python.rst

Python

The Python library for the AMD Inference Server allows you to communicate with the server using Python.

Install the Python library

The Python library is built and installed in the development container as part of the regular CMake build. To install it outside Docker or in different containers, you need to install a precompiled wheel. Currently, these wheels are not available on Pip and must be built locally.

Build wheels

You can build wheels for the Python library to make a precompiled package that can be installed in any Linux host, container or environment. It is recommended to perform the following steps on a fresh clone of the inference server repository. These instructions assume you're only building wheels for x86_64 Linux with CPython.

# generate a Dockerfile that defines an image for building wheels
./docker/generate.py --cibuildwheel --base-image=quay.io/pypa/manylinux2014_x86_64 --base-image-type yum

# build the image. You should add some suffix to differentiate this image
# from the regular image
./amdinfer dockerize --suffix="-ci"

# this will build an image with the name $(whoami)/amdinfer-dev-ci:latest
# if you're not building wheels on the same host, you will need to upload
# this image to a Docker registry

# on a host where your image exists or can be pulled
export CIBW_MANYLINUX_X86_64_IMAGE=$(whoami)/amdinfer-dev-ci:latest

pip install cibuildwheel

# you can edit pyproject.toml to control which wheels to build or just use the defaults

cibuildwheel --platform linux

# your built wheels will be in ./wheelhouse

After following these instructions, your built wheels will be in ./wheelhouse/. The names on the wheels indicate the Python version they are compatible with. For example, cp37 in the name indicates that it's compatible with CPython 3.7. You can install these wheels in a virtual environment, Conda environment, a container or on a bare host.

pip install <path/to/wheel>

API

.. automodule:: amdinfer
    :imported-members:
    :undoc-members: