biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[V0 Deprecation][Models] Remove all V0 condition for mm embeddings merge (#25331)

Browse Source 
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: isotr0py <2037008807@qq.com>
This commit is contained in:
Isotr0py
2025-09-29 14:09:18 +08:00
committed by GitHub
parent 65ecb4f134
commit bd51f78e39
42 changed files with 13 additions and 809 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
vllm/model_executor/models/phi4mm.py
View File

@@ -35,8 +35,7 @@ from vllm.utils.tensor_schema import TensorSchema, TensorShape
from .idefics2_vision_model import Idefics2VisionTransformer
from .interfaces import MultiModalEmbeddings, SupportsLoRA, SupportsMultiModal
from .phi4mm_audio import AudioEmbedding
from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn, maybe_prefix,
merge_multimodal_embeddings)
from .utils import AutoWeightsLoader, WeightsMapper, flatten_bn, maybe_prefix
# <|endoftext10|> (see vocab.json in hf model)
_IMAGE_PLACEHOLDER_TOKEN_ID = 200010
@@ -1174,35 +1173,6 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
return multimodal_embeddings
def get_input_embeddings_v0(
self,
input_ids: torch.Tensor,
image_input: Optional[Phi4MMImagePixelInputs] = None,
audio_input: Optional[Phi4MMAudioFeatureInputs] = None,
) -> torch.Tensor:
audio_projection_mode = 'speech'
inputs_embeds = self.get_input_embeddings(input_ids)
if image_input is not None:
image_embeds = self._process_image_input(image_input)
inputs_embeds = merge_multimodal_embeddings(
input_ids,
inputs_embeds,
image_embeds,
placeholder_token_id=_IMAGE_PLACEHOLDER_TOKEN_ID,
)
audio_projection_mode = 'vision'
if audio_input is not None:
audio_embeds = self._process_audio_input(
audio_input, audio_projection_mode=audio_projection_mode)
inputs_embeds = merge_multimodal_embeddings(
input_ids,
inputs_embeds,
audio_embeds,
placeholder_token_id=_AUDIO_PLACEHOLDER_TOKEN_ID,
)
return inputs_embeds
def forward(
self,
input_ids: torch.Tensor,
@@ -1214,22 +1184,6 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
if intermediate_tensors is not None:
inputs_embeds = None
# NOTE: In v1, inputs_embeds is always generated at model runner from
# `get_multimodal_embeddings` and `get_input_embeddings`, this
# condition is only for v0 compatibility.
elif inputs_embeds is None:
image_input = self._parse_and_validate_image_input(**kwargs)
audio_input = self._parse_and_validate_audio_input(**kwargs)
if image_input is None and audio_input is None:
inputs_embeds = None
else:
inputs_embeds = self.get_input_embeddings_v0(
input_ids,
image_input=image_input,
audio_input=audio_input)
input_ids = None
hidden_states = self.model(
input_ids,
positions,