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

35
vllm/model_executor/models/glm4v.py
View File

@@ -43,7 +43,7 @@ from vllm.utils.tensor_schema import TensorSchema, TensorShape
from .chatglm import ChatGLMBaseModel, ChatGLMModel
from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
SupportsMultiModal, SupportsPP)
from .utils import flatten_bn, merge_multimodal_embeddings
from .utils import flatten_bn, isin_list
class GLMVImagePixelInputs(TensorSchema):
@@ -607,28 +607,6 @@ class GLM4VForCausalLM(ChatGLMBaseModel, SupportsLoRA, SupportsPP,
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=input_ids,
inputs_embeds=inputs_embeds,
multimodal_embeddings=multimodal_embeddings,
placeholder_token_id=[
self.config.boi_token_id,
self.config.pad_token_id,
self.config.eoi_token_id,
],
)
return inputs_embeds
def forward(
self,
input_ids: torch.Tensor,
@@ -644,8 +622,15 @@ class GLM4VForCausalLM(ChatGLMBaseModel, SupportsLoRA, SupportsPP,
# 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=isin_list(input_ids, [
self.config.boi_token_id,
self.config.pad_token_id,
self.config.eoi_token_id,
]),
)
input_ids = None
hidden_states = self.transformer(input_ids, positions,