Welcome to MQBench’s documentation!

What is MQBench?

MQBench is short for Model Quantization Benchmark, an open-source Python package for evalutaing quantization algorithoms.

MQBench tries to solve two long-neglected problems in the quantization community, namely reproducibility and deployability. For reproducibility, we setup a unified set of training hyper-parameters and run all algorithms under this setting. For deployability, we ensure every algorithms can be deployed, this includes the folding of Batch Normalization layers and corresponding hardware quantizer design.

Key Features in MQBench:

• Implementation with the lastest Pytorch-1.8 tracing techniques, easy to extend.

• Support for 6 quantization-aware training algorithms and is easily extendable.

• Support 5 hardware platforms.

Contents:

• Quantization Algorithm
  • Post-training Quantization v.s. Quantization-aware Training
  • QAT Algorithms
  • PTQ Algorithms
• Quantization Hardware
  • Overview
  • Academic Setup
  • TensorRT Setup
  • SNPE Setup
  • TVM Setup
  • ACL Setup
  • FBGEMM Setup
• Benchmark
  • Running experiments in MQBench
  • How to self-implement a quantization algorithm
  • How to self-implement a hardware configuration