Scalify: scale propagation for efficient low-precision LLM training
Fuente:
arXiv
Guardado en:
| Autores principales: | , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2024
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866914885386895360 |
|---|---|
| author | Balança, Paul Hosegood, Sam Luschi, Carlo Fitzgibbon, Andrew |
| author_facet | Balança, Paul Hosegood, Sam Luschi, Carlo Fitzgibbon, Andrew |
| contents | 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 https://github.com/graphcore-research/jax-scalify |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2407_17353 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Scalify: scale propagation for efficient low-precision LLM training Balança, Paul Hosegood, Sam Luschi, Carlo Fitzgibbon, Andrew Machine Learning 68T07 I.2.7 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 https://github.com/graphcore-research/jax-scalify |
| title | Scalify: scale propagation for efficient low-precision LLM training |
| topic | Machine Learning 68T07 I.2.7 |
| url | https://arxiv.org/abs/2407.17353 |