NeMo-Run

[!IMPORTANT] NeMo-Run is still in active development and this is a pre-release. The API is subject to change without notice while in pre-release. First official release will be 0.1.0 and will be included in NeMo FW 24.09 as well.

NeMo-Run is a powerful tool designed to streamline the configuration, execution, and management of machine learning experiments across various computing environments. NeMo-Run has three core responsibilities:

1. Configuration
2. Execution
3. Management

To learn more, click on each link. This represents the typical order that Nemo-Run users follow for setting up and launching experiments.

• NeMo-Run
  • Why Use Nemo-Run?
  • Install NeMo-Run
  • Get Started
  • Design Philosophy and Inspiration
  • Pythonic
  • Modular
  • Opinionated but Flexible
  • Set Up Once and Scale Easily
  • Tutorials
    • Hello world
  • Contribute to NeMo-Run
  • FAQs

Why Use Nemo-Run?

Please see this detailed guide for reasons to use Nemo-Run.

Install NeMo-Run

To install the project, use the following command:

pip install git+https://github.com/NVIDIA/NeMo-Run.git

Make sure you have pip installed and configured properly.

Get Started

To get started with Nemo-Run, follow these three steps based on the core responsibilities mentioned above. For this example, we’ll showcase a pre-training example in Nemo 2.0 using Llama3.

1. Configure your function:

   from nemo.collections import llm
   partial_func = llm.llama3_8b.pretrain_recipe(name="llama3-8b", ckpt_dir="/path/to/store/checkpoints", num_nodes=1, num_gpus_per_node=8)

2. Define your Executor:

   import nemo_run as run
   # Local executor
   local_executor = run.LocalExecutor()

3. Run your experiment:

   run.run(partial_func, executor=local_executor, name="llama3_8b_pretraining")

Design Philosophy and Inspiration

In building NeMo-Run, we drew inspiration from and relied on the following primary libraries. We would like to extend our gratitude for their work.

• Fiddle
• TorchX
• Skypilot
• XManager
• Fabric and Paramiko
• Rich
• Jinja

Apart from these, we also build on other libraries. A full list of dependencies can be found in pyproject.toml.

NeMo-Run was designed keeping the following principles in mind:

Pythonic

In NeMo-Run, you can build and configure everything using Python, eliminating the need for multiple combinations of tools to manage your experiments. The only exception is when setting up the environment for remote execution, where we rely on Docker.

Modular

The decoupling of task and executor allows you to form different combinations of execution units with relative ease. You configure different remote environments once, and you can reuse it across a variety of tasks in a Pythonic way.

Opinionated but Flexible

NeMo-Run is opinionated in some places, like storing of metadata information for experiments in a particular manner. However, it remains flexible enough to accommodate most user experiments.

Set Up Once and Scale Easily

While it may take some time initially for users to become familiar with NeMo-Run concepts, the tool is designed to scale experimentation in a fluid and easy manner.

Tutorials

Hello world

The hello_world tutorial series provides a comprehensive introduction to NeMo-Run, demonstrating its capabilities through a simple example. The tutorial covers:

• Configuring Python functions using Partial and Config classes.
• Executing configured functions locally and on remote clusters.
• Visualizing configurations with graphviz.
• Creating and managing experiments using run.Experiment.

You can find the tutorial series below:

• Part 1.
• Part 2.
• Part 3.

Contribute to NeMo-Run

Please see the contribution guide to contribute to NeMo Run.

FAQs

Please find a list of frequently asked questions here.