chunked/nltk_chunking/_deepspeed/chunk_93.txt

Depending on the CPU and/or NVMe memory available, you can offload both the optimizer states and parameters, just one of them, or none.