[andyluo7]

Current combined AMD evidence - September 30, 2026, 22:53 UTC

The integration at c30e63d04dc9a54c5c4b43235eb2ce535fe68df3 now has two complete AMD passes at the identical commit. Each build has 65/65 jobs finished, every exit 0 and non-soft, with no retries. Both include all five NIGHTLY leaves.

Build	READY	MERGE	Full result
AMD #13061	28/28, 34m01.983s	30/30, 34m26.568s	65/65 passed
AMD #13064	28/28, 33m58.742s	30/30, 35m09.532s	65/65 passed

Phase time runs from the earliest phase job start to the latest finish; queue and image-build time are excluded. Both READY and MERGE are below 40 minutes in both observations.

Repeat Accuracy evaluates all 1,088 Seed-TTS samples, mean WER 0.3088705089, with zero request, PCM or ASR failures. Daily-Omni evaluates all 1,197 samples, accuracy 0.7218045113, with zero HTTP or parse failures. Both result JSON files are downloaded and SHA-256 verified against Buildkite metadata. Full CosyVoice passes all 3 cases in 2,317.40s, including both reference voice-clone cases.

These results exercise the combined c30 tree. They establish combined implementation and interaction coverage; they do not establish a second standalone run at a different source PR SHA. Earlier 024 results remain separate evidence.

CUDA #16446 at c30 still has a Documentation Whisper-small transcript failure against the unchanged 0.8 gate; WAV/JSON retention is qualified, and the exact-model replay remains pending. Intel #9379 failed during a dependency download before runtime testing. Source PR checks and maintainer review remain separate merge gates.

This validation PR remains draft; source review belongs to the individual repair PRs. The retained sections below describe earlier observations and diagnostic limits.

Retained validation history

Purpose

Repeat the exact current ROCm stability integration head c30e63d04dc9a54c5c4b43235eb2ce535fe68df3, base 27a6321a6311870fe68c2c9c13b694f2612ff779, with no added source commit. Source review and repair mapping remain in #8344 and its individual source PRs.

First observation

AMD #13061 has all 28 READY and all 30 MERGE jobs passed, every exit 0 non-soft, in 34m01.983s / 34m26.568s. All six executor leaves finish in 3-7 minutes; AuK/AR/Pi0.5, Ulysses, Breeze, Stable Audio and the remaining blocking models pass. All five NIGHTLY leaves pass and the full build is 65/65 finished, every exit 0 non-soft, with no retries. Accuracy passes 2 / 3,059.90s; Seed-TTS evaluates 1,088, WER 0.3152781520 and all configured checks pass. Full CosyVoice passes 3 / 2,303.31s. Daily-Omni evaluates 1,197 at accuracy 0.7218045113; both Daily-Omni and Seed-TTS JSON artifacts are SHA-256 verified, with zero HTTP/parse/request/PCM/ASR failures.

The earlier024 revision already has two full AMD correctness passes (#13053/#13060), but those cannot count as a repeat for this newer c30 revision. This branch includes AR routing #8346, CPU reference threads #8347 and Documentation WAV/JSON #8348.

Acceptance

• All READY, MERGE and five NIGHTLY model leaves must finish with zero exits; inspect soft failures and retries individually.
• Repeat READY and MERGE below 40 minutes at this identical head. Timing excludes queue/image-build duration.
• Retain actual numerical, RNG, output ownership, voice-conditioning and accuracy assertions.
• CUDA Documentation audio retention and the earlier CUDA generation/ASR mismatch remain separate gates. Historical HIP/node incidents still require runner telemetry.

Keep this PR draft for validation only. Require maintainer approval before merging source changes; no automatic merge.

Cross-platform check limits

Intel #9379 at this c30 commit fails during image build while downloading PyPI pyelftools metadata (connection timeout after three retries); no XPU runtime ran. Same-c30 CUDA #16449 has 13 completed non-H100 jobs and eight H100 jobs reserved/unassigned. Source #8348 CUDA #16444 passes20/20 and retains Documentation artifacts, while integrated CUDA #16446 repeats the Whisper-small transcript-tail failure; all ten diagnostic artifacts upload. Independent exact-WAV ASR supports a decoder artifact but does not replace the unchanged 0.8 CI gate.

[mergify]

⚠️ The sha of the head commit of this PR conflicts with #8344. Mergify cannot evaluate rules on this PR. Once #8344 is merged or closed, Mergify will resume processing this PR. ⚠️

[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]

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

• buildkite/vllm-omni
• buildkite/vllm-omni-intel-ci
• 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.