theluyuan/DiffSynth-Studio
1
0
Fork 0
mirror of https://github.com/modelscope/DiffSynth-Studio.git synced 2026-03-18 22:08:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

bugfix

Browse Source
This commit is contained in:
Artiprocher
2025-03-10 18:25:23 +08:00
parent e757013a14
commit 718b45f2af
2 changed files with 10 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
diffsynth/models/wan_video_dit.py
View File

@@ -24,8 +24,14 @@ except ModuleNotFoundError:
SAGE_ATTN_AVAILABLE = False
def flash_attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, num_heads: int):
if FLASH_ATTN_3_AVAILABLE:
def flash_attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, num_heads: int, compatibility_mode=False):
if compatibility_mode:
q = rearrange(q, "b s (n d) -> b n s d", n=num_heads)
k = rearrange(k, "b s (n d) -> b n s d", n=num_heads)
v = rearrange(v, "b s (n d) -> b n s d", n=num_heads)
x = F.scaled_dot_product_attention(q, k, v)
x = rearrange(x, "b n s d -> b s (n d)", n=num_heads)
elif FLASH_ATTN_3_AVAILABLE:
q = rearrange(q, "b s (n d) -> b s n d", n=num_heads)
k = rearrange(k, "b s (n d) -> b s n d", n=num_heads)
v = rearrange(v, "b s (n d) -> b s n d", n=num_heads)

4
diffsynth/models/wan_video_image_encoder.py
View File

@@ -260,7 +260,7 @@ class SelfAttention(nn.Module):
q, k, v = self.to_qkv(x).chunk(3, dim=-1)
# compute attention
x = flash_attention(q, k, v, num_heads=self.num_heads)
x = flash_attention(q, k, v, num_heads=self.num_heads, compatibility_mode=True)
# output
x = self.proj(x)
@@ -371,7 +371,7 @@ class AttentionPool(nn.Module):
k, v = self.to_kv(x).chunk(2, dim=-1)
# compute attention
x = flash_attention(q, k, v, num_heads=self.num_heads)
x = flash_attention(q, k, v, num_heads=self.num_heads, compatibility_mode=True)
x = x.reshape(b, 1, c)
# output