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

50
vllm/model_executor/models/glm4_1v.py
View File

@@ -71,7 +71,6 @@ from vllm.multimodal.processing import (BaseMultiModalProcessor,
from vllm.multimodal.profiling import BaseDummyInputsBuilder
from vllm.platforms import _Backend
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.config import uses_mrope
from vllm.utils.tensor_schema import TensorSchema, TensorShape
from ..layers.activation import SiluAndMul
@@ -80,8 +79,7 @@ from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
from .qwen2_vl import (_create_qwen2vl_field_factory,
apply_rotary_pos_emb_vision)
from .utils import (AutoWeightsLoader, WeightsMapper,
init_vllm_registered_model, maybe_prefix,
merge_multimodal_embeddings)
init_vllm_registered_model, maybe_prefix)
from .vision import get_vit_attn_backend, run_dp_sharded_mrope_vision_model
logger = init_logger(__name__)
@@ -1552,32 +1550,6 @@ class Glm4vForConditionalGeneration(nn.Module, SupportsMultiModal,
multimodal_embeddings += video_embeddings
return multimodal_embeddings
def get_input_embeddings_v0(
self,
input_ids: torch.Tensor,
image_input: Optional[Glm4vImageInputs] = None,
video_input: Optional[Glm4vVideoInputs] = 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_id,
)
if video_input is not None:
video_embeds = self._process_video_input(video_input)
inputs_embeds = merge_multimodal_embeddings(
input_ids,
inputs_embeds,
video_embeds,
placeholder_token_id=self.config.video_token_id,
)
return inputs_embeds
def forward(
self,
input_ids: torch.Tensor,
@@ -1604,26 +1576,6 @@ class Glm4vForConditionalGeneration(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:
if uses_mrope(self.config):
assert positions.ndim == 2 and positions.size(0) == 3, (
"multimodal section rotary embedding requires "
f"(3, seq_len) positions, but got {positions.size()}")
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=input_ids,
positions=positions,