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

43
vllm/model_executor/models/nano_nemotron_vl.py
View File

@@ -35,8 +35,7 @@ from vllm.model_executor.models.nemotron_h import NemotronHForCausalLM
from vllm.model_executor.models.radio import RadioModel
from vllm.model_executor.models.utils import (flatten_bn,
init_vllm_registered_model,
maybe_prefix,
merge_multimodal_embeddings)
isin_list, maybe_prefix)
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
MultiModalKwargs, MultiModalKwargsItems,
@@ -1096,8 +1095,8 @@ class NemotronH_Nano_VL_V2(nn.Module, HasInnerState, IsHybrid,
return modalities
def get_multimodal_embeddings(
self, **kwargs: object) -> Optional[MultiModalEmbeddings]:
def get_multimodal_embeddings(self,
**kwargs: object) -> MultiModalEmbeddings:
# Validate the multimodal input keyword arguments
modalities = self._parse_and_validate_multimodal_inputs(**kwargs)
if modalities is None:
@@ -1121,30 +1120,6 @@ class NemotronH_Nano_VL_V2(nn.Module, HasInnerState, IsHybrid,
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):
context_token_ids = [
token_id for token_id in (self.img_context_token_id,
self.video_context_token_id)
if token_id is not None
]
assert len(context_token_ids) >= 1
inputs_embeds = merge_multimodal_embeddings(
input_ids,
inputs_embeds,
multimodal_embeddings,
context_token_ids,
)
return inputs_embeds
def get_language_model(self) -> torch.nn.Module:
return self.language_model
@@ -1163,9 +1138,17 @@ class NemotronH_Nano_VL_V2(nn.Module, HasInnerState, IsHybrid,
# NOTE: In v1, inputs_embeds is always generated at model runner, this
# condition is for v0 compatibility.
elif inputs_embeds is None:
context_token_ids = [
token_id for token_id in (self.img_context_token_id,
self.video_context_token_id)
if token_id is not 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, context_token_ids),
)
input_ids = None
hidden_states = self.language_model(