Parallel Multi-Task Training

distributed
#1

I’m trying to parallelize multi-task training in PyTorch Lightning(have different threads compute losses for each task). More precisely:

  • Parallelizing the forward function of components that can be executed in parallel
  • Parallelizing the computation of multiple losses

I want to know if PyTorch currently supports it or has anyone tried/implemented this before?

#2

I’m not very familiar with lightning but in the PyTorch distributed package we don’t currently have a framework that is exactly for this paradigm. Distributed Data Parallel will parallelize the forward function and each loss is computed locally on each node, but this is for a replicated model and not a multi-task model as I believe you are asking for. A way to accomplish this on your own would be to use Distributed RPC Framework — PyTorch master documentation and handle the parallelization and tensor communication on your own via remote procedure calls