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

[Bugfix] Check dimensions of multimodal embeddings in V1 (#15816)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-04-01 00:01:35 +08:00
committed by GitHub
parent e5ef4fa99a
commit 09e974d483
14 changed files with 98 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/models/gemma3_mm.py
View File

@@ -613,7 +613,7 @@ class Gemma3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP,
def _process_image_input(
self,
image_input: Gemma3ImageInputs,
) -> tuple[torch.Tensor, ...]:
) -> list[torch.Tensor]:
assert self.vision_tower is not None
pixel_values = image_input["pixel_values"]
@@ -625,7 +625,9 @@ class Gemma3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP,
)
image_embeds = self.multi_modal_projector(image_features)
return image_embeds.split(num_patches.tolist())
return [
e.flatten(0, 1) for e in image_embeds.split(num_patches.tolist())
]
def get_multimodal_embeddings(
self, **kwargs: object) -> Optional[MultiModalEmbeddings]: