vd-kuznetsov / qtools

A convenient module for working with quantization of models

About

A convenient module for working with quantization of models.

The project is based on PyTorch's FX/Eager technology.

The report on the third homework assignment can be found at triton_info.md.

Install

git clone https://github.com/vd-kuznetsov/qtools.git

cd qtools
conda create --name qtools python==3.8
conda activate qtools

poetry install
pre-commit install
pre-commit run -a

Usage

Recommended actions:

1. Launch the MLflow server, check if your address matches the one specified in the config

2. python commands.py

The second point can be run separately:

• python commands.py mode=train - model training

  • Downloads a dataset using DVC from GDrive

• python commands.py mode=infer - quantizes the trained model with a single line of code

  • By default, the weights of the model are used after training