Skip to content

README.md

Latest commit

 

History

History
44 lines (30 loc) · 1.76 KB

README.md

File metadata and controls

44 lines (30 loc) · 1.76 KB

🚧 This project is currently under construction and untested. 🚧

Tinker API on HPC

An orchestration toolkit for running SkyRL on HPCs to get a Tinker API

Disclaimer: This project is not affiliated with Tinker by Thinking Machines Lab, it merely supports using the public Tinker API.

Prerequisites

  • Active HPC account (and project workspace, depending on the HPC).
  • Local Unix-like environment (Linux, macOS, or WSL) with bash and SSH configured.
  • Python environment with the tinker SDK installed.

Setup Instructions

Configure the Environment

cp config-template.env config.env and update HPC_USER and HPC_PROJECT with your specific credentials.

Run the Initialization Script

Execute the local setup script from your terminal: ./local_setup.sh

This script will:

  • Create the necessary directory structure on HPC.
  • Transfer the required Slurm and bash scripts.
  • Submit a background job to the cpuonly partition to build the Apptainer image and cache Python dependencies.

Usage Instructions

Start the Server and Tunnel

Run the orchestration script: ./local_run.sh

This script submits the GPU job via Slurm, polls the queue for the allocated compute node, establishes a local port forward (default: 8000), and tails the remote logs.

Wait until the terminal outputs "Uvicorn running on http://0.0.0.0:8000".

Execute Client Code

In a separate terminal, run your ML workloads or the provided example: python client_example.py

Teardown

When you are finished, return to the terminal running ./local_run.sh and press Ctrl+C. The script will automatically trigger a cleanup function to cancel the Slurm job and terminate the SSH tunnel, freeing the HPC resources.