Skip to content
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

FlashAttention needs contiguous gradients #548

Merged
merged 2 commits into from
Nov 29, 2022
Merged

FlashAttention needs contiguous gradients #548

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5922cdd
FlashAttention needs contiguous gradients
fmassa Nov 28, 2022
7ee1305
Fix lint
fmassa Nov 29, 2022
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
Next Next commit
FlashAttention needs contiguous gradients 
This was not visible before because the test was creating tensors of the wrong dtype, which when casted to the right dtype would yield contiguous tensors
  • Loading branch information
@fmassa
fmassa committed Nov 28, 2022
commit 5922cdd2a101461cbf86219d375f04e9a464246a
2 changes: 1 addition & 1 deletion tests/test_mem_eff_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -605,7 +605,7 @@ def test_backward(

grad_out = torch.ones_like(out)
if grad_out_contiguous is False:
grad_out = torch.tensor([1.0], device=device)[None, None, :].expand_as(out)
grad_out = torch.tensor([1.0], dtype=query.dtype, device=device)[None, None, :].expand_as(out)

out.backward(grad_out)
del out
Expand Down
2 changes: 1 addition & 1 deletion xformers/ops/memory_efficient_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -624,7 +624,7 @@ def _backward(cls, ctx, grad, saved_tensors):

assert grad.dtype in cls.SUPPORTED_DTYPES
cls._flash_attn_backward(
grad.reshape(ctx.kernel_output_shape),
grad.reshape(ctx.kernel_output_shape).contiguous(),
q,
k,
v,
Expand Down