Skip to content

fix(conv1d): prevent gradient error by removing in-place ops #82

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
ricardocardn wants to merge 2 commits into
base: main
Choose a base branch
from
Open

fix(conv1d): prevent gradient error by removing in-place ops #82

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c98181e
fix(conv1d): avoid in-place ops in conv1d_step to preserve gradient c…
Mar 31, 2025
d732255
fix(conv1d): avoid in-place ops in conv1d_step to preserve gradient c…
Mar 31, 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
26 changes: 9 additions & 17 deletions xlstm/components/conv.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,27 +29,19 @@ def conv1d_step(
) -> tuple[torch.Tensor, torch.Tensor]:
"""
B: batch size
S: sequence length
S: sequence length (must equal 1 in step mode)
D: feature dimension
KS: kernel size
Args:
x (torch.Tensor): (B, S, D)
conv_state (torch.Tensor): (B, KS, D)
conv1d_weight (torch.Tensor): (KS, D)
"""
assert (
x.shape[0] == conv_state.shape[0]
), f"x has batch size {x.shape[0]} but conv_state has batch size {conv_state.shape[0]}"
assert (
x.shape[2] == conv_state.shape[2]
), f"x has feature dimension {x.shape[2]} but conv_state has feature dimension {conv_state.shape[2]}"
assert x.shape[1] == 1, f"x has sequence length {x.shape[1]} but it should be 1"
conv_state.copy_(torch.roll(conv_state, shifts=-1, dims=1))
conv_state[:, -1:, :] = x
y = torch.sum(conv_state * conv1d_weight, dim=1, keepdim=True)
assert x.shape[1] == 1, "x must have a sequence length equal to 1"
new_conv_state = torch.roll(conv_state, shifts=-1, dims=1).clone()
new_conv_state[:, -1:, :] = x

y = torch.sum(new_conv_state * conv1d_weight, dim=1, keepdim=True)
if conv1d_bias is not None:
y += conv1d_bias
return y, conv_state
y = y + conv1d_bias

return y, new_conv_state


class CausalConv1d(nn.Module):
Expand Down