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

[Core] More fixes to MultiModalEmbeddings type handling (#19715)

Browse Source 
Signed-off-by: Russell Bryant <rbryant@redhat.com>
This commit is contained in:
Russell Bryant
2025-06-18 18:48:29 -04:00
committed by GitHub
parent 04fefe7c9a
commit 14fdd21d39
35 changed files with 71 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/model_executor/models/qwen2_5_omni_thinker.py
View File

@@ -805,7 +805,8 @@ class Qwen2_5OmniThinkerForConditionalGeneration(
multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
) -> torch.Tensor:
inputs_embeds = self.language_model.get_input_embeddings(input_ids)
if multimodal_embeddings is not None:
if multimodal_embeddings is not None \
and len(multimodal_embeddings) != 0:
# TODO (ywang96): support overlapping modalitiy embeddings so that
# `use_audio_in_video` will work on V1.
@@ -845,7 +846,7 @@ class Qwen2_5OmniThinkerForConditionalGeneration(
multimodal_embeddings: Optional[NestedTensors] = None,
) -> torch.Tensor:
inputs_embeds = self.language_model.get_input_embeddings(input_ids)
if multimodal_embeddings is None:
if multimodal_embeddings is None or len(multimodal_embeddings) == 0:
return inputs_embeds
for embeddings, modality in multimodal_embeddings: