Skip to content

Fix dtype casting in swinv2 and swinv2sr to allow non-FP32 inference #31589

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
amyeroberts merged 8 commits into from
Jun 26, 2024
Merged

Fix dtype casting in swinv2 and swinv2sr to allow non-FP32 inference #31589

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
d87a5e2
Fix dtype casting in modeling_swin2sr to allow non-FP32 inference
aliencaocao Jun 25, 2024
631174c
Fix formattting
aliencaocao Jun 25, 2024
2b83074
Fix for swinv2 too
aliencaocao Jun 25, 2024
bfe3ff8
Update src/transformers/models/swin2sr/modeling_swin2sr.py
aliencaocao Jun 25, 2024
3ddf1e9
Update src/transformers/models/swinv2/modeling_swinv2.py
aliencaocao Jun 25, 2024
606f1d3
Add FP16 tests for swin2sr and swinv2
aliencaocao Jun 25, 2024
1a2a92e
[run_slow] swin2sr, swinv2
aliencaocao Jun 26, 2024
266a6e1
[run_slow] swin2sr, swinv2
aliencaocao Jun 26, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions src/transformers/models/swin2sr/modeling_swin2sr.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -301,6 +301,9 @@ def __init__(self, config, dim, num_heads, window_size, pretrained_window_size=[
relative_coords_table = (
torch.sign(relative_coords_table) * torch.log2(torch.abs(relative_coords_table) + 1.0) / math.log2(8)
)
relative_coords_table = relative_coords_table.to(
next(self.continuous_position_bias_mlp.parameters()).dtype
) # set to same dtype as mlp weight
self.register_buffer("relative_coords_table", relative_coords_table, persistent=False)

# get pair-wise relative position index for each token inside the window
Expand Down
3 changes: 3 additions & 0 deletions src/transformers/models/swinv2/modeling_swinv2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -492,6 +492,9 @@ def __init__(self, config, dim, num_heads, window_size, pretrained_window_size=[
relative_coords_table = (
torch.sign(relative_coords_table) * torch.log2(torch.abs(relative_coords_table) + 1.0) / math.log2(8)
)
relative_coords_table = relative_coords_table.to(
next(self.continuous_position_bias_mlp.parameters()).dtype
) # set to same dtype as mlp weight
self.register_buffer("relative_coords_table", relative_coords_table, persistent=False)

# get pair-wise relative position index for each token inside the window
Expand Down