tensorplay.distributed.distributed_c10d.all_to_all

tensorplay.distributed.distributed_c10d.all_to_all(output_tensor_list: List[TensorBase], input_tensor_list: List[TensorBase], group=None, async_op: bool = False)[source]

Scatters a list of tensors to ranks and collects one from each.

Mirrors ProcessGroupNCCL::alltoall: per-rank splits are the tensor numels, executed as one grouped send/recv exchange.