Why dim = 3 of cross_attention increase

[HuynhKEn]

I am a beginner, so I tried to fix this issue. I don't know if it has any problems. Please help me with this issue.

[iriscxy]

Hi, I'm not sure what your problem is. Can you show the bug report? By squeeze dim=3 we change the vector shape from [12,64,16,1,1024] to [12,64,16,1024] according to the comment.