Apache TVM is a compiler stack for deep learning systems. It is designed to close the gap between productivity-focused deep learning frameworks and performance and efficiency-focused hardware backends. TVM works with deep learning frameworks to provide end-to-end compilation to different backends.

This version of TVM enables the use of Bayesian Optimization as a search strategy in MetaSchedule.

We recommend following this guide or the steps below:

1. Clone and build TVM-Bayesian-Optimization

   $ git clone --recursive https://github.com/felix-ro/TVM-Bayesian-Optimization
   $ cd TVM-Bayesian-Optimization
   $ mkdir build
   $ cp cmake/config.cmake build/
   $ cd build
   # Now configure the `config.cmake` file in the build directory
   $ cmake ..
   $ make -j4

2. Standard TVM Python dependencies

   $ pip install numpy decorator attrs tornado psutil 'xgboost>=1.1.0' cloudpickle bayesian-optimization

3. Add TVM to python path

   $ export PYTHONPATH=/path-to-tvm-unity/python:$PYTHONPATH

General usage examples with ready-to-use scripts can be found here.

For a brief overview of the results, see the slides here (fyi Safari sometimes fails to render them). For a more in-depth discussion of the project, see the report here.