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

25
vllm/model_executor/models/midashenglm.py
View File

@@ -54,8 +54,7 @@ from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.configs.midashenglm import DashengConfig
from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
from .utils import (AutoWeightsLoader, init_vllm_registered_model,
maybe_prefix, merge_multimodal_embeddings)
from .utils import AutoWeightsLoader, init_vllm_registered_model, maybe_prefix
_Tuple2 = Union[int, tuple[int, int], Sequence[int]]
@@ -744,21 +743,6 @@ class MiDashengLMModel(nn.Module, SupportsMultiModal, SupportsPP):
return []
return self._process_audio_input(audio_input)
def get_input_embeddings(
self,
input_ids: torch.Tensor,
multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
) -> torch.Tensor:
inputs_embeds = self.decoder.get_input_embeddings(input_ids)
if multimodal_embeddings and len(multimodal_embeddings) > 0:
inputs_embeds = merge_multimodal_embeddings(
input_ids,
inputs_embeds,
multimodal_embeddings,
self.config.audio_token_id,
)
return inputs_embeds
def forward(
self,
input_ids: torch.Tensor,
@@ -771,8 +755,11 @@ class MiDashengLMModel(nn.Module, SupportsMultiModal, SupportsPP):
inputs_embeds = None
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=input_ids == self.config.audio_token_id,
)
input_ids = None
return self.decoder.model(input_ids,