Foundation Model Stack

Foundation Model Stack (fms) is a collection of components developed out of IBM Research used for development, inference, training, and tuning of foundation models leveraging PyTorch native components.

Optimizations

In FMS, we aim to bring the latest optimizations for pre-training/inference/fine-tuning to all of our models. A few of these optimizations include, but are not limited to:

• fully compilable models with no graph breaks
• full tensor-parallel support for all applicable modules developed in fms
• training scripts leveraging FSDP
• state of the art light-weight speculators for improving inference performance

Usage

FMS is currently being deployed in Text Generation Inference Server

Repositories

• foundation-model-stack: Main repository for which all fms models are based
• fms-extras: New features staged to be integrated with foundation-model-stack
• fms-fsdp: Pre-Training Examples using FSDP wrapped foundation models
• fms-hf-tuning: Basic Tuning scripts for fms models leveraging SFTTrainer