PyTorch Evaluates Sharding Representation For Extensibility

On December 8, 2025, the author analyzes sharding representation trade-offs between JAX and PyTorch, arguing JAX's NamedSharding is effectively closed while PyTorch's DTensor Placement list is more extensible. The piece details why mesh-dim, imperative placements enable custom, invertible sharding transformations and gives practical examples (uneven sharding, deferred reductions, view operations). It recommends targeted expressivity and limited use of local_map to preserve DTensor correctness.