wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-01-11 14:50:49 +08:00
Code Issues Actions 11 Packages Projects Releases Wiki Activity

Update model_multitalk.py

Browse Source
This commit is contained in:
kijai 2025-11-26 18:15:12 +02:00
parent af4d412e67
commit 64a984177f
1 changed files with 8 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
comfy/ldm/wan/model_multitalk.py
View File

@ -16,30 +16,30 @@ def calculate_x_ref_attn_map(visual_q, ref_k, ref_target_masks, split_num=8):
x_ref_attnmap = torch.zeros(B, H, x_seqlens, device=visual_q.device, dtype=visual_q.dtype)
chunk_size = min(max(x_seqlens // split_num, 1), x_seqlens)
for i in range(0, x_seqlens, chunk_size):
end_i = min(i + chunk_size, x_seqlens)
attn_chunk = visual_q[:, :, i:end_i] @ ref_k.permute(0, 2, 3, 1) # B, H, chunk, ref_seqlens
# Apply softmax
attn_max = attn_chunk.max(dim=-1, keepdim=True).values
attn_chunk = (attn_chunk - attn_max).exp()
attn_sum = attn_chunk.sum(dim=-1, keepdim=True)
attn_chunk = attn_chunk / (attn_sum + 1e-8)
# Apply mask and sum
masked_attn = attn_chunk * ref_target_mask
x_ref_attnmap[:, :, i:end_i] = masked_attn.sum(-1) / (ref_target_mask.sum() + 1e-8)
del attn_chunk, masked_attn
# Average across heads
x_ref_attnmap = x_ref_attnmap.mean(dim=1) # B, x_seqlens
x_ref_attn_maps.append(x_ref_attnmap)
del visual_q, ref_k
return torch.cat(x_ref_attn_maps, dim=0)
def get_attn_map_with_target(visual_q, ref_k, shape, ref_target_masks=None, split_num=2):