xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-31 00:27:12 +08:00
Code Issues Packages Projects Releases Wiki Activity

Remove unnecessary CUDA sync of qwen image and video preprocess (#22792)

Browse Source 
Signed-off-by: cyy <cyyever@outlook.com>
Signed-off-by: Yuanyuan Chen <cyyever@outlook.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
This commit is contained in:
Yuanyuan Chen 2025-08-13 21:07:28 +08:00 committed by GitHub
parent fceafaf582
commit 6772bb0f7d
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 8 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/model_executor/models/qwen2_5_vl.py
View File

@ -976,10 +976,12 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
image_embeds = self.visual(pixel_values, grid_thw=grid_thw_list) image_embeds = self.visual(pixel_values, grid_thw=grid_thw_list)
# Split concatenated embeddings for each image item. # Split concatenated embeddings for each image item.
# Using prod on grid_thw_list instead of grid_thw.prod avoids CUDA sync
merge_size = self.visual.spatial_merge_size 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( def _process_video_input(
self, self,
@ -998,9 +1000,11 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
# Split concatenated embeddings for each video item. # Split concatenated embeddings for each video item.
merge_size = self.visual.spatial_merge_size merge_size = self.visual.spatial_merge_size
sizes = grid_thw.prod(-1) // merge_size // merge_size # Using prod on grid_thw_list instead of grid_thw.prod avoids CUDA sync
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: def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
mm_input_by_modality = {} mm_input_by_modality = {}