xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-26 06:23:05 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix glm4.1v video_grid_thw tensor shape scheme (#21744)

Browse Source 
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py 2025-07-28 19:26:49 +08:00 committed by GitHub
parent 65e8466c37
commit 0ae970ed15
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 3 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/model_executor/models/glm4_1v.py
View File

@ -126,7 +126,6 @@ class Glm4vVideoPixelInputs(TensorSchema):
- np: Number of patches
- ctpp: Number of channels * temporal_patch_size *
patch_size * patch_size
- nv: Number of videos
- f: Number of frames
- g: Grid dimensions (3 for grid_t which is usually 1 for processed
video, grid_h, grid_w)
@ -134,8 +133,7 @@ class Glm4vVideoPixelInputs(TensorSchema):
type: Literal["pixel_values_videos"] = "pixel_values_videos"
pixel_values_videos: Annotated[torch.Tensor, TensorShape("np", "ctpp")]
# video_metadata: Union[list[VideoMetadata], list[dict]]
video_grid_thw: Annotated[torch.Tensor, TensorShape("nv", "f", 3)]
video_grid_thw: Annotated[torch.Tensor, TensorShape("f", 3)]
class Glm4vVideoEmbeddingInputs(TensorSchema):
@ -143,14 +141,14 @@ class Glm4vVideoEmbeddingInputs(TensorSchema):
Dimensions:
- p: Number of video patches across all frames
- h: Hidden size (must match language model backbone)
- n: Number of videos
- f: Number of frames
- g: Grid dimensions (3 for grid_t which is usually 1 for processed
video, grid_h, grid_w)
"""
type: Literal["video_embeds"] = "video_embeds"
video_embeds: Annotated[torch.Tensor, TensorShape("p", "h")]
video_grid_thw: Annotated[torch.Tensor, TensorShape("n", 1, 3)]
video_grid_thw: Annotated[torch.Tensor, TensorShape("f", 3)]
Glm4vVideoInputs = Union[Glm4vVideoPixelInputs, Glm4vVideoEmbeddingInputs]
@ -1348,7 +1346,6 @@ class Glm4vForConditionalGeneration(nn.Module, SupportsMultiModal,
return Glm4vVideoPixelInputs(
type="pixel_values_videos",
# video_metadata=video_metadata,
pixel_values_videos=pixel_values_videos,
video_grid_thw=video_grid_thw,
)