Repository navigation
[CI][Qwen3-Omni] Compare structured-config sampling defaults against vLLM normalization - #7742
Merged
NickCao merged 1 commit intoSep 18, 2026
Conversation
…vLLM normalization test_structured_multistage_config_reaches_runtime has failed on every main merge build since it landed with vllm-project#6849 (15444, 15507, 15510, 15523), always at the sampling-defaults assertion. The code2wav stage in vllm_omni/deploy/qwen3_omni_moe.yaml is greedy (temperature 0.0) and spells the disabled top_k as -1; vLLM v0.29.0 SamplingParams normalizes greedy sampling to top_k=0 (and top_p=1.0, min_p=0.0), so the runtime object never equals the raw YAML literal. Keep the YAML literals as the deploy-owned expectation, check that the resolver carries them verbatim into the typed stage config, and compare the runtime StageClient defaults against SamplingParams built from the same values so vLLM's own normalization is applied on both sides. Signed-off-by: Yueqian Lin <linyueqian@outlook.com>
|
This PR appears to be related to model: qwen-omni. Model owners: @amy-why-3459 @yenuo26 @NickCao Routing: @amy-why-3459 via semantic router; @yenuo26 via CI owner, CODEOWNERS; @NickCao via CODEOWNERS @linyueqian, 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. |
hsliuustc0106
added
CI/CD
NickCao
approved these changes
Sep 17, 2026
NickCao
left a comment
NickCao
left a comment
Collaborator
There was a problem hiding this comment.
LGTM, comparing against the normalized configuration instead of the raw literals matches the test's intention.
NickCao
enabled auto-merge (squash)
September 17, 2026 18:18
Collaborator
passed, 1 passed, 2 deselected, 19 warnings in 450.33s (0:07:30) |
tzhouam
added a commit
that referenced
this pull request
Sep 21, 2026
Brings in 65 main commits (825ab20..ae3880f). Motivation: the three CI failures left on Buildkite #3060 are main drift, not rebase regressions: - MiniCPM-o 4.5 seed-tts perf (ready + nightly): the identical request ("Tim Tebow ..." item, 30.2 s of silence, response never reaching response.done, audio_past_key_values reset at 1500) also fails on main's own builds #3049 and #3052, and passes on main from #3054 on, after [CI/Build][MiniCPM-o] Fix the Nightly tests (#7758, 6df62d5), which the rebase branch did not carry. - Nemotron VoiceChat native-duplex e2e: main marked the test skip in the same #7758 (the pipeline declares no duplex_plugin, so /v1/realtime falls through to vLLM's speech-to-text handler and rejects session.update); main's build #3054 skipped it, ours ran it. Conflict resolutions: - tests/e2e/offline_inference/test_qwen3_omni.py: main's side. #7742/#7803 compare the deploy-YAML literal (top_k -1) verbatim and normalize the runtime check through SamplingParams, which supersedes 3dee645's top_k=0 expectation. - vllm_omni/diffusion/diffusion_kv/model_runner_backend.py: both sides dropped. Main (#7166) removed the scheduler_config.max_num_seqs override; the rebase removed the kv_caches list because f2aad6aa70's init_kv_cache returns the caches as a dict. - vllm_omni/benchmarks/patch/patch.py: main's Video-MME helpers kept together with the rebase's get_samples(args, tokenizer, **kwargs) signature; both delegate paths forward **kwargs. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
mlaneuville
pushed a commit
to mlaneuville/vllm-omni
that referenced
this pull request
Sep 22, 2026
…vLLM normalization (vllm-project#7742) Signed-off-by: Yueqian Lin <linyueqian@outlook.com> Signed-off-by: Matthieu Laneuville <matthieu.laneuville@surf.nl>
khairulkabir1661
pushed a commit
to khairulkabir1661/vllm-omni
that referenced
this pull request
Sep 25, 2026
…vLLM normalization (vllm-project#7742) Signed-off-by: Yueqian Lin <linyueqian@outlook.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Purpose
Turn the merge-pipeline step
Omni · Qwen3-Omni Testgreen again.tests/e2e/offline_inference/test_qwen3_omni.py::test_structured_multistage_config_reaches_runtime(added in #6849) has failed on every main build that ran the step since it landed: 15444, 15507, 15510 and 15523, each at the same assertion (test_qwen3_omni.py:131,assert all(...)over the sampling defaults), with the other 12 tests in the step passing.Closes #7732.
Root cause
The code2wav stage in
vllm_omni/deploy/qwen3_omni_moe.yamlis greedy (temperature: 0.0) and spells the disabledtop_kas-1. vLLM v0.29.0SamplingParams.__post_init__normalizes greedy sampling totop_k=0,top_p=1.0,min_p=0.0(and quietly accepts-1as the legacy disable sentinel), so the runtimeSamplingParamsheld by the stage client never equals the raw YAML literal-1. The test compared the runtime object against the YAML literals directly, which fails for that one field. The thinker and talker stages match because none of their deploy values are touched by normalization.This is a test expectation bug; the deploy values do reach the runtime.
Change
stage.model_config.default_sampling_params), where the pipeline only adds its owndetokenize/stop_token_idsconstraints.SamplingParams(**expected)built from the same values, so vLLM's own normalization applies on both sides and the check stays correct if that normalization changes again.Test-only change; no runtime code touched.
Validation
ruff checkandruff format --checkpass on the file.advanced_modeland only runs in the merge pipeline (H100, 2 cards), so the PR lane cannot exercise it. The root cause is read directly from vLLM v0.29.0vllm/sampling_params.py(greedy branch in__post_init__) and from the four failing merge logs; a maintainer with H100 access can confirm with: