CrossAttention IN Generator

[Roserland]

I've tested the Generator in your model with vanilla attention module, but it brings OOM errors when the BEV_Feature is in big size.

I noticed that you may have used some SparseAttention transformer variants (mentioned in your paper), so whether these Sparse Attention modules like LSH or others, are used as CrossAttention in your decoder layers? From my view, using LSH attention as cross-attention is unusual.

Thank you for your excellent work!