Cuda Installer for Tensorflow and Pytorch.

LAST UPDATED: (09/04/2024)

• Finally made a script.
• There is currently no way for both Tensorflow and Torch to use the same local cuda install and Python 3.12
• If you would like to use the same cuda version for both, use Python 3.8-3.11

Keeping the list updated

• Submit a pull request chaning the pacman_cuda.json
• Open an issue if you encounter any bugs with the script.

Changes

Python 3.12 support

• Script(Recent):
  • Installs cuda 12.3, the only cuda version supported for Tensorflow in Python 3.12.
  • For Torch, cuda has to be installed within the virtual environment.
• Script(Compatible):
  • Installs cuda 11.8, this is the last time Tensorflow and Torch shared a cuda version.

---

Script

curl -O https://raw.githubusercontent.com/cartersusi/pacman_cuda/main/install
chmod +x install
./install

DIY

❗❗❗DO NOT USE YAY OR GIT❗❗❗

My experiences from the first 10+ times using yay and git

• 30+ minute gcc compile times ✅
• Linker Errors ✅
• Auto-updates & Version Mismatches ✅
• Nvidia doesn't like you ✅
• They actually hate you ✅

---

Current Compatability

https://www.tensorflow.org/install/source#gpu
https://pytorch.org/get-started/locally/

Version	Python version	Compiler	Build tools	cuDNN	CUDA
tensorflow-2.17.0	3.9-3.12	Clang 17.0.6	Bazel 6.5.0	8.9	12.3
tensorflow-2.13.0	3.8-3.11	Clang 16.0.0	Bazel 5.3.0	8.6	11.8
Pytorch(Stable)	3.8+				11.8, 12.1

---

Torch typically bundles pre-compiled CUDA binaries and does not require the system Cuda install.

# Current:
pip install torch torchvision torchaudio

---

1. Update and download nvidia drivers ('nvidia' and 'nvidia-dkms' are interchangeable, no need to replace your 'nvidia' package if it is already installed)

sudo pacman -Syu nvidia-dkms opencl-nvidia nvidia-utils nvidia-settings curl

2. Download and install gcc12

curl -O https://archive.archlinux.org/packages/g/gcc12/gcc12-12.3.0-6-x86_64.pkg.tar.zst
curl -O https://archive.archlinux.org/packages/g/gcc12-libs/gcc12-libs-12.3.0-6-x86_64.pkg.tar.zst
sudo pacman -U gcc12-12.3.0-6-x86_64.pkg.tar.zst  gcc12-libs-12.3.0-6-x86_64.pkg.tar.zst

3. Download and install CUDA and cuDNN

curl -O https://archive.archlinux.org/packages/c/cuda/cuda-12.3.2-1-x86_64.pkg.tar.zst
curl -O https://archive.archlinux.org/packages/c/cudnn/cudnn-8.9.7.29-1-x86_64.pkg.tar.zst
sudo pacman -U cuda-12.3.2-1-x86_64.pkg.tar.zst cudnn-8.9.7.29-1-x86_64.pkg.tar.zst

4. Update /etc/pacman.conf to exclude cuda and cudnn

• Uncomment the line "#IgnorePkg =", then add cuda and cudnn

IgnorePkg = cuda cudnn 

Common Tensorflow error

# ERROR: libdevice not found at ./libdevice.10.bc 
export XLA_FLAGS=--xla_gpu_cuda_data_dir=/opt/cuda

Common Docker error

sudo nvim /etc/nvidia-container-runtime/config.toml # change no-cgroups = false, save

sudo systemctl restart docker
sudo docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi

Links for Cuda 11.8

gcc11:

• https://archive.archlinux.org/packages/g/gcc11/gcc11-11.3.0-5-x86_64.pkg.tar.zst
• https://archive.archlinux.org/packages/g/gcc11-libs/gcc11-libs-11.3.0-5-x86_64.pkg.tar.zst

cuda:

• https://archive.archlinux.org/packages/c/cuda/cuda-11.8.0-1-x86_64.pkg.tar.zst

cudnn:

• https://archive.archlinux.org/packages/c/cudnn/cudnn-8.6.0.163-1-x86_64.pkg.tar.zst