Imitation Learning with Horovod and Open MPI HPC Libraries

post by william_schram on Feb 19, 2019

I’m attempting to run the imitation learning library Atari-Reset from openai on Github.

This repo requires installation of Horovod on GPU and Cuda-aware Open MPI. Would it be possible for you to write a installation guide for this stack? It gets confusing when Open MPI asks for the locations of the ucx-cuda and cuda locations (See ‘Building CUDA-aware Open MPI’).

I have been able to install the libraries with CPU support, but I want to get the GPU compute power of my Lambda Quad.

post by sabalaba on Feb 20, 2019

Hi William, I’ll let the team that develops Lambda Stack know that you’d like CUDA aware Open MPI / Horovod integrated into Lambda Stack.

post by william_schram on Feb 20, 2019

That would be an even better outcome! These tools are particularly valuable to your customers who are looking at clusters of Lambda machines for reinforcement / imitation learning tasks. I look forward to future updates to the Lambda Stack.

post by geppy on Mar 17, 2019

Hi,

Do you have an ETA for this implementation? Thank you.

Geppy

post by william_schram on Apr 11, 2019

Ditto, looking to see a formal timeline communicated.

post by william_schram on Oct 15, 2019

Has there been any discussion on the timeline to implement this request to Lambda stack? Or reach out to the open source developers at Uber?