[Bugfix] revert PR#1758 which introduced the accuracy problem of qwen3-omni#2009
Merged
Conversation
Signed-off-by: Rein Yang <ruiruyang2@gmail.com>
R2-Y
changed the title
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f903af9324
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
Comment on lines
+196
to
+198
| self.layer0_embed_buffer[:batch_size].copy_(layer0_embed) | ||
| pos_all_layers, current_input = self.code_predictor( | ||
| layer0_code, self.layer0_embed_buffer[:batch_size], last_talker_hidden |
There was a problem hiding this comment.
layer0_embed_buffer is allocated as a plain CPU tensor and then passed directly into self.code_predictor(...); during normal CUDA talker decode, layer0_embed/last_talker_hidden are on GPU, so this introduces mixed-device tensors and can fail at the first torch.cat in Qwen3OmniMoeTalkerCodePredictor.forward. This makes the decode path crash for GPU inference unless everything is forced to CPU.
Useful? React with 👍 / 👎.
Comment on lines
+1126
to
+1128
| logits_cpu = logits.cpu() | ||
| logits_cpu[:, self.suppressed_tokens] = -1e9 | ||
| logits = logits_cpu.to(logits.device) |
There was a problem hiding this comment.
This block copies every talker logits tensor to CPU and back on each decode step, then applies the same suppression again on GPU. The extra host round-trip (logits.cpu() + .to(...)) forces synchronization and moves full-vocab logits across PCIe/host memory, causing a significant throughput regression in production decoding while providing no correctness benefit beyond the following in-place GPU assignment.
Useful? React with 👍 / 👎.
yiliu30
pushed a commit
to yiliu30/vllm-omni-fork
that referenced
this pull request
Mar 20, 2026
…3-omni (vllm-project#2009) Signed-off-by: Rein Yang <ruiruyang2@gmail.com> Signed-off-by: yiliu30 <yi4.liu@intel.com>
hsliuustc0106
added a commit
to hsliuustc0106/vllm-omni-skills
that referenced
this pull request
Mar 22, 2026
### vllm-omni-audio-tts - Source: [PR #2059](vllm-project/vllm-omni#2059) - [BugFix][Qwen3TTS] CodePredictor CudaGraph Pool - Changes: - Bug fix: [BugFix][Qwen3TTS] CodePredictor CudaGraph Pool ### vllm-omni-perf - Source: [PR #2059](vllm-project/vllm-omni#2059) - [BugFix][Qwen3TTS] CodePredictor CudaGraph Pool - Changes: - Bug fix: [BugFix][Qwen3TTS] CodePredictor CudaGraph Pool ### vllm-omni-api - Source: [PR #2058](vllm-project/vllm-omni#2058) - [Bugfix] Fix Fish Speech and CosyVoice3 online serving - missing is_comprehension and broken model detection - Changes: - Bug fix: [Bugfix] Fix Fish Speech and CosyVoice3 online serving - missing is_comprehension and broken model detection ### vllm-omni-contrib - Source: [PR #2045](vllm-project/vllm-omni#2045) - [Voxtral] Improve example ### vllm-omni-cicd - Source: [PR #2045](vllm-project/vllm-omni#2045) - [Voxtral] Improve example ### vllm-omni-api - Source: [PR #2042](vllm-project/vllm-omni#2042) - [bugfix] /chat/completion doesn't read extra_body for diffusion model - Changes: - Bug fix: [bugfix] /chat/completion doesn't read extra_body for diffusion model ### vllm-omni-perf - Source: [PR #2042](vllm-project/vllm-omni#2042) - [bugfix] /chat/completion doesn't read extra_body for diffusion model - Changes: - Bug fix: [bugfix] /chat/completion doesn't read extra_body for diffusion model ### vllm-omni-contrib - Source: [PR #2038](vllm-project/vllm-omni#2038) - [Doc] Update docs and dockerfiles for rebase of vllm v0.18.0 ### vllm-omni-serving - Source: [PR #2037](vllm-project/vllm-omni#2037) - [Rebase] Rebase to vllm v0.18.0 ### vllm-omni-contrib - Source: [PR #2037](vllm-project/vllm-omni#2037) - [Rebase] Rebase to vllm v0.18.0 ### vllm-omni-api - Source: [PR #2037](vllm-project/vllm-omni#2037) - [Rebase] Rebase to vllm v0.18.0 ### vllm-omni-cicd - Source: [PR #2037](vllm-project/vllm-omni#2037) - [Rebase] Rebase to vllm v0.18.0 ### vllm-omni-cicd - Source: [PR #2032](vllm-project/vllm-omni#2032) - [CI] Change Bagel online test environment variable `VLLM_TEST_CLEAN_GPU_MEMORY` to `0` ### vllm-omni-cicd - Source: [PR #2031](vllm-project/vllm-omni#2031) - [CI] Fix test. - Changes: - Bug fix: [CI] Fix test. ### vllm-omni-cicd - Source: [PR #2017](vllm-project/vllm-omni#2017) - [CI] [ROCm] Setup `test-ready.yml` and `test-merge.yml` ### vllm-omni-cicd - Source: [PR #2014](vllm-project/vllm-omni#2014) - [Test] Implement mock HTTP request handling in benchmark CLI tests ### vllm-omni-perf - Source: [PR #2014](vllm-project/vllm-omni#2014) - [Test] Implement mock HTTP request handling in benchmark CLI tests ### vllm-omni-serving - Source: [PR #2012](vllm-project/vllm-omni#2012) - [Fixbug][Perf] Qwen3-omni: code predictor with re-prefill + SDPA and eliminate decode hot-path CPU round-trips - Changes: - Bug fix: [Fixbug][Perf] Qwen3-omni: code predictor with re-prefill + SDPA and eliminate decode hot-path CPU round-trips ### vllm-omni-image-gen - Source: [PR #2012](vllm-project/vllm-omni#2012) - [Fixbug][Perf] Qwen3-omni: code predictor with re-prefill + SDPA and eliminate decode hot-path CPU round-trips - Changes: - Bug fix: [Fixbug][Perf] Qwen3-omni: code predictor with re-prefill + SDPA and eliminate decode hot-path CPU round-trips ### vllm-omni-perf - Source: [PR #2012](vllm-project/vllm-omni#2012) - [Fixbug][Perf] Qwen3-omni: code predictor with re-prefill + SDPA and eliminate decode hot-path CPU round-trips - Changes: - Bug fix: [Fixbug][Perf] Qwen3-omni: code predictor with re-prefill + SDPA and eliminate decode hot-path CPU round-trips ### vllm-omni-serving - Source: [PR #2009](vllm-project/vllm-omni#2009) - [Bugfix] revert PR#1758 which introduced the accuracy problem of qwen3-omni - Changes: - Bug fix: [Bugfix] revert PR#1758 which introduced the accuracy problem of qwen3-omni ### vllm-omni-image-gen - Source: [PR #2007](vllm-project/vllm-omni#2007) - [Bugfix]Fix bug of online server can not return mutli images - Changes: - Bug fix: [Bugfix]Fix bug of online server can not return mutli images - Additions: - Qwen-Image-Layered - Qwen-Image-Layered - Qwen-Image-Layered ### vllm-omni-api - Source: [PR #2007](vllm-project/vllm-omni#2007) - [Bugfix]Fix bug of online server can not return mutli images - Changes: - Bug fix: [Bugfix]Fix bug of online server can not return mutli images ### vllm-omni-cicd - Source: [PR #1998](vllm-project/vllm-omni#1998) - [CI] Split BAGEL tests into dummy/real weight tiers (L2/L3) ### vllm-omni-serving - Source: [PR #1985](vllm-project/vllm-omni#1985) - [Perf] [Qwen3-TTS] Keep audio_codes and last_talker_hidden on GPU to eliminate per-step sync stalls - Changes: - Performance improvement: [Perf] [Qwen3-TTS] Keep audio_codes and last_talker_hidden on GPU to eliminate per-step sync stalls ### vllm-omni-audio-tts - Source: [PR #1985](vllm-project/vllm-omni#1985) - [Perf] [Qwen3-TTS] Keep audio_codes and last_talker_hidden on GPU to eliminate per-step sync stalls - Changes: - Performance improvement: [Perf] [Qwen3-TTS] Keep audio_codes and last_talker_hidden on GPU to eliminate per-step sync stalls ### vllm-omni-perf - Source: [PR #1985](vllm-project/vllm-omni#1985) - [Perf] [Qwen3-TTS] Keep audio_codes and last_talker_hidden on GPU to eliminate per-step sync stalls - Changes: - Performance improvement: [Perf] [Qwen3-TTS] Keep audio_codes and last_talker_hidden on GPU to eliminate per-step sync stalls ### vllm-omni-serving - Source: [PR #1984](vllm-project/vllm-omni#1984) - [CI] [ROCm] Bugfix device environment issue - Changes: - Bug fix: [CI] [ROCm] Bugfix device environment issue ### vllm-omni-api - Source: [PR #1984](vllm-project/vllm-omni#1984) - [CI] [ROCm] Bugfix device environment issue - Changes: - Bug fix: [CI] [ROCm] Bugfix device environment issue ### vllm-omni-serving - Source: [PR #1982](vllm-project/vllm-omni#1982) - [Fix] Fix slow hasattr in CUDAGraphWrapper.__getattr__ - Changes: - Bug fix: [Fix] Fix slow hasattr in CUDAGraphWrapper.__getattr__ ### vllm-omni-cicd - Source: [PR #1982](vllm-project/vllm-omni#1982) - [Fix] Fix slow hasattr in CUDAGraphWrapper.__getattr__ - Changes: - Bug fix: [Fix] Fix slow hasattr in CUDAGraphWrapper.__getattr__ ### vllm-omni-api - Source: [PR #1979](vllm-project/vllm-omni#1979) - [Bugfix] Fix config misalignment between offline and online diffusion inference (Wan2.2, Qwen-Image series) - Changes: - Bug fix: [Bugfix] Fix config misalignment between offline and online diffusion inference (Wan2.2, Qwen-Image series) - Additions: - `/v1/chat/completions` ### vllm-omni-perf - Source: [PR #1979](vllm-project/vllm-omni#1979) - [Bugfix] Fix config misalignment between offline and online diffusion inference (Wan2.2, Qwen-Image series) - Changes: - Bug fix: [Bugfix] Fix config misalignment between offline and online diffusion inference (Wan2.2, Qwen-Image series) ### vllm-omni-contrib - Source: [PR #1976](vllm-project/vllm-omni#1976) - [skip ci][Docs] Update WeChat QR code (fix filename case) - Changes: - Bug fix: [skip ci][Docs] Update WeChat QR code (fix filename case) ### vllm-omni-contrib - Source: [PR #1974](vllm-project/vllm-omni#1974) - [Docs] Update WeChat QR code for community support ### vllm-omni-cicd - Source: [PR #1945](vllm-project/vllm-omni#1945) - Fix Base voice clone streaming quality and stop-token crash - Changes: - Bug fix: Fix Base voice clone streaming quality and stop-token crash ### vllm-omni-cicd - Source: [PR #1938](vllm-project/vllm-omni#1938) - [Test] L4 complete diffusion feature test for Bagel models - Changes: - New feature: [Test] L4 complete diffusion feature test for Bagel models ### vllm-omni-perf - Source: [PR #1938](vllm-project/vllm-omni#1938) - [Test] L4 complete diffusion feature test for Bagel models - Changes: - New feature: [Test] L4 complete diffusion feature test for Bagel models ### vllm-omni-perf - Source: [PR #1934](vllm-project/vllm-omni#1934) - Fix OmniGen2 transformer config loading for HF models - Changes: - Bug fix: Fix OmniGen2 transformer config loading for HF models ### vllm-omni-audio-tts - Source: [PR #1930](vllm-project/vllm-omni#1930) - [Bug][Qwen3TTS][Streaming] remove dynamic initial chunk and only compute on initial request ### vllm-omni-perf - Source: [PR #1930](vllm-project/vllm-omni#1930) - [Bug][Qwen3TTS][Streaming] remove dynamic initial chunk and only compute on initial request ### vllm-omni-audio-tts - Source: [PR #1926](vllm-project/vllm-omni#1926) - [Misc] removed qwen3_tts.py as it is out-dated ### vllm-omni-contrib - Source: [PR #1920](vllm-project/vllm-omni#1920) - [Docs] Add Wan2.1-T2V as supported video generation models - Changes: - New feature: [Docs] Add Wan2.1-T2V as supported video generation models ### vllm-omni-video-gen - Source: [PR #1915](vllm-project/vllm-omni#1915) - [Bugfix] fix helios video generate use cpu device - Changes: - Bug fix: [Bugfix] fix helios video generate use cpu device ### vllm-omni-perf - Source: [PR #1915](vllm-project/vllm-omni#1915) - [Bugfix] fix helios video generate use cpu device - Changes: - Bug fix: [Bugfix] fix helios video generate use cpu device ### vllm-omni-audio-tts - Source: [PR #1913](vllm-project/vllm-omni#1913) - [Optim][Qwen3TTS][CodePredictor] support torch.compile with reduce-overhead and dynamic False ### vllm-omni-perf - Source: [PR #1913](vllm-project/vllm-omni#1913) - [Optim][Qwen3TTS][CodePredictor] support torch.compile with reduce-overhead and dynamic False ### vllm-omni-api - Source: [PR #1908](vllm-project/vllm-omni#1908) - [Entrypoint][Refactor] vLLM-Omni Entrypoint Refactoring ### vllm-omni-perf - Source: [PR #1908](vllm-project/vllm-omni#1908) - [Entrypoint][Refactor] vLLM-Omni Entrypoint Refactoring ### vllm-omni-contrib - Source: [PR #1908](vllm-project/vllm-omni#1908) - [Entrypoint][Refactor] vLLM-Omni Entrypoint Refactoring ### vllm-omni-serving - Source: [PR #1908](vllm-project/vllm-omni#1908) - [Entrypoint][Refactor] vLLM-Omni Entrypoint Refactoring ### vllm-omni-cicd - Source: [PR #1908](vllm-project/vllm-omni#1908) - [Entrypoint][Refactor] vLLM-Omni Entrypoint Refactoring ### vllm-omni-image-gen - Source: [PR #1900](vllm-project/vllm-omni#1900) - [Feat] support HSDP for Flux family - Changes: - New feature: [Feat] support HSDP for Flux family ### vllm-omni-contrib - Source: [PR #1900](vllm-project/vllm-omni#1900) - [Feat] support HSDP for Flux family - Changes: - New feature: [Feat] support HSDP for Flux family ### vllm-omni-distributed - Source: [PR #1898](vllm-project/vllm-omni#1898) - [Feature]: Remove some useless `hf_overrides` in yaml - Changes: - New feature: [Feature]: Remove some useless `hf_overrides` in yaml ### vllm-omni-quantization - Source: [PR #1898](vllm-project/vllm-omni#1898) - [Feature]: Remove some useless `hf_overrides` in yaml - Changes: - New feature: [Feature]: Remove some useless `hf_overrides` in yaml ### vllm-omni-cicd - Source: [PR #1898](vllm-project/vllm-omni#1898) - [Feature]: Remove some useless `hf_overrides` in yaml - Changes: - New feature: [Feature]: Remove some useless `hf_overrides` in yaml ### vllm-omni-perf - Source: [PR #1898](vllm-project/vllm-omni#1898) - [Feature]: Remove some useless `hf_overrides` in yaml - Changes: - New feature: [Feature]: Remove some useless `hf_overrides` in yaml ### vllm-omni-contrib - Source: [PR #1890](vllm-project/vllm-omni#1890) - [NPU] Upgrade to v0.17.0 ### vllm-omni-contrib - Source: [PR #1889](vllm-project/vllm-omni#1889) - Add `Governance` section - Changes: - New feature: Add `Governance` section ### vllm-omni-distributed - Source: [PR #1881](vllm-project/vllm-omni#1881) - [Feat] Support T5 Tensor Parallelism - Changes: - New feature: [Feat] Support T5 Tensor Parallelism ### vllm-omni-cicd - Source: [PR #1881](vllm-project/vllm-omni#1881) - [Feat] Support T5 Tensor Parallelism - Changes: - New feature: [Feat] Support T5 Tensor Parallelism
5 tasks
clodaghwalsh17
pushed a commit
to clodaghwalsh17/nm-vllm-omni-ent
that referenced
this pull request
May 12, 2026
…3-omni (vllm-project#2009) Signed-off-by: Rein Yang <ruiruyang2@gmail.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
2 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.
PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.
Purpose
revert PR#1758 which introduced the accuracy problem of qwen3-omni
Test Plan
Test Result
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model. Please runmkdocs serveto sync the documentation editions to./docs.BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)
@hsliuustc0106 @amy-why-3459 @yenuo26 @LJH-LBJ