Why not perform layer fusion in the downstream models?

[hyz-xmaster]

I observed that you don't utilize this function to do structural re-parameterization for the RepViT-backbone in downstream tasks, such as detection and RepViT-SAM, even though it could considerably improve the inference speed of these models.

I would appreciate it if you could share your thoughts on this choice.