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

25
vllm/model_executor/models/qwen_vl.py
View File

@@ -45,7 +45,7 @@ from vllm.utils.tensor_schema import TensorSchema, TensorShape
from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
SupportsMultiModal, SupportsPP)
from .qwen import QWenBaseModel, QWenModel
from .utils import flatten_bn, merge_multimodal_embeddings
from .utils import flatten_bn
class QwenImagePixelInputs(TensorSchema):
@@ -756,21 +756,6 @@ class QwenVLForConditionalGeneration(QWenBaseModel, SupportsPP, SupportsLoRA,
vision_embeddings = self._process_image_input(image_input)
return vision_embeddings
def get_input_embeddings(
self,
input_ids: torch.Tensor,
multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
) -> torch.Tensor:
inputs_embeds = self.transformer.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.transformer.visual.image_pad_id)
return inputs_embeds
def forward(
self,
input_ids: torch.Tensor,
@@ -786,8 +771,12 @@ class QwenVLForConditionalGeneration(QWenBaseModel, SupportsPP, SupportsLoRA,
# condition is for v0 compatibility.
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.transformer.visual.image_pad_id,
)
input_ids = None
hidden_states = self.transformer(input_ids, positions,