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/pixtral.py
View File

@@ -50,8 +50,7 @@ from vllm.transformers_utils.tokenizer import (MistralTokenizer,
from vllm.utils.tensor_schema import TensorSchema, TensorShape
from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
from .utils import (flatten_bn, init_vllm_registered_model, maybe_prefix,
merge_multimodal_embeddings)
from .utils import flatten_bn, init_vllm_registered_model, maybe_prefix
from .vision import VisionEncoderInfo, resolve_visual_encoder_outputs
try:
@@ -433,22 +432,6 @@ class PixtralForConditionalGeneration(nn.Module, SupportsMultiModal,
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.vision_args.image_token_id,
)
return inputs_embeds
def forward(
self,
input_ids: torch.Tensor,
@@ -465,8 +448,11 @@ class PixtralForConditionalGeneration(nn.Module, SupportsMultiModal,
# 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.vision_args.image_token_id,
)
input_ids = None
hidden_states = self.language_model.model(input_ids,