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

33
vllm/model_executor/models/qwen3_omni_moe_thinker.py
View File

@@ -1163,6 +1163,39 @@ class Qwen3OmniMoeThinkerProcessingInfo(
def get_supported_mm_limits(self) -> Mapping[str, int | None]:
return {"audio": None, "image": None, "video": None}
def get_mm_max_tokens_per_item(
self,
seq_len: int,
mm_counts: Mapping[str, int] | None = None,
) -> Mapping[str, int] | None:
mm_counts = mm_counts or {}
requested_modalities = {m for m, c in mm_counts.items() if c > 0}
mm_max_tokens: dict[str, int] = {}
if requested_modalities & {"image", "video"}:
vl_tokens = Qwen2_5_VLProcessingInfo.get_mm_max_tokens_per_item(
self,
seq_len=seq_len,
mm_counts=mm_counts,
)
mm_max_tokens.update(
{
m: vl_tokens[m]
for m in ["image", "video"]
if m in requested_modalities
}
)
if "audio" in requested_modalities:
audio_tokens = Qwen2AudioProcessingInfo.get_mm_max_tokens_per_item(
self,
seq_len=seq_len,
mm_counts=mm_counts,
)
mm_max_tokens["audio"] = audio_tokens["audio"]
return mm_max_tokens
Qwen3OmniMoeThinkerDummyInputsBuilder = Qwen2_5OmniThinkerDummyInputsBuilder