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

46
vllm/model_executor/models/llava_onevision.py
View File

@@ -30,8 +30,7 @@ from .llava_next import (BaseLlavaNextMultiModalProcessor, LlavaNextLikeConfig,
LlavaNextProcessingInfo)
from .siglip import SiglipVisionModel
from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
init_vllm_registered_model, maybe_prefix,
merge_multimodal_embeddings)
init_vllm_registered_model, maybe_prefix)
# For profile run
_MAX_FRAMES_PER_VIDEO = 16
@@ -850,33 +849,6 @@ class LlavaOnevisionForConditionalGeneration(nn.Module, SupportsMultiModal,
return multimodal_embeddings
def get_input_embeddings_v0(
self,
input_ids: torch.Tensor,
image_input: Optional[LlavaOnevisionImagePixelInputs] = None,
video_input: Optional[LlavaOnevisionVideoPixelInputs] = None,
) -> torch.Tensor:
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=self.config.image_token_index,
)
if video_input is not None:
video_embeds = self._process_video_pixels(video_input)
inputs_embeds = merge_multimodal_embeddings(
input_ids,
inputs_embeds,
video_embeds,
placeholder_token_id=self.config.video_token_index,
)
return inputs_embeds
def forward(
self,
input_ids: torch.Tensor,
@@ -894,22 +866,6 @@ class LlavaOnevisionForConditionalGeneration(nn.Module, 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)
video_input = self._parse_and_validate_video_input(**kwargs)
if image_input is None and video_input is None:
inputs_embeds = None
else:
inputs_embeds = self.get_input_embeddings_v0(
input_ids,
image_input=image_input,
video_input=video_input)
input_ids = None
hidden_states = self.language_model.model(input_ids,
positions,
intermediate_tensors,