frozenleaves/vllm
1
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2025-10-20 14:53:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Model] Remove unnecessary CUDA sync of Qwen2VL image and video preprocess (#24334)

Browse Source 
Signed-off-by: Win <chatcharinsang@gmail.com>
Co-authored-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Chatcharin Sangbutsarakum
2025-09-08 13:11:12 +07:00
committed by GitHub
parent 8a46602606
commit 60f0843ef8
1 changed files with 8 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/model_executor/models/qwen2_vl.py
View File

@ -1218,6 +1218,7 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
grid_thw = image_input["image_grid_thw"]
assert grid_thw.ndim == 2
grid_thw_list = grid_thw.tolist()
if image_input["type"] == "image_embeds":
image_embeds = image_input["image_embeds"]
@ -1227,15 +1228,17 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
# Split concatenated embeddings for each image item.
merge_size = self.visual.spatial_merge_size
sizes = grid_thw.prod(-1) // merge_size // merge_size
sizes = (torch.tensor(grid_thw_list, dtype=torch.long).prod(-1) //
(merge_size * merge_size)).tolist()
return image_embeds.split(sizes.tolist())
return image_embeds.split(sizes)
def _process_video_input(
self, video_input: Qwen2VLVideoInputs) -> tuple[torch.Tensor, ...]:
grid_thw = video_input["video_grid_thw"]
assert grid_thw.ndim == 2
grid_thw_list = grid_thw.tolist()
if video_input["type"] == "video_embeds":
video_embeds = video_input["video_embeds"]
@ -1245,9 +1248,10 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
# Split concatenated embeddings for each video item.
merge_size = self.visual.spatial_merge_size
sizes = grid_thw.prod(-1) // merge_size // merge_size
sizes = (torch.tensor(grid_thw_list, dtype=torch.long).prod(-1) //
(merge_size * merge_size)).tolist()
return video_embeds.split(sizes.tolist())
return video_embeds.split(sizes)
def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
modalities = {}