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

37
vllm/model_executor/models/nemotron_vl.py
View File

@@ -38,7 +38,7 @@ from vllm.transformers_utils.tokenizer import AnyTokenizer
from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
SupportsMultiModal, SupportsPP)
from .utils import (AutoWeightsLoader, flatten_bn, init_vllm_registered_model,
maybe_prefix, merge_multimodal_embeddings)
maybe_prefix)
IMG_START = '<img>'
IMG_END = '</img>'
@@ -576,20 +576,24 @@ class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP,
self,
input_ids: torch.Tensor,
multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
*,
is_multimodal: Optional[torch.Tensor] = None,
handle_oov_mm_token: bool = False,
) -> torch.Tensor:
inputs_embeds = self.language_model.get_input_embeddings(input_ids)
if multimodal_embeddings is not None \
and len(multimodal_embeddings) != 0:
context_token_ids = [self.img_context_token_id]
assert len(context_token_ids) >= 1
if multimodal_embeddings is not None and len(
multimodal_embeddings) > 0:
self._set_visual_token_mask(input_ids)
inputs_embeds = merge_multimodal_embeddings(
input_ids,
inputs_embeds,
multimodal_embeddings,
context_token_ids,
)
return inputs_embeds
# This is to satisfy the type checker for each overload
if multimodal_embeddings is None or is_multimodal is None:
return super().get_input_embeddings(input_ids)
return super().get_input_embeddings(
input_ids,
multimodal_embeddings=multimodal_embeddings,
is_multimodal=is_multimodal,
handle_oov_mm_token=handle_oov_mm_token,
)
def forward(
self,
@@ -608,8 +612,11 @@ class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, 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=input_ids == self.img_context_token_id,
)
input_ids = None
forward_kwargs = {