[MM] Keep more metadata tensors on CPU (#52827)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
@@ -431,20 +431,20 @@ def _create_field_factory(
|
||||
image_embeds=MultiModalFieldConfig.flat_from_sizes(
|
||||
"image", image_embed_grid_sizes
|
||||
),
|
||||
image_grid_thw=MultiModalFieldConfig.batched("image"),
|
||||
image_grid_thw=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
|
||||
# OV2 first-class MM kwarg: per-patch (t,h,w)
|
||||
# positions required by the 3-D vision RoPE.
|
||||
patch_positions=MultiModalFieldConfig.flat_from_sizes(
|
||||
"image", image_pixel_grid_sizes
|
||||
"image", image_pixel_grid_sizes, keep_on_cpu=True
|
||||
),
|
||||
pixel_values_videos=MultiModalFieldConfig.flat_from_sizes(
|
||||
"video", video_patch_sizes
|
||||
),
|
||||
video_grid_thw=MultiModalFieldConfig.flat_from_sizes(
|
||||
"video", video_num_frames
|
||||
"video", video_num_frames, keep_on_cpu=True
|
||||
),
|
||||
patch_positions_videos=MultiModalFieldConfig.flat_from_sizes(
|
||||
"video", video_patch_sizes
|
||||
"video", video_patch_sizes, keep_on_cpu=True
|
||||
),
|
||||
video_num_frames=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
|
||||
frame_timestamps=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
|
||||
|
||||
Reference in New Issue
Block a user