[andyluo7]

Purpose

Draft validation-only stack for the three current AMD fixes. Do not merge this PR; review and merge the focused source PRs individually.

Source PR	Repair
#8341	Keep the NVIDIA seeded-exponential Triton path off ROCm and preserve ordered Torch RNG draws
#8343	Consolidated CosyVoice3 structured full-payload and live per-request conditioning repair; includes #8262
#8463	Route MOSS reference-encoder varlen attention through AITER/upstream FlashAttention on ROCm, including AITER's grad-enabled return_lse contract

Current revision

• Base: ee8fdab1dc26ea8f09d7910cdc98a17c4b0ea02a
• Head: 28dc8571c9545ca6bd80f906b02fd5ae3f8586b2
• Seven focused commits; previously merged AMD repairs are not carried in this integration branch.
• The MOSS production file is byte-identical to [Bugfix] Route MOSS reference attention through ROCm FlashAttention #8463 head 50180bc9451b82b58d8032664e5ea692d53712d5.
• git diff --check, Python compilation, and all applicable changed-file hooks pass.
• No thresholds, quarantine policy, or Buildkite job selection changed.

Preserved evidence

AMD #13236 at the preceding integration head finished all 63 jobs with zero hard failures and one unrelated quarantined R2-01 soft failure. Qwen Function Expansion, Documentation, AITER-on, Accuracy, Seed-TTS, and full CosyVoice all passed.

That run contained the initial MOSS ROCm dispatch but predated the AITER autograd return_lse correction. It is preserved as Qwen/CosyVoice interaction evidence and does not qualify the corrected MOSS implementation.

Exact-head validation

AMD #13238 at this exact integration head is terminal:

• 63/63 jobs finished with zero hard failures and one known quarantined R2-01 soft failure.
• R2-01 completed with 806 passed, 372 skipped, and 5076 deselected. All four MOSS reference-attention cases and all six seeded-exponential variants passed; only the seven pre-existing quarantined Qwen3-TTS __nv_rintf cases failed.
• Qwen3 Function Expansion and AITER-on Smoke passed. Accuracy passed both cases in 2761.45 seconds.
• CosyVoice3 SDPA smoke passed, and the full CosyVoice matrix passed all three cases in 2324.90 seconds.
• No retry chain was needed.

CUDA #16630 had one failure among 15 H100 entrypoint cases: the synchronous Qwen3-Omni realtime test timed out opening its WebSocket after the server had reached application startup. The other 14 cases passed, the same 15-test shard passed in builds #16620, #16621, and #16628, and none of this stack changes WebSocket or entrypoint code. This is classified as a transient server/handshake failure rather than a demonstrated regression. The individual source-PR CUDA builds #16626, #16627, and #16629 passed. Intel #9543 passed. NPU #8369 is non-required, with 21/22 script jobs passed and its final Diffusion A5 job reserved awaiting a worker.

AI assistance: Codex rebuilt the validation stack from current main, preserved patch equivalence, and prepared the exact-head CI gate. Human review belongs to the focused source PRs.

[hsliuustc0106]

This PR touches tests/model_executor/, .buildkite/amd/, tests/diffusion/, tests/e2e/, vllm_omni/model_executor/ (30 files). Based on CODEOWNERS coverage of the changed files, the most-related reviewers appear to be:

@yenuo26 @NickCao @congw729

Could one of you take a look when you get a chance? Thanks!

[hsliuustc0106]

Ready for full review when draft status is removed. This remains a draft at c30e63d04dc9; no full code review was performed here.

[hsliuustc0106]

@andyluo7 this PR is labeled ready + high priority, but CI is failing on the latest commit:

• buildkite/vllm-omni
• buildkite/vllm-omni-npu-ci

Could you please take a look and push an update to get CI green? Once the checks pass we can proceed with review/merge. Thanks!

[andyluo7]

Closing as an obsolete diagnostic/evidence-only PR. Current main AMD merge builds #13240, #13241, and #13242 are effectively green, and this stale stacked branch should not merge as-is. If the underlying issue still reproduces, the relevant source change should be extracted into a focused PR rebased on current main.