Build the Tensorflow with GPU Support from the source.
I bought a new computing machine that has RTX2070 GPU in Jan 2019. But the RTX20x0 series don't run on current tensorflow_gpu python package because it was builded for CUDA9 and RTX20x0 only supports up-to-date CUDA10. So I have to build tensorflow from source code.
To create the tensorflow_gpu package that supports RTX20x0.
-
Install the nvidia driver to your machine. CUDA is not needed because we'll use nvidia-docker2
-
Install the docker-ce, the docker-compose and the nvidia-docker2
-
Download TensorRT5 package from https://developer.nvidia.com/tensorrt
-
Modify Dockerfile. Change versions of libraries and options of tensorflow build configure.
-
Run docker-compose
$ cd Build_Tensorflow_GPU
$ docker-compose build
That download build libraries, checkout TF sources from git, and build them. The build process on the Dockerfile will be faild. But you don't need to worry about it.
I know this is ugly workaround to call bazel build twice. But it works and our goal is not to write a perfect Dockerfile. Our goal is to make the tensorflow_gpu package that supports CUDA10 and be optimized ... and make a lot of money by using it (:-p.
- Enter the docer machine and run build command manually. And package it.
$ docker-compose run nivid /bin/bash
# cd ~/tensorflow
# bazel build //tensorflow/tools/pip_package:build_pip_package --config=opt
...
INFO: Elapsed time: 1843.411s, Critical Path: 168.01s
INFO: 3364 processes: 3364 local.
INFO: Build completed successfully, 4486 total actions
# ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/
don't exit from the bash now.
- Copy the package form the docker-container to the Host machine.
$ docker ps
... ... ... NAMES
... ... ... tensorflow_nvidia_run_67e67b06722a
$ docker cp tensorflow_nvidia_run_67e67b06722a:/tmp/tensorflow_pkg/tensorflow-1.13.0rc0-cp36-cp36m-linux_x86_64.whl ./
- Terminate the docker container.
bye.