higgsfield-ai/higgsfield

Higgsfield is an open-source framework for training very large AI models, specifically the kind that have billions or even trillions of parameters, across multiple computers at once. Models this size, often called Large Language Models, are too large to fit on a single machine, so Higgsfield handles the coordination work of splitting training across many GPU-equipped servers simultaneously. The tool acts as both a GPU workload manager and a training framework. It handles giving users access to compute nodes, maintains a queue so multiple experiments do not interfere with each other, and uses techniques from PyTorch for distributing model weights across machines. Teams can train massive models without writing all the coordination code from scratch. One of the core problems it addresses is environment setup. Rather than debugging mismatched library versions across different servers, Higgsfield installs everything consistently using Docker on each node. Configuration is also simplified: instead of hundreds of command-line arguments or complex YAML files, you define an experiment as a short Python function and the tool generates the necessary deployment workflows automatically. The GitHub integration is central to how it operates. Once a project is set up, pushing code to GitHub triggers automatic deployment to your configured nodes. You then monitor and launch experiments through the GitHub Actions interface, and checkpoints are saved as training runs proceed. The code example in the README shows how training a 70-billion-parameter model can be expressed in a few dozen lines of Python. The project requires Ubuntu servers with SSH access and has been tested on Azure, LambdaLabs, and FluidStack. The README includes a setup guide covering node initialization and environment configuration, along with a tutorial on data loading, optimization, model saving, and monitoring.