Skip to content

performance(ring para): change sdpa backend to flash attention #267

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
JasonHe-WQ wants to merge 1 commit into
base: v1
Choose a base branch
from
+10 −7
Open

performance(ring para): change sdpa backend to flash attention #267

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
17 changes: 10 additions & 7 deletions diffsynth_engine/layers/attention/backends/sdpa.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@
AttentionType,
)

_scaled_dot_product_efficient_attention = torch.ops.aten._scaled_dot_product_efficient_attention
_scaled_dot_product_flash_attention = torch.ops.aten._scaled_dot_product_flash_attention


class SDPABackend(AttentionBackend):
Expand Down Expand Up @@ -88,6 +88,12 @@ def forward_with_lse(
attn_metadata: AttentionMetadata | None = None,
**kwargs,
) -> tuple[torch.Tensor, torch.Tensor]:
if attn_mask is not None:
raise NotImplementedError(
"SDPA forward_with_lse does not support attn_mask; "
"masked Ring attention is not supported"
)

query = rearrange(query, "b s n d -> b n s d")
key = rearrange(key, "b s n d -> b n s d")
value = rearrange(value, "b s n d -> b n s d")
Expand All @@ -96,18 +102,15 @@ def forward_with_lse(
key = torch.repeat_interleave(key, self.num_kv_groups, dim=1)
value = torch.repeat_interleave(value, self.num_kv_groups, dim=1)

seq_len = query.shape[2]
output, lse = _scaled_dot_product_efficient_attention(
output, lse = _scaled_dot_product_flash_attention(
query,
key,
value,
attn_bias=attn_mask,
compute_log_sumexp=True,
dropout_p=0.0,
is_causal=self.causal,
return_debug_mask=False,
scale=self.softmax_scale,
)[:2]

output = rearrange(output, "b n s d -> b s n d")
# the returned lse is padded but not restored, so we need to slice it
lse = lse[:, :, :seq_len]
return output, lse