tractoray 0.0.8

Run ray on tracto.ai

Tractoray

The tool for running Ray clusters on Tracto.ai. Allows you to easily deploy and manage Ray clusters within YT infrastructure.

Features

• Launch Ray clusters with configurable resources
• Support native ray dashboard and ray client
• Flexible Docker image configuration

Installation

Install tractoray with Ray CLI and all required dependencies:

pip install -U "tractoray[ray]"

Usage

Basic Commands

To use tractoray, you need to specify the working directory, for example your homedir //home/<login>/tractoray.

Start a cluster:

tractoray start --workdir //your/cypress/path --node-count 2

return an instruction to connect to the cluster:

Check cluster status:

tractoray status --workdir //your/yt/path

also supports output in JSON format:

tractoray status --workdir //your/cypress/path --format json

Stop the cluster:

tractoray stop --workdir //your/cypress/path

For detailed information about task submission, log reading, and other operations, please check the ray status command output.

Using Custom Docker Images

You have two options for Docker images:

1. Use the default image as base (recommended):

   FROM cr.eu-north1.nebius.cloud/e00faee7vas5hpsh3s/tractoray/default:2025-04-10-21-17-47-7f93a1500
   
   # Add your dependencies
   RUN pip install your-package
   

2. Build from scratch:

   • Install tractoray via pip
   • Make sure to use the same version as in your local environment and all necessary dependencies for CUDA and infiniband

   FROM python:3.9
   
   RUN pip install tractoray==<your-local-version>
   # Add other dependencies
   

The default image includes all necessary dependencies and configurations for Ray cluster operation for machine learning tasks. Using it as a base image is recommended to ensure compatibility.

Environment Variables

• YT_LOG_LEVEL: Set logging level

Limitations

• Some Ray CLI options, such as ray status, may not function properly due to Ray authentication constraints. It is recommended to use Ray dashboard and Ray SDK instead.
• Ray Serve is not supported.
• Observability features are disabled by default. You can enable and configure them in your custom image.