Install.md

Pre-requisites

The main prerequisites are:

• CUDA Toolkit 12.4 (versions greater than 12.4 might also work)
• build-essential: This is needed for torch-memory-saver
• uv: We use the uv + ray integration to easily manage dependencies in multi-node training.
• python 3.12
• ray 2.43.0

Once installed, configure ray to use uv with

export RAY_RUNTIME_ENV_HOOK=ray._private.runtime_env.uv_runtime_env_hook.hook

Installation & Verification

Ensure you are in the project root directory where pyproject.toml and uv.lock are located:

cd MARS-SQL/Mars-train

We rely on uv to manage the environment and lock dependencies strictly. You can verify your setup by performing a "dry run" which initializes Ray within a uv-managed environment.

1. Base Installation Dry Run

Run the following command to verify that the base environment (Ray, Torch, etc.) can be installed and initialized correctly:

uv run --isolated --frozen python -c 'import ray; ray.init(); print("Success!")'

2. SQL Support Dry Run

If you require SQL functionalities (e.g., text-to-sql tasks), we have defined an optional dependency group named sql. Run the following command to verify the environment with SQL extras:

uv run --isolated --extra sql --frozen python -c 'import ray; ray.init(); print("Success!")'