README.md

This repository is deprecated

Torch-UCC plugin has been merged into PyTorch as one of the native ProcessGroup implementations (please refer to pytorch/pytorch#79918), which will be the single source of Torch-UCC (i.e., ProcessGroupUCC). Please check out PyTorch repository (https://github.com/pytorch/pytorch). Some code points for reference:

• ProcessGroupUCC: https://github.com/pytorch/pytorch/blob/master/torch/csrc/distributed/c10d/ProcessGroupUCC.cpp, https://github.com/pytorch/pytorch/blob/master/torch/csrc/distributed/c10d/ProcessGroupUCC.hpp)
• UCC Utiliies: https://github.com/pytorch/pytorch/blob/master/torch/csrc/distributed/c10d/UCCUtils.cpp, https://github.com/pytorch/pytorch/blob/master/torch/csrc/distributed/c10d/UCCUtils.hpp
• UCC Comms Tracing: https://github.com/pytorch/pytorch/blob/master/torch/csrc/distributed/c10d/UCCTracing.cpp, https://github.com/pytorch/pytorch/blob/master/torch/csrc/distributed/c10d/UCCTracing.hpp

PyTorch plugin for UCC

This repo implements PyTorch Process Group API for UCC as a third-party plugin.

Requirements

• PyTorch
• UCX
• UCC

License

This repo is released under the MIT license. Please see the LICENSE file for more information.

Contributing

We actively welcome your pull requests! Please see CONTRIBUTING.md and CODE_OF_CONDUCT.md for more info.

# Build
UCX_HOME=<PATH_TO_UCX> UCC_HOME=<PATH_TO_UCC> WITH_CUDA=<PATH_TO_CUDA> python setup.py install

UCX_HOME required, specifies path to UCX installation directory

UCC_HOME required, specifies path to UCC installation directory

WITH_CUDA optional, if WITH_CUDA=no is set then only CPU tensors are supported

Run

Configuration variables

Name	Values	Description
TORCH_UCC_ALLGATHER_BLOCKING_WAIT	0 or 1	Sets behavior of wait function for CUDA Allgather. Async collective in PyTorch
TORCH_UCC_ALLREDUCE_BLOCKING_WAIT	0 or 1	Sets behavior of wait function for CUDA Allreduce.
TORCH_UCC_ALLTOALL_BLOCKING_WAIT	0 or 1	Sets behavior of wait function for CUDA Alltoall.
TORCH_UCC_BCAST_BLOCKING_WAIT	0 or 1	Sets behavior of wait function for CUDA Bcast.

export LD_LIBRARY_PATH=<PATH_TO_UCX>/lib:<PATH_TO_UCC>/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)