README.md

JIT compile c++ code from python

This way is recommended for prototyping

python example_reduce_jit.py

Compile and install c++ code from python

pip install -e .
python example_reduce.py

Run rudimentary tests

The tests use the JIT compile solution

python tests.py

Benchmark the different implementations

python bench.py

Compile c++ code with cmake

mkdir build
cd build
cmake -DCMAKE_PREFIX_PATH=<LIBTORCH_PATH> ..
./example_reduce

you might need to reference your cuda library

Save model from Python and load it from C++

python save-torch-model.py
cd build
# compile C++ if needed
./use-torch-model ../model.pt