Also, tip on weights sharding: as a generic rule of thumb you can use full replication for all variable dimensions except the last one, which should be sharded across the "model" axis

→ View original post on X — @fchollet,