Skip to content

Accelerate Cross Attention Computing #2

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
zhangtao0408 wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Accelerate Cross Attention Computing #2

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
e6ec8a5
Refactor rope_apply function to simplify tensor processing and improv…
TaoZhang-Work Nov 1, 2025
2ffd232
Refactor rope_apply function to simplify tensor processing and improv…
TaoZhang-Work Nov 1, 2025
1a976fa
Update attention.py to correct layout parameter from "BNSD" to "BSND"…
TaoZhang-Work Nov 1, 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
11 changes: 2 additions & 9 deletions wan/distributed/xdit_context_parallel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,21 +27,14 @@ def pad_freqs(original_tensor, target_len):
return padded_tensor


@amp.autocast(enabled=False)
def rope_apply(x, grid_sizes, freqs_list):
"""
x: [B, L, N, C].
grid_sizes: [B, 3].
freqs: [M, C // 2].
"""
s, n, c = x.size(1), x.size(2), x.size(3)
output = []
for i, (f, h, w) in enumerate(grid_sizes.tolist()):
x_i = x[i, :s].reshape(1, s, n, c)
cos, sin = freqs_list[i]
x_i = rotary_position_embedding(x_i, cos, sin, rotated_mode="rotated_interleaved", fused=True)
output.append(x_i)
return torch.cat(output).float()
cos, sin = freqs_list[0]
return rotary_position_embedding(x, cos, sin, rotated_mode="rotated_interleaved", fused=True)


def usp_dit_forward(
Expand Down
2 changes: 1 addition & 1 deletion wan/modules/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -161,7 +161,7 @@ def attention(
opt_mode="manual", op_type="ascend_laser_attention", layout="BNSD")
else:
out = attention_forward(q, k, v,
opt_mode="manual", op_type="fused_attn_score", layout="BNSD")
opt_mode="manual", op_type="fused_attn_score", layout="BSND")
return out.to(qtype)
elif FLASH_ATTN_2_AVAILABLE or FLASH_ATTN_3_AVAILABLE:
return flash_attention(
Expand Down
11 changes: 2 additions & 9 deletions wan/modules/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,21 +39,14 @@ def rope_params(max_seq_len, dim, theta=10000):
return freqs


@amp.autocast(enabled=False)
def rope_apply(x, grid_sizes, freqs_list):
"""
x: [B, L, N, C].
grid_sizes: [B, 3].
freqs: [M, C // 2].
"""
s, n, c = x.size(1), x.size(2), x.size(3)
output = []
for i, (f, h, w) in enumerate(grid_sizes.tolist()):
x_i = x[i, :s].reshape(1, s, n, c)
cos, sin = freqs_list[i]
x_i = rotary_position_embedding(x_i, cos, sin, rotated_mode="rotated_interleaved", fused=True)
output.append(x_i)
return torch.cat(output).float()
cos, sin = freqs_list[0]
return rotary_position_embedding(x, cos, sin, rotated_mode="rotated_interleaved", fused=True)


class WanRMSNorm(nn.Module):
Expand Down