Metadata-Version: 2.3
Name: tractoray
Version: 0.0.10
Summary: Run ray on tracto.ai
Author: TractoAI team
Requires-Python: >=3.10,<4.0
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Requires-Dist: ytsaurus-client (>=0.13.24)
Project-URL: Homepage, https://tracto.ai
Description-Content-Type: text/markdown

# Tractoray

The tool for running [Ray](https://www.ray.io/) clusters on [Tracto.ai](https://tracto.ai/). Allows you to easily deploy and manage Ray clusters within Tracto.ai infrastructure.

## Features

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

To see all the configuration options for Ray, please check the output of the `tractoratay start --help` command.

## Installation

Install `tractoray` with Ray CLI and all required dependencies:
```bash
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:
```bash
tractoray start --workdir //your/cypress/path --node-count 2
```

An output of the command contains an instruction to connect to the cluster.

Check the cluster status:
```bash
tractoray status --workdir //your/yt/path
```

Also JSON format is supported:
```bash
tractoray status --workdir //your/cypress/path --format json
```

Stop the cluster:
```bash
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):
   ```dockerfile
   FROM cr.eu-north1.nebius.cloud/e00faee7vas5hpsh3s/tractoray/default:2025-06-12-16-59-42-9a2ce5611
   
   # 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 install all necessary dependencies for CUDA and infiniband.
   ```dockerfile
   FROM python:3.12
   
   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.

## Connect to the Ray Head Node terminal

It's possible to connect to the Ray head node terminal using `yt run-job-shell` command or using web terminal in the UI. You can find the instructions and commands for connecting in the output of `tractoratay status` command.

## Ray Debugging

Currently, only the [legacy Ray debugger](https://docs.ray.io/en/latest/ray-observability/user-guides/debug-apps/ray-debugging.html) is supported.

How to use it on Tracto.ai:
1. Start your Ray cluster with `RAY_DEBUG=legacy` environment variable and `--ray-debugger-external` option: `tractoray start --workdir //your/cypress/path --ray-head-params="--ray-debugger-external" --ray-worker-params="--ray-debugger-external" --env-var "RAY_DEBUG=legacy"`
2. Connect to the Ray head node terminal using `yt run-job-shell` or web UI (instructions and necessary links can be found in the output of `ray start` and `ray status`).
3. Run `ray debug` in the terminal on the head node to start the Ray Debugger.

You can read more about debugging Ray tasks in the official [Ray documentation](https://docs.ray.io/en/latest/ray-observability/user-guides/debug-apps/ray-debugging.html).

## 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 or run Ray CLI on the head node.
- Ray Serve is not supported.
- Observability features are disabled by default. You can enable and configure them in your custom image.
- Only legacy Ray Debugger is supported.

