Skip to content

Linux.md

Latest commit

 

History

History
75 lines (63 loc) · 3.69 KB

Linux.md

File metadata and controls

75 lines (63 loc) · 3.69 KB
Raw

Linux support — Soporte Linux

[EN] This project runs on Linux besides Windows. Every .bat has a same-named .sh equivalent, and scripts/launcher.py picks the right one per platform. [ES] El proyecto funciona en Linux además de en Windows. Cada .bat tiene su equivalente .sh con el mismo nombre, y scripts/launcher.py elige el que toca según la plataforma.

[EN] Requirements / [ES] Requisitos

  • Linux 64-bit (tested on Nobara/Fedora 44, should work on Ubuntu/Debian/Arch).
  • NVIDIA GPU RTX 20xx / GTX 16xx or newer, driver 580+, 8 GB VRAM (4 GB for Anima and SDXL in NF4, 12 GB for LTX-2.3, FLUX.2 Klein 9B, Ideogram 4 and SDXL in BF16).
  • Git, Python 3.13 (Install_LoRAlab-TrainerStudio.sh tries to install it via apt/dnf/pacman, sudo asked), CUDA Toolkit with nvcc (only needed to build SageAttention), and ffmpeg (needed for MiniMax-H3 video clips).

[EN] Install / [ES] Instalación

git clone https://github.com/AcademiaSD/AcademiaSD_LoRAlab-TrainerStudio.git
cd AcademiaSD_LoRAlab-TrainerStudio
chmod +x *.sh code/*.sh
./Install_LoRAlab-TrainerStudio.sh   # one shared venv: PyTorch cu130, Diffusers, Transformers, PEFT, bitsandbytes, Flask...
./Install_Triton_SageAtten220.sh     # optional: Triton from PyPI + SageAttention 2.2.0 built from source

[EN] Use / [ES] Uso

./Start_LoRAlab-TrainerStudio.sh     # launcher at http://127.0.0.1:4990
./Update_LoRAlab-TrainerStudio.sh    # git pull equivalent (keeps models, datasets, settings)

Clicking a trainer card opens its code/Run_LoRAlab-*.sh in a new terminal (konsole → gnome-terminal → xfce4-terminal → xterm); closing that window stops the trainer, same as start does with the .bat on Windows. Over SSH (no graphical terminal) the trainer runs detached and logs to settings/trainer_console.log. Trainers can also be started directly:

./code/Run_LoRAlab-Krea2.sh
./code/Run_LoRAlab-LTX23.sh
./code/Run_LoRAlab-MiniMaxH3.sh
./code/Run_LoRAlab-QwenImage21.sh
./code/Run_LoRAlab-ZImage.sh
./code/Run_LoRAlab-Anima.sh
./code/Run_LoRAlab-Klein9B.sh
./code/Run_LoRAlab-Ideogram4.sh
./code/Run_LoRAlab-SDXL.sh

On a headless server, enable 🌐 Remote access in the launcher (or edit settings/network.json) to open the launcher and trainers from another device, with a password or behind your own reverse proxy: see Remote access in the README. To use your own conda/uv environment instead of venv/, install requirements.txt in it and start with LORALAB_PYTHON=/path/to/env/bin/python ./Start_LoRAlab-TrainerStudio.sh.

[EN] Notes / [ES] Notas

  • SageAttention 2.2.0 is not on PyPI (only 1.x there) and its Windows .whl files don't work on Linux, so it is built from source (thu-ml/SageAttention@v2.2.0). The build needs nvcc + gcc + ninja and takes several minutes; torch 2.14 requires -std=c++20, which the script passes via CXX/NVCC_APPEND_FLAGS. If the build fails, the PyPI 1.x version is installed instead, with an explicit warning that it is not 2.2.0.
  • GUI/launcher.json is shared between platforms: on Linux the launcher maps each Run_LoRAlab-*.bat entry to its .sh. No Windows file was modified in behavior; all Python changes are os.name == "nt" branches.
  • Out of scope: tools/*.bat model converters (advanced use, Windows-only for now).
  • If a trainer stops with an error, its terminal stays open until you press Enter, like the pause at the end of the .bat files.
  • Browse buttons open the system folder dialog through tkinter; without it (e.g. no python3.13-tk package) they open the web interface's own folder browser instead.
  • Edit the .sh files with LF line endings (.gitattributes enforces it on checkout).