Skip to content

Interpolate fix on cuda for large output tensors #10067

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
merged 4 commits into from
Dec 2, 2024
Merged

Interpolate fix on cuda for large output tensors #10067

Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
35085ea
Workaround for upscale with large output tensors.
pcuenca Dec 1, 2024
73177df
Fix scale when output_size is given
pcuenca Dec 1, 2024
e527813
Style
pcuenca Dec 1, 2024
b2e1c2f
Merge branch 'main' into interpolate-fix
sayakpaul Dec 2, 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
8 changes: 8 additions & 0 deletions src/diffusers/models/upsampling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -165,6 +165,14 @@ def forward(self, hidden_states: torch.Tensor, output_size: Optional[int] = None
# if `output_size` is passed we force the interpolation output
# size and do not make use of `scale_factor=2`
if self.interpolate:
# upsample_nearest_nhwc also fails when the number of output elements is large
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Do we need to check for channels_first or channels_last memory layout?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

My assumption is it will behave the same way for any non-contiguous layouts, and that calling .contiguous() on a contiguous tensor will be a no-op. I'm also mimicking the same structure as in lines 161-163 above.

# https://github.com/pytorch/pytorch/issues/141831
scale_factor = (
2 if output_size is None else max([f / s for f, s in zip(output_size, hidden_states.shape[-2:])])
)
if hidden_states.numel() * scale_factor > pow(2, 31):
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Consider keeping pow(2, 31) in a variable and then using it.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I feel that pow(2, 31) is easy to understand and provides explicit documentation about what's happening. Using another level of indirection wouldn't help, unless we are concerned about performance, in which case I can use a constant.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Along the same lines as #10067 (comment).

hidden_states = hidden_states.contiguous()

if output_size is None:
hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest")
else:
Expand Down
Loading