep-sinusoid.md
docs/operators/h3-token-refiner.md
docs/operators/h3-tp-adaln-3mod.md
docs/operators/h3-trajectory-artifact.md
examples/models/minimax_h3/README.md
examples/models/minimax_h3/candidate.sbatch
examples/models/minimax_h3/diagnose-block.sbatch
examples/models/minimax_h3/full-model-ensemble-rule.json
examples/models/minimax_h3/full-model-thresholds.json
examples/models/minimax_h3/generation-compare.sbatch
examples/models/minimax_h3/generation-contract.json
examples/models/minimax_h3/generation-inputs.json
examples/models/minimax_h3/generation-prompt.txt
examples/models/minimax_h3/generation-resume-check.sbatch
examples/models/minimax_h3/generation.sbatch
examples/models/minimax_h3/reference-manifest.json
examples/models/minimax_h3/reference.sbatch
rl_engine/_C.pyi
rl_engine/backends/cuda/model_specific/__init__.py
rl_engine/backends/cuda/model_specific/minimax_h3/__init__.py
rl_engine/backends/cuda/model_specific/minimax_h3/adaln_modulation.py
rl_engine/backends/cuda/model_specific/minimax_h3/adaln_projection.py
rl_engine/backends/cuda/model_specific/minimax_h3/adaln_row_gather.py
rl_engine/backends/cuda/model_specific/minimax_h3/det_linear.py
rl_engine/backends/cuda/model_specific/minimax_h3/final_adaln_out.py
rl_engine/backends/cuda/model_specific/minimax_h3/gate_residual.py
rl_engine/backends/cuda/model_specific/minimax_h3/projection.py
rl_engine/backends/cuda/model_specific/minimax_h3/rmsnorm.py
rl_engine/backends/cuda/model_specific/minimax_h3/sp_norm_adaln.py
rl_engine/backends/cuda/model_specific/minimax_h3/timestep_mlp.py
rl_engine/backends/cuda/model_specific/minimax_h3/timestep_sinusoid.py
rl_engine/backends/cuda/model_specific/minimax_h3/tp_adaln_projection.py
rl_engine/backends/cuda/model_specific/minimax_h3/ws2_comm.py
rl_engine/backends/shared/triton/gemm/h3_projection.py
rl_engine/models/minimax_h3/arch_fingerprint.json
rl_engine/models/minimax_h3/fingerprint.py
rl_engine/models/minimax_h3/operator_trace.py
rl_engine/models/minimax_h3/pinned_checkpoint.json
rl_engine/models/minimax_h3/runtime.py
rl_engine/models/minimax_h3/trajectory.py
rl_engine/models/minimax_h3/transition.py
rl_engine/reference/minimax_h3/__init__.py
rl_engine/reference/minimax_h3/adaln_projection.py
rl_engine/reference/minimax_h3/adaln_row_gather.py
rl_engine/reference/minimax_h3/block.py
rl_engine/reference/minimax_h3/final_adaln_out.py
rl_engine/reference/minimax_h3/fixed_order.py
rl_engine/reference/minimax_h3/gate_residual.py
rl_engine/reference/minimax_h3/projection.py
rl_engine/reference/minimax_h3/rmsnorm.py
rl_engine/reference/minimax_h3/timestep_mlp.py
rl_engine/reference/minimax_h3/timestep_sinusoid.py
rl_engine/runtime/registry.py
rl_engine/validation/models/h3_block.py
rl_engine/validation/models/h3_cases.py
rl_engine/validation/models/h3_chain.py
rl_engine/validation/models/h3_full_manifest.json
rl_engine/validation/models/h3_full_model.py
rl_engine/validation/models/h3_generation_resume.py
rl_engine/validation/models/h3_manifest.json
rl_engine/validation/models/h3_provider.py
rl_engine/validation/models/h3_refiner.py
rl_engine/validation/models/h3_report.py
rl_engine/validation/models/h3_weights.py
rl_engine/validation/models/h3_ws2.py
rl_engine/validation/operators/operator_inputs.py
rl_engine/validation/operators/operator_specs.py
tests/build/test_build_platform_collectives.py
tests/models/minimax_h3/conftest.py
tests/models/minimax_h3/test_h3_adaln_gate_residual.py
tests/models/minimax_h3/test_h3_adaln_modulation.py
tests/models/minimax_h3/test_h3_adaln_projection.py
tests/models/minimax_h3/test_h3_adaln_row_gather.py
tests/models/minimax_h3/test_h3_arch_fingerprint.py
tests/models/minimax_h3/test_h3_benchmark_timing.py
tests/models/minimax_h3/test_h3_chain_replay_cli.py
tests/models/minimax_h3/test_h3_cli.py
tests/models/minimax_h3/test_h3_conditioning_e2e.py
tests/models/minimax_h3/test_h3_det_linear.py
tests/models/minimax_h3/test_h3_final_adaln_out.py
tests/models/minimax_h3/test_h3_full_model.py
tests/models/minimax_h3/test_h3_generation_resume.py
tests/models/minimax_h3/test_h3_native_validation.py
tests/models/minimax_h3/test_h3_one_block.py
tests/models/minimax_h3/test_h3_operator_trace.py
tests/models/minimax_h3/test_h3_projection_rows.py
tests/models/minimax_h3/test_h3_report.py
tests/models/minimax_h3/test_h3_rmsnorm.py
tests/models/minimax_h3/test_h3_runtime_integration.py
tests/models/minimax_h3/test_h3_sp_norm_adaln.py
tests/models/minimax_h3/test_h3_timestep_mlp.py
tests/models/minimax_h3/test_h3_timestep_sinusoid.py
tests/models/minimax_h3/test_h3_token_refiner.py
tests/models/minimax_h3/test_h3_tp_adaln.py
tests/models/minimax_h3/test_h3_trajectory_artifact.py
tests/models/minimax_h3/test_h3_weights.py
tests/models/minimax_h3/test_h3_ws2_processes.py
tests/models/minimax_h3/test_prepare_h3_weights.py
tests/runtime/test_dispatch.py
tests/validation/operators/test_ws1_gtest_gpu.py
tools/migration/README.md
tools/migration/layout.json
tools/validation/models/h3_bitwise_reuse.py
tools/validation/models/h3_block_diagnostic.py
tools/validation/models/h3_block_replay.py
tools/validation/models/h3_block_ulp_sweep.py
tools/validation/models/h3_candidate_prepare.py
tools/validation/models/h3_chain_replay.py
tools/validation/models/h3_dual_trajectory.py
tools/validation/models/h3_evidence.py
tools/validation/models/h3_evidence_integrity.py
tools/validation/models/h3_fingerprint.py
tools/validation/models/h3_first_step_diagnostic.py
tools/validation/models/h3_full_charts.py
tools/validation/models/h3_full_compare.py
tools/validation/models/h3_full_ensemble_rule.py
tools/validation/models/h3_full_generation.py
tools/validation/models/h3_full_model.py
tools/validation/models/h3_full_provider_spread.py
tools/validation/models/h3_full_transformer.py
tools/validation/models/h3_generation_adopt.py
tools/validation/models/h3_generation_compare.py
tools/validation/models/h3_operator_trace.py
tools/validation/models/h3_prior_art.py
tools/validation/models/h3_projection_bench.py
tools/validation/models/h3_projection_charts.py
tools/validation/models/h3_refiner_report.py
tools/validation/models/h3_runtime_report.py
tools/validation/models/h3_trajectory.py
tools/validation/models/h3_ws2_evidence.py
tools/validation/models/plot_h3_bitwise_reuse.py
tools/validation/models/plot_h3_block.py
tools/validation/models/plot_h3_evidence.py
tools/validation/models/plot_h3_prior_art.py
tools/weights/prepare_h3_weights.py
tools/weights/snapshot_h3_checkpoint.py

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

…fixed-order Triton GEMM Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 650e2148c6f0e5dcdef136537e0e62f745020273) Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit d099c0b)

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit b41f868)

…il-closed test instead of a skip Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 74e5b83)

…l keys Review finding (RL-Align#568): jointly permuting x and g changed dW/dbias because the token axis was reduced in call order. All projection ops (audio in/out, text context, Q/K/V) now take an optional logical_keys [T,2|3] (negative first column = padding); rows are reduced in ascending key order with padding dropped, active keys must be unique. Key format follows RL-Align#562. Adds the reviewer's repro plus permutation/padding tests for every row and Q/K/V. Co-authored-by: Yitao Wang <75578057+walker-ai@users.noreply.github.com> Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit d2941f0)

…ready ordered rows Follows the canonical-row design of RL-Align#562: the Q/K/V node of the block graph supplies (batch row, position) keys so its dW is reduced in logical order, and rows that already arrive in canonical order are reduced in place (same bytes, no gather of x and g). Co-authored-by: Yitao Wang <75578057+walker-ai@users.noreply.github.com> Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 0da0583)

…block 0 node-by-node and 50-block/generation gates Squashes the one-block branch (its row-by-row history: timestep sinusoid, RMSNorm, AdaLN rows, sequence/tensor parallel norm rows, final AdaLN, one block, 50-block qualification) onto the current test-h3 so the pull request carries a reviewable, signed history. RFC RL-Align#420. Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

…pull request Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit eaba5efe3df971ea9d1a969f53e48f09e09ed176) Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 7da0574)

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit b45190d08fc033c7e63ad821d3918eb32dd454f6) Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit b5536d6)

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit ad1925045b522772bf56f8468b23cca47f04156b) Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit f9041e8)

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit b44540e)

…ody) Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit eab901b)

…g provider, LoRA gate and CPS transition RFC RL-Align#420 rows h3_runtime_modes, h3_ablation_overlay, diffusers_runtime_adapter_h3, h3_training_provider and h3_lora_gradient_gate, plus the layout-invariant CPS transition the training provider uses. full_model_execution gains native_boundaries so a run can keep selected boundary modules on provider arithmetic (ablation). Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 65d7fff)

… weight loader, tests, doc Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit d12a6bb26566fae58a309e76916af9aae767d487) Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit b316869)

…ith H100 evidence Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 0734a32ee0c427790a244c7f83e826ba9fe66393) Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit c2c2270)

JSON and figure: runtime modes, LoRA gradients against FP64, CPS transition against the scheduler formula and verl-omni, and the 33B model under off / strict / audit / ablations. Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 3f0bbe6)

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 1e3a2ea)

… and lint Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit d9b0c91)

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 81515ae)

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 67af796)

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 2c08c25)

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 58b724b)

…age import Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit e452447)

….path Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

…#567 RL-Align#568 RL-Align#570 RL-Align#571 RL-Align#574 RL-Align#575 Test summaries (tests.json), runner outputs and figures measured on NVIDIA B200 (sm100), driver 580, CUDA 13.0, torch 2.13.0+cu130, MiniMax-H3@42ed227, clean trees at the PR heads eab901b, 3cf51bf, 6de320d, eb7840b, af3dcae and 4f7c5d1. Figures use B200 axis labels. No host names, paths or scheduler identifiers. Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

…jectory diag summary - h3-fingerprint-trace-trajectory-h100/tests.json (77 passed, eab901b) - environment.json for RL-Align#574/RL-Align#575 H100 adapter reproducibility pins - dual_trajectory_diag_summary.json (H100/B200 diag strict 0/16 on seeds 11/22/33) Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

…with the H100-pinned vllm-omni and verl-omni Same heads (af3dcae, 4f7c5d1), with vllm-omni 12e92800 and verl-omni efb9dc7f as recorded in the H100 environment files, replacing the first B200 run that used newer checkouts. Test summaries record the dependency revisions. Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

fusheng-ji requested review from EthanZero2Hero, Flink-ddd, KJLdefeated, bitborne, inaniloquentee and ryankert01 as code owners October 22, 2026 16:58

This was referenced Oct 22, 2026

[WS1][CUDA][MiniMax-H3] checkpoint mapper and the vLLM-Omni rollout adapter (RFC #420 stack 7/7) #575

Open

[RFC][MiniMax-H3][CUDA/ROCm] WS1/WS2 kernel roadmap, ablation matrix and integration plan #420

Open

fusheng-ji and others added 21 commits October 22, 2026 22:54

feat(h3): projection rows Hopper evidence, charts and lint

a6533fa

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit b41f868)

test(h3): make the single-unit audio_output shard case an explicit fa…

e109e16

…il-closed test instead of a skip Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 74e5b83)

docs(h3): do not link to the fingerprint page, it lives in a sibling …

3cf51bf

…pull request Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

style(h3): black/isort/flake8 for the foundation rows

bf06181

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit b44540e)

fix(h3): type the trajectory step body so mypy accepts StepRecord(**b…

ea1687f

…ody) Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit eab901b)

feat(h3): token refiner Hopper evidence, charts and lint

7ccd4a0

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 1e3a2ea)

feat(h3): full transformer and dual trajectory Hopper evidence charts…

d7c1952

… and lint Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit d9b0c91)

test(h3): runtime report covers 925 and 4253 tokens and tolerates OOM

51ad1db

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 81515ae)

style(h3): lint refiner report tool

cd1b0fd

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 67af796)

style(h3): black for the full-transformer chart tool

8acada2

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 2c08c25)

fusheng-ji added 4 commits October 22, 2026 22:54

test(h3): runtime report prints OOM tracebacks and can skip ablations

f63494e

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit 58b724b)

fix(h3): satisfy mypy in the refiner report and the operator-trace st…

6de320d

…age import Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

style(h3): docstring length in the full-transformer chart tool

eb7840b

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com> (cherry picked from commit e452447)

test(h3): load the verl-omni scheduler by path instead of editing sys…

af3dcae

….path Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

fusheng-ji force-pushed the feat/h3-runtime-integration branch from 58b724b to af3dcae Compare October 22, 2026 21:04

Flink-ddd added platform: cuda Specific optimizations or bugs in NVIDIA graphics cards (such as FlashInfer, TMA optimizations) new-model MiniMax-H3 WS1 labels Oct 23, 2026

fusheng-ji added 5 commits October 25, 2026 14:26

Merge test-h3 into feat/h3-one-block

0da990e

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

Merge local/h3merge/h3-one-block into feat/h3-projection-rows

09fcd9c

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

Merge local/h3merge/h3-projection-rows into feat/h3-token-refiner

3803588

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

Merge local/h3merge/h3-token-refiner into feat/h3-full-transformer

f069b29

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

Merge local/h3merge/h3-full-transformer into feat/h3-runtime-integration

7b284f6

Signed-off-by: Wenbo Ji <36562829+fusheng-ji@users.noreply.github.com>

Navigation Menu

Repository navigation

[WS1][CUDA][MiniMax-H3] runtime modes, ablation overlay, Diffusers adapter, training provider, LoRA gate (RFC #420 stack 6/7) - #574

[WS1][CUDA][MiniMax-H3] runtime modes, ablation overlay, Diffusers adapter, training provider, LoRA gate (RFC #420 stack 6/7)#574
fusheng-ji wants to merge 30 commits into
RL-Align:test-h3from
fusheng-ji:feat/h3-runtime-integration

fusheng-ji commented Oct 22, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Oct 22, 2026 •

edited

Loading

Review skipped

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

fusheng-ji commented Oct 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

Prior art & reuse decision

Results (H100 / Hopper)

Results (B200 / Blackwell)

Tests

Scope & Known limitations

Uh oh!

coderabbitai Bot commented Oct 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review skipped

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

fusheng-ji commented Oct 22, 2026 •

edited

Loading

coderabbitai Bot commented Oct 22, 2026 •

edited

Loading