Higgsfield provides a fault-tolerant, GPU-centric orchestration and ML framework for large models, with multi-node training via PyTorch/ZeRO-3 and integration with GitHub actions. It includes setup, distributed training examples, and a quick-start workflow to deploy experiments on managed nodes.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
Higgsfield is a fault-tolerant, highly scalable GPU orchestration, and a machine learning framework designed for training models with billions to trillions of parameters, such as Large Language Models (LLMs).
How it works
It allocates compute resources (nodes) to users for training tasks and supports ZeRO-3 Deepspeed API and fully sharded data parallel API of PyTorch. It provides a framework for initiating, executing, and monitoring training on allocated nodes, and maintains a queue for running experiments to manage resource contention. It integrates with GitHub and GitHub Actions for CI/CD of ML development.
Getting started
Install: ```bash $ pip install higgsfield==0.0.3
Getting started sections mention setup guides and tutorials, including:
- Setup quick start for initializing the project, environment, git configuration, and running first experiments.
- Tutorials covering working with distributed models, preparing data, optimizing parameters, saving models, and monitoring.
## Recent releases
Latest release: v0.0.4-rc (2024-03-23) with changes including:
- fix(action-builder): default field gen by @higgsfield in https://github.com/higgsfield-ai/higgsfield/pull/33
- add: specify invoker version if needed by @higgsfield in https://gith
## Traction
Stars: 4015
Forks: 685
Open issues: 9
## Caveats
License: Apache-2.0
Created: 2018-05-26
Last push: 2024-05-25
Language: Jupyter Notebook
Topics include cluster-management, deep-learning, distributed, llama, llama2, llm, machine-learning, mlops, pytorch