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/hyperclovax_vision.py
View File

@@ -45,8 +45,8 @@ from vllm.sequence import IntermediateTensors
from .clip import CLIPVisionModel
from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
from .siglip import SiglipVisionModel
from .utils import (AutoWeightsLoader, init_vllm_registered_model,
maybe_prefix, merge_multimodal_embeddings)
from .utils import (AutoWeightsLoader, init_vllm_registered_model, isin_list,
maybe_prefix)
from .vision import get_vision_encoder_info
EOT = "<|endofturn|>"
@@ -691,7 +691,7 @@ class HCXVisionForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
def get_multimodal_embeddings(
self,
**kwargs: Unpack[HCXVisionMultimodalInputs],
) -> Optional[MultiModalEmbeddings]:
) -> MultiModalEmbeddings:
multimodal_embeddings = list()
if kwargs.get("pixel_values_images") is not None:
@@ -736,26 +736,6 @@ class HCXVisionForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
multimodal_embeddings.append(_multimodal_embeddings_videos)
return multimodal_embeddings
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,
placeholder_token_id=[
self.config.image_token_id,
self.config.video_token_id,
],
)
return inputs_embeds
def forward(
self,
input_ids: torch.Tensor,
@@ -771,8 +751,13 @@ class HCXVisionForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
# condition is for v0 compatibility.
elif inputs_embeds is None:
multimodal_embeddings = self.get_multimodal_embeddings(**kwargs)
inputs_embeds = self.get_input_embeddings(input_ids,
multimodal_embeddings)
inputs_embeds = self.get_input_embeddings(
input_ids,
multimodal_embeddings,
is_multimodal=isin_list(
input_ids,
[self.config.image_token_id, self.config.video_token_id]),
)
input_ids = None
hidden_states = self.language_model.model(input_ids,
positions,