Cuda-and-Tensorflow-on-Linux-with-Compute-Capability-3.0: A repository from flashlan

Cuda ad Tensorflow on Linux with Compute Capability 3.0

tensorflow 2.3 | cuda 10.1 | cudnn 7.6.5 | bazel 3.1.0 | Python3.7 | GCC 8.3.0:

GCC (8.3)

gcc-8 is in Debian Buster.

Add buster repositories

sudo apt -y install gcc-8 g++-8

sudo ln -s /usr/bin/gcc-8 /usr/local/cuda-10.1/bin/gcc

on source directory

patch < patch.patch

Change Default Gcc

sudo update-alternatives --query gcc
sudo update-alternatives --install /usr/bin/gcc gcc /home/sandman/opt/gcc-8.2.0/bin/gcc 1
sudo update-alternatives --install /usr/bin/g++ g++ /home/sandman/opt/gcc-8.2.0/bin/g++ 1

(to recover)

sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 2
sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 2

to clear cache if gcc --version mismatch

`hash -r`

Nvidia Driver

Need a patch to compile driver on kernels > 5.10

cd ~
mkdir NVIDIA
cd NVIDIA
wget -O inttf-nvidia-patcher.sh https://nvidia.if-not-true-then-false.com/patcher/inttf-nvidia-patcher.sh
chmod +x inttf-nvidia-patcher.sh
./inttf-nvidia-patcher.sh -v 418.113

sudo service lightdm stop

Cuda (10.1)


sudo ./cuda_10.1.243_418.87.00_linux.run --override
sudo ln -s /home/sandman/opt/gcc-8.2.0/bin/gcc /usr/local/cuda-10.1/bin/gcc
sudo ln -s /home/sandman/opt/gcc-8.2.0/bin/gcc-ar /usr/local/cuda-10.1/bin/ar
sudo ln -s /home/sandman/opt/gcc-8.2.0/bin/gcc-ranlib /usr/local/cuda-10.1/bin/ranlib
sudo ln -s /home/sandman/opt/gcc-8.2.0/bin/gcc-nm /usr/local/cuda-10.1/bin/nm

nano ~/.bashrc
export PATH=/usr/local/cuda-10.1/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda-10.1/lib64:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/usr/local/cuda-10.1/include:$LD_LIBRARY_PATH
source ~/.bashrc
sudo ldconfig

Cudnn (7.6.2)

download cudnn

cd Downloads/
tar -xzvf cudnn-7.6.5-cuda10.0_0.tgz
sudo cp cuda/include/cudnn*.h /usr/local/cuda/include
sudo cp -P cuda/lib/libcudnn* /usr/local/cuda/lib64
sudo chmod a+r /usr/local/cuda/include/cudnn*.h /usr/local/cuda/lib64/libcudnn*

TensorFlow

sudo apt install python3.7-full python3.7 python3.7-venv python3.7-dev
virtualenv --python=python3.7 ~/dev/.env37
source ~/dev/.env37/bin/activate
which python
python --version

sudo apt install software-properties-common -y
sudo apt-add-repository ppa:deadsnakes/ppa
sudo nano /etc/apt/sources.list.d/deadsnakes-ubuntu-ppa-mantic.list

change mantic to focal

sudo apt update
sudo dpkg -i bellsoft-jdk8u382+6-linux-amd64.deb
sudo update-alternatives --config java

Install Bazel

wget https://github.com/bazelbuild/bazel/releases/download/3.1.0/bazel-3.1.0-installer-linux-x86_64.sh
chmod +x bazel-3.1.0-installer-linux-x86_64.sh
./bazel-3.1.0-installer-linux-x86_64.sh --user
shell
export PATH=~/bin:$PATH
bazel --version

(3.1.0)

Build Tensorflow

git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow
git checkout r2.3
TF_UNOFFICIAL_SETTING=1 ./configure
cuda=y clang=N 
bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
pip install --user /tmp/tensorflow_pkg/tensorflow-*.whl

Jupyter lab

pip install jupyterlab
jupyter lab

TODO: conpile tensorflow with Tensorrt support

CMake

download and put on ~/dev/.cmake

./cmake-3.12.3-Linux-x86_64.sh

y and n

export PATH=~/dev/.cmake/bin:$PATH

remove CMakeCache.txt em caso de erros

FIX RIGHT Version (cuda capabilities 3.0 -> Gtx 870m)

tensorflow/tensorflow#46653

This is to confirm that the following recipe solved this problem in my case with tensorflow 2.3.2 + cuda 10.1 + cudnn 7.6.5 + bazel 3.10:

disable XLA (not supported with compute capability 3.0) by removing the line build:xla in .tf_configure.bazelrc and adding:

build --define=with_xla_support=false 
build --action_env TF_ENABLE_XLA=0

for tensorflow > 2.2 the code changed and it is necessary to define TF_EXTRA_CUDA_CAPABILITIES in bazel command line:

--copt=-DTF_EXTRA_CUDA_CAPABILITIES=3.0

flashlan/Cuda-and-Tensorflow-on-Linux-with-Compute-Capability-3.0