Is there any approach in pytorch to achieve distributed network connection between CPU prediction and GPU training?

distributed
1

Out team is planning to use CPUs from multiple computers to do network prediction and data production, and then use a single GPU server to do network training. Is there any method we can use in torch.distributed package that can help us with this situation?

2

torch.distributed.rpc should be able to help. Here is a list of tutorials.

The use case looks similar to the following two examples:

  1. https://pytorch.org/tutorials/intermediate/rpc_tutorial.html#distributed-reinforcement-learning-using-rpc-and-rref
  2. https://pytorch.org/tutorials/intermediate/rpc_async_execution.html#batch-processing-cartpole-solver
3

Thank you very much, I’ll look into them carefully!