TensorFlow Wheels

This repository contains Python Linux wheels for TensorFlow. Most of the wheels are compiled using modified settings from the Archlinux PKGBUILDs. The build script is build-tf2-gpu-avx2-mkl.sh, which builds the optimised TensorFlow wheel with TensorRT support.

Each release page also has the checksums of the attached files.

manylinux_2_28_x86_64 wheels

The following wheels were compiles in the manylinux_2_28_x86_64 container described below. These are required to have more recent CUDA compatibility since Nvidia stopped releasing updates for CentOS 7 based releases.

TF	Python	GPU	CUDA	cuDNN	AVX2	MKL/oneDNN	TensorRT	Links
2.16.2	3.9-3.12	5.2-9.0	12.6	9.4	✔️	✔️	10.4	Release

manylinux2014_x86_64 wheels

The following wheels were compiled in the manylinux2014_x86_64 container described below. These should have better glibc compatibility than the Ubuntu 20.04 ones.

TF	Python	GPU	CUDA	cuDNN	AVX2	MKL/oneDNN	TensorRT	Links
2.14.1, 2.15.0	3.8-3.12	5.2-9.0	12.3	8.9	✔️	✔️	8.6	Release
2.13.0	3.8-3.11	5.2-9.0	12.2	8.9	✔️	✔️	8.6	Release
2.12.0	3.8-3.11	5.2-9.0	12.1	8.9	✔️	✔️	8.6	Release
2.7.4, 2.8.4, 2.9.3, 2.10.1, 2.11.0	3.7-3.11	5.2-9.0	11.8	8.8	✔️	✔️	8.5	Release
2.7.4	3.6	5.2-9.0	11.8	8.8	✔️	✔️	8.5	Release

Ubuntu 20.04 wheels

The following wheels were compiled on an Ubuntu 20.04 system

TF	Python	GPU	CUDA	cuDNN	AVX2	MKL/oneDNN	TensorRT	Links
2.11.0 (nightly)	3.11	5.2-9.0	11.8	8	✔️	✔️	8	Release
2.11.0	3.7-3.10	5.2-9.0	11.8	8	✔️	✔️	8	Release
2.10.0	3.10	5.2-8.7	11.7	8	✔️	✔️	8	Release
2.10.0	3.9	5.2-8.7	11.7	8	✔️	✔️	8	Release
2.10.0	3.8	5.2-8.7	11.7	8	✔️	✔️	8	Release
2.4.0	3.9	7.x	11.2	8	✔️	✔️	❌	Release
2.3.0	3.8	7.x	11.0	8	✔️	✔️	7	Release
2.4.0 commit	3.8	❌	❌	❌	❌	❌	❌	Release
2.2.0	3.8	7.x	10.2	7	✔️	✔️	7	Release
2.1.0	3.8	7.x	10.2	7	✔️	✔️	6	Release
1.15.0	3.8	7.x	10.2	7	✔️	✔️	6	Release
1.14.1	3.8	7.x	10.2	7	✔️	✔️	6	Release

Docker containers

The Dockerfiles under docker/ are based on manylinux* and can be built with the following command, from within the docker/ directory:

docker build -t tf_build -f Dockerfile.<manylinux_ver> .

The container can then be run like so:

docker run --gpus all -it --rm --tmpfs /tmp:exec -v /path/to/tensorflow:/build -u $(id -u):$(id -g) -e USER=$(id -u) tf_build

Then, you can run the build script:

cd /path/to/tensorflow
bash build-tf2-gpu-avx2-mkl.sh -p 12

The other scripts assume a directory structure as follows:

.../
    tensorflow/
        build-tf2-gpu-avx2-mkl.sh
    text/
        build-text.sh
    io/
        build-io.sh
    wheels/
        tensorflow/
            *.whl
        keras/
            *.whl
    sanity-check.sh

Patches

The patches/ directory contains several patches that fix bugs and compatibility issues between TensorFlow and different versions of Python. The full subdirectory contains the complete diff used for each version of TensorFlow and versions of Python.