Scalify: scale propagation for efficient low-precision LLM training

Paul Balanca, Samuel Hosegood, Carlo Luschi, Andrew W Fitzgibbon

Keywords: llm, fp8, training, low precision

TL;DR: Automated end-to-end scale propagation in neural-net computational graphs for efficient and simplified low-precision LLM training.

Abstract:

Low-precision formats such as float8 have been introduced in machine learning accelerated hardware to improve computational efficiency for large language models training and inference. Nevertheless, adoption by the ML community has been slowed down by the complex, and sometimes brittle, techniques required to match higher precision training accuracy. In this work, we present Scalify, a end-to-end scale propagation paradigm for computational graphs, generalizing and formalizing existing tensor scaling methods. Experiment results show that Scalify supports out-of-the-box float8 matrix multiplication and gradients representation, as well as float16 optimizer state storage. Our JAX
implementation of Scalify is open-sourced at [github.com/graphcore-research/jax-scalify](https://github.com/graphcore-research/jax-scalify).

Submission Number: 24