Need some help/clarity on installing

[philm001]

Hello all,

First I want to apologize because I know that this doesn't really belong here but rather on the forums themselves. But I have posted on the forums and still waiting to hear back. I sent messages to mods but haven't heard from them and there are certain boards that I am locked out of that I would post to.

I need to ask questions about NeMo installation because I want to learn how to use it. But I am finding it difficult getting help from the community. My only option here is to post on the actual issues tab. Again I apologize but I am not sure what else to do.

I am currently learning how to use the NeMo framework. I have things setup for a SLURM cluster. I have Nemo-curator installed and I am currently going through the tutorials. (Shout out to Ryan for his great assistance). While I am going through those tutorials. I want to finish setting up NeMo on my cluster. I hope to begin Nemo tutorials by mid-month. When I start them, I want to run the tutorials under the SLURM cluster.

Now just for some background. I am at the point of running NeMo-Curator. I am able to run it under a single machine and the tutorials seem pretty basic. Do I need the cluster, maybe not. But I would like to learn how to properly set things up because in the not so distant future I will be upgrading my hardware to do my training much better.

As of right now, I do not have NVidia GPUs but that will soon change. So using the cluster is two-fold. Mainly to speed up the process a little bit. And to use my existing resources to learn NeMo before going big. From one of the discussions here, I know it is possible to run NeMo in CPU-only mode. I wouldn't be working with large models just yet. I would start off with Bart (< 1 billion parameters) as my "learning model".

So I essentially have 2 questions about installing everything. I want to make sure that I get the workflow right the first time around.

1. Now we come to the setup. NeMo says that I need to use it with NeMo-run and I can use this script to run on the SLURM cluster. However, NeMo-Curator says that I should run using the Nvidia Nemo Framework Launcher to run under SLURM. With these two, I am wondering which one is the correct approach? Or do I have to use them separate depending on if I want to use NeMo for training (NeMo-run) or for curating data (Framework Launcher)

2. Since I am going to be running on the SLURM cluster, I am assuming that each node will need access to the library. There are 2 options here. 1. Install Docker container or 2. Install via pip. For a cluster setup, which one makes the most amount of sense? I don't need latest and greatest, I just need a simple way to get NeMo v2 up and running on the cluster.

Please let me know your feedback and I look forward to your replies.

[philm001]

Hello all,

I think that after much consideration, I am going to go with NeMo-run since this supports NeMo 2.0. The framework laucher only support NeMo 1.0 at this time. I would like to switch over to the launcher once it gets updated

I also think that at least for the NeMo-curator aspect, I will be using the Pip installer. Once I get more comfortable with using NeMo, may switch over to the docker container because I do like having everything contained into one space. I need to familiarize myself with working with docker containers and I don't want to stretch myself too thin.