mirrors/segment-anything
1
0
Fork 0
mirror of https://github.com/facebookresearch/segment-anything.git synced 2025-06-03 14:59:27 +08:00
Code Issues Packages Projects Releases Wiki Activity

Update on "fixing image_encoder to work with cuda_graphs"

Browse Source 
Summary: the combination of tensors on multiple devices in get_rel_pos
was preventing cuda graphs from correctly optimizing things

Test Plan:

Reviewers:

Subscribers:

Tasks:

Tags:

[ghstack-poisoned]
This commit is contained in:
Charles Hernandez 2023-05-30 20:37:47 +00:00
parent 9aacb82524
commit 51bc7a2e0b
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
segment_anything/modeling/image_encoder.py
View File

@ -315,8 +315,8 @@ def get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor
rel_pos_resized = rel_pos
# Scale the coords with short length if shapes for q and k are different.
q_coords = (torch.arange(q_size).to(rel_pos.device)[:, None] * max(k_size / q_size, 1.0))
k_coords = (torch.arange(k_size).to(rel_pos.device)[None, :] * max(q_size / k_size, 1.0))
q_coords = (torch.arange(q_size, device=rel_pos.device)[:, None] * max(k_size / q_size, 1.0))
k_coords = (torch.arange(k_size, device=rel_pos.device)[None, :] * max(q_size / k_size, 1.0))
relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
return rel_pos_resized[relative_coords.long()]