For any proposed (op, sample inputs, sharding prop rule: i.e. input -> output placements), you can rule out if this is a valid sharding prop rule with:

1. Run the op on full_tensor inputs, to get the reference full_tensor output
2. Get the local_tensor inputs, according to the input placements
3. Run the local op
4. Redistribute local outputs to Replicate, according to the output placements
5. Check full_tensor shapes & numerics against reference

This runs this validation against OpInfo entries for aten ops with registered single-dim rules, enumerating all the strategies and replacing ShardPlaceholder -> Shard.

A future extension could be: by exhaustively enumerating all single-dim strategies on R/S/P, we could check if any prop rules are potentially missing. However this can see false positives (e.g. Partial with zero tensors), so likely that shouldn't be a hard error, and maybe shouldn't be a "test".

checked by turning on single-dim rules in

• mm:

  pytorch/torch/distributed/tensor/_ops/_matrix_ops.py

  Line 353 in 5787448

  # @register_single_dim_strategy(aten.mm.default)

• pointwise ops:

  pytorch/torch/distributed/tensor/_ops/_pointwise_ops.py

  Lines 837 to 839 in 5787448

  	# register_single_dim_strategy(
  	# op, schema_info=RuntimeSchemaInfo(static_kwargkey=["out"])
  	# )(single_mesh_dim_pointwise_strategy)

• cat:

  pytorch/torch/distributed/tensor/_ops/_tensor_ops.py

  Line 828 in 5787448

  # @register_single_dim_strategy(aten.cat.default, RuntimeSchemaInfo(1, needs_pytree=True))