CUDA backend is 馃敟 Are there any plans to get FSDP2 (per param arbitrary mesh sharding) in?
CUDA backend is 馃敟
Are there any plans to get FSDP2 (per param arbitrary mesh sharding) in?