Xilinx/brevitas

What it solves

Brevitas is designed to reduce the memory and computational requirements of neural networks through quantization. It provides tools to convert high-precision floating-point models into lower-precision formats, making them more efficient for deployment on hardware.

How it works

It is a PyTorch library that implements quantized versions of common deep learning layers (such as QuantConv2d, QuantLSTM, and QuantMultiheadAttention). It supports two primary methods of quantization:

• Post-Training Quantization (PTQ): Quantizing a model after it has already been trained.
• Quantization-Aware Training (QAT): Incorporating quantization effects during the training process to maintain accuracy.

Users can individually tune quantization settings for different tensors, including inputs, weights, bias, and outputs, adjusting parameters like bit-width and scale granularity.

Who it’s for

This library is intended for researchers and developers working on neural network optimization and deployment, specifically those using PyTorch.

Highlights

• Support for both PTQ and QAT workflows.
• Quantized implementations of standard PyTorch layers (Conv1d/2d, RNN, LSTM, MultiheadAttention).
• Granular control over quantization settings for weights and activations.
• Compatible with Windows, Linux, and macOS.