chunked/nltk_chunking/_fsdp/chunk_27.txt

By sharding the model parameters, optimizer and gradient states, and even offloading them to the CPU when they're inactive, FSDP can reduce the high cost of large-scale training.