There are two ways to save and load models with FSDP. The 5th FSDP tutorial goes through a notebook with one method — full_state_dict. This is a unique model-saving process that puts together models up to 20B and optimizers on the rank 0 CPU memory. We’ll walk you through each step of the process. If you’re familiar with loading models via regular PyTorch, you’ll recognize the similarities it shares with loading models with FSDP. With the fullstate_save_policy, you can alter offload to CPU for C

L13.9.2 Saving and Loading Models in PyTorch