• Choose Manjaro Arch based linux distrbution and more friendly for starters
• See: https://bishwarup-paul.medium.com/why-i-love-manjaro-144841982ece
• The installation is quite straight forward with iso file
• Can update the network address to manual later
• Common commands:
  • sudo pacman -Syu
  • sudo pacman -Ql packagename

• No good option for the time being, tried nomachine, anydesk, tigervnc, all without luck

• code-server: basically vscode via browser
• pamac build code-server
• settings.json location: /home/cs/.local/share/code-server/User/settings.json

• Mount NFS directory: /mnt/sample-pool/nfs
• Use systemd to auto mount on startup
• sudo systemd-escape -p --suffix=mount '/mnt/nfs': Where= must be matched to the unit name
• sudo systemctl enable mnt-nfs.mount --now

# /etc/systemd/system/mnt-nfs.mount
[Unit]
Description=NFS from TrueNAS
After=network.target

[Mount]
What=192.168.x.x:/mnt/path/nfs
Where=/mnt/nfs
Type=nfs
Options=_netdev,auto

[Install]
WantedBy=multi-user.target

• https://wiki.manjaro.org/index.php/Networking

sudo systemctl mask suspend.target hibernate.target sleep.target
sudo systemctl unmask suspend.target hibernate.target sleep.target

sudo pamac build rstudio-server-bin
sudo useradd --system rstudio-server
sudo rstudio-server verify-installation
sudo systemctl start rstudio-server

In rstudio can use py_config() to check who the python path is determined, e.g.

• by conda
• by environment variable RETICULATE_PYTHON_FALLBACK

Then login with your default linux user

sudo pacman -S tensorflow-cuda python-tensorflow-cuda

Sometimes, you may want to benchmark again CPU, can use Sys.setenv(CUDA_VISIBLE_DEVICES = -1) to disable CUDA support

NOTE:

• ArchLinux update is rolling based, which makes the dependencies in relation to CUDA unstable; in this case, conda is preferred.
• However, conda doesn't work well with tensorflow[and-cuda] for 2.11-2.15
• So, elder version tensorflow-gpu==2.10.0 is more stable, can test again when tensorflow 2.16 is released

Quick start with conda:

conda_create(env_name, python_version = "3.10")  # python 3.10 is latest version compatible with tensorflow 2.10
use_condaenv(condaenv = env_name)
reticulate::conda_install(envname = env_name, packages = c("tensorflow-gpu==2.10.0"))
reticulate::import("tensorflow")
tf$sysconfig$get_build_info()

reticulate's conda can be used in shell after init, e.g. ~/.local/share/r-miniconda/bin/conda init zsh

• In rstudio, can combine use_virtualenv and py_install(..., pip = TRUE) to install pytorch with CUDA support
• then, set CUDA version via sys.setenv accordingly
• then, install mlverse/torch
• If certain .so files cannot be found, just use find and copy them over

sudo pacman -Sc
sudo pacman -Qdt
sudo pacman -Rns $(pacman -Qtdq)
sudo journalctl --vacuum-size=50M

• See: M1 Mac Dev Machine Setup
• install-dependencies-osx shall be install-dependencies-osx-arch
• soci error
  • brew install soci
  • update the CMakeLists.txt from set(SOCI_LIBRARY_DIR "${RSTUDIO_TOOLS_SOCI}/build/lib") to set(SOCI_LIBRARY_DIR "/opt/homebrew/lib")
  • actually the above may be avoided using by install the dependences in dependencies/common/install-soci
• dependencies/common/install-soci
• sudo cp /etc/pam.d/cups /etc/pam.d/rstudio
• In order to open the GWT Development Mode, need to login to the desktop GUI

• conda active, then Jupyter Lab
• Jupyter Extensions:
  • jupyter-lsp-jupyterlab-lsp
  • jupyterlab-jupytext (For Rmd and qmd support)

conda install r-recommended r-irkernel
R -e 'IRkernel::installspec()'

conda install -y -c apple tensorflow-deps
python -m pip install tensorflow-macos
python -m pip install tensorflow-metal