/vast-ai-setup

Primary LanguagePython

Dev GPU setup with vast.ai

This is a collection of helper scripts and instructions to setup a GPU machine on vast.ai for deep learning development. I use this setup for my personal projects and research.

My workflow

  1. Rent a machine. rent.py script helps you find a machine with the desired GPU with my default configuration (e.g. docker image, ssh, and some sensible filtering of machines). See gpu_names.json for the list of GPUs.

    # Pick a machine with the desired GPU (e.g. RTX 3000 series)
python rent.py "RTX 30.*"

  2. Machine starts. The machine will setup according to the setup.sh script. Modify to suit your needs (e.g. personal dotfiles and git setup). This usually takes about 5 minutes.

  3. Connect to machine. rent.py will wait for the machine to be ready, then output the ssh command to connect to the machine (with ssh or vscode).

    ssh ssh://root@123.123.123.123:12312
# or
code-insiders --remote ssh-remote+root@123.123.123.123:12312 /root

  4. Clone your project. Clone your project and start developing. Make sure to enable SSH agent forwarding to clone by SSH.

  5. Install project dependencies. Install your project dependencies (e.g. pip install -r requirements.txt).

Setup

  1. Install dependencies. The only non-standard dependency is requests used by the Vast CLI.

    pip install requests

  2. Vast AI setup. You need to upload your SSH public key to Vast AI to connect to the machines. You also need to configure your Vast API key. Check out the Vast CLI documentation for more information.