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

26
vllm/model_executor/models/gemma3_mm.py
View File

@@ -37,8 +37,7 @@ from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
SupportsMultiModal, SupportsPP)
from .siglip import SiglipVisionModel
from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
init_vllm_registered_model, maybe_prefix,
merge_multimodal_embeddings)
init_vllm_registered_model, maybe_prefix)
logger = init_logger(__name__)
@@ -588,22 +587,6 @@ class Gemma3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP,
return self._process_image_input(image_input)
def get_input_embeddings(
self,
input_ids: torch.Tensor,
multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
) -> torch.Tensor:
inputs_embeds = self.language_model.get_input_embeddings(input_ids)
if multimodal_embeddings is not None \
and len(multimodal_embeddings) != 0:
inputs_embeds = merge_multimodal_embeddings(
input_ids,
inputs_embeds,
multimodal_embeddings,
self.config.image_token_index,
)
return inputs_embeds
def forward(self,
input_ids: torch.Tensor,
positions: torch.Tensor,
@@ -618,8 +601,11 @@ class Gemma3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP,
elif inputs_embeds is None:
vision_embeddings = self.get_multimodal_embeddings(**kwargs)
inputs_embeds = self.get_input_embeddings(input_ids,
vision_embeddings)
inputs_embeds = self.get_input_embeddings(
input_ids,
vision_embeddings,
is_multimodal=input_ids == self.config.image_token_index,
)
if (vision_embeddings is not None) and len(vision_embeddings) != 0:
kwargs = self.prepare_attn_masks(
input_ids,