[inaniloquentee]

Scope

Draft, partial item 27 of KLR #81 / RL-Kernel #514. Adds an opt-in request-mode gate around the existing native joint attention plan and autograd adapter. Tracks contiguous prefill/decode positions per request, enforces model/cache version transitions, and checks training/recompute Q, canonical KV, sink, logical selection, BF16 O and FP32 LSE byte-for-byte. Native operator failures cannot advance request state. No numerical GPU kernel has been copied or added.

Base and prerequisites

Started from latest main 8588158650aff58c9c1f527674de94bad06c3d10. This branch currently contains prerequisite #629 (including #628) and #626 (including #625/#624/#618), which are still open. Target is dsv4-rocm; it trails that main by 22 commits. Review this commit separately as d46ad78 until prerequisites land.

Verification

MI300X/gfx942 in the ljj ROCm container: 605 passed, 17 warnings across operator contracts and DeepSeek-V4 tests with pinned aiter forward/wrapper and Megatron backward sources; Ruff passed. New tests cover request sequencing, stale versions, provider failure, byte-preserving recompute under reordered rows, native Graph replay, and the native autograd path.

Remaining before item 27 signoff

• Wire actual Megatron and vLLM model call sites, then verl and vime; the present engine modules only re-export the opt-in gate.
• Combine live C4/C128/Indexer/recent-window cache state and external writer/read synchronization with owned logical KV metadata; prove stale scale and version barriers.
• Validate prefill/decode/recompute and whole-engine Graph against the same 0731 checkpoint and exact logical inputs. CP/DCP and PP belong to separate parallel signoff.

The current container exposes one GPU, and its model directory is empty. This PR must remain draft until the missing integration and acceptance evidence are supplied. It does not claim full model or two-node consistency.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

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

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

[inaniloquentee]

Item-27 framework-integration gate checked on the ljj MI300X container:

• Installed vLLM is 0.26.0; its model_executor/models has no DeepSeek V4 model entry. The ljj/models directory contains no checkpoint files.
• Referenced upstream vLLM #60442, #60443 and #60454 remain open (not merged). #60442 changes NVIDIA FlashMLA, so it is not an executable ROCm companion by itself.
• The existing RL-Kernel vLLM attention_runtime.py is an opt-in re-export, not a model call-site hook. This PR's request-mode/Graph tests validate the shared plan, not live vLLM or verl/vime serving.

Before item 27 can leave draft: pin a compatible V4 vLLM/ROCm implementation and model checkpoint, wire the model call site to the shared plan, then validate prefill/decode/recompute, KV/O/LSE, Graph, and weight-version barriers in both frameworks. No V2 monkeypatch should be treated as a V4 integration.