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

Please see this detailed guide for reasons to use 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.

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")

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:

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.

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.

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.

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.

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.

Please see the contribution guide to contribute to NeMo Run.

Please find a list of frequently asked questions here.