Skip to content

Improve performance of NPU FA #12260

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 2 commits into from
Aug 30, 2025
Merged

Improve performance of NPU FA #12260

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
89e83e6
Improve performance of NPU FA
Aug 30, 2025
6853fa3
Merge branch 'main' into npu
a-r-r-o-w Aug 30, 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
9 changes: 6 additions & 3 deletions src/diffusers/models/attention_dispatch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -955,12 +955,13 @@ def _native_npu_attention(
dropout_p: float = 0.0,
scale: Optional[float] = None,
) -> torch.Tensor:
return npu_fusion_attention(
query, key, value = (x.transpose(1, 2).contiguous() for x in (query, key, value))
out = npu_fusion_attention(
query,
key,
value,
query.size(2), # num_heads
input_layout="BSND",
query.size(1), # num_heads
input_layout="BNSD",
pse=None,
scale=1.0 / math.sqrt(query.shape[-1]) if scale is None else scale,
pre_tockens=65536,
Expand All @@ -969,6 +970,8 @@ def _native_npu_attention(
sync=False,
inner_precise=0,
)[0]
out = out.transpose(1, 2).contiguous()
return out


# Reference: https://github.com/pytorch/xla/blob/06c5533de6588f6b90aa1655d9850bcf733b90b4/torch_xla/experimental/custom_kernel.py#L853
Expand Down