[FayeSpica]

Contributor intent (required human answers)

1. What outcome do you want to achieve?

Upgrade the Ascend CI base image to v0.32.0 and keep CI working.

2. What is the highest-leverage variable for that outcome?

Compatibility between Omni and the vLLM v0.32.0 / vLLM-Ascend stack.

3. How will this work move that variable?

Update the base image, installation examples, CI dependencies and KV-cache test fixture.

4. Whose customer or user needs does this address, and what are those needs?

Ascend users and contributors who need working inference and reliable CI on v0.32.0.

Purpose

Fix #9139.

Upgrade the Ascend base image from v0.31.0 to v0.32.0 in both NPU Dockerfiles and the A2/A3/A5/310P installation examples.

• Add pytest-xdist to the CI image to support the default --dist=loadgroup option.
• Set decode_context_parallel_size=1 in the contiguous-KV test fixture. Ascend now compares this field with an integer; the unset MagicMock field raises TypeError.

Test Plan

vLLM Version: v0.32.0.

vLLM-Omni Commit: da572cddb38383deae9caef33b8d490e292ef071.

vLLM-Ascend Image: quay.io/atlas-ci/vllm-ascend:v0.32.0.

• Contiguous-KV fixture: python -m pytest -s -v tests/platforms/npu/test_contiguous_kv_cache.py -m 'core_model and npu and A3 and omni' --run-level core_model
• HunyuanImage-3.0 TP4: python -m pytest -s -v tests/e2e/online_serving/test_hunyuan_image.py -m 'core_model and diffusion and npu' --run-level core_model
• Qwen3-TTS real-weight smoke on the published Ascend A3 base: python examples/offline_inference/text_to_speech/qwen3_tts/end2end.py -q CustomVoice --num-prompts 1 --output-dir /tmp/val/tts/out
• Qwen3-Omni real-weight smoke on the published Ascend A3 base: python examples/offline_inference/qwen3_omni/end2end.py --query-type use_audio --output-wav /tmp/val/omni/out --stage-init-timeout 600
• Wan2.2 real-weight smoke on the published Ascend A3 base: python examples/offline_inference/text_to_video/text_to_video.py --model Wan-AI/Wan2.2-TI2V-5B-Diffusers --prompt "Two anthropomorphic cats in boxing gear fight on a spotlighted stage." --height 480 --width 832 --num-frames 17 --num-inference-steps 8 --fps 16 --output /tmp/val/wan/t2v_out.mp4

Test Result

Contiguous-KV: 3 passed.
HunyuanImage-3.0-Instruct TP4: 1 passed in 119.02s; real weights, 512x512, 2 steps.
Qwen3-TTS: exit 0, no traceback; WAV 4.56s, 24 kHz;
  peak 0.777, non-silent 71.8%; graph capture and ACL replay logged.
Qwen3-Omni: exit 0, no traceback; WAV 20.54s, 24 kHz;
  peak 0.416, non-silent 59.7%; correctly identified Mary Had a Little Lamb.
Wan2.2 TI2V-5B: exit 0, no traceback; decodable MP4, 17 frames, 832x480;
  pixel std 77.3, mean inter-frame delta 16.11.

AI Assistance

Codex assisted with code, regression checks and this draft. Human review remains required.

BEFORE SUBMITTING: read CONTRIBUTING.md and run the precheck-pr skill with the code agent for a self-check against project conventions.

Note

To avoid agent spam, vLLM-Omni limits contributors without write access to 5 open PRs. If this blocks well-intentioned critical work, please contact a committer to request bypass list access.
(anything written below this line will be removed by GitHub Actions)

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-20T06:13:42.603562Z	0c39cb3	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[vllm-omni-review-bot]

This PR matches CODEOWNERS paths: /docs/, /tests/.

Code owners: @david6666666 @Gaohan123 @hsliuustc0106

Routing: @david6666666 via CODEOWNERS; @Gaohan123 via CODEOWNERS; @hsliuustc0106 via CODEOWNERS

@FayeSpica, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[yenuo26]

The NPU release install snippet now pulls vLLM-Ascend v0.32.0 and still clones Omni v0.30.0. Those two do not share a transformers range.

[yenuo26]

The NPU release tab now pulls quay.io/atlas-ci/vllm-ascend:v0.32.0 (and -a3 / -a5 / -310p), but the same snippet still runs git clone -b v0.30.0 at line 56.

Omni v0.30.0 pins transformers>=5.13.0,<5.15 for vLLM 0.30.0. This checkout requires transformers>=5.16.1,<5.20.0 for vLLM 0.32.0, and that <5.15 cap conflicts with the 5.16+ APIs. pip install -e . inside this image pulls transformers back below 5.15.

Move these v0.32.0 image exports into the "NPU from main" section, which clones current main. Keep the release snippet on the Ascend image that matches the v0.30.0 checkout.

[FayeSpica]

Thanks! Reverted the installation doc changes in b530b92.

[vllm-omni-review-bot]

Omni ReviewBot triage note

Automated triage of commit b530b9244099 produced:

• Priority: high. Prompt maintainer attention is suggested.

These are automated triage suggestions only — the final decision belongs to the maintainers.

[yenuo26]

LGTM