[Refactor][OutputProcessor 2/3]: OmniRequestOutput should inherit RequestOutput and no nested wrap-up by bowieshi · Pull Request #5146 · vllm-project/vllm-omni

bowieshi · 2026-07-16T09:03:58Z

PLEASE FILL IN THE PR DESCRIPTION HERE.

Purpose

OmniRequestOutput taks RequestOutput as a field. However, RequestOutput is a class that contains the output of a LLM output. Our OmniRequestOutput is designed to be a unified output class for both LLM and diffusion models, which should inherit from RequestOutput instead of containing it as a field. This will allow us to use polymorphism and treat OmniRequestOutput as a RequestOutput when needed, while still having the additional fields for diffusion models.

Related code snippets:
vllm_omni/outputs.py: line 62 ~ line 111
before modification:

@dataclass
class OmniRequestOutput:
    request_id: str = ""
    finished: bool = True
    # Pipeline stage fields
    stage_id: int | None = None
    final_output_type: str = "text"
    request_output: RequestOutput | None = None
    # Diffusion model fields
    images: list[Image.Image] = field(default_factory=list)
    prompt: OmniPromptType | None = None
    latents: torch.Tensor | None = None
    trajectory_latents: torch.Tensor | None = None
    trajectory_timesteps: torch.Tensor | None = None
    trajectory_log_probs: torch.Tensor | None = None
    trajectory_decoded: list | None = None
    metrics: dict[str, Any] = field(default_factory=dict)
    _multimodal_output: dict[str, Any] = field(default_factory=dict)
    _custom_output: dict[str, Any] = field(default_factory=dict)
    # profiling data
    stage_durations: dict[str, float] = field(default_factory=dict)
    # memory usage info
    peak_memory_mb: float = 0.0
    # error handling
    error: str | None = None
    error_status_code: int | None = None
    error_type: str | None = None

After modification:

@dataclass
class OmniRequestOutput(RequestOutput):
    # --- Inherited RequestOutput attributes (redeclared as fields) ---
    request_id: str = ""
    prompt: OmniPromptType | None = None
    prompt_token_ids: list[int] | None = None
    prompt_logprobs: Any = None
    outputs: list[CompletionOutput] = field(default_factory=list)
    finished: bool = True
    metrics: Any = field(default_factory=dict)
    lora_request: Any = None
    encoder_prompt: str | None = None
    encoder_prompt_token_ids: list[int] | None = None
    num_cached_tokens: int | None = None
    kv_transfer_params: dict[str, Any] | None = None

    # --- Pipeline stage fields ---
    stage_id: int | None = None
    final_output_type: str = "text"

    # --- Diffusion model fields ---
    images: list[Image.Image] = field(default_factory=list)
    latents: torch.Tensor | None = None
    trajectory_latents: torch.Tensor | None = None
    trajectory_timesteps: torch.Tensor | None = None
    trajectory_log_probs: torch.Tensor | None = None
    trajectory_decoded: list | None = None
    _multimodal_output: dict[str, Any] = field(default_factory=dict)
    _custom_output: dict[str, Any] = field(default_factory=dict)

    # profiling data
    stage_durations: dict[str, float] = field(default_factory=dict)

    # memory usage info
    peak_memory_mb: float = 0.0

    # error handling
    error: str | None = None
    error_status_code: int | None = None
    error_type: str | None = None

Test Plan

vLLM Version: 0.25.0

vLLM-Omni Commit: 4c40955

Test Result

pytest -s -v -m "core_model and cpu"

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
--- Running Summary
===================================================================== 4156 passed, 13 skipped, 2189 deselected, 70 warnings in 446.35s (0:07:26) ======================================================================
sys:1: DeprecationWarning: builtin type swigvarlink has no __module__ attribute

BEFORE SUBMITTING: read CONTRIBUTING.md and run the precheck-pr skill with the code agent for a self-check against project conventions.
(anything written below this line will be removed by GitHub Actions)

hsliuustc0106 · 2026-07-17T00:08:53Z

    # --- Diffusion model fields ---
    images: list[Image.Image] = field(default_factory=list)
    latents: torch.Tensor | None = None
    trajectory_latents: torch.Tensor | None = None
    trajectory_timesteps: torch.Tensor | None = None
    trajectory_log_probs: torch.Tensor | None = None
    trajectory_decoded: list | None = None
    _multimodal_output: dict[str, Any] = field(default_factory=dict)
    _custom_output: dict[str, Any] = field(default_factory=dict)

are all of these needed? the output should only be related to different modalities: text, audio, video, image, action...

bowieshi · 2026-07-17T01:04:58Z

    # --- Diffusion model fields ---
    images: list[Image.Image] = field(default_factory=list)
    latents: torch.Tensor | None = None
    trajectory_latents: torch.Tensor | None = None
    trajectory_timesteps: torch.Tensor | None = None
    trajectory_log_probs: torch.Tensor | None = None
    trajectory_decoded: list | None = None
    _multimodal_output: dict[str, Any] = field(default_factory=dict)
    _custom_output: dict[str, Any] = field(default_factory=dict)

are all of these needed? the output should only be related to different modalities: text, audio, video, image, action...

Yes, this should be refactored to align with #4922 definition for better readability and usablity. But this will inflence a lot of files because this is the front-end part. This PR I dealt with some structure problems and I will refactor this in latter PR.

hsliuustc0106 · 2026-07-18T13:11:45Z

+    encoder_prompt: str | None = None
+    encoder_prompt_token_ids: list[int] | None = None
+    num_cached_tokens: int | None = None
+    kv_transfer_params: dict[str, Any] | None = None


multi_modal_placeholders is part of RequestOutput.__init__, but it is not redeclared as a dataclass field here. Since OmniRequestOutput is now encoded by msgspec as a dataclass, dynamic attrs copied from the wrapped RequestOutput are dropped on connector serialization; I reproduced multi_modal_placeholders becoming None after OmniMsgpackEncoder/Decoder. Please make this a real field and copy/round-trip it like the other RequestOutput content attrs.

I am not sure whether multi_modal_placeholders is still necessary. I found upstream VLLM already removed it in commit vllm-project/vllm@f9ac19204. But our vllm-omni serialization.py still has multi_modal_placeholders. This is the only file in our vllm-omni repo has multi_modal_placeholders. If multi_modal_placeholders not used in the future, should we remove it in serialization.py?

hsliuustc0106 · 2026-07-18T13:13:04Z

            request_output=request_output,
-            finished=True,
        )
+        out.finished = True


This now changes the inner RequestOutput.finished semantics. _copy_content_from() copies request_output.finished, but then this overwrites it with True; because request_output now returns self, callers that still check output.request_output.finished will see True for unfinished stage outputs. Please preserve the wrapped RequestOutput.finished value here, or only force the wrapper-level flag through a separate field that does not affect the RequestOutput view.

I checked again. I found that it seems from_pipeline is a dead code. Only test code uses it and no production code will call it. Should I delete this function?

good for me

hsliuustc0106 · 2026-07-20T02:49:17Z

@bjf-frz @fhfuih

bjf-frz · 2026-07-20T12:39:00Z

+_dataclass_init = OmniRequestOutput.__init__
+
+
+def _omni_request_output_init(self: OmniRequestOutput, *args: Any, request_output: Any = None, **kwargs: Any) -> None:
+    _dataclass_init(self, *args, **kwargs)
+    if request_output is not None:
+        self._copy_content_from(request_output)
+
+
+OmniRequestOutput.__init__ = _omni_request_output_init  # type: ignore[method-assign]


this snippet looks less elegant and a little bit weird.

fixed. Remove property and logic of mutating __init__.

hsliuustc0106

Requesting changes for the two compatibility issues in the inline comments:

multi_modal_placeholders is part of the RequestOutput contract but is not a real OmniRequestOutput dataclass field, so it is dropped by msgspec/dataclass connector serialization.
from_pipeline() overwrites the copied RequestOutput.finished value with True; since request_output now aliases self, callers checking output.request_output.finished can see unfinished stage outputs as finished.

Please fix these before merging.

Signed-off-by: bowieshi <aoibosh@connect.hku.hk>

bowieshi · 2026-08-10T10:47:29Z

please paste your local tests pass on previous failures


=== Job timing summary ===                                                                                                                                                 
  Diffusion_Bagel_Test  19m 28s  OK                                                                                                                                        
  Diffusion_FLUX.2_Klein_Test  5m 5s  OK                                                                                                                                   
  Diffusion_LoRA_Test  5m 53s  OK                                                                                                                                          
  Diffusion_Qwen_Image_Edit_Test  12m 33s  FAILED (exit 1)       PASSED

======================================================================= 3 passed, 14 warnings in 560.87s (0:09:20) =======================================================================
                                                                                                          
  Diffusion_Qwen_Image_Layered_Test  6m 57s  FAILED (exit 1)    PASSED
======================================================================= 1 passed, 14 warnings in 382.80s (0:06:22) =======================================================================                                                                                                              
  Diffusion_Qwen_Image_Test  5m 43s  OK                                                                                                                                    
  Diffusion_Wan22_Test  11m 22s  FAILED (exit 1)       
====================================================================== 2 passed, 14 warnings in 620.59s (0:10:20) =======================================================================                                                                                                                    
  Diffusion_Z_Image_Test  9m 57s  OK                                                                                                                                       
  Distributed_Test  17m 53s  OK                                                                                                                                            
  Entrypoint_Test_with_H100  38m 8s  FAILED (exit 1)                          
======================================================= 10 passed, 1 skipped, 1322 deselected, 16 warnings in 2294.19s (0:38:14) ========================================================                                                                                             
  Entrypoint_Test_with_L4  46m 20s  OK    
  Omni_MiniCPM-o_4.5_Duplex_Test  11m 16s  OK                                        
  Omni_MiniCPM-o_4.5_Test  21m 50s  FAILED (exit 1)     
=============================================================== 4 passed, 9 deselected, 14 warnings in 1580.73s (0:26:20) ===============================================================                             
  Omni_Qwen3-Omni_Test  19m 9s  FAILED (exit 1)    
====================================================================== 7 passed, 24 warnings in 1665.63s (0:27:45) ======================================================================                                  
  Simple_Diffusion_Test  20m 1s  TIMED OUT                                           
  Simple_Engine_Entrypoints_Test  7m 9s  FAILED (exit 1)                             
  Simple_Model_Executor_Test  8m 34s  OK                                             
  Simple_Other_Test  14m 56s  OK          
  TTS_Higgs-Audio-V3_Test  2m 56s  FAILED (exit 1)                                   
========================================================= 2 passed, 3 skipped, 7 deselected, 14 warnings in 1001.54s (0:16:41) ==========================================================
  TTS_Qwen3-TTS_Base_Test  21m 38s  OK    
  TTS_Qwen3-TTS_CustomVoice_Test  27m 11s  OK                                        
  TTS_VoxCPM2_Test  17m 7s  OK            
Total wall time: 5h 51m 6s (22 jobs)      
Failed jobs: 9/22

local test for L3, 7 failed jobs fixed and passed. 2 failed job can not be tested in local (possible GPU error).

Signed-off-by: bowieshi <aoibosh@connect.hku.hk>

hsliuustc0106 · 2026-08-12T06:15:55Z

fix conflicts and let's get it merged

Signed-off-by: bowieshi <aoibosh@connect.hku.hk>

…n weekly CI main's fcf1085 (vllm-project#5146) made OmniRequestOutput inherit RequestOutput rather than nest one, so `.request_output` no longer exists. The e2e file and the offline example both still read it and raised AttributeError before producing any audio; the branch rebased past the refactor without migrating. Add the weekly Gepard step the `TO CONFIRM` note in test-merge.yml was asking for, and drop the note. The CI image carries no NanoCodec, so the step installs nemo_toolkit[tts] with the README's overrides first; without a step anywhere these tests ran in no tier at all, which is how the stale accessor survived. Also document `VLLM_USE_FLASHINFER_SAMPLER=0` for hosts whose CUDA toolkit cannot build the sampler kernels, matching the Qwen3-TTS and JoyAI-VL notes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Signed-off-by: Mengjie Zhao <zmj0129@gmail.com>

Blocking fixes from the maintainer review on vllm-project#3422. 1. Stale output accessor (offline example + offline e2e suite) `OmniRequestOutput` has subclassed `RequestOutput` since vllm-project#5146, so the nested `.request_output` no longer exists and both files crashed with `AttributeError` *after* synthesizing correctly. They were the only two users left in the tree, and `tests/docs/test_example_output_accessors.py` -- a cpu/core_model guard -- was already failing on them, so this was red at L1, not only in the GPU lane. Each item yielded by `Omni.generate` is itself an `OmniRequestOutput`, so read `.outputs` off it directly. Note `OmniRequestOutput` is *not* iterable: the `moss_tts_nano` example cited in review iterates it and would raise `TypeError`, so this follows `gepard/end2end.py` instead, which has the same two-level indexing. 2. Online requests could not influence Vevo2's sampling, including `seed` Vevo2 samples inside Amphion's `inference_ar_and_fm`, reading its knobs from `additional_information`; the `SamplingParams` the dummy AR scheduler carries only drive the forced-EOS sampler. `Vevo2Adapter.build` never copied anything across, so `/v1/audio/speech` seeds were inert and `default_sampling_params.seed` was a dead setting. `build()` runs before the shared path writes `request.seed` onto `SamplingParams`, so the `moss_tts.py` precedent would have wired the deploy default rather than the request seed. Read `request.seed` directly and fall back to the deploy default; also forward top_k / top_p / temperature / flow_matching_steps from `extra_params`. 3. Seeded requests leaked global RNG state across a batch The save/restore `finally` sat after `yield chunk`, so it only ran when the generator was resumed on a later `forward()` step -- leaving the process RNG seeded while batch siblings ran, and never returning to the original state under nested save/restore. Seeding now happens in a `_seeded_rng` context manager scoped to the inference call alone, never spanning a yield. Also in this change: - `forward()` creates every generator for a step before advancing any, and `_create_stream_gen` validates eagerly, so one malformed row no longer aborts a step that already consumed its siblings' chunks. The remaining raise-on-inference-failure is documented against the missing per-row error channel in `OmniOutput`. - `init_vevo2_checkpoint.py` is idempotent and repairs in place, re-raises instead of warning on a failed placeholder write, fails loudly on a missing AR sub-checkpoint, and verifies every root artifact before reporting success. The CI lane gates the download on a sub-checkpoint the initialiser never writes and always runs the initialiser. - The `LlamaConfig` positional shim is a scoped context manager instead of a permanent process-wide patch, and rejects duplicate positional/keyword arguments. Dropped the `rope_scaling` backfill: it never fired, and on transformers>=5 `rope_scaling` is a write-through property over `rope_parameters`, so forcing it to None would clear the rope config. - Deploy YAML and recipe restate the memory budget subtractively, matching `worker/base.py`: the pipeline is subtracted from `gpu_memory_utilization`, not stacked on top of it. - Recipe verification numbers re-attributed to the paths that produced them, and the broken `#vevo2-license-considerations` anchor fixed. - `run_server.sh` and the `end2end.py` docstring document the one-time init step and use `hf download`. - Tests: streaming lifecycle (waveform -> sentinel -> EOS -> cleanup, and the runaway-generation guard), RNG hygiene, eager validation, adapter seed plumbing, shim scoping, online seed reproducibility, a 400-not-5xx assertion, output duration bounds, and an offline batch test that actually distinguishes isolation from reproducibility. 18 -> 33 CPU unit tests. Signed-off-by: Yash Jain <yash@metica.com>

Blocking fixes from the maintainer review on vllm-project#3422. 1. Stale output accessor (offline example + offline e2e suite) `OmniRequestOutput` has subclassed `RequestOutput` since vllm-project#5146, so the nested `.request_output` no longer exists and both files crashed with `AttributeError` *after* synthesizing correctly. They were the only two users left in the tree, and `tests/docs/test_example_output_accessors.py` -- a cpu/core_model guard -- was already failing on them, so this was red at L1, not only in the GPU lane. Each item yielded by `Omni.generate` is itself an `OmniRequestOutput`, so read `.outputs` off it directly. Note `OmniRequestOutput` is *not* iterable: the `moss_tts_nano` example cited in review iterates it and would raise `TypeError`, so this follows `gepard/end2end.py` instead, which has the same two-level indexing. 2. Online requests could not influence Vevo2's sampling, including `seed` Vevo2 samples inside Amphion's `inference_ar_and_fm`, reading its knobs from `additional_information`; the `SamplingParams` the dummy AR scheduler carries only drive the forced-EOS sampler. `Vevo2Adapter.build` never copied anything across, so `/v1/audio/speech` seeds were inert and `default_sampling_params.seed` was a dead setting. `build()` runs before the shared path writes `request.seed` onto `SamplingParams`, so the `moss_tts.py` precedent would have wired the deploy default rather than the request seed. Read `request.seed` directly and fall back to the deploy default; also forward top_k / top_p / temperature / flow_matching_steps from `extra_params`. 3. Seeded requests leaked global RNG state across a batch The save/restore `finally` sat after `yield chunk`, so it only ran when the generator was resumed on a later `forward()` step -- leaving the process RNG seeded while batch siblings ran, and never returning to the original state under nested save/restore. Seeding now happens in a `_seeded_rng` context manager scoped to the inference call alone, never spanning a yield. Also in this change: - `forward()` creates every generator for a step before advancing any, and `_create_stream_gen` validates eagerly, so one malformed row no longer aborts a step that already consumed its siblings' chunks. The remaining raise-on-inference-failure is documented against the missing per-row error channel in `OmniOutput`. - `init_vevo2_checkpoint.py` is idempotent and repairs in place, re-raises instead of warning on a failed placeholder write, fails loudly on a missing AR sub-checkpoint, and verifies every root artifact before reporting success. The CI lane gates the download on a sub-checkpoint the initialiser never writes and always runs the initialiser. - The `LlamaConfig` positional shim is a scoped context manager instead of a permanent process-wide patch, and rejects duplicate positional/keyword arguments. Dropped the `rope_scaling` backfill: it never fired, and on transformers>=5 `rope_scaling` is a write-through property over `rope_parameters`, so forcing it to None would clear the rope config. - Deploy YAML and recipe restate the memory budget subtractively, matching `worker/base.py`: the pipeline is subtracted from `gpu_memory_utilization`, not stacked on top of it. - Recipe verification numbers re-attributed to the paths that produced them, and the broken `#vevo2-license-considerations` anchor fixed. - `run_server.sh` and the `end2end.py` docstring document the one-time init step and use `hf download`. - Tests: streaming lifecycle (waveform -> sentinel -> EOS -> cleanup, and the runaway-generation guard), RNG hygiene, eager validation, adapter seed plumbing, shim scoping, online seed reproducibility, a 400-not-5xx assertion, output duration bounds, and an offline batch test that actually distinguishes isolation from reproducibility. 18 -> 33 CPU unit tests. Signed-off-by: Yash Jain <yashjainjain1704@gmail.com>

…lm-project#5146 examples/offline_inference/text_to_speech/qwen3_tts/word_timestamps.py was deleted in vllm-project#4795 when word timestamps became a serving-path pipeline stage, re-added by vllm-project#5146 in a rebase, and later edited by vllm-project#6200. On main it fails at import: it imports `align` from vllm_omni.utils.forced_aligner, which no longer exists, and calls build_forced_aligner_config with the pre-vllm-project#4795 signature. Nothing else references the file; the README in the same directory already points to `vllm-omni serve --forced-aligner`. Co-authored-by: Claude Signed-off-by: Tianyao Wu <rayroy31@gmail.com>

Blocking fixes from the maintainer review on vllm-project#3422. 1. Stale output accessor (offline example + offline e2e suite) `OmniRequestOutput` has subclassed `RequestOutput` since vllm-project#5146, so the nested `.request_output` no longer exists and both files crashed with `AttributeError` *after* synthesizing correctly. They were the only two users left in the tree, and `tests/docs/test_example_output_accessors.py` -- a cpu/core_model guard -- was already failing on them, so this was red at L1, not only in the GPU lane. Each item yielded by `Omni.generate` is itself an `OmniRequestOutput`, so read `.outputs` off it directly. Note `OmniRequestOutput` is *not* iterable: the `moss_tts_nano` example cited in review iterates it and would raise `TypeError`, so this follows `gepard/end2end.py` instead, which has the same two-level indexing. 2. Online requests could not influence Vevo2's sampling, including `seed` Vevo2 samples inside Amphion's `inference_ar_and_fm`, reading its knobs from `additional_information`; the `SamplingParams` the dummy AR scheduler carries only drive the forced-EOS sampler. `Vevo2Adapter.build` never copied anything across, so `/v1/audio/speech` seeds were inert and `default_sampling_params.seed` was a dead setting. `build()` runs before the shared path writes `request.seed` onto `SamplingParams`, so the `moss_tts.py` precedent would have wired the deploy default rather than the request seed. Read `request.seed` directly and fall back to the deploy default; also forward top_k / top_p / temperature / flow_matching_steps from `extra_params`. 3. Seeded requests leaked global RNG state across a batch The save/restore `finally` sat after `yield chunk`, so it only ran when the generator was resumed on a later `forward()` step -- leaving the process RNG seeded while batch siblings ran, and never returning to the original state under nested save/restore. Seeding now happens in a `_seeded_rng` context manager scoped to the inference call alone, never spanning a yield. Also in this change: - `forward()` creates every generator for a step before advancing any, and `_create_stream_gen` validates eagerly, so one malformed row no longer aborts a step that already consumed its siblings' chunks. The remaining raise-on-inference-failure is documented against the missing per-row error channel in `OmniOutput`. - `init_vevo2_checkpoint.py` is idempotent and repairs in place, re-raises instead of warning on a failed placeholder write, fails loudly on a missing AR sub-checkpoint, and verifies every root artifact before reporting success. The CI lane gates the download on a sub-checkpoint the initialiser never writes and always runs the initialiser. - The `LlamaConfig` positional shim is a scoped context manager instead of a permanent process-wide patch, and rejects duplicate positional/keyword arguments. Dropped the `rope_scaling` backfill: it never fired, and on transformers>=5 `rope_scaling` is a write-through property over `rope_parameters`, so forcing it to None would clear the rope config. - Deploy YAML and recipe restate the memory budget subtractively, matching `worker/base.py`: the pipeline is subtracted from `gpu_memory_utilization`, not stacked on top of it. - Recipe verification numbers re-attributed to the paths that produced them, and the broken `#vevo2-license-considerations` anchor fixed. - `run_server.sh` and the `end2end.py` docstring document the one-time init step and use `hf download`. - Tests: streaming lifecycle (waveform -> sentinel -> EOS -> cleanup, and the runaway-generation guard), RNG hygiene, eager validation, adapter seed plumbing, shim scoping, online seed reproducibility, a 400-not-5xx assertion, output duration bounds, and an offline batch test that actually distinguishes isolation from reproducibility. 18 -> 33 CPU unit tests. Signed-off-by: Yash Jain <yashjainjain1704@gmail.com>

…uestOutput and no nested wrap-up (vllm-project#5146) Signed-off-by: Boao Shi <aoibosh@connect.hku.hk> Signed-off-by: bowieshi <aoibosh@connect.hku.hk> Signed-off-by: princepride <wangzhipeng628@gmail.com> Signed-off-by: R2-Y <ruiruyang2@gmail.com> Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com> Signed-off-by: Shi Boao Bowie <aoibosh@connect.hku.hk> Co-authored-by: 汪志鹏 <wangzhipeng628@gmail.com> Co-authored-by: ruirui(rein) yang <73573651+R2-Y@users.noreply.github.com> Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com> Co-authored-by: Gao Han <hgaoaf@connect.ust.hk> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

(#7233) Signed-off-by: Tianyao Wu <rayroy31@gmail.com>

bowieshi requested review from Gaohan123, alex-jw-brooks, hsliuustc0106, princepride, tzhouam, xuechendi, yuanheng-zhao and ywang96 as code owners July 16, 2026 09:03

hsliuustc0106 added the refactor refactoring for better code scalability and quality label Jul 17, 2026

hsliuustc0106 added the core related to core module: cache, scheduler, engine, worker, modelrunner label Jul 17, 2026 — with ChatGPT Codex Connector

Gaohan123 added this to the v0.26.0 milestone Jul 18, 2026

hsliuustc0106 reviewed Jul 18, 2026

View reviewed changes

bjf-frz reviewed Jul 20, 2026

View reviewed changes

hsliuustc0106 previously requested changes Jul 22, 2026

View reviewed changes

bowieshi requested review from Isotr0py, RuixiangMa, SamitHuang, Sy0307, ZJY0516, ZeldaHuang, amy-why-3459, david6666666, gcanlin, linyueqian, lishunyang12 and wtomin as code owners July 22, 2026 03:19

bowieshi added 2 commits August 10, 2026 18:18

fix CI error

0d7baa3

Signed-off-by: bowieshi <aoibosh@connect.hku.hk>

merge: 解决冲突并合并远程更新

a7966ee

Signed-off-by: bowieshi <aoibosh@connect.hku.hk>

bowieshi requested a review from Gaohan123 August 10, 2026 10:47

bowieshi and others added 4 commits August 10, 2026 18:55

fix no found tts2code2wav_token_only

9316f61

Signed-off-by: bowieshi <aoibosh@connect.hku.hk>

fix test_vae_decode_parallelism

87bbfce

Signed-off-by: bowieshi <aoibosh@connect.hku.hk>

fix pre-commit test_minicpmo_4_5_async_chunk.py

3c17811

Signed-off-by: bowieshi <aoibosh@connect.hku.hk>

Merge branch 'main' into feature/refactor-2

932d0f5

hsliuustc0106 added merge-test label to trigger buildkite merge test CI and removed ready label to trigger buildkite CI labels Aug 10, 2026

bowieshi added 2 commits August 12, 2026 14:39

fix conflic

6392b6b

Signed-off-by: bowieshi <aoibosh@connect.hku.hk>

Merge branch 'main' into feature/refactor-2

6ce63b8

hsliuustc0106 merged commit fcf1085 into vllm-project:main Aug 12, 2026
4 of 5 checks passed

bowieshi changed the title ~~[Refactor][OutputProcessor 2/8]: OmniRequestOutput should inherit RequestOutput and no nested wrap-up~~ [Refactor][OutputProcessor 2/3]: OmniRequestOutput should inherit RequestOutput and no nested wrap-up Aug 12, 2026

congw729 mentioned this pull request Aug 13, 2026

[Bug]: Nightly / CI failed - AttributeError: 'OmniRequestOutput' object has no attribute 'request_output' breaks 3 CUDA jobs (X2I Doc Test, LTX accuracy, VAE decode parallelism) after #5146 refactor #6133

Closed

linyueqian mentioned this pull request Aug 13, 2026

[Model] Add Gepard-1.0 native-AR FSQ/NanoCodec TTS (offline inference) #5666

Merged

3 of 6 tasks

MrlixiangWE mentioned this pull request Aug 13, 2026

[Bugfix] Stop reading the removed OmniRequestOutput.request_output accessor #6172

Merged

yashkgp mentioned this pull request Aug 22, 2026

[Bug]: moss_tts_nano offline example iterates OmniRequestOutput and raises TypeError #6479

Closed

Sy0307 mentioned this pull request Aug 22, 2026

[RFC][Tracking] Full-Duplex Serving Roadmap and PR Status #6494

Open

twu3202 mentioned this pull request Sep 8, 2026

[Misc] Remove stale offline word_timestamps example resurrected by #5146 #7233

Merged

hsliuustc0106 pushed a commit that referenced this pull request Sep 26, 2026

[Misc] Remove stale offline word_timestamps example resurrected by #5146

d418004

(#7233) Signed-off-by: Tianyao Wu <rayroy31@gmail.com>

Conversation

bowieshi commented Jul 16, 2026

Purpose

Test Plan

Test Result

Uh oh!

hsliuustc0106 commented Jul 17, 2026

Uh oh!

bowieshi commented Jul 17, 2026

Uh oh!

hsliuustc0106 Jul 18, 2026

Choose a reason for hiding this comment

Uh oh!

bowieshi Jul 20, 2026

Choose a reason for hiding this comment

Uh oh!

hsliuustc0106 Jul 18, 2026

Choose a reason for hiding this comment

Uh oh!

bowieshi Jul 20, 2026

Choose a reason for hiding this comment

Uh oh!

hsliuustc0106 Jul 20, 2026

Choose a reason for hiding this comment

Uh oh!

hsliuustc0106 commented Jul 20, 2026

Uh oh!

bjf-frz Jul 20, 2026

Choose a reason for hiding this comment

Uh oh!

bowieshi Jul 27, 2026

Choose a reason for hiding this comment

Uh oh!

hsliuustc0106 left a comment

Choose a reason for hiding this comment

Uh oh!

bowieshi commented Aug 10, 2026

Uh oh!

hsliuustc0106 commented Aug 12, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants