Skip to content

perf: optimize CausalConv3d for wan autoencoders #12800

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

Open
c8ef wants to merge 2 commits into
base: main
Choose a base branch
from
+10 −6
Open

perf: optimize CausalConv3d for wan autoencoders #12800

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
411cfee
perf: optimize CasualConv3d for wan autoencoders
c8ef Dec 6, 2025
a9cdb5f
revise
c8ef Dec 6, 2025
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
16 changes: 10 additions & 6 deletions src/diffusers/models/autoencoders/autoencoder_kl_wan.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -162,16 +162,20 @@ def __init__(
)

# Set up causal padding
self._padding = (self.padding[2], self.padding[2], self.padding[1], self.padding[1], 2 * self.padding[0], 0)
self.padding = (0, 0, 0)
self.temporal_padding = 2 * self.padding[0]
# Keep spatial padding, remove temporal padding from conv layer
self.padding = (0, self.padding[1], self.padding[2])
Comment on lines +165 to +167
Copy link
Member

@sayakpaul sayakpaul Dec 7, 2025

Choose a reason for hiding this comment

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

This bit feels a little confusing to me, TBH.

We're assiging a scalar to temporal_padding and then a 3-member tuple to padding. I would have expected temporal_padding to be a 3-member tuple, unless I am missing something obvious.

Perhaps, you could help me understand this?

Copy link
Contributor Author

@c8ef c8ef Dec 7, 2025

Choose a reason for hiding this comment

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

Although the standard Conv module includes a padding attribute, it does not support causal padding in the temporal dimension. Previously, we manually removed all internal padding and relied explicitly on F.pad. In this implementation, we apply manual padding only to the temporal dimension when necessary, while retaining the module's native padding for spatial (H/W) dimensions. This could be faster because the internal padding may use better optimizations.


def forward(self, x, cache_x=None):
padding = list(self._padding)
if cache_x is not None and self._padding[4] > 0:
b, c, _, h, w = x.size()
padding = self.temporal_padding
if cache_x is not None and self.temporal_padding > 0:
cache_x = cache_x.to(x.device)
x = torch.cat([cache_x, x], dim=2)
padding[4] -= cache_x.shape[2]
x = F.pad(x, padding)
padding -= cache_x.shape[2]
# Manually pad time dimension
if padding > 0:
x = torch.cat([x.new_zeros(b, c, padding, h, w), x], dim=2)
return super().forward(x)


Expand Down