feat(rocm): add end-to-end Vime rollout/training attention ablation matrix

[Flink-ddd]

Summary

• Add the ROCm end-to-end counterpart of the production/RL-Kernel Attention matrix introduced in feat(alignment): add cross-configuration framework #230.
• Execute the P/P, P/R, R/P, and R/R matrix through real Vime orchestration, covering both vLLM rollout and Megatron training.
• Keep FFN and Logp fixed at P/P, so Attention is the only implementation dimension changed between cells.
• Require runtime readback evidence from both frameworks and fail closed when execution falls back, selects the wrong route, or cannot prove the strict ROCm backend.
• Remove the previous operator-only micro-probe runner and all checked-in generated result artifacts, including results.json.

Ablation matrix

P selects the framework-native implementation. R selects the RL-Kernel ROCm implementation.

Case	Megatron training	vLLM rollout
P/P	Framework-native	Framework-native
P/R	Framework-native	RL-Kernel AITER/CK
R/P	RL-Kernel AITER/CK	Framework-native
R/R	RL-Kernel AITER/CK	RL-Kernel AITER/CK

Every cell starts in a fresh process and uses the same model, checkpoint, prompt data, seeds, token limits, rollout count, and pre-update state.

Implementation

• Add rl_engine.integrations.rocm_ablation as the end-to-end matrix runner and runtime evidence validator.
• Add a runnable Qwen3-8B Vime example under examples/vime_qwen3_8b_rocm_ablation.
• Extend the Attention binding layer with platform-aware strict ROCm runtime selection.
• Add the Megatron ROCm training route using the strict AITER/CK Attention core and RCCL AG/RS transport.
• Add the vLLM ROCm rollout route with paged-KV reconstruction and strict AITER/CK execution.
• Extend runtime and registry metadata so ROCm platform, backend, core, and schedule identities can be validated from executed readbacks.
• Add CPU-safe end-to-end matrix and framework-adapter contract coverage to the default CI workflow.

Runtime acceptance boundary

A subprocess exit code of zero is not sufficient for a passing cell.

Each cell must emit executed Attention readbacks from both Megatron training and vLLM rollout. For every R side, the runner requires evidence for:

• runtime_platform=rocm
• Semantic backend rlkernel.attention.deterministic.v1
• Actual runtime rlkernel.rocm.attention.aiter_ck_ag_rs.v1
• The strict AITER/CK core
• The fixed no-Split-KV ROCm schedule
• No native, PyTorch-reference, Triton, CUDA, or fallback execution

The runner also rejects missing, malformed, stale, duplicated, or case-inconsistent readbacks.

Replay invariants

The matrix freezes the command and replay-defining environment before the first cell. Only the Attention case selection and case-local output paths may change.

An executable run requires identifiable model, distributed checkpoint, and prompt inputs. It also enforces NUM_ROLLOUT=1, ensuring every cell begins from the same pre-update checkpoint state.

Running the matrix

Review the resolved launch contract without allocating a GPU run:

python examples/vime_qwen3_8b_rocm_ablation/run.py \
  --output-dir /tmp/rocm-attention-ablation \
  -- bash /path/to/vime/scripts/run-qwen3-8B-rocm.sh

Execute all four cells:

python examples/vime_qwen3_8b_rocm_ablation/run.py \
  --run \
  --output-dir /tmp/rocm-attention-ablation \
  -- bash /path/to/vime/scripts/run-qwen3-8B-rocm.sh

Individual cells can be selected with repeatable --case arguments while debugging. The final hardware acceptance run should execute the complete four-cell matrix.

Results and artifacts

Runtime logs and unmodified framework readbacks are written to the requested external output directory. The aggregate report is a human-readable summary.md.

No generated benchmark or result JSON is checked into the repository.

Validation
Added end-to-end matrix, replay-freezing, readback-validation, and failure-boundary tests.
Extended Attention binding and framework-runtime adapter coverage for ROCm.
Added the ROCm end-to-end tests to the default CI workflow.
Source formatting, Python compilation/AST validation, shell syntax validation, and the isolated matrix/readback dry-run contract passed locally.
A full physical MI300X Vime matrix remains the hardware acceptance step before moving the PR out of Draft.

Reference: #230

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Team

Run ID: a483ac14-e29c-465a-a1f2-369714b7eb7e

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

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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