The Torch-UCC plugin is a research prototype that enables collective communication over XCCL for distributed PyTorch applications that load the plugin at application runtime. The XCCL interface is a non-standard API with a corresponding reference implementation used to guide design decisions for the UCC project.
The torch-ucc plugin is licensed as:
In order to contribute to torch-ucc, please sign up with an appropriate Contributor Agreement. Follow these instructions when submitting contributions and changes.
Required packages:
- PyTorch
- UCX
- XCCL
UCX_HOME=<PATH_TO_UCX> UCC_HOME=<PATH_TO_XCCL> python setup.py clean --all install
export LD_LIBRARY_PATH=<PATH_TO_UCX>/lib:<PATH_TO_XCCL>/lib:$LD_LIBRARY_PATH
python example.py
import torch
import torch.distributed as dist
import torch_ucc
....
dist.init_process_group('ucc', rank=comm_rank, world_size=comm_size)
....
dist.all_to_all_single(recv_tensor, send_tensor)