How to make tensorflow, pytorch, gpu, cuda

How to setup environment for Machine Learning

Window

그래픽 드라이버(최신), CUDA toolkit (12.3 x -> 12.1), pytorch, python (3.12 x -> 3.8.10 downgrade) (2024. Jan)

Ubuntu

요약하자면

그래픽 드라이버, CUDA toolkit(11.8), ~~cuDNN~~, tensorflow
그래픽 드라이버, CUDA toolkit any (12.2 > 11.8), ~~cuDNN~~, pytorch
tensorflow, pytorch를 같이 까는 것은 안된다고 한다

주의할 점

tensorflow는 호환성이 아주 중요하다
파일 다운로드, 압축 풀기, 경로에 복사해서 위치시키기, 환경 변수에 넣어주기가 모두 잘 되어야 한다.

cuda 11.8, cuDNN 8.xx, tensorflow 2.13인데 GPU 장치는 인식되지만 모델학습이 안된다

model.fit(train_images, train_labels, epochs=10, batch_size=64)  # 모델을 학습시킵니다. Need cuDNN libraries

설치된 그래픽 드라이버, CUDA toolkit version, cuDNN version, pytorch version 정보는 아래와 같이 확인한다.

nvidia-smi # 그래픽 드라이버
nvcc --version # CUDA
cat cudnn.h | grep CUDNN_MAJOR -A 2 # cuDNN

cuDNN은 설치 경로에 복사되어 있는 지 확인한다 컴퓨터 > usr > local > cuda-x.x > include > cudnn*.h

아카이브에서 .deb 파일을 다운받고 dpkg했을 때 /var/cudnn-local-repo-ubuntu2004-8.9.5.29 폴더가 생성되고 그 하위에 또 .deb들이 생성된 것이 보인다. > 원하는 동작이 아니다 다시 아카이브에서, .tar를 받고 압축을 푼 후(우클릭 > 압축풀기), 필요한 경로(cuda toolkit 설치 경로)에 복사/붙여넣기 한다.

복사해서 붙여 넣는 방법

압축 푼 폴더 내에서 터미널을 열고

# cuda 
sudo cp /home/ko/Downloads/cudnn-linux-x86_64-8.9.3.28_cuda11-archive/include/cudnn*.h /usr/local/cuda/include/
sudo cp /home/ko/Downloads/cudnn-linux-x86_64-8.9.3.28_cuda11-archive/lib/libcudnn* /usr/local/cuda/lib64/
# cuda-11
sudo cp /home/ko/Downloads/cudnn-linux-x86_64-8.9.3.28_cuda11-archive/include/cudnn*.h /usr/local/cuda-11/include/
sudo cp /home/ko/Downloads/cudnn-linux-x86_64-8.9.3.28_cuda11-archive/lib/libcudnn* /usr/local/cuda-11/lib64/
# cuda-11.8
sudo cp /home/ko/Downloads/cudnn-linux-x86_64-8.9.3.28_cuda11-archive/include/cudnn*.h /usr/local/cuda-11.8/include/
sudo cp /home/ko/Downloads/cudnn-linux-x86_64-8.9.3.28_cuda11-archive/lib/libcudnn* /usr/local/cuda-11.8/lib64/
# cuda-12
sudo cp /home/ko/Downloads/cudnn-linux-x86_64-8.9.4.25_cuda12-archive/include/cudnn*.h /usr/local/cuda-12/include/
sudo cp /home/ko/Downloads/cudnn-linux-x86_64-8.9.4.25_cuda12-archive/lib/libcudnn* /usr/local/cuda-12/lib64/
# cuda-12.1
sudo cp /home/ko/Downloads/cudnn-linux-x86_64-8.9.4.25_cuda12-archive/include/cudnn*.h /usr/local/cuda-12.1/include/
sudo cp /home/ko/Downloads/cudnn-linux-x86_64-8.9.4.25_cuda12-archive/lib/libcudnn* /usr/local/cuda-12.1/lib64/
# cuda-12.2
sudo cp /home/ko/Downloads/cudnn-linux-x86_64-8.9.4.25_cuda12-archive/include/cudnn*.h /usr/local/cuda-12.2/include/
sudo cp /home/ko/Downloads/cudnn-linux-x86_64-8.9.4.25_cuda12-archive/lib/libcudnn* /usr/local/cuda-12.2/lib64/

환경변수 (path)에 libr64, include를 추가해준다(임시)

export LD_LIBRARY_PATH=/usr/local/cuda-11.8/lib64:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/usr/local/cuda-11.8/include:$LD_LIBRARY_PATH
echo $LD_LIBRARY_PATH # check registered paths

터미널을 다시 열면, 도로 안된다. 환경변수 파일을 수정(계속 유지됨)

nano ~/.bashrc
.bashrc의 맨 아래 추가한다.

export PATH=/usr/local/cuda-11.8/bin:/usr/local/cuda-11.8/NsightCompute-2023.1:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-11.8/lib64:/usr/local/cuda-11.8/include:$LD_LIBRARY_PATH

적용한다

  source ~/.bashrc

확인한다

  echo $LD_LIBRARY_PATH
  echo $PATH

참고 중인 블로그 https://normal-engineer.tistory.com/356
Install graphic driver

Install CUDA
- Tensorflow requires CUDA Toolkit 11.8 https://developer.nvidia.com/cuda-11-8-0-download-archive

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-ubuntu2004.pin
sudo mv cuda-ubuntu2004.pin /etc/apt/preferences.d/cuda-repository-pin-600
wget https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda-repo-ubuntu2004-11-8-local_11.8.0-520.61.05-1_amd64.deb
sudo dpkg -i cuda-repo-ubuntu2004-11-8-local_11.8.0-520.61.05-1_amd64.deb
sudo cp /var/cuda-repo-ubuntu2004-11-8-local/cuda-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get -y install cuda

기존의 것을 삭제한 후 다시 깔았다면 nvidia-smi에서 library mismatch 오류가 날 수 있다. 재부팅하면 된다.

CUDA toolkit release note https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#
Install cuDNN
- manula https://docs.nvidia.com/deeplearning/cudnn/install-guide/index.html#installlinux-deb
- archive https://developer.nvidia.com/rdp/cudnn-archive
- .deb말고(힘듬) .tar를 받자(쉬움)
Check properly installed
- files > 다른 위치 > 컴퓨터 > usr > local > cuda > include > cudnn*

Install (.deb 오히려 좀 더 귀찮음)

CUDA toolkit에 포함되어 있다고 한다. ~~안되는데~~
Go to official site https://developer.nvidia.com/rdp/cudnn-download
Download proper version
- Local Installers for Windows and Linux, Ubuntu(x86_64, armsbsa)
- Local Installer for Ubuntu20.04 x86_64 (Deb)
- CUDA 11.8

GPG key

sudo cp /var/cudnn-local-repo-ubuntu2004-8.9.5.29/cudnn-local-C1AC07B2-keyring.gpg /usr/share/keyrings/

depackage

sudo dpkg -i cudnn-local-repo-ubuntu2004-8.9.5.29_1.0-1_amd64.deb

apt update
```
sudo apt-get update
```

install

sudo apt-get install cudnn-local-repo-ubuntu2004-8.9.5.29

add environment variable

sudo find /usr -name cuda
# outputs
# /usr/include/thrust/system/cuda
# /usr/local/cuda-11.8/targets/x86_64-linux/include/thrust/system/cuda
# /usr/local/cuda-11.8/targets/x86_64-linux/include/cuda
# /usr/local/cuda
# /usr/share/doc/cuda
# /usr/share/doc/libthrust-dev/examples/cuda
export PATH=/usr/local/cuda-11.8/bin:/usr/local/cuda-11.8/NsightCompute-2023.1:$PATH

$disto = ubuntu2004 x.x.x = 8.9.5.29 X.Y = 11.8

Installed location

dpkg -L cudnn-local-repo-ubuntu2004-8.9.5.29
# /.
# /etc
# /etc/apt
# /etc/apt/sources.list.d
# /etc/apt/sources.list.d/cudnn-local-ubuntu2004-8.9.5.29.list
# /usr
# /usr/share
# /usr/share/doc
# /usr/share/doc/cudnn-local-repo-ubuntu2004-8.9.5.29
# /usr/share/doc/cudnn-local-repo-ubuntu2004-8.9.5.29/changelog.Debian.gz
# /var
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29/C1AC07B2.pub
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29/InRelease
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29/Local.md5
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29/Local.md5.gpg
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29/Packages
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29/Packages.gz
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29/Release
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29/Release.gpg
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29/cudnn-local-C1AC07B2-keyring.gpg
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29/libcudnn8-dev_8.9.5.29-1+cuda11.8_amd64.deb
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29/libcudnn8-samples_8.9.5.29-1+cuda11.8_amd64.deb
# /var/cudnn-local-repo-ubuntu2004-8.9.5.29/libcudnn8_8.9.5.29-1+cuda11.8_amd64.deb

installed location is 컴퓨터 > var > cudnn-local~~~

add cudnn-local repository on system (using package manager)

sudo cp /var/cudnn-local-repo-ubuntu2004-8.9.5.29/cudnn-local-C1AC07B2-keyring.gpg /usr/share/keyrings/
echo "deb [signed-by=/usr/share/keyrings/cudnn-local-C1AC07B2-keyring.gpg] file:/var/cudnn-local-repo-ubuntu2004-8.9.5.29/ /" | sudo tee /etc/apt/sources.list.d/cudnn-local-ubuntu2004-8.9.5.29.list
sudo apt-get update
sudo apt-get install libcudnn8

TensorRT install and merge to Tensorflow
- download https://developer.nvidia.com/nvidia-tensorrt-8x-download
- TensorRT 8.6 GA for Ubuntu 20.04 and CUDA 11.0, 11.1, 11.2, 11.3, 11.4, 11.5, 11.6, 11.7 and 11.8 DEB local repo Package
dpkg
저장소 키 업데이트 sudo cp /var/nv-tensorrt-local-repo-ubuntu2004-8.6.1-cuda-11.8/nv-tensorrt-local-D7BB1B18-keyring.gpg /usr/share/keyrings/
Tar version

Version

https://robot9710.tistory.com/29

https://jjuke-brain.tistory.com/entry/GPU-%EC%84%9C%EB%B2%84-%EC%82%AC%EC%9A%A9%EB%B2%95-CUDA-PyTorch-%EB%B2%84%EC%A0%84-%EB%A7%9E%EC%B6%94%EA%B8%B0-%EC%B4%9D%EC%A0%95%EB%A6%AC

https://www.tensorflow.org/install/pip?hl=ko

https://jjuke-brain.tistory.com/entry/GPU-%EC%84%9C%EB%B2%84-%EC%82%AC%EC%9A%A9%EB%B2%95-CUDA-PyTorch-%EB%B2%84%EC%A0%84-%EB%A7%9E%EC%B6%94%EA%B8%B0-%EC%B4%9D%EC%A0%95%EB%A6%AC

Share on

Twitter Facebook LinkedIn

How to make tensorflow, pytorch, gpu, cuda

Window

Ubuntu

복사해서 붙여 넣는 방법

Share on

You may also enjoy

CUDA example code 1. main.cpp, kernel.cu, kernel.cuh 구성으로 빌드하기, ubuntu

CMake

Tesseract OCR and OpenCV in Ubuntu

Qt for Linux