wangbo/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2026-05-08 08:12:34 +08:00
Code Issues Actions 18 Packages Projects Releases Wiki Activity

Restore ViT spatial order after windowed attention.

Browse Source
This commit is contained in:
Talmaj Marinc 2026-03-13 20:37:03 +01:00 committed by Talmaj Marinc
parent b202f842af
commit e00ae62907
1 changed files with 3 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
comfy/text_encoders/qwen_vl.py
View File

@ -425,4 +425,7 @@ class Qwen2VLVisionTransformer(nn.Module):
hidden_states = block(hidden_states, position_embeddings, cu_seqlens_now, optimized_attention=optimized_attention)
hidden_states = self.merger(hidden_states)
# Potentially important for spatially precise edits. This is present in the HF implementation.
reverse_indices = torch.argsort(window_index)
hidden_states = hidden_states[reverse_indices, :]
return hidden_states