Tensor learning compiler binary distribution package.

We use github action to build wheel and conda packages nightly.

Checkout .github/workflows

1. Build docker images

./docker/build_image.sh <CONTAINER_TYPE>

CONTAINER_NAME: Type of the docker container used to build wheels, e.g., (cpu|cu100|cu101|cu102|cu111)

2. Checkout tvm and sync version

git clone https://github.com/apache/tvm --recursive
# synchronize the package version
python common/sync_package.py [tlcpack|tlcpack-nightly]

The nightly will point to the latest main, tlcpack will point to a stable build hashtag defined in common/sync_package.py

3. Build tlcpack manylinux wheels.

./docker/bash.sh [docker-image] ./wheel/build_manylinux_wheel.sh --cuda none

To build wheels for a specific CUDA version, for example, CUDA 10.1, run

./docker/bash.sh [docker-image] ./wheel/build_manylinux_wheel.sh --cuda 10.1

The docker image is built in step 1 and needs to match the cuda version.

4. Get the wheels

The wheels are now available in

./tvm/python/repaired_wheels