I've changed your training code to get optimal rotation matrix only by setting groupsize for weight and activations to 32. However, I consistently encounter CUDA out of memory error during forward pass. I personally think that this is due to large activations memory required during forward pass. I 've also included your FSDP config file. Could you give me an explanation for this?
I've changed your training code to get optimal rotation matrix only by setting groupsize for weight and activations to 32. However, I consistently encounter CUDA out of memory error during forward pass. I personally think that this is due to large activations memory required during forward pass. I 've also included your FSDP config file. Could you give me an explanation for this?