Welcome to the MPTorch documentation!

MPTorch is a PyTorch-based framework that is designed to simulate the use of custom/mixed precision arithmetic in PyTorch, in particular for DNN training workflows. It offers quantization support for various number formats (fixed-point, floating-point and block floating-point based representations) and reimplements the underlying computations of commonly used DNN layers (in CNN and Transformer-based models) using user-specified formats for each operation (e.g. addition, multiplication). All operations are internally done using IEEE-754 32-bit floating-point arithmetic, with the results rounded to the specified format.

Note

This project is under active development.

Contents:

• Usage
  • Installation
  • Basic usage example
• Tutorial
• Examples
• API Reference
  • Number formats
  • Quantization
  • Quantization Formats
  • Modules
  • Functional
  • cuBLAS Acceleration

Indices and tables

• Index

• Module Index

• Search Page