NOTE: PyTorch now includes the distributed RPC framework (https://pytorch.org/docs/stable/rpc.html). Most tasks can now be achieved using RPC. If you need a simpler MPI-like alternative, diffdist
is for you!
diffdist
is a python library for pytorch. It extends the default functionality of torch.autograd
and adds support for differentiable communication between processes. This enables backpropagation to work in distributed settings and makes it super easy to use distributed model parallelism! diffdist
achieves that by simply implementing the backward passes for most common communication primitives of torch.distributed
. Processes that communicate during the forward pass, will automatically communicate during the backward pass to exchange gradients.
After installing pytorch
install simply using:
$ pip install diffdist
You can find examples on how to use diffdist
in diffdist/testing.py
GNU GPLv3