amazon-science/fmcore

amazon-science/fmcore

Description: Running Foundation Models at every scale, on every modality. Includes reinforcement tuning, supervised fine-tuning, distillation, inference, upcoming research.

Language: Python

License: Apache-2.0

Stars: 7

Forks: 2

Open issues: 8

Created: 2025-01-08T05:09:24Z

Pushed: 2026-01-26T07:51:31Z

Default branch: main

Fork: no

Archived: no

README: 

fmcore is a specialized toolkit that empowers AI scientists to break new ground by simplifying large-scale experimentation with massive Foundation Models and datasets.

A primary bottleneck in Foundation Model research is implementation overhead. With fmcore, scientists can rapidly prototype new innovations in hours instead of weeks, accelerating the path to new research breakthroughs or user experiences.

Key features:

• Easy scaling of model training and inference (see examples).
• Standardized interfaces for parameter tuning and evaluation.
• Built-in support for distributed computing and Foundation Model parallelism.

Installation

The minimal fmcore package can be installed from PyPI:

pip install fmcore

To get all features, we recommend installing in a new Conda environment:

conda create -n fmcore python=3.11 --yes
conda activate fmcore
pip install uv
uv pip install "fmcore[all]"

License

This project is licensed under the Apache-2.0 License.

Contributing to fmcore

See [CONTRIBUTING](CONTRIBUTING.md#security-issue-notifications) for more information.