GPU requirements on Docker

[leandermaben]

Hello,

I noticed in the setup that the docker containers are expected to have GPU access. This is verified with nvidia-smi.

I have 2 questions:

1. Is GPU access required to be able to properly evaluate on this benchmark?
2. How should a mac user with a non-nvidia gpu go about the setup?

Thanks in advance!

[deepakn97]

Hi @leandermaben

Apologies for the late reply. Thank you for opening this issue.

TL;DR: GPU access is not required for some of the tasks in the benchmark.

To answer your questions:

1. GPU access is not required to evaluate on some of the tasks of this benchmark.
2. Mac users can skip installing Nvidia drivers and the GPU related arguments as no dedicated GPUs are available.

Caveat:
Although GPU access is not required, the baseline scripts for some compute-intensive tasks, such as Language Modeling, RL tasks use multi-gpu setups to enable fast training and evaluation runs. Libraries like DeepSpeed and Jax on Metal are not thoroughly tested, so it may cause issues.

Next Steps:
For now, I recommend skipping the NVIDIA-specific installation steps and trying to run some tasks that do not require GPUs. Some examples of these tasks are: BattleOfSexes, PrisonersDilemma, Blotto, HousePricePrediction, 3SATTime. This is to check that the basic setup works for you.

We will eventually integrate the relevant changes to enable the running of all tasks on Macs and machines with AMD GPUs.

[leandermaben]

Hey @deepakn97, thanks for the detailed response. This clears things up. I was wondering if you have a slack or discord forum for discussion. I am currently looking into integrating MLGym benchmarks into OpenHands (https://github.com/All-Hands-AI/OpenHands/tree/main) and a discussion forum would help if I run into snags.

[deepakn97]

Hey @leandermaben, That's sound amazing!
We do have a Discord channel. The link to join it is here.

I will keep this issue open to keep a track of the required changes in MLGym. Thank you for bringing this to our attention.