A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bugfix] fix Qwen2.5-Omni processor output mapping (#23058)
Signed-off-by: double7 <33449816+DoubleVII@users.noreply.github.com> Co-authored-by: 杨森 <yangsen.double7@bytedance.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
double7 committed
Aug 18, 2025 at 13:09 UTC
9f1c6422549d37eee22bfa4dbadaaa91d95e98ba
1 file changed
+5
vllm/model_executor/models/qwen2_5_omni_thinker.py
+5
@@ -88,6 +88,11 @@ def _qwen2_5_omni_thinker_field_config(hf_inputs: Mapping[str, torch.Tensor]):
88
video_grid_thw = hf_inputs.get("video_grid_thw", torch.empty((0, 3)))
89
video_grid_sizes = video_grid_thw.prod(-1)
90
91
+ # vllm use `second_per_grid_ts` to compute multimodal rotary embedding
92
+ video_second_per_grid = hf_inputs.get("video_second_per_grid", None)
93
+ if video_second_per_grid is not None:
94
+ hf_inputs["second_per_grid_ts"] = video_second_per_grid
95
+
96
return dict(
97
input_audio_features=MultiModalFieldConfig.flat_from_sizes(
98
"audio", audio_feature_lengths, dim=1),