biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[BUGFIX]Fix Qwen-Omni models audio max_token_per_item estimation error leading to encoder_cache_size is 0 (#35994)

Browse Source 
Signed-off-by: Miao, Avery <avery.miao@intel.com>
This commit is contained in:
Avery Miao
2026-03-06 01:16:29 +08:00
committed by GitHub
parent 6a895197fa
commit e998fa76b9
3 changed files with 86 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
vllm/model_executor/models/qwen2_audio.py
View File

@@ -179,6 +179,26 @@ class Qwen2AudioProcessingInfo(BaseProcessingInfo):
def get_supported_mm_limits(self) -> Mapping[str, int | None]:
return {"audio": None}
def get_mm_max_tokens_per_item(
self,
seq_len: int,
mm_counts: Mapping[str, int] | None = None,
) -> Mapping[str, int]:
mm_counts = mm_counts or {}
if mm_counts.get("audio", 0) <= 0:
return {}
feature_extractor = self.get_feature_extractor()
chunk_length = min(feature_extractor.chunk_length, 30)
audio_len = int(chunk_length * feature_extractor.sampling_rate)
hop_length = feature_extractor.hop_length
max_mel_seq_len = audio_len // hop_length
input_lengths = torch.tensor([max_mel_seq_len], dtype=torch.long)
_, output_lengths = _get_feat_extract_output_lengths(input_lengths)
return {"audio": int(output_lengths.item())}
class Qwen2AudioDummyInputsBuilder(BaseDummyInputsBuilder[Qwen2AudioProcessingInfo]):
def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: