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

[Bugfix] Merge MM embeddings by index instead of token IDs (#16229)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: NickLucche <nlucches@redhat.com>
Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: NickLucche <nlucches@redhat.com>
Co-authored-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Cyrus Leung
2025-09-27 16:15:12 +08:00
committed by GitHub
parent 176173989a
commit 27d7638b94
80 changed files with 966 additions and 1139 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/model_executor/models/phi4mm.py
View File

@@ -1151,7 +1151,6 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
modalities = self._parse_and_validate_multimodal_inputs(**kwargs)
if not modalities:
return []
return None
# The result multimodal_embeddings is tuple of tensors, with each
# tensor corresponding to a multimodal data item (image or video).
@@ -1175,19 +1174,6 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
return multimodal_embeddings
def get_input_embeddings(
self,
input_ids: torch.Tensor,
multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
) -> torch.Tensor:
inputs_embeds = self.model.embed_tokens(input_ids)
if multimodal_embeddings is not None and len(
multimodal_embeddings) != 0:
inputs_embeds = merge_multimodal_embeddings(
input_ids, inputs_embeds, multimodal_embeddings,
[_IMAGE_PLACEHOLDER_TOKEN_ID, _AUDIO_PLACEHOLDER_TOKEN_ID])
return inputs_embeds
def get_input_embeddings_v0(
self,
input_ids: torch.Tensor,