transformers
e1d33d25 - Take TP and EP meshes in apply_tensor_parallelism

Commit
9 days ago
Take TP and EP meshes in apply_tensor_parallelism Drop the separate shard_mesh argument. Parameters shard on the EP mesh when one is given, otherwise on the TP mesh, and the dispatch style still receives both meshes explicitly. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Author
Parents
Loading