QPyTorch

In this fork from QPyTorch, I am experimenting with updating the quantization operations implemented for CPU such that they are registered for TrochScript according to the official PyTorch documentation for extending TorchScript with custom C++ operators instead of using PYBIND11.

This fork also includes an experiment on creating a basic Torch module and compiling it down using Torch-MLIR (details can be found here).

Installation

In addition to the basic requirements for QPyTorch, packages below are needed to run the experiment:

• wheel
• numpy
• An installation of Torch-MLIR

Torch-MLIR Examperiment

First make sure you have the changes similar to this PR, and then

1. Build the custom operators for qtorch_ops by creating some directory named build and running the CMake command cmake -DCMAKE_PREFIX_PATH="$(python -c 'import torch.utils; print(torch.utils.cmake_prefix_path)')" $PATH_TO_QUANT-CPU_DIR in it. Where PATH_TO_QUANT-CPU_DIR is the path to QPyTorch/qtroch/quant/quant_cpu.
2. Run make to create the shared library libqtorch_ops.so.
3. Build qtorch package using pip install . command from the top-level directory in QPyTorch.
4. Setup Python environment for Torch-MLIR using the command below:

export PYTHONPATH=`pwd`/build/tools/torch-mlir/python_packages/torch_mlir:`pwd`/examples

1. Run the Torch-MLIR experiment script in examples/torch-mlir_experiment.py