xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-24 12:05:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Model] Remove unnecessary CUDA sync of GLM-4.1V image and video preprocess (#24332)

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 13:10:54 +07:00 committed by GitHub
parent 61aa4b2901
commit 8a46602606
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/glm4_1v.py
View File

@ -1429,6 +1429,7 @@ class Glm4vForConditionalGeneration(nn.Module, SupportsMultiModal,
self, image_input: Glm4vImageInputs) -> tuple[torch.Tensor, ...]:
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"].type(self.visual.dtype)
@ -1443,13 +1444,15 @@ class Glm4vForConditionalGeneration(nn.Module, SupportsMultiModal,
image_embeds = self.visual(pixel_values,
grid_thw=grid_thw.tolist())
merge_size = self.visual.spatial_merge_size
sizes = grid_thw.prod(-1) // merge_size // merge_size
return image_embeds.split(sizes.tolist())
sizes = (torch.tensor(grid_thw_list, dtype=torch.long).prod(-1) //
(merge_size * merge_size)).tolist()
return image_embeds.split(sizes)
def _process_video_input(
self, video_input: Glm4vVideoInputs) -> 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"].type(self.visual.dtype)
@ -1466,8 +1469,9 @@ class Glm4vForConditionalGeneration(nn.Module, SupportsMultiModal,
grid_thw=grid_thw.tolist())
# Split concatenated embeddings for each video item.
merge_size = self.visual.spatial_merge_size
sizes = grid_thw.prod(-1) // merge_size // merge_size
return video_embeds.split(sizes.tolist())
sizes = (torch.tensor(grid_thw_list, dtype=torch.long).prod(-1) //
(merge_size * merge_size)).tolist()
return video_embeds.split(sizes)
def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
mm_input_by_modality = {}