[Feature] : Support disaggregated inference pipeline for Qwen3_TTS#1161
Conversation
![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: 5b17195c6c
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".
| else: | ||
| break |
There was a problem hiding this comment.
When omni_connector is enabled and the head waiting request has no initial chunk, the scheduler executes break and exits the waiting loop entirely. That means subsequent waiting requests (including non-connector or already-ready ones) are not scheduled in this cycle, creating head‑of‑line blocking and potential throughput collapse if the first request’s upstream chunk is delayed. This is a behavior change from placeholder scheduling and can stall unrelated traffic until the first chunk arrives. Consider skipping only that request (e.g., move it to skipped_waiting_requests) instead of breaking the whole waiting loop.
Useful? React with 👍 / 👎.
| if finished: | ||
| return { | ||
| "code_predictor_codes": [], | ||
| "codec_streaming": codec_streaming, | ||
| "finished": torch.tensor(True, dtype=torch.bool), |
There was a problem hiding this comment.
The async-chunk adapter returns early when request.is_finished() is true, emitting an empty code_predictor_codes payload and skipping any audio_codes present in the same step. If the request becomes "finished" on the step that produces the last codec frame (common in AR decoding), the final frame is discarded, which will truncate the downstream SpeechTokenizer input. Consider checking/forwarding audio_codes even when finished, or only treating the stop step as finished when no frame is produced.
Useful? React with 👍 / 👎.
|
Good job! How about performance for the current implementation? |
|
@tsdocode Could you please help review? Your vLLM implementation performance is impressive. So I think we can co-work in this PR. Or after we have an initial disaggregated implementation, you may integrate your optimizations based on this implementation in a following-up PR. |
| except Exception: | ||
| fps = None | ||
| if fps is not None and fps > 0: | ||
| self.codec_frame_rate_hz = fps |
There was a problem hiding this comment.
Here config seems to be too model-specific. How about moving it into configuration_qwen3_tts.py like:
@property
def codec_frame_rate_hz(self):
...And in model.py, maybe consider the following generic config:
if self.codec_frame_rate_hz is None:
self.codec_frame_rate_hz = getattr(self.hf_config, "codec_frame_rate_hz", None)
| # per-request additional_information. | ||
| self.talker_mtp_output_key = "audio_codes" | ||
|
|
||
| self.model = Qwen3Model(vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")) |
There was a problem hiding this comment.
One small thing I noticed is RoPE embedding of Talker model is multimodal rope embedding, using the the original Qwen3Model RoPe will cause a slightly difference in the model output which can accumulate error for code predictor model
There was a problem hiding this comment.
I will verify it later. Thanks :D
|
This is my first time looking at vllm-omni so please correct me if I'm wrong, after reading the PR my understanding is:
My question is:
|
1.PagedAttention / PagedKVCache is now supported. However, I still need to test Continuous Batching. I'll test it once my new machine arrives, but you are welcome to test it yourself in the meantime—just modify the settings in the stage config. |
The preliminary test results on 5090 are as follows:
Are there any other metrics worth paying attention to? I can conduct further testing when I am free. And welcome more comments on implementation. I will add more descriptions of design tomorrow as possible. |
Good job! I think we just need one or some correctness test like holding the random state to see if the output is difference between original implementation and the new one. |
Sy0307
force-pushed
the
dev/tts_disaggregation
branch
2 times, most recently
from
aaee00b to
38bad84
Compare
|
Please fix the precommit, thanks |
|
Please modify your PR title. |
Sy0307
changed the title
|
I've fixed some bugs and cleaned up the code. It's now ready for review. Feel free to provide any feedback or raise any questions, and I'll respond and make fixes when I have time. Additionally, I apologize for the previous RTF statistics, which seem to have some errors. I introduced a large number of samples, but some of the generated audio was incorrect, causing the results to deviate from normal values. My preliminary measurements show that the 2-stage RTF should be around 0.30. I will supplement the testing later. |
gcanlin
left a comment
gcanlin
left a comment
There was a problem hiding this comment.
Could we remove modeling_qwen3_tts.py now?
Yes, I have removed legacy code using one single stage for Qwen3_TTS. I also removed qwen3_tts.py as it is out-dated. |
There was a problem hiding this comment.
Pull request overview
Adds a vLLM-native, disaggregated (2-stage) inference pipeline for Qwen3-TTS, enabling Stage-0 AR “Talker” codec generation and Stage-1 “Code2Wav” waveform decoding (including async_chunk streaming integration).
Changes:
- Introduces new Qwen3-TTS stage-specific model implementations: AR Talker + local-KV-cache residual code predictor + Code2Wav decoder.
- Adds/updates stage configs and stage-input processor to stream codec frames across stages via connectors (async_chunk).
- Updates runner/scheduler/connector + OpenAI speech endpoint plumbing to support codec-token prompting and per-chunk metadata propagation.
Reviewed changes
Copilot reviewed 18 out of 18 changed files in this pull request and generated 17 comments.
Show a summary per file
| File | Description |
|---|---|
vllm_omni/worker/gpu_model_runner.py |
Generalizes GPU-side talker_mtp handling for TTS and adjusts preprocess/embed plumbing for stage execution. |
vllm_omni/model_executor/stage_input_processors/qwen3_tts.py |
New async_chunk adapter that windows codec frames and packs them for Stage-1 decoding. |
vllm_omni/model_executor/stage_configs/qwen3_tts.yaml |
Updates default Qwen3-TTS stage config to a 2-stage async_chunk pipeline with shared-memory connector bindings. |
vllm_omni/model_executor/stage_configs/qwen3_tts_talker_speech_tokenizer_async_chunk.yaml |
Adds a dedicated async_chunk stage config variant for Talker→SpeechTokenizer decoding. |
vllm_omni/model_executor/models/registry.py |
Switches registry entries to stage-specific Qwen3-TTS architectures (Talker / Code2Wav). |
vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_tokenizer.py |
Adds optional feature-extractor loading and relaxes encode typing for stage-specific usage. |
vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_talker.py |
New vLLM-native AR Talker model with preprocess/postprocess and GPU-side talker_mtp residual prediction. |
vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_code_predictor_vllm.py |
New lightweight residual code predictor with an independent local KV cache and prefill/decode logits APIs. |
vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_code2wav.py |
New Stage-1 “decode-only” model that turns codec windows into waveform via SpeechTokenizer. |
vllm_omni/model_executor/models/qwen3_tts/qwen3_tts.py |
Removes the previous HF-style monolithic generation wrapper. |
vllm_omni/model_executor/models/qwen3_tts/processing_qwen3_tts.py |
Removes HF processor wrapper (no longer used by the new vLLM-native pipeline). |
vllm_omni/model_executor/models/qwen3_tts/modeling_qwen3_tts.py |
Removes the large HF reference modeling implementation. |
vllm_omni/model_executor/models/qwen3_tts/configuration_qwen3_tts.py |
Adjusts dummy vision token IDs and stage-specific config behavior (mrope stripping for Code2Wav). |
vllm_omni/entrypoints/openai/serving_speech.py |
Switches TTS requests to prompt_token_ids placeholders + prompt-length estimation; adds streaming chunk concat. |
vllm_omni/entrypoints/async_omni.py |
Ensures next-stage placeholder prompt length is at least 1 token. |
vllm_omni/distributed/omni_connectors/adapter.py |
Updates chunk ingestion to support Code2Wav header + codec code windows and forwards per-chunk metadata. |
vllm_omni/core/sched/omni_generation_scheduler.py |
async_chunk-aware scheduling: only pull new chunks once prior tokens are consumed; propagate per-step additional_information. |
vllm_omni/config/model.py |
Adds codec_frame_rate_hz plumbing from HF config into Omni model config. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| "Qwen3TTSModelForGeneration", | ||
| "qwen3_tts_code2wav", | ||
| "Qwen3TTSCode2Wav", | ||
| ), |
There was a problem hiding this comment.
The registry drops Qwen3TTSForConditionalGeneration in favor of stage-specific Qwen3TTSTalkerForConditionalGeneration/Qwen3TTSCode2Wav, but there are still repo configs referencing the old architecture (e.g. vllm_omni/platforms/npu/stage_configs/qwen3_tts.yaml uses model_arch: Qwen3TTSForConditionalGeneration). Those configs will no longer resolve; please update them (or provide an alias) to avoid broken non-GPU stage configs.
| ), | |
| ), | |
| # Backward-compatible alias for configs still using the generic Qwen3 TTS arch name. | |
| "Qwen3TTSForConditionalGeneration": ( | |
| "qwen3_tts", | |
| "qwen3_tts_talker", | |
| "Qwen3TTSTalkerForConditionalGeneration", | |
| ), |
| """Decode one new token for residual group `generation_step` (1..Q-1).""" | ||
| self._maybe_init_kv_cache(input_ids.device) | ||
| assert self._kv_cache is not None | ||
| bsz = int(input_ids.shape[0]) | ||
| if generation_step <= 0: | ||
| raise ValueError("generation_step must be >= 1 for decode_logits") |
There was a problem hiding this comment.
decode_logits is documented as accepting generation_step in the range (1..Q-1), but the implementation indexes self.lm_head[generation_step] and embed_idx = generation_step - 1. With the current prefill+decode calling pattern, passing generation_step == len(self.lm_head) would raise an IndexError. Please either tighten the docstring to the actually supported range and/or add an explicit upper-bound check to fail with a clear error message.
| """Decode one new token for residual group `generation_step` (1..Q-1).""" | |
| self._maybe_init_kv_cache(input_ids.device) | |
| assert self._kv_cache is not None | |
| bsz = int(input_ids.shape[0]) | |
| if generation_step <= 0: | |
| raise ValueError("generation_step must be >= 1 for decode_logits") | |
| """Decode one new token for residual group `generation_step` (1..len(self.lm_head)-1). | |
| Note: | |
| - generation_step == 0 is reserved for prefill_logits (uses lm_head[0]). | |
| - Valid decode steps are 1 through len(self.lm_head) - 1, inclusive. | |
| """ | |
| self._maybe_init_kv_cache(input_ids.device) | |
| assert self._kv_cache is not None | |
| bsz = int(input_ids.shape[0]) | |
| max_step = len(self.lm_head) - 1 | |
| if generation_step < 1 or generation_step > max_step: | |
| raise ValueError( | |
| f"generation_step must be in [1, {max_step}] for decode_logits; " | |
| f"got {generation_step}" | |
| ) |
| def _load_audio_to_np(self, x: str) -> tuple[np.ndarray, int]: | ||
| import librosa | ||
|
|
||
| if self._is_url(x): | ||
| with urlopen(x) as resp: | ||
| audio_bytes = resp.read() | ||
| with io.BytesIO(audio_bytes) as f: | ||
| audio, sr = sf.read(f, dtype="float32", always_2d=False) | ||
| elif self._is_probably_base64(x): |
There was a problem hiding this comment.
_load_audio_to_np fetches arbitrary URLs via urlopen with no timeout, size limit, or allowlisting. Since ref_audio can come from API inputs (Base task), this is a concrete SSRF / DoS risk (e.g., internal network access or huge downloads). Consider disabling URL fetching by default, adding an explicit allowlist, enforcing https, and setting timeouts + maximum download size.
| # Must use prompt_token_ids (not text prompt): the AR Talker | ||
| # operates on codec tokens; text token IDs exceed codec vocab. | ||
| # model.preprocess replaces all embeddings, so placeholder value | ||
| # is irrelevant -- but length must match to avoid excess padding. | ||
| tts_params = self._build_tts_params(request) | ||
| prompt_text = self._build_tts_prompt(request.input) | ||
| ph_len = self._estimate_prompt_len(tts_params) | ||
| prompt = { | ||
| "prompt": prompt_text, | ||
| "prompt_token_ids": [1] * ph_len, | ||
| "additional_information": tts_params, | ||
| } |
There was a problem hiding this comment.
tests/entrypoints/openai_api/test_serving_speech.py currently calls OmniOpenAIServingSpeech._build_tts_prompt(), but this method was removed and TTS prompting is now done via prompt_token_ids. This will break unit tests (and any downstream callers relying on the old helper); either restore a compatibility wrapper or update the tests/callers to the new prompt_token_ids-based flow.
| gpu_memory_utilization: 0.3 | ||
| distributed_executor_backend: "mp" | ||
| max_num_batched_tokens: 512 | ||
| max_model_len: 4096 | ||
| # Stage-0 emits flattened codec codes via async_chunk connector. | ||
| custom_process_next_stage_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_tts.talker2code2wav_async_chunk | ||
| default_sampling_params: | ||
| temperature: 0.9 | ||
| top_k: 50 | ||
| max_tokens: 4096 | ||
| seed: 42 | ||
| detokenize: false | ||
| repetition_penalty: 1.05 | ||
| stop_token_ids: [2150] | ||
|
|
||
| - stage_id: 1 | ||
| stage_type: llm | ||
| runtime: | ||
| devices: "0" | ||
| max_batch_size: 1 | ||
| engine_args: | ||
| model_stage: code2wav | ||
| model_arch: Qwen3TTSCode2Wav | ||
| hf_overrides: | ||
| architectures: [Qwen3TTSCode2Wav] | ||
| # Stage-1 has no main checkpoint weights (SpeechTokenizer is loaded from | ||
| # `speech_tokenizer/` lazily). Avoid probing for model.safetensors. | ||
| load_format: dummy | ||
| worker_type: generation | ||
| scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler | ||
| enforce_eager: true | ||
| trust_remote_code: true | ||
| async_scheduling: false | ||
| enable_prefix_caching: false | ||
| engine_output_type: audio | ||
| gpu_memory_utilization: 0.2 | ||
| distributed_executor_backend: "mp" | ||
| # Must be >= num_code_groups * (codec_left_context_frames + codec_chunk_frames). | ||
| max_num_batched_tokens: 8192 | ||
| # async_chunk appends windows per step; max_model_len must cover accumulated stream. | ||
| max_model_len: 32768 | ||
| engine_input_source: [0] | ||
| final_output: true | ||
| final_output_type: audio |
There was a problem hiding this comment.
This async_chunk stage config defines runtime.connectors.connector_of_shared_memory, but neither stage declares output_connectors / input_connectors to bind that named connector to the edge (unlike model_executor/stage_configs/qwen3_tts.yaml, which explicitly maps to_stage_1 / from_stage_0). If the runtime expects explicit connector mapping, the extra codec streaming settings may not be applied; consider adding the connector mappings for stage 0/1 for consistency.
| head, | ||
| ) | ||
| except Exception: | ||
| pass |
There was a problem hiding this comment.
'except' clause does nothing but pass and there is no explanatory comment.
| pass | |
| # Codec statistics logging is best-effort; failures must not break decoding. | |
| logger.debug("Failed to compute or log Code2Wav codec statistics.", exc_info=True) |
| if arr.size > 0: | ||
| wav_candidates.append(arr) | ||
| return | ||
| except Exception: |
There was a problem hiding this comment.
'except' clause does nothing but pass and there is no explanatory comment.
| if arr.size > 0: | ||
| wav_candidates.append(arr) | ||
| return | ||
| except Exception: |
There was a problem hiding this comment.
'except' clause does nothing but pass and there is no explanatory comment.
| return int(x.numel()) | ||
| if isinstance(x, list): | ||
| return int(len(x)) | ||
| except Exception: |
There was a problem hiding this comment.
'except' clause does nothing but pass and there is no explanatory comment.
| except Exception: | |
| except Exception: | |
| # If length computation fails for any reason, treat it as zero-length. |
| info_update["talker_prefill_offset"] = int(offset + span_len) | ||
| else: | ||
| # Subsequent prefill chunk: slice from our own running offset. | ||
| if not isinstance(prompt_embeds_cpu, torch.Tensor) or prompt_embeds_cpu.ndim != 2: |
There was a problem hiding this comment.
This statement is unreachable.
|
I ran this end-to-end on H200s and the online serving path works well! A few things need fixing before merge though. Critical:
Other suggestions to consider:
Runtime testingComparison against the old monolithic single-stage pipeline (same hardware, same model):
RTF = latency / audio duration, below 1.0 is faster than real-time. Audio durations vary slightly between runs due to stochastic generation (temperature 0.9). The old pipeline sits at ~1.0x RTF (barely real-time). The disaggregated pipeline runs 0.56-0.80x RTF, with larger speedups on longer text thanks to the async chunk streaming between stages. The PR also adds instruction/emotion control which the old pipeline doesn't support (returns 400 on the instructions field). One recurring warning: |
|
|
||
| num_computed_tokens = request.num_computed_tokens | ||
| required_tokens = max(len(request.prompt_token_ids) - num_computed_tokens, 1) | ||
| required_tokens = len(request.prompt_token_ids) - num_computed_tokens |
There was a problem hiding this comment.
Why do we need this judgment? Or can we abstract it into a function?
There was a problem hiding this comment.
In async_chunk mode (the Code2Wav stage of qwen3-tts and qwen3-omni), prompt_token_ids is dynamically appended chunk by chunk by the upstream component. When the upstream has not produced a new chunk yet, len(prompt_token_ids) == num_computed_tokens, so required_tokens should be 0, meaning “no new data to process yet.” In this case, the scheduler should skip this request (i.e., continue) and check again in the next iteration.
If we keep max(..., 1), then when there is no new chunk it will still force scheduling one placeholder token, causing the model to run inference repeatedly on already-processed data and resulting in incorrect audio output.
There was a problem hiding this comment.
I think this is unnecessary. In our design, if no new block is produced upstream, the request should not execute this round of scheduling.
There was a problem hiding this comment.
On the other hand, when enabling or disabling async_chunk, we need to minimize intrusive modifications to the scheduler.
There was a problem hiding this comment.
Maybe we should move it to adapter.py or other part by using other method? If max(...,1) make efforts then Qwen3_TTS will be frozen when receive a request. Your thought?
There was a problem hiding this comment.
After talking with @amy-why-3459 , this part of TTS will follow #951. WIP.
Signed-off-by: Sy03 <1370724210@qq.com>
|
Tested E2E offline inference on the latest revision (b6c1928 + 8fbf458). The disaggregated pipeline works, both stages initialize fine, codec streaming via SharedMemoryConnector is functional, and Code2Wav produces intelligible audio at RTF=0.76. Nice work! I did run into a few issues with the offline example ( First, the example only passes 1 SamplingParams but the 2-stage pipeline expects 2 (one per stage), so it throws Second, the SamplingParams in the example is missing Third, the example uses text prompts ( The first two are quick fixes. The third one just needs the example to use |
…put handling in offline test Signed-off-by: Sy03 <1370724210@qq.com>
Fix offline tests by following your advice and thanks a lot. |
|
Re-tested with 5530aaa. The updated end2end.py works correctly now, got clean 4.72s audio with CUDA graphs enabled. LGTM! |
…wen3_tts Signed-off-by: Sy03 <1370724210@qq.com>
Signed-off-by: Sy03 <1370724210@qq.com>
…l runners to prevent mutation Signed-off-by: Sy03 <1370724210@qq.com>
Issue AnalysisDue to a previous modification ( vLLM v0.16 enables
Otherwise, the following error will occur: a SolutionTo avoid the output holding mutable references and prevent the held objects from being corrupted, we make the following modification in b6e6972: req_ids_output_copy = self.input_batch.req_ids.copy()
req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy() |
| enforce_eager: true | ||
| trust_remote_code: true | ||
| async_scheduling: false | ||
| enable_prefix_caching: false |
There was a problem hiding this comment.
BTW, you could disable async schedule here. I have met this issue before. And I found that Qwen3-Omni have disabled it in stage-2. So we didn't add the copy in model runner. But copying may be more general.
cc @amy-why-3459 @tzhouam
There was a problem hiding this comment.
Yes, copying here follows how vllm do.vllm code
There was a problem hiding this comment.
Oh, sorry. I missed async_scheduling: false. So even if you have disabled async_scheduling, the key error still happened?
There was a problem hiding this comment.
This error happened in qwen 2.5 omni CI test. And in qwen2_5_omni_ci.yaml we do not have async_scheduling: false (BTW, I found qwen2_5_omni.yaml has such setting.
Signed-off-by: Sy03 <1370724210@qq.com>
| @@ -0,0 +1,92 @@ | |||
| async_chunk: true | |||
There was a problem hiding this comment.
What's the difference between qwen3_tts_talker_speech_tokenizer_async_chunk.yaml and qwen3_tts.yaml? And this PR makes async chunk on by default. But from what I knew, we haven't supported offline async chunk. Could we? Or, this PR also takes this feature in?
There was a problem hiding this comment.
qwen3_tts_talker_speech_tokenizer_async_chunk.yaml should be the 2stage config files for preliminary test to distinguish with qwen3_tts.yaml. But I should delete it now.
Not support offline async chunk but I am working on this related issue #1193 . I think it will be implemented in another PR.
…speech tokenizer async chunk. Signed-off-by: Sy03 <1370724210@qq.com>
* [Frontend][Model] Support batch request with refined OmniDiffusionReq… (#797) Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com> * [Model]: add FLUX.1-dev model (#853) * [BugFix] ignore mm data from stages to async omni (#954) Signed-off-by: dengyunyang <584797741@qq.com> * Revert "[BugFix] ignore mm data from stages to async omni" (#1023) * [Bugfix] Modify output to model_runner_output (#1026) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [Feature] Support cache-dit for Wan 2.2 inference (#1021) Signed-off-by: samithuang <285365963@qq.com> Signed-off-by: Samit <285365963@qq.com> * [Doc]Format profiling doc (#993) Signed-off-by: lishunyang <lishunyang12@163.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * [Hardware] Support platforms and plugin system (#774) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [Core]: KV Cache Transfer Encapsulation (#979) Signed-off-by: princepride <wangzhipeng628@gmail.com> Signed-off-by: 汪志鹏 <wangzhipeng628@gmail.com> * [Test]Delete skip mark for amd ci test and fix CI failure (#927) Signed-off-by: wangyu31577 <wangyu31577@hundsun.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: wangyu31577 <wangyu31577@hundsun.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Bugfix][Doc]Specify Qwen3-TTS model name for each task type (#1036) Signed-off-by: Kyle Huang <yellowsea@gmail.com> * [Misc] pin version of fa3-fwd (#1051) Signed-off-by: zjy0516 <riverclouds.zhu@qq.com> * [CI] [ROCm] Add more AMD CI tests (#1039) Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> * [Bugfix] fix qwen image layerd in dummy run (#1027) Signed-off-by: zjy0516 <riverclouds.zhu@qq.com> * [BugFix] Fix noisy output without setting a seed in Qwen Image (#1043) Signed-off-by: natureofnature <wzliu@connect.hku.hk> * [bugfix] remove vllm speech route (#1060) Signed-off-by: linyueqian <linyueqian@outlook.com> * [Debug] Update GLM-Image Pipeline (#1049) Co-authored-by: root <root@hk01dgx028.cm.cluster> * [Diffusion][Bugfix] Fix the flash_attn backends selection logic (#983) Signed-off-by: mxuax <mxuax@connect.ust.hk> Signed-off-by: XU Mingshi <91017482+mxuax@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [BugFix] Fix the accuracy issue of multimodal input. (#1020) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com> Co-authored-by: Rein Yang <ruiruyang2@gmail.com> * [Bugfix] Set VaeImageProcessor `do_convert_rgb` True (#1032) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [feat]: adapt batch request for flux (#1028) Signed-off-by: wuzhongjian wuzhongjian_yewu@cmss.chinamobile.com * [CI] Change Qwen3 Omni stage placement strategy (#1072) Signed-off-by: ZeldaHuang <hzm414167@alibaba-inc.com> * [BugFix] Fix to use correct attn backend (#1038) Signed-off-by: Divyansh Singhvi <divyanshsinghvi@gmail.com> * [Perf] Qwen3 Omni talker mtp optimization (#1005) Signed-off-by: ZeldaHuang <hzm414167@alibaba-inc.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Wan2.2] Optimize memory usage with conditional transformer loading (#980) Signed-off-by: Lin, Fanli <fanli.lin@intel.com> Signed-off-by: Samit <285365963@qq.com> Co-authored-by: Samit <285365963@qq.com> * [Feat] Support XPU Backend in vLLM-Omni (#191) Signed-off-by: Fanli Lin <fanli.lin@intel.com> Signed-off-by: Fanli Lin <fanli0116@gmail.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * [Fix] stabilize diffusion images LoRA E2E across CI drift (#1075) Signed-off-by: dongbo910220 <1275604947@qq.com> * [Bugfix][Test] Re-enable the log simple tests (#1065) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [Bugfix] pr conflict fix, bugfix ignore mm data from stages to async omni (#1025) Signed-off-by: dengyunyang <584797741@qq.com> * [Doc][Bagel] Add BAGEL-7B-MoT documentation and edit the default stage configuration (#987) Signed-off-by: Ding Zuhao <e1583181@u.nus.edu> Signed-off-by: jzz <e1583181@u.nus.edu> * [Fix] Increase max wait time for server readiness to accommodate model loading (#1089) Signed-off-by: Andy Zhou <46011930+AndyZhou952@users.noreply.github.com> * [Benchmark] Add vLLM-Omni Omni model online benchmark (#780) Signed-off-by: wangyu31577 <wangyu31577@hundsun.com> Signed-off-by: wangyu <53896905+yenuo26@users.noreply.github.com> Co-authored-by: wangyu31577 <wangyu31577@hundsun.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Bugfix] Remove Mooncake/Yuanrong connector import warning (#1091) Signed-off-by: natureofnature <wzliu@connect.hku.hk> * fix: UnboundLocalError for role in streaming audio/image responses (#784) Signed-off-by: Pierre Le Guen <26087574+PierreLeGuen@users.noreply.github.com> * [Misc] update wechat image (#1096) * [Feature] Support DiT Layerwise (Blockwise) CPU Offloading (#858) Signed-off-by: yuanheng <jonathan.zhaoyh@gmail.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * [BugFix] Modify max_tokens and modify the log and fix #1103 (#1097) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [BugFix] Fix modulate_index shape error in Qwen-Image-Edit Task (#1100) Signed-off-by: mxuax <mxuax@connect.ust.hk> Signed-off-by: XU Mingshi <91017482+mxuax@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Platform] Add supports_torch_inductor interface (#1108) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [BugFix] Fix Qwen3 Omni talker mtp torch.compile startup error (#1104) Signed-off-by: ram16g <anlianfengjie@163.com> Signed-off-by: ZeldaHuang <hzm414167@alibaba-inc.com> Co-authored-by: ram16g <anlianfengjie@163.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Bugfix] fix request_id of image generation in api server (#1112) Signed-off-by: zjy0516 <riverclouds.zhu@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Perf]: CFG parallel abstraction (#851) Signed-off-by: Didan Deng <33117903+wtomin@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [BugFix] Fix Qwen3 TTS 0.6B profile run hang (#995) (#1082) * [CI] [ROCm] Quick fix amd ci (#1116) Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> * [Bugfix] fix benchmark audio timing error and add benchmark test (#1109) Signed-off-by: wangyu31577 <wangyu31577@hundsun.com> Co-authored-by: wangyu31577 <wangyu31577@hundsun.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Bugfix][Qwen3TTS] Load speaker_id/voices from model configuration (#1079) Signed-off-by: pablo <juanz9312@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com> * [NPU] Align with GPUModelRunner (#1114) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [FEATURE] /v1/images/edit interface (#1101) Signed-off-by: dengyunyang <584797741@qq.com> * [Bugfix] Fix NPU SDPA attention mask shape and semantics (#1031) Signed-off-by: gcanlin <canlinguosdu@gmail.com> Co-authored-by: muziyuhui666 <111362884+muziyuhui666@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [TeaCache]: Add Coefficient Estimation (#940) Signed-off-by: princepride <wangzhipeng628@gmail.com> Signed-off-by: 汪志鹏 <wangzhipeng628@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [CI]: Bagel E2E Smoked Test (#1074) Signed-off-by: princepride <wangzhipeng628@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Misc] Bump version to 0.14.0 (#1128) Signed-off-by: Roger Wang <hey@rogerw.io> * [Doc] First stable release of vLLM-Omni (#1129) Signed-off-by: Roger Wang <hey@rogerw.io> * [Misc] Align error handling with upstream vLLM v0.14.0 (#1122) Signed-off-by: anna <lee.anna@navercorp.com> Co-authored-by: anna <lee.anna@navercorp.com> * [Feature] add Tensor Parallelism to LongCat-Image(-Edit) (#926) Signed-off-by: Rustam Khadipash <16683750+hadipash@users.noreply.github.com> * [CI] Temporarily remove slow tests. (#1143) Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> Signed-off-by: princepride <wangzhipeng628@gmail.com> Co-authored-by: princepride <wangzhipeng628@gmail.com> * [CI] Refactor test_sequence_parallel.py and add a warmup run for more accurate performance stat (#1165) Signed-off-by: mxuax <mxuax@connect.ust.hk> Signed-off-by: XU Mingshi <91017482+mxuax@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * Dev/rebase v0.15.0 (#1159) Signed-off-by: Taichang Zhou <tzhouam@connect.ust.hk> Signed-off-by: tzhouam <tzhouam@connect.ust.hk> Signed-off-by: princepride <wangzhipeng628@gmail.com> Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com> * Docs update paper link (#1169) Signed-off-by: hsliu <liuhongsheng4@huawei.com> Signed-off-by: hsliu_ustc <hsliu_ustc@noreply.gitcode.com> Co-authored-by: hsliu_ustc <hsliu_ustc@noreply.gitcode.com> * [Debug] Clear Dockerfile.ci to accelerate build image (#1172) Signed-off-by: tzhouam <tzhouam@connect.ust.hk> * [Debug] Correct Unreasonable Long Timeout (#1175) Signed-off-by: tzhouam <tzhouam@connect.ust.hk> * [Doc]Fix - Align with repo. (#1176) Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> * [Bugfix][Qwen-Image-Edit] Add a warning log for none negative_prompt (#1170) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [Bugfix] fix qwen image oom (#1168) Signed-off-by: zjy0516 <riverclouds.zhu@qq.com> * [Hardware] Disable compile of diffusion on XPU (#1148) Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com> * [Doc] Fix vLLM version in user docs (#1179) Signed-off-by: Yuanheng Zhao <jonathan.zhaoyh@gmail.com> * [Refactor] Refactor async chunk and fix the shape mismatch issue (#1151) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com> * bugfix: /images/edits endpoint fails pipeline data format check (#1141) Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Perf] resolving prolonged `cudastreamsynchronize` execution in z image processing (#1105) Signed-off-by: erfgss <97771661+erfgss@users.noreply.github.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * [Bugfix] modify RTF use audio_e2e/audio_duration (#1157) Signed-off-by: wangyu31577 <wangyu31577@hundsun.com> Co-authored-by: wangyu31577 <wangyu31577@hundsun.com> * [Doc] Highlight paper & slides. (#1186) Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> * [chore] Remove zmq context initialize (#1187) Signed-off-by: xiedeyantu <czjourney@163.com> * [NPU] Update Dockerfile and docs for v0.14.0 (#671) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [Bugfix] E2E metric incorrect qwen3-omni with async chunk feature (#1018) Signed-off-by: Junhong Liu <98734602+LJH-LBJ@users.noreply.github.com> Signed-off-by: Junhong Liu <ljh_lbj@163.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Doc] opt doc (#1118) Signed-off-by: David Chen <530634352@qq.com> * [Bugfix] Fix tp+sp accuracy, incorrect process group mapping (#1178) Signed-off-by: David Chen <530634352@qq.com> * [Feature] Enable use_audio_in_video for Qwen 3 Omni Online (#1198) Signed-off-by: tzhouam <tzhouam@connect.ust.hk> * [Bugfix] async_chunk rebase v0.15.0 (#1195) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com> * [feature]: support flux cache_dit (#1145) Co-authored-by: Jiangyun Zhu <riverclouds.zhu@qq.com> * [CI] Add CI branch coverage calculation, fix statement coverage results and add log before test for buildkite log group (#1120) Signed-off-by: wangyu31577 <wangyu31577@hundsun.com> Co-authored-by: wangyu31577 <wangyu31577@hundsun.com> * [Wan 2.2][Diffusion] Add TP Support (#964) Signed-off-by: weichen <calvin_zhu0210@outlook.com> * [Hardware] [Feat] Setup platform dependent package installation (#1046) Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> Co-authored-by: PopSoda2002 <zhouhp.me@gmail.com> Co-authored-by: gcanlin <canlinguosdu@gmail.com> * [XPU] Fix XPU UTs for basic coverage (#1164) Signed-off-by: Yan Ma <yan.ma@intel.com> * [Test] Add BuildKite test-full script for full CI. (#867) Signed-off-by: wangyu31577 <wangyu31577@hundsun.com> Co-authored-by: wangyu31577 <wangyu31577@hundsun.com> * [Refactor] Reuse upstream Qwen3MoeSparseMoeBlock (#1202) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [Bugfix] Fix wan2.2 ti2v (#1221) Signed-off-by: mxuax <mxuax@connect.ust.hk> Signed-off-by: XU Mingshi <91017482+mxuax@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Bugfix] Fix '--max-generated-image-size' cli args type (#1249) Signed-off-by: ApsarasX <apsarax@outlook.com> * [Bugfix] Ensure seed=0 is correctly handled in image edit (#1248) Signed-off-by: ApsarasX <apsarax@outlook.com> * [Docs] Add example image download step to Image-To-Video examples (#1258) Signed-off-by: lishunyang <lishunyang12@163.com> * [Bugfix] Fix padding bug in 12Hz tokenizer ConvTranspose1d decode (#1241) Signed-off-by: linyueqian <linyueqian@outlook.com> * [bugfix] Fix multimodal_output property to check completion outputs where audio data is attached (#1203) Signed-off-by: linyueqian <linyueqian@outlook.com> * [Doc] Update QA relevant to quantization (#1257) Signed-off-by: lishunyang <lishunyang12@163.com> * [Bugfix] Fix Doc link Rrror (#1263) Signed-off-by: lishunyang <lishunyang12@163.com> * Process-Scoped GPU Memory Accounting (#1204) Signed-off-by: Divyansh Singhvi <divyanshsinghvi@gmail.com> * [ComfyUI]: ComfyUI integration (#1113) Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com> * fix: add diffusion offload args to OmniConfig group instead of serve_parser (#1271) Signed-off-by: Chenguang ZHENG <645327136@qq.com> * [Doc] Adding models/pipelines/features Tutorial (#1196) Signed-off-by: Didan Deng <33117903+wtomin@users.noreply.github.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Co-authored-by: dongbo910220 <32610838+dongbo910220@users.noreply.github.com> * [CI] Add env variable check for nightly CI (#1281) Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> * [CI] Add pytest markers to current tests and update the doc. (#577) Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Diffusion][Perf] Remove Redundant Communication Cost by Refining SP Hook Design (#1275) Signed-off-by: mxuax <mxuax@connect.ust.hk> Signed-off-by: XU Mingshi <91017482+mxuax@users.noreply.github.com> * [Feature] Opt metrics structure (#891) Signed-off-by: Junhong Liu <98734602+LJH-LBJ@users.noreply.github.com> Signed-off-by: Junhong Liu <ljh_lbj@163.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Test] Add example test cases for omni online (#1086) Signed-off-by: wangyu31577 <wangyu31577@hundsun.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Signed-off-by: yenuo26 <410167048@qq.com> Co-authored-by: wangyu31577 <wangyu31577@hundsun.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * [CI] Reduce the time for Diffusion Sequence Parallelism Test (#1283) Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> * [Model] SupportHunyuanImage3 Diffusion Model in vllm-omni (#1085) Signed-off-by: Semmer2 <semmer@live.cn> * [Chore] Update copyright year. (#1256) Signed-off-by: lishunyang <lishunyang12@163.com> * [feature]: support Flux.1-dev CFG-Parallel (#1269) * [Bugfix] Fix 'NoneType' AttributeError in stable-diffusion model detect (#1254) Signed-off-by: Yan Ma <yan.ma@intel.com> * [Doc] Update Qwen3-TTS docs for consistency with Omni examples (#1226) Signed-off-by: linyueqian <linyueqian@outlook.com> Signed-off-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Fix]Ensure HuggingFace downloads complete before initialization. (#1213) Signed-off-by: zhou zhuoxin <zhouzhuoxin1508@outlook.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * [BugFix] Fixed the issue where ignore_eos was not working. (#1286) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com> * [Test] Add e2e tests for Qwen3-TTS speech endpoint (#1206) Signed-off-by: linyueqian <linyueqian@outlook.com> Signed-off-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com> * [Feat]: support VAE patch parallelism (#756) Signed-off-by: dongbo910220 <1275604947@qq.com> Co-authored-by: hsliuustc0106 <liuhongsheng4@huawei.com> * [CI] Disable Qwen3-TTS E2E Test in pipeline.yml (#1306) Signed-off-by: Gao Han <hgaoaf@connect.ust.hk> * [Misc] Add per-request generator_device to online image gen and edit (#1183) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [Bagel]: Support TP (#1293) Signed-off-by: princepride <wangzhipeng628@gmail.com> * [Bugfix] Fix image edit RoPE crash when explicit height/width are provided (#1265) Signed-off-by: lishunyang <lishunyang12@163.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Doc] Sync (#1216) Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> * [Bugfix] fix precision issues of qwen3-omni when enable async_chunk without system prompt (#1288) Signed-off-by: Rein Yang <ruiruyang2@gmail.com> * [Debug] Add trigger to concurrent stage init (#1274) Signed-off-by: tzhouam <tzhouam@connect.ust.hk> * [Bugfix][Qwen3-TTS] Fix task type (#1317) Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com> * Unifying CLI Argument Naming Style (#1309) Signed-off-by: Didan Deng <33117903+wtomin@users.noreply.github.com> * [Bugfix][Qwen3-TTS] Preserve original model ID in omni_snapshot_download (#1318) * [CI] Run nightly tests. (#1333) Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> * [Feature]: FP8 Quantization Support for DiT (#1034) Signed-off-by: lishunyang <lishunyang12@163.com> Signed-off-by: SYLAR <125541396+lishunyang12@users.noreply.github.com> * Fix yield token metrics and opt metrics record stats (#1292) * [Test] L2 & L3 Test Case Stratification Design for Omni Model (#1272) Signed-off-by: wangyu31577 <wangyu31577@hundsun.com> Signed-off-by: yenuo26 <410167048@qq.com> Signed-off-by: wangyu <53896905+yenuo26@users.noreply.github.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: wangyu31577 <wangyu31577@hundsun.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Pref] Support Qwen3 Omni code2wav batch infernce with async chunk (#1246) Signed-off-by: ZeldaHuang <hzm414167@alibaba-inc.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Signed-off-by: Ziming Huang <1520787127@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * update qwen3-omni & qwen2.5-onmi openai client (#1304) Signed-off-by: Rein Yang <ruiruyang2@gmail.com> * [Feature] Support Wan2.2 T2V and I2V Online Serving with OpenAI /v1/videos API (#1073) Signed-off-by: samithuang <285365963@qq.com> Signed-off-by: Samit <285365963@qq.com> Signed-off-by: SamitHuang <285365963@qq.com> Co-authored-by: Flora Feng <4florafeng@gmail.com> * [Feature] add Tensor Parallelism to SD_3.5 (#1336) Signed-off-by: GG-li <3226868735@qq.com> * [Feature]async scheduling to overlap chunk IO and compute (#951) Signed-off-by: CHEN <116010019@link.cuhk.edu.cn> Co-authored-by: Bhanu068 <voutharoja.bhanu06@gmail.com> Co-authored-by: Gao Han <gaohan19@huawei.com> * [Bugfix] reused metrics to modify the API Server token statistics in Stream Response (#1301) Signed-off-by: John Liu BUAA <liukecheng97@gmail.com> * Refactor CPU Offloading Backend Pattern (#1223) Signed-off-by: yuanheng <jonathan.zhaoyh@gmail.com> Signed-off-by: Yuanheng Zhao <jonathan.zhaoyh@gmail.com> Signed-off-by: Samit <285365963@qq.com> Co-authored-by: Samit <285365963@qq.com> * [DOC] Doc for CI test - Details about five level stucture and some other files. (#1167) Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> Co-authored-by: yenuo26 <410167048@qq.com> * [Bugfix] remove Tongyi-MAI/Z-Image-Turbo related test from L2 ci (#1348) Signed-off-by: dengyunyang <584797741@qq.com> * [Misc] wechat image update (#1354) Signed-off-by: David Chen <530634352@qq.com> * [Misc] Support WorkerWrapperBase and CustomPipeline for Diffusion Worker (#764) Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com> * [Feature][Bugfix] Add CFG feature to Bagel (#1310) Signed-off-by: Ding Zuhao <e1583181@u.nus.edu> Signed-off-by: 汪志鹏 <wangzhipeng628@gmail.com> Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com> * [Feature]: Diffusion sleep to use process level memory calculation (#1276) Signed-off-by: Divyansh Singhvi <divyanshsinghvi@gmail.com> Signed-off-by: 汪志鹏 <wangzhipeng628@gmail.com> Signed-off-by: dsinghvi <divyanshsinghvi@gmail.com> Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com> * change qwen3-omni open cudagraph by default (#1352) Signed-off-by: Rein Yang <ruiruyang2@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [XPU] Update Bagel's flash_attn_varlen_func to fa utils (#1295) Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com> * [Test] Add Omni Model Performance Benchmark Test (#1321) Signed-off-by: yenuo26 <410167048@qq.com> Signed-off-by: wangyu <53896905+yenuo26@users.noreply.github.com> * [BugFix]: Revert utils change (#1369) Signed-off-by: princepride <wangzhipeng628@gmail.com> * [Rebase] Rebase to vllm v0.16.0 (#1357) Signed-off-by: Taichang Zhou <tzhouam@connect.ust.hk> Signed-off-by: tzhouam <tzhouam@connect.ust.hk> Signed-off-by: princepride <wangzhipeng628@gmail.com> Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Isotr0py <Isotr0py@outlook.com> Co-authored-by: ZJY0516 <zhu.jiangyun@foxmail.com> Co-authored-by: youkaichao <youkaichao@gmail.com> * [Test] Fix expansion and example test case for qwen3-omni (#1358) Signed-off-by: yenuo26 <410167048@qq.com> * [v0.16.0][BUG FIX]Fix hunyuan MOE after update to 0.16.0 (#1401) Signed-off-by: Chendi Xue <chendi.xue@intel.com> * [0.16.0] remove cuda hard-code for Hunyuan Image3 (#1402) Signed-off-by: Chendi Xue <chendi.xue@intel.com> * [XPU] Add XPU Dockerfile and related docs (#1162) Signed-off-by: Yan Ma <yan.ma@intel.com> Signed-off-by: Daniel Huang <daniel1.huang@intel.com> Co-authored-by: Daniel Huang <daniel1.huang@intel.com> * [Bugfix] Fix Hardcoded Datatypes in Z-image (#1393) Signed-off-by: Alex Brooks <albrooks@redhat.com> * [Feature] : Support disaggregated inference pipeline for Qwen3_TTS (#1161) Signed-off-by: Sy03 <1370724210@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Feature] Add automated PR reviewer bot with GLM integration (#1424) Signed-off-by: hsliu <liuhongsheng4@huawei.com> Signed-off-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> * [Misc] Add Qwen2.5-Omni-3B model support to Gradio demo (#1382) Signed-off-by: UsamaKenway <usamakenway@gmail.com> * [misc] Feature/pr reviewer auto trigger&update model (#1431) Signed-off-by: hsliu <liuhongsheng4@huawei.com> Signed-off-by: Claude Opus 4.6 <noreply@anthropic.com> Signed-off-by: Hunter Liu <hunter@liu.sh> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> * Revert "[misc] Feature/pr reviewer auto trigger&update model" (#1432) * [Doc] Update GPU installation commands (#1434) * [ROCM] [CI] fix dockerfile.rocm to support nightly build and also fix amd ci v0.16.0rc1 (#1380) Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> * [Feature][BAGEL] Combine multi-branch cfg into a single batch to accelerate inference. (#1429) Signed-off-by: Ding Zuhao <e1583181@u.nus.edu> Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com> * [Feat]: add ASCII art logo for vLLM-Omni (#1430) * [Bug] [Bagel] Fix kv transfer bug (#1437) Signed-off-by: Ding Zuhao <e1583181@u.nus.edu> Co-authored-by: Wang Zhipeng: princepride <wangzhipeng628@gmail.com> * [CI] Set L2 & L3 tests running conditions. (#1344) Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> * [Feature] vLLM-Omni RDMA connector (#1019) Signed-off-by: natureofnature <wzliu@connect.hku.hk> * [Minor][Refactor] Pass seq_token_counts explicitly (#1425) Signed-off-by: gcanlin <canlinguosdu@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Misc] Extend Diffusion Benchmark script to other backends (#875) Signed-off-by: NickLucche <nlucches@redhat.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Feature] Support Stage Based Deployment CLI (#939) Signed-off-by: wuhang <wuhang6@huawei.com> Signed-off-by: princepride <wangzhipeng628@gmail.com> Signed-off-by: wuhang <whlbx@hotmail.com> Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * [Doc] Optimize vLLM-Omni metrics documentation (#1311) Signed-off-by: Junhong Liu <98734602+LJH-LBJ@users.noreply.github.com> Signed-off-by: Junhong Liu <ljh_lbj@163.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Bugfix] Forward all vllm-omni serve command parameters to model (#985) Signed-off-by: Junhong Liu <98734602+LJH-LBJ@users.noreply.github.com> Signed-off-by: Junhong Liu <ljh_lbj@163.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Doc]: Add bagel single/multi node usage with mooncake document (#1450) * [Qwen3TTS][Feat] Code2Wav batched decoding (#1426) Signed-off-by: pablo <pablo@agigo.ai> Co-authored-by: pablo <pablo@agigo.ai> * [CI] Remove overwhelming debug log (#1463) Signed-off-by: tzhouam <tzhouam@connect.ust.hk> * [Misc] update wechat image (#1464) Signed-off-by: David Chen <530634352@qq.com> * [Doc] Refine Diffusion Tutorial Documents (#1305) Signed-off-by: Didan Deng <33117903+wtomin@users.noreply.github.com> * [Bugfix] Robust Audio Data Handling in _create_audio_choice (#1222) Signed-off-by: Junhong Liu <98734602+LJH-LBJ@users.noreply.github.com> * [Bugfix]: Fix merging updated additional information to ensure dict type (#1296) Signed-off-by: Shijin Zhang <75300765+Dovis01@users.noreply.github.com> * [Model]Add new nextstep_1(Diffusion) model(only T2I) (#612) Signed-off-by: Dong Wang <dongw2019@gmail.com> Signed-off-by: sniper35 <dongw2019@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Bugfix] Add TTS configuration options (#1177) Signed-off-by: Yanick Schraner <yanick.schraner@bs.ch> * [Debug] Multi-Request for Qwen 3 Omni use_audio_in_video (#1433) Signed-off-by: tzhouam <tzhouam@connect.ust.hk> * [Bugfix] Fix case-sensitive task_type matching in Qwen3TTSModelForGeneration (#1455) Signed-off-by: Sangchun Ha <seomk9896@gmail.com> * [BugFix] process request.num_cached_tokens if it equals to the initial value (#1468) Signed-off-by: Junhong Liu <98734602+LJH-LBJ@users.noreply.github.com> Co-authored-by: Gao Han <gaohan19@huawei.com> * [Bugfix] Fix SDPA attention mask dtype and shape (Fix #857) (#1349) Signed-off-by: jader <yjader@foxmail.com> * [Test] Reduce Perf test case and fix modify stage config (#1449) Signed-off-by: yenuo26 <410167048@qq.com> * [NPU] Upgrade to v0.16.0 (#1375) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [CI] Update Dockerfile for vllm-omni CI image and remove obsolete dep… (#1491) Signed-off-by: tzhouam <tzhouam@connect.ust.hk> * [Fix][Chore] Qwen3-TTS Modeling Minor Code Sanity Improvements (#1482) Signed-off-by: yuanheng <jonathan.zhaoyh@gmail.com> * [Bugfix] Fix tuple/list KV cache extraction crash (#1405) Signed-off-by: junuxyz <216036880+junuxyz@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Doc] format lora related docs for the user's end (#1009) Signed-off-by: AndyZhou952 <jzhoubc@connect.ust.hk> Signed-off-by: Andy Zhou <46011930+AndyZhou952@users.noreply.github.com> * [Feature] Support Wan2.2 output with irregular shapes (#1279) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [Misc] Migrate L1 tests to use pytest-mock (#1315) Signed-off-by: Yuanheng Zhao <jonathan.zhaoyh@gmail.com> Signed-off-by: yuanheng <jonathan.zhaoyh@gmail.com> * [Bugfix] Fix LoRA Scaling on Active Adapters (#1421) Signed-off-by: Alex Brooks <albrooks@redhat.com> * [Bugfix] fix record audio generated frame in offline infer (#1312) Signed-off-by: Junhong Liu <98734602+LJH-LBJ@users.noreply.github.com> Signed-off-by: Junhong Liu <ljh_lbj@163.com> * [Model] Support OmniGen2 (#513) Signed-off-by: Yupu <feng.yu.pu0330@gmail.com> * [Bugfix][Qwen3TTS] (#1289) Signed-off-by: pablo <juanz9312@gmail.com> Co-authored-by: Gao Han <gaohan19@huawei.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * Use pull through cache image for H100 pool (#1518) Signed-off-by: Kevin H. Luu <khluu000@gmail.com> * [ROCm] [CI] [Docker] Point to use the latest vLLM v0.16.0 stable version (#1500) Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> * [Bugfix] fix offline text_to_image error from #1009 (#1515) Signed-off-by: David Chen <530634352@qq.com> * [XPU] Enable FLASH_ATTN on XPU (#1332) Signed-off-by: Yan Ma <yan.ma@intel.com> * Revert gpu_1 job to use regular image (#1521) Signed-off-by: Kevin H. Luu <khluu000@gmail.com> * [Chore] remove unused logger in omni_diffusion (#531) (#1509) Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com> Co-authored-by: Gao Han <gaohan19@huawei.com> * [Qwen3TTS][Feat] Streaming output (#1438) Signed-off-by: pablo <pablo@agigo.ai> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: pablo <pablo@agigo.ai> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Bugfix] Race condition in MultiprocExecutor when concurent access to Scheduler (#1448) Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Doc][Test][Misc] ComfyUI test, more screenshot, and code cleaning (#1435) Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com> Signed-off-by: Samit <285365963@qq.com> Co-authored-by: Samit <285365963@qq.com> * [Performance]Qwen3-Omni performance optimization (#1378) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com> * [Feature] Support HSDP for diffusion models (#1339) Signed-off-by: gcanlin <canlinguosdu@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [CI] fixed CI timeout (#1460) Signed-off-by: zhumingjue <zhumingjue@huawei.com> Signed-off-by: zhumingjue138 <zhumingjue@huawei.com> * [Bugfix] Use uds for zmq address if not set --stage-id (#1522) Signed-off-by: wuhang <wuhang6@huawei.com> * [BugFix] Restore talker's config (#1524) Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Canlin Guo <961750412@qq.com> * [XPU] fix qwen_omni after rebase to v0.16.0 (#1416) Signed-off-by: Chendi Xue <chendi.xue@intel.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Platform] Enable layerwise offload on all hardware (#1492) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * diffusion: enable VAE patch parallel for SD3.5 (#1428) Signed-off-by: dongbo910220 <1275604947@qq.com> * [Perf] GLM Image (#920) Signed-off-by: JaredforReal <w13431838023@gmail.com> Signed-off-by: Jared Wen <w13431838023@gmail.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [skip ci][Doc] add design docs for async chunk in qwen3-omni (#962) Signed-off-by: Rein Yang <ruiruyang2@gmail.com> * feat(qwen3-tts): Add CUDA Graph support for speech tokenizer decoder (#1205) Signed-off-by: xulusjb <fdukeshik@gmail.com> Co-authored-by: xulusjb <fdukeshik@gmail.com> * [New Model]: XiaomiMiMo/MiMo-Audio-7B-Instruct support (#750) Signed-off-by: wangyu31577 <wangyu31577@hundsun.com> Signed-off-by: 齐保元 <qibaoyuan@xiaomi.com> Signed-off-by: hsliu <liuhongsheng4@huawei.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Signed-off-by: GG-li <3226868735@qq.com> Signed-off-by: Sihao Li <111170255+GG-li@users.noreply.github.com> Signed-off-by: XU Mingshi <91017482+mxuax@users.noreply.github.com> Signed-off-by: mxuax <mxuax@connect.ust.hk> Signed-off-by: Baoyuan Qi <qibaoyuan@126.com> Signed-off-by: gcanlin <canlinguosdu@gmail.com> Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn> Signed-off-by: wuzhongjian <wuzhongjian_yewu@cmss.chinamobile.com> Signed-off-by: dongbo910220 <1275604947@qq.com> Signed-off-by: dongbo910220 <32610838+dongbo910220@users.noreply.github.com> Signed-off-by: Jiangyun Zhu <riverclouds.zhu@qq.com> Signed-off-by: Junhong Liu <98734602+LJH-LBJ@users.noreply.github.com> Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> Signed-off-by: baoyuan qi <qibaoyuan@126.com> Signed-off-by: tzhouam <tzhouam@connect.ust.hk> Signed-off-by: zjy0516 <riverclouds.zhu@qq.com> Signed-off-by: Prajwal A <prajwalanagani@gmail.com> Signed-off-by: Shijin Zhang <75300765+Dovis01@users.noreply.github.com> Signed-off-by: 丁宁 <nndding@gmail.com> Signed-off-by: SHIJIN ZHANG <75300765+Dovis01@users.noreply.github.com> Signed-off-by: dingning<dingning7@xiaomi.com> Signed-off-by: dingning <dingning7@xiaomi.com> Signed-off-by: dingning <dingning@xiaomi.com> Co-authored-by: wangyu <53896905+yenuo26@users.noreply.github.com> Co-authored-by: wangyu31577 <wangyu31577@hundsun.com> Co-authored-by: Zhang Shijin <zhangshijin@xiaomi.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Sihao Li <111170255+GG-li@users.noreply.github.com> Co-authored-by: XU Mingshi <91017482+mxuax@users.noreply.github.com> Co-authored-by: Canlin Guo <canlinguosdu@gmail.com> Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn> Co-authored-by: JohnJan <wuzhongjian_yewu@cmss.chinamobile.com> Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com> Co-authored-by: dongbo910220 <32610838+dongbo910220@users.noreply.github.com> Co-authored-by: Jiangyun Zhu <riverclouds.zhu@qq.com> Co-authored-by: Junhong Liu <ljh_lbj@163.com> Co-authored-by: TJian <tunjian.tan@embeddedllm.com> Co-authored-by: shijin zhang <zsj1364226740@gmail.com> Co-authored-by: Zhou Taichang <tzhouam@connect.ust.hk> Co-authored-by: root <root@hk01dgx028.cm.cluster> Co-authored-by: Prajwal A <34590600+LawJarp-A@users.noreply.github.com> Co-authored-by: Shijin Zhang <75300765+Dovis01@users.noreply.github.com> Co-authored-by: dingning <dingning7@xiaomi.com> Co-authored-by: ning ding <nndding@gmail.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * [Feature]: Native GGUF Quantization Support for DiT (#1285) Signed-off-by: David Chen <530634352@qq.com> Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn> Signed-off-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn> Co-authored-by: Isotr0py <2037008807@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * Add benchmark for `v1/audio/speech` non-streaming (#1408) Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> * [Version] Auto generate version using `setuptool_scm` (#1224) Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> * [Feat] : Support Async chunk cleanup (#1087) Signed-off-by: Sy03 <1370724210@qq.com> * [Profiler] Support online profiling (#1136) Signed-off-by: gcanlin <canlinguosdu@gmail.com> Signed-off-by: Canlin Guo <961750412@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Nicolò Lucchesi <nicolo.lucchesi@gmail.com> * [Bugfix] Fix redundant finished req status updating on OmniGenerationScheduler (#1510) Signed-off-by: shijin zhang <75300765+Dovis01@users.noreply.github.com> Co-authored-by: 齐保元 <qibaoyuan@xiaomi.com> * [XPU][NPU][ROCM] enable cpu_offloading flag for non_cuda (#1488) Signed-off-by: gcanlin <canlinguosdu@gmail.com> Signed-off-by: Chendi Xue <chendi.xue@intel.com> Co-authored-by: gcanlin <canlinguosdu@gmail.com> * [Chore] Cleanup dead code in GGUF DiT code path (#1533) Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn> * [Doc] Update installation instructions for vllm 0.16.0 (#1505) Signed-off-by: tzhouam <tzhouam@connect.ust.hk> * [Doc] [skip ci]Sync. (#1363) Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> Co-authored-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com> * [CI][skip ci]Update H100 image link based on #1518 (#1538) Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> * Fix no embed text spk tokens (#1540) Signed-off-by: Junhong Liu <98734602+LJH-LBJ@users.noreply.github.com> * [Debug] Merge vllm pull 35368 (#1534) Signed-off-by: tzhouam <tzhouam@connect.ust.hk> * [Docs] update async chunk docs diagram [skip ci] (#1530) Signed-off-by: Rein Yang <ruiruyang2@gmail.com> * fix(qwen3-tts): fix Base ICL voice clone producing corrupted audio (#1554) Signed-off-by: linyueqian <linyueqian@outlook.com> * [NPU][Bugfix] Align GPU side and recover qwen3-tts (#1564) Signed-off-by: gcanlin <canlinguosdu@gmail.com> * [BugFix] Fix unexpected crash when init OmniDiffusion (#1562) Signed-off-by: Semmer2 <semmer@live.cn> * [CI] Modify some CI test cases to run on L4 environment to reduce H100 resource usage. (#1543) Signed-off-by: yenuo26 <410167048@qq.com> Signed-off-by: wangyu <53896905+yenuo26@users.noreply.github.com> * [BugFix]: fix a lot of bug (#1565) Signed-off-by: princepride <wangzhipeng628@gmail.com> * feat: add HyperCLOVAX-SEED-Omni-8B support Model files: - vllm_omni/diffusion/models/hyperclovax_vision/: vision decoder pipeline (HyperCLOVAXVisionPipeline) using flow matching diffusion + VisionTransformer - vllm_omni/diffusion/models/hyperclovax_audio/: audio decoder pipeline (HyperCLOVAXAudioPipeline) using Unit-BigVGAN codec - vllm_omni/model_executor/stage_input_processors/hyperclovax_seed_omni.py: thinker2vision_decoder and thinker2audio_decoder — extract discrete tokens from LLM output; truncate/pad vision codes to 729 (27x27) for decoder Registry: - vllm_omni/diffusion/registry.py: register HyperCLOVAXVisionPipeline and HyperCLOVAXAudioPipeline with post-process functions Stage config: - vllm_omni/model_executor/stage_configs/hcx_omni.yaml: 3-stage config Stage 0: LLM thinker (TP=4, GPUs 0-3), Stage 1: vision decoder (GPU 4), Stage 2: audio decoder (GPU 5) Bug fixes for HyperCLOVAX compatibility: - diffusion/request.py: add extra dict field to OmniDiffusionRequest so vision_tokens/audio_tokens from stage input processors reach the pipeline - entrypoints/async_omni_diffusion.py: extract OmniTokensPrompt.additional_information into OmniDiffusionRequest.extra before creating request - entrypoints/omni_stage.py: skip empty engine inputs (text-only requests where thinker2vision_decoder/thinker2audio_decoder return []) - entrypoints/async_omni.py: handle skipped sentinel in _process_single_result so text-only requests complete without crashing on Stage 1/2 * fix: correct decoder params and HCX porting fixes - hcx_omni.yaml: guidance_scale 3.5→0.75, num_inference_steps 30→50 (matches OmniServe production defaults; 3.5 caused over-amplified autoguidance → shrunken/degraded output images) - omni_stage.py: skip empty engine inputs for text-only requests - async_omni_diffusion.py: extract OmniTokensPrompt.additional_information into OmniDiffusionRequest.extra (audio_tokens/vision_tokens) - registry.py: HCX Omni diffusion model registration fix Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * feat: HyperCLOVAX-SEED-Omni-8B stage pipeline and entrypoint fixes * fix: change guidance_scale from 9.0 to 0.75 (autoguidance scale, OmniServe default) * feat: add audio decoder Stage 2 to hcx_omni pipeline - Wire HyperCLOVAXAudioPipeline as Stage 2 in hcx_omni.yaml - GPU 5 assigned for audio decoder (Unit-BigVGAN / NCCosybigvganDecoder) - Add runtime edge 0->2 (thinker -> audio decoder) - Implement post-generation PCM chunk streaming for audio output (4800 samples / 200ms per SSE event @ 24kHz, int16 base64-encoded) Refs: github.com/vllm-project/vllm-omni/pull/869 (already incorporated) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix: vllm version compatibility for HyperCLOVAX audio decoder startup - config/model.py: try/except fallback for AttentionBackendEnum import (vllm.v1.attention.backends.registry absent in older vllm builds) - pipeline_hyperclovax_audio.py: return actual named_parameters() from load_weights() when using MAR checkpoint so diffusers_loader strict check passes (weights loaded eagerly in __init__ via MAR extraction) - qwen3_omni_moe_thinker.py, qwen2_5_omni_thinker.py: try/except stubs for check_interleaved_audio_video and merge_interleaved_embeddings which are absent in older vllm qwen2_5_omni_thinker; these symbols are only exercised by Qwen models, not HyperCLOVAX Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix: add edge 1→2 and correct model key in hcx_omni.yaml Stage 2 - Add runtime edge from:1 to:2 (required for Stage-2 connector init; without it AsyncOrchestrator cannot route to audio decoder at runtime) - Change model_subdir to model for Stage-2 engine_args to match total-poc working reference config Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix: audio S2S output - handle diffusion outputs in _create_audio_choice HyperCLOVAXAudioPipeline (diffusion) stores audio in multimodal_output directly (OmniRequestOutput.from_diffusion), not in outputs[0].multimodal_output like LLM pipelines. Fix three locations: 1. _create_audio_choice (non-streaming): use omni_outputs.multimodal_output when final_res.outputs is empty (diffusion path). 2. Streaming audio path: same fix for _final_res.outputs[0]. 3. Both loops (for output in final_res.outputs): fall back to single synthetic choice at index 0 when outputs list is empty. 4. Handle bytes audio output from HyperCLOVAXAudioPipeline post-process (returns WAV bytes, not tensors like Qwen3-Omni). Also fixes audio input (A2T) regression: skip diffusion prompt extraction when mm_data has audio content (added in previous session). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix: parse WAV bytes with soundfile for uniform PCM chunk streaming HyperCLOVAXAudioPipeline returns WAV bytes including 44-byte header. The previous byte-offset splitting included the header in the first chunk, corrupting it. Fix: parse with soundfile to get float32 PCM, then convert to int16 chunks uniformly regardless of source type (bytes or tensor). Verified: 136 audio chunks x 200ms = 27.04s audio streamed correctly. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * feat: zero-shot TTS with speaker embedding from input audio - serving_chat.py: extract last input_audio base64 from request messages and inject as ref_audio_b64 into engine_prompt dict - thinker2audio_decoder: read ref_audio_b64 from prompt and pass as ref_audio_tokens to Stage 2 (HyperCLOVAXAudioPipeline) - hcx_omni.yaml: switch Stage 2 to NCZSCosybigvganDecoder.mar (zero-shot) which uses ECAPA-TDNN speaker encoder instead of finetuned ID lookup Pipeline: input audio -> ECAPA-TDNN -> speaker embedding -> BigVGAN synthesis matching the voice characteristics of the original speaker. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * feat: wire audio decoder Stage 2 to hcx_omni pipeline and fix S2S flow - Add Stage 2 (HyperCLOVAXAudioPipeline / NCZSCosybigvganDecoder) to hcx_omni.yaml with GPU 5, gpu_memory_utilization 0.4, edge 0->2 from thinker - Fix thinker2audio_decoder: correct audio token range (128606-135167), remap to [0, 6561) for BigVGAN input, handle empty token case gracefully - Fix pipeline_hyperclovax_audio.py post_process_func signature and incorporate PR#869 BUG FIX patches for stable audio generation * fix: use finetuned audio decoder and fix transformers_modules deserialization - hcx_omni.yaml: switch Stage 2 from NCZSCosybigvganDecoder (zero-shot, ECAPA-TDNN) to NCCosybigvganDecoder (finetuned, nn.Embedding speaker id). Zero-shot decoder required ref_audio (mel spectrogram) which is unavailable for text-only requests and incompatible with finetuned decoder path. - pipeline_hyperclovax_audio.py: guard ref_audio processing with 'not self.bigvgan.finetune' — finetuned decoder has no ECAPA-TDNN encoder, so passing ref_audio bytes would crash with 'expected 100 channels'. - omni_stage.py: add HuggingFace modules cache (~/.cache/huggingface/modules) to sys.path before queue.get_nowait() in try_collect(). Stage-0 pickles outputs containing custom classes from transformers_modules (trust_remote_code), but the API server process doesn't have this path, causing deserialization failures that silently drop Stage-0 outputs. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix: restore zero-shot speaker cloning with fallback for text-only requests - hcx_omni.yaml: revert to NCZSCosybigvganDecoder.mar (zero-shot ECAPA-TDNN) for voice-preserving S2S synthesis. NCCosybigvganDecoder used a fixed integer speaker_id and lost the input speaker's voice. - pipeline_hyperclovax_audio.py: add zero-mel fallback branch for finetune=False + ref_audio=None case. When a text-only request arrives (no input audio → no ref_audio), ECAPA-TDNN receives a zero mel tensor [1, num_mels, 64] instead of crashing with 'expected 100 channels'. S2S requests always have ref_audio so the zero-shot cloning path is unchanged. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * feat: add stage config yaml for HCX audio decoder Signed-off-by: Hyunjoon Jeong <hyunjoon.jeong@navercorp.com> * feat: add HyperCLOVAX-SEED-Omni 8B model as vllm-omni executor Signed-off-by: Hyunjoon Jeong <hyunjoon.jeong@navercorp.com> * feat: add HCX audio decoder pipeline Signed-off-by: Hyunjoon Jeong <hyunjoon.jeong@navercorp.com> * fix: modify exception for HCX audio decoder (GAN) Signed-off-by: Hyunjoon Jeong <hyunjoon.jeong@navercorp.com> * fix: default temperature set to 0, and pipeline model evaluation mode Signed-off-by: Hyunjoon Jeong <hyunjoon.jeong@navercorp.com> --------- Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com> Signed-off-by: dengyunyang <584797741@qq.com> Signed-off-by: gcanlin <canlinguosdu@gmail.com> Signed-off-by: samithuang <285365963@qq.com> Signed-off-by: Samit <285365963@qq.com> Signed-off-by: lishunyang <lishunyang12@163.com> Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Signed-off-by: princepride <wangzhipeng628@gmail.com> Signed-off-by: 汪志鹏 <wangzhipeng628@gmail.com> Signed-off-by: wangyu31577 <wangyu31577@hundsun.com> Signed-off-by: Kyle Huang <yellowsea@gmail.com> Signed-off-by: zjy0516 <riverclouds.zhu@qq.com> Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> Signed-off-by: natureofnature <wzliu@connect.hku.hk> Signed-off-by: linyueqian <linyueqian@outlook.com> Signed-off-by: mxuax <mxuax@connect.ust.hk> Signed-off-by: XU Mingshi <91017482+mxuax@users.noreply.github.com> Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com> Signed-off-by: wuzhongjian wuzhongjian_yewu@cmss.chinamobile.com Signed-off-by: ZeldaHuang <hzm414167@alibaba-inc.com> Signed-off-by: Divyansh Singhvi <divyanshsinghvi@gmail.com> Signed-off-by: Lin, Fanli <fanli.lin@intel.com> Signed-off-by: Fanli Lin <fanli.lin@intel.com> Signed-off-by: Fanli Lin <fanli0116@gmail.com> Signed-off-by: dongbo910220 <1275604947@qq.com> Signed-off-by: Ding Zuhao <e1583181@u.nus.edu> Signed-off-by: jzz <e1583181@u.nus.edu> Signed-off-by: Andy Zhou <46011930+AndyZhou952@users.noreply.github.com> Signed-off-by: wangyu <53896905+yenuo26@users.noreply.github.com> Signed-off-by: Pierre Le Guen <26087574+PierreLeGuen@users.noreply.github.com> Signed-off-by: yuanheng <jonathan.zhaoyh@gmail.com> Signed-off-by: ram16g <anlianfengjie@163.com> Signed-off-by: Didan Deng <33117903+wtomin@users.noreply.github.com> Signed-off-by: pablo <juanz9312@gmail.com> Signed-off-by: Roger Wang <hey@rogerw.io> Signed-off-by: anna <lee.anna@navercorp.com> Signed-off-by: Rustam Khadipash <16683750+hadipash@users.noreply.github.com> Signed-off-by: Alicia <115451386+congw729@users.noreply.github.com> Signed-off-by: Taichang Zhou <tzhouam@connect.ust.hk> Signed-off-by: tzhouam <tzhouam@connect.ust.hk> Signed-off-by: hsliu <liuhongsheng4@huawei.com> Signed-off-by: hsliu_ustc <hsliu_ustc@noreply.gitcode.com> Signed-off-by: zhenwei-intel <zhenwei.liu@intel.com> Signed-off-by: Yuanheng Zhao <jonathan.zhaoyh@gmail.com> Signed-off-by: erfgss <97771661+erfgss@users.noreply.github.com> Signed-off-by: xiedeyantu <czjourney@163.com> Signed-off-by: Junhong Liu <98734602+LJH-LBJ@users.noreply.github.com> Signed-off-by: Junhong Liu <ljh_lbj@163.com> Signed-off-by: David Chen <530634352@qq.com> Signed-off-by: weichen <calvin_zhu0210@outlook.com> Signed-off-by: Yan Ma <yan.ma@intel.com> Signed-off-by: ApsarasX <apsarax@outlook.com> Signed-off-by: Chenguang ZHENG <645327136@qq.com> Signed-off-by: yenuo26 <410167048@qq.com> Signed-off-by: Semmer2 <semmer@live.cn> Signed-off-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com> Signed-off-by: zhou zhuoxin <zhouzhuoxin1508@outlook.com> Signed-off-by: Gao Han <hgaoaf@connect.ust.hk> Signed-off-by: Rein Yang <ruiruyang2@gmail.com> Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com> Signed-off-by: SYLAR <125541396+lishunyang12@users.noreply.github.com> Signed-off-by: Ziming Huang <1520787127@qq.com> Signed-off-by: SamitHuang <285365963@qq.com> Signed-off-by: GG-li <3226868735@qq.com> Signed-off-by: CHEN <116010019@link.cuhk.edu.cn> Signed-off-by: John Liu BUAA <liukecheng97@gmail.com> Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com> Signed-off-by: dsinghvi <divyanshsinghvi@gmail.com> Signed-off-by: Chendi Xue <chendi.xue@intel.com> Signed-off-by: Daniel Huang <daniel1.huang@intel.com> Signed-off-by: Alex Brooks <albrooks@redhat.com> Signed-off-by: Sy03 <1370724210@qq.com> Signed-off-by: Claude Opus 4.6 <noreply@anthropic.com> Signed-off-by: UsamaKenway <usamakenway@gmail.com> Signed-off-by: Hunter Liu <hunter@liu.sh> Signed-off-by: NickLucche <nlucches@redhat.com> Signed-off-by: wuhang <wuhang6@huawei.com> Signed-off-by: wuhang <whlbx@hotmail.com> Signed-off-by: pablo <pablo@agigo.ai> Signed-off-by: Shijin Zhang <75300765+Dovis01@users.noreply.github.com> Signed-off-by: Dong Wang <dongw2019@gmail.com> Signed-off-by: sniper35 <dongw2019@gmail.com> Signed-off-by: Yanick Schraner <yanick.schraner@bs.ch> Signed-off-by: Sangchun Ha <seomk9896@gmail.com> Signed-off-by: jader <yjader@foxmail.com> Signed-off-by: junuxyz <216036880+junuxyz@users.noreply.github.com> Signed-off-by: AndyZhou952 <jzhoubc@connect.ust.hk> Signed-off-by: Yupu <feng.yu.pu0330@gmail.com> Signed-off-by: Kevin H. Luu <khluu000@gmail.com> Signed-off-by: zhumingjue <zhumingjue@huawei.com> Signed-off-by: zhumingjue138 <zhumingjue@huawei.com> Signed-off-by: JaredforReal <w13431838023@gmail.com> Signed-off-by: Jared Wen <w13431838023@gmail.com> Signed-off-by: xulusjb <fdukeshik@gmail.com> Signed-off-by: 齐保元 <qibaoyuan@xiaomi.com> Signed-off-by: Sihao Li <111170255+GG-li@users.noreply.github.com> Signed-off-by: Baoyuan Qi <qibaoyuan@126.com> Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn> Signed-off-by: wuzhongjian <wuzhongjian_yewu@cmss.chinamobile.com> Signed-off-by: dongbo910220 <32610838+dongbo910220@users.noreply.github.com> Signed-off-by: Jiangyun Zhu <riverclouds.zhu@qq.com> Signed-off-by: baoyuan qi <qibaoyuan@126.com> Signed-off-by: Prajwal A <prajwalanagani@gmail.com> Signed-off-by: 丁宁 <nndding@gmail.com> Signed-off-by: SHIJIN ZHANG <75300765+Dovis01@users.noreply.github.com> Signed-off-by: dingning<dingning7@xiaomi.com> Signed-off-by: dingning <dingning7@xiaomi.com> Signed-off-by: dingning <dingning@xiaomi.com> Signed-off-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com> Signed-off-by: Canlin Guo <961750412@qq.com> Signed-off-by: shijin zhang <75300765+Dovis01@users.noreply.github.com> Signed-off-by: Hyunjoon Jeong <hyunjoon.jeong@navercorp.com> Signed-off-by: Hyunjoon Jeong <with1015@unist.ac.kr> Co-authored-by: Zeyu Huang | 黃澤宇 <11222265+fhfuih@users.noreply.github.com> Co-authored-by: JohnJan <wuzhongjian_yewu@cmss.chinamobile.com> Co-authored-by: dengyunyang <584797741@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: Canlin Guo <canlinguosdu@gmail.com> Co-authored-by: Samit <285365963@qq.com> Co-authored-by: SYLAR <125541396+lishunyang12@users.noreply.github.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com> Co-authored-by: wangyu <53896905+yenuo26@users.noreply.github.com> Co-authored-by: wangyu31577 <wangyu31577@hundsun.com> Co-authored-by: kYLe <yellowsea@gmail.com> Co-authored-by: Jiangyun Zhu <riverclouds.zhu@qq.com> Co-authored-by: TJian <tunjian.tan@embeddedllm.com> Co-authored-by: NATURE <wzliu@connect.hku.hk> Co-authored-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com> Co-authored-by: Zhou Taichang <tzhouam@connect.ust.hk> Co-authored-by: root <root@hk01dgx028.cm.cluster> Co-authored-by: XU Mingshi <91017482+mxuax@users.noreply.github.com> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com> Co-authored-by: Rein Yang <ruiruyang2@gmail.com> Co-authored-by: Ziming Huang <hzm414167@alibaba-inc.com> Co-authored-by: dsinghvi <divyanshsinghvi@gmail.com> Co-authored-by: Fanli Lin <fanli.lin@intel.com> Co-authored-by: dongbo910220 <32610838+dongbo910220@users.noreply.github.com> Co-authored-by: Ding Zuhao <e1583181@u.nus.edu> Co-authored-by: Andy Zhou <46011930+AndyZhou952@users.noreply.github.com> Co-authored-by: Pierre LE GUEN <26087574+PierreLeGuen@users.noreply.github.com> Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com> Co-authored-by: Yuanheng Zhao <54058983+yuanheng-zhao@users.noreply.github.com> Co-authored-by: ram16g <anlianfengjie@163.com> Co-authored-by: Didan Deng <33117903+wtomin@users.noreply.github.com> Co-authored-by: Markus / Mark <46672778+marksverdhei@users.noreply.github.com> Co-authored-by: Juan Pablo Zuluaga <46724788+JuanPZuluaga@users.noreply.github.com> Co-authored-by: muziyuhui666 <111362884+muziyuhui666@users.noreply.github.com> Co-authored-by: Roger Wang <hey@rogerw.io> Co-authored-by: ceanna93 <fairyanna@naver.com> Co-authored-by: anna <lee.anna@navercorp.com> Co-authored-by: Rustam Khadipash <16683750+hadipash@users.noreply.github.com> Co-authored-by: Alicia <115451386+congw729@users.noreply.github.com> Co-authored-by: hsliu_ustc <hsliu_ustc@noreply.gitcode.com> Co-authored-by: liuzhenwei <zhenweiliu@habana.ai> Co-authored-by: erfgss <97771661+erfgss@users.noreply.github.com> Co-authored-by: Jensen <czjourney@163.com> Co-authored-by: Junhong Liu <ljh_lbj@163.com> Co-authored-by: weichen <calvin_zhu0210@outlook.com> Co-authored-by: PopSoda2002 <zhouhp.me@gmail.com> Co-authored-by: Yan Ma <yan.ma@intel.com> Co-authored-by: ApsarasX <apsarax@outlook.com> Co-authored-by: Chenguang Zheng <645327136@qq.com> Co-authored-by: Jiaping Wu <53215702+ElleElleWu@users.noreply.github.com> Co-authored-by: zhou zhuoxin <zhouzhuoxin1508@outlook.com> Co-authored-by: Gao Han <gaohan19@huawei.com> Co-authored-by: rein yang <73573651+R2-Y@users.noreply.github.com> Co-authored-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com> Co-authored-by: Flora Feng <4florafeng@gmail.com> Co-authored-by: Sihao Li <111170255+GG-li@users.noreply.github.com> Co-authored-by: ChenWenjing <54166744+Shirley125@users.noreply.github.com> Co-authored-by: Bhanu068 <voutharoja.bhanu06@gmail.com> Co-authored-by: John Liu BUAA <liukecheng97@gmail.com> Co-authored-by: yenuo26 <410167048@qq.com> Co-authored-by: knlnguyen1802 <knlnguyen1802@gmail.com> Co-authored-by: liuzhenwei <zhenwei.liu@intel.com> Co-authored-by: Isotr0py <Isotr0py@outlook.com> Co-authored-by: ZJY0516 <zhu.jiangyun@foxmail.com> Co-authored-by: youkaichao <youkaichao@gmail.com> Co-authored-by: Chendi.Xue <chendi.xue@intel.com> Co-authored-by: Daniel Huang <daniel1.huang@intel.com> Co-authored-by: Alex Brooks <albrooks@redhat.com> Co-authored-by: Sy03 <1370724210@qq.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: UsamaKenway <56207634+UsamaKenway@users.noreply.github.com> Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com> Co-authored-by: wuhang <wuhang6@huawei.com> Co-authored-by: pablo <pablo@agigo.ai> Co-authored-by: SHIJIN ZHANG <75300765+Dovis01@users.noreply.github.com> Co-authored-by: Dong W <89223086+sniper35@users.noreply.github.com> Co-authored-by: Yanick Schraner <yanick.schraner@gmail.com> Co-authored-by: Sangchun Ha <seomk9896@naver.com> Co-authored-by: 亦瑾 <76905040+yJader@users.noreply.github.com> Co-authored-by: junuxyz <216036880+junuxyz@users.noreply.github.com> Co-authored-by: Yupu <feng.yu.pu0330@gmail.com> Co-authored-by: Kevin H. Luu <khluu000@gmail.com> Co-authored-by: zhumingjue138 <zhumingjue@huawei.com> Co-authored-by: Canlin Guo <961750412@qq.com> Co-authored-by: Jared Wen <w13431838023@gmail.com> Co-authored-by: Xu Lu <572605156@qq.com> Co-authored-by: xulusjb <fdukeshik@gmail.com> Co-authored-by: Baoyuan Qi <qibaoyuan@xiaomi.com> Co-authored-by: Zhang Shijin <zhangshijin@xiaomi.com> Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn> Co-authored-by: shijin zhang <zsj1364226740@gmail.com> Co-authored-by: Prajwal A <34590600+LawJarp-A@users.noreply.github.com> Co-authored-by: dingning <dingning7@xiaomi.com> Co-authored-by: ning ding <nndding@gmail.com> Co-authored-by: Isotr0py <2037008807@qq.com> Co-authored-by: Nicolò Lucchesi <nicolo.lucchesi@gmail.com> Co-authored-by: Ting FU <futing10@huawei.com> Co-authored-by: developer-account <irteam@vllm-omni-dev-0.vllm-omni-dev.p-nb13557.svc.cluster.local> Co-authored-by: Hyunjoon Jeong <hyunjoon.jeong@navercorp.com>
…llm-project#1161) Signed-off-by: Sy03 <1370724210@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
Support disaggregated inference pipeline for Talker and SpeechTokenizer.
qwen3_tts_talker_ar.py: Stage-0 main model. Implements vLLM-native AR Talker that autoregressively generates layer-0 codec tokens step by step. At each step, it uses the embedded code_predictor to complete residual codebooks (1..Q-1) to form complete audio_codes. Supports preprocess/postprocess/talker_mtp, covering three task types: CustomVoice, VoiceDesign, and Base.
qwen3_tts_code_predictor_vllm.py: Residual code prediction sub-model. A lightweight Transformer that maintains its own _LocalPredictorKVCache (independent of vLLM engine KV cache), providing prefill_logits and decode_logits two-stage interfaces, called by talker_mtp at each step.
qwen3_tts_code2wav.py: Stage-1 model (standard mode). Receives frame-aligned codec tokens and decodes them into waveforms through SpeechTokenizer. Supports both streaming (with left context concatenation) and non-streaming modes. Pure generation phase with no logits/sampling.
qwen3_tts_disaggregated.py: Stage-1 model (async_chunk streaming mode). A phase-aware disaggregated wrapper that completes codec-to-waveform decoding in the preprocess stage, adapting to vLLM's async_chunk streaming scheduler.
Related issues:
#938
#976
Thanks for advice from @JuanPZuluaga and @gcanlin .
CC @gcanlin for preliminary code review.
Purpose
Test Plan
e2e tests in test_qwen3_tts.py & self e2e tests.
Test Result
.venv/bin/python -m pytest tests/e2e/online_serving/test_qwen3_tts.py -v -p no:skip ============================ test session starts ============================ platform linux -- Python 3.12.3, pytest-9.0.2, pluggy-1.6.0 -- /home/sy03/bpftime_sy03/vllm-omni/.venv/bin/python cachedir: .pytest_cache rootdir: /home/sy03/bpftime_sy03/vllm-omni configfile: pyproject.toml plugins: cov-7.0.0, anyio-4.12.1 collected 6 items tests/e2e/online_serving/test_qwen3_tts.py::TestQwen3TTSCustomVoice::test_speech_english_basic PASSED [ 16%] tests/e2e/online_serving/test_qwen3_tts.py::TestQwen3TTSCustomVoice::test_speech_chinese_basic PASSED [ 33%] tests/e2e/online_serving/test_qwen3_tts.py::TestQwen3TTSCustomVoice::test_speech_different_voices PASSED [ 50%] tests/e2e/online_serving/test_qwen3_tts.py::TestQwen3TTSCustomVoice::test_speech_binary_response_not_utf8_error PASSED [ 66%] tests/e2e/online_serving/test_qwen3_tts.py::TestQwen3TTSAPIEndpoints::test_list_voices_endpoint PASSED [ 83%] tests/e2e/online_serving/test_qwen3_tts.py::TestQwen3TTSAPIEndpoints::test_models_endpoint PASSED [100%]Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model.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)