🚜 Tractorun

Tractorun is a powerful tool for distributed ML operations on the Tracto.ai platform. It helps manage and run workflows across multiple nodes with minimal changes in the user's code:

• Training and fine-tuning models. Use Tractorun to train models across multiple compute nodes efficiently.
• Offline batch inference. Perform fast and scalable model inference.
• Running arbitrary GPU operations, ideal for any computational tasks that require distributed GPU resources.

How it works

Built on top of Tracto.ai, Tractorun is responsible for coordinating distributed machine learning tasks. It has out-of-the-box integrations with PyTorch and Jax, also it can be easily used for any other training or inference framework.

Key advantages:

• No need to manage your cloud infrastructure, such as configuring Kubernetes cluster, or managing GPU and Infiniband drivers. Tracto.ai solves all these infrastructure problems for you.
• No need to coordinate distributed processes. Tractorun handles it based on the training configuration: the number of nodes and GPUs used.

Key features:

• Simple distributed task setup, just specify the number of nodes and GPUs.
• Convenient ways to run and configure: CLI, YAML config, and Python SDK.
• A range of powerful capabilities, including sidecars for auxiliary tasks and transparent mounting of local files directly into distributed operations.
• Integration with the Tracto.ai platform: use datasets and checkpoints stored in the Tracto.ai storage, build pipelines with Tractorun, MapReduce, Clickhouse, Spark, and more.

Getting started

To use these examples, you'll need a Tracto account. If you don't have one yet, please sign up at tracto.ai.

Install tractorun into your python3 environment:

pip install --upgrade tractorun

Configure the client to work with your cluster:

mkdir ~/.yt
cat <<EOF > ~/.yt
"proxy"={
  "url"="$YT_PROXY";
};
"token"="$YT_TOKEN";
EOF

Please put your actual Tracto.ai cluster address to $YT_PROXY and your token to $YT_TOKEN.

How to try

Run an example script:

tractorun \
    --yt-path "//tmp/$USER/tractorun_getting_started" \
    --bind-local './examples/pytorch/lightning_mnist_ddp_script/lightning_mnist_ddp_script.py:/lightning_mnist_ddp_script.py' \
    --bind-local-lib ./tractorun \
    --docker-image ghcr.io/tractoai/tractorun-examples-runtime:2025-02-10-16-14-27 \
    python3 /lightning_mnist_ddp_script.py

How to run

CLI

tractorun --help

or with yaml config

tractorun --run-config-path config.yaml

You can find a relevant examples:

• CLI arguments example.
• YAML config example.

Python SDK

SDK is convenient to use from Jupyter notebooks for development purposes.

You can find a relevant example in the repository.

WARNING: the local environment should be equal to the remote docker image on the TractoAI platform to use SDK.

• This requirement is met in Jupyter Notebook on the Tracto.ai platform.
• For local use, it is recommended to run the code locally in the same container as specified in the docker_image parameter in tractorun

How to adapt code for tractorun

CLI

1. Wrap all training/inference code to a function.
2. Initiate environment and Toolbox by from tractorun.run.prepare_and_get_toolbox

An example of adapting the mnist training from the PyTorch repository: https://github.com/tractoai/tractorun/tree/main/examples/adoptation/mnist_simple/cli

SDK

1. Wrap all training/inference code to a function with a toolbox: tractorun.toolbox.Toolbox parameter.
2. Run this function by tractorun.run.run.

An example of adapting the mnist training from the PyTorch repository: https://github.com/tractoai/tractorun/tree/main/examples/adoptation/mnist_simple/sdk

Features

Toolbox

tractorun.toolbox.Toolbox provides extra integrations with the Tracto.ai platform:

• Preconfigured client by toolbox.yt_client
• Basic checkpoints by toolbox.checkpoint_manager
• Control over the operation description in the UI by toolbox.description_manager
• Access to coordination information by toolbox.coordinator

Toolbox page provides an overview of all available toolbox components.

Coordination

Tractorun always sets following environment variables in each process:

• MASTER_ADDR - the address of the master node
• MASTER_PORT - the port of the master node
• WORLD_SIZE - the total number of processes
• NODE_RANK - the unique id of the current node (job in terms of Tracto.ai)
• LOCAL_RANK - the unique id of the current process on the current node
• RANK - the unique id of the current process across all nodes

Backends

Backends configure tractorun to work with a specific ML framework.

Tractorun supports multiple backends:

• Tractorch for PyTorch
  • examples
• Tractorax for Jax
  • examples
• Generic
  • non-specialized backend, can be used as a basis for other backends

Backend page provides an overview of all available backends.

Options and settings

Options reference page provides an overview of all available options for tractorun, explaining their purpose and usage. Options can be defined by:

• CLI parameters
• yaml config
• python options

More information

• Examples
• More examples in Jupyter Notebooks