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

93
vllm/model_executor/models/qwen3_vl.py
View File

@@ -68,7 +68,6 @@ 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 import is_list_of
from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
@@ -82,8 +81,7 @@ from .qwen2_5_vl import (Qwen2_5_VisionAttention,
from .qwen2_vl import Qwen2VLProcessingInfo
from .qwen3 import Qwen3ForCausalLM, Qwen3Model
from .utils import (AutoWeightsLoader, PPMissingLayer, WeightsMapper,
_merge_multimodal_embeddings, maybe_prefix,
merge_multimodal_embeddings)
_merge_multimodal_embeddings, maybe_prefix)
from .vision import get_vit_attn_backend, run_dp_sharded_mrope_vision_model
logger = init_logger(__name__)
@@ -1464,75 +1462,6 @@ class Qwen3VLForConditionalGeneration(nn.Module, SupportsMultiModal,
return inputs_embeds
def get_input_embeddings_v0(
self,
input_ids: torch.Tensor,
image_input: Optional[Qwen2_5_VLImageInputs] = None,
video_input: Optional[Qwen2_5_VLVideoInputs] = None,
) -> torch.Tensor:
inputs_embeds = self.get_input_embeddings(input_ids)
if self.use_deepstack:
visual_dim = inputs_embeds.shape[-1]
deepstack_input_embeds = None
if image_input is not None or video_input is not None:
deepstack_input_embeds = torch.zeros_like(
inputs_embeds).unsqueeze(1).repeat(
1, self.deepstack_num_level, 1).flatten(1)
if image_input is not None:
image_embeds = self._process_image_input(image_input)
if self.use_deepstack:
image_embeds = torch.cat(image_embeds)
image_embeds, image_embeds_multiscale = image_embeds.split(
[visual_dim, visual_dim * self.deepstack_num_level],
dim=-1)
deepstack_input_embeds = merge_multimodal_embeddings(
input_ids,
deepstack_input_embeds,
image_embeds_multiscale,
placeholder_token_id=self.config.image_token_id,
)
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)
if self.use_deepstack:
video_embeds = torch.cat(video_embeds)
video_embeds, video_embeds_multiscale = video_embeds.split(
[visual_dim, visual_dim * self.deepstack_num_level],
dim=-1)
deepstack_input_embeds = merge_multimodal_embeddings(
input_ids,
deepstack_input_embeds,
video_embeds_multiscale,
placeholder_token_id=self.config.video_token_id,
)
inputs_embeds = merge_multimodal_embeddings(
input_ids,
inputs_embeds,
video_embeds,
placeholder_token_id=self.config.video_token_id,
)
if self.use_deepstack and deepstack_input_embeds is not None:
deepstack_input_embeds = deepstack_input_embeds.view(
inputs_embeds.shape[0], self.deepstack_num_level,
visual_dim).permute(1, 0, 2).contiguous()
self._set_deepstack_input_embeds(deepstack_input_embeds)
return inputs_embeds
def forward(
self,
input_ids: torch.Tensor,
@@ -1568,26 +1497,6 @@ class Qwen3VLForConditionalGeneration(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
if self.use_deepstack and inputs_embeds is not None and get_pp_group(
).is_first_rank:
deepstack_input_embeds = self._get_deepstack_input_embeds(