LoRA training on Linux

Troubleshooting

These are some issues I had while trying to run the LoRA training repo on Arch Linux.
This is based on this other guide, so use that as a base and use this guide if you have trouble installing xformers or some message saying CUDA couldn't be found.

Keep in mind everything below should be done after activating the sd-scripts venv. The activate thingy is located in a different directory from what you'll see on Windows guides. After creating the venv you can run . venv/bin/activate to activate it.

There's also a section at end describing step by step what I did to get it to run. Of course, you may run into totally different issues depending on your environment, so it may not work for you.

xformers

First of all, make sure you have CUDA installed.

The instructions on the sd-scripts repo for installing xformers only work on Windows. You should be able to simply run pip install xformers --prefer-binary, so try doing that first. If you run into a CUTLASS submodule not found error, you'll need to clone the xformers repo and compile it yourself:

⎗

git clone https://github.com/facebookresearch/xformers/
cd xformers
git submodule update --init --recursive
pip install --verbose --no-deps -e .
pip install -r requirements.txt
pip install functorch==0.2.1 ninja bitsandbytes
pip install -U --pre triton

Credits to this Github comment.

After that, make sure it's actually installed with python3 -c 'import xformers'. If you don't see an error, that means it's installed.

CUDA not found

First of all, make sure you have CUDA installed.

⎗
✓
CUDA_SETUP: WARNING! libcudart.so not found in any environmental path. Searching /usr/local/cuda/lib64...
UserWarning: WARNING: The following directories listed in your path were found to be non-existent: {PosixPath('/usr/local/cuda/lib64')}
WARNING: No libcudart.so found! Install CUDA or the cudatoolkit package (anaconda)!

If you have CUDA installed and still see this message, this means bitsandbytes couldn't find your CUDA installation. For some reason bitsandbytes is dumb and even though everything else is able to find CUDA, it is not. I looked at the code and bitsandbytes is so fucking weird it looks at ALL environment variables for a directory containing libcudart.so, so you just have to figure out where libcudart.so is on your system and set whatever variable you want to that.

On Arch Linux,libcudart.so is located at /opt/cuda/targets/x86_64-linux/lib/, so you launch training with:

⎗

1	FUCK_BITSANDBYTES=/opt/cuda/targets/x86_64-linux/lib/ accelerate launch --num_cpu_threads_per_process 12 lora_train_command_line.py

What I did

Here's step by step of what I had to do for it to run on Arch Linux. There's a troubleshooting section after this with more detailed info about xformers and CUDA not being found.

Install CUDA: sudo pacman -S cuda on Arch. You'll have to figure out the exact command if you're not on Arch.
Clone the repo and cd: git clone https://github.com/kohya-ss/sd-scripts.git && cd sd-scripts
Download Python script: curl -O https://raw.githubusercontent.com/derrian-distro/LoRA_Easy_Training_Scripts/main/lora_train_command_line.py

Edit script to configure directories, model, etc. Here's how mine looks like:

⎗

1
2
3

self.base_model: string = "/media/ssd/stable-diffusion-webui/models/Stable-diffusion/nai-anime-full-pruned.ckpt"
self.img_folder: string = "/media/ssd/sd-scripts/input"
self.output_folder: string = "/media/ssd/sd-scripts/output"

Create venv and activate it: python -m venv --system-site-packages venv && . venv/bin/activate

Install stuff:

⎗
✓
pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116
pip install --upgrade -r requirements.txt

Install xformers:

⎗

cd ..
git clone https://github.com/facebookresearch/xformers/
cd xformers
git submodule update --init --recursive
pip install --verbose --no-deps -e .
pip install -r requirements.txt
pip install functorch==0.2.1 ninja bitsandbytes
pip install -U --pre triton

Make sure it's actually installed with python3 -c 'import xformers'. You'll get an error if it isn't.

Setup your input directory as described in the base guide. Should look something like this:

⎗

input
└── 10_neneneji
    ├── 100038048.jpg
    ├── 100038048.txt
    ├── 100038050.jpg
    ├── 100038050.txt
    ...

You can try running it now with accelerate launch --num_cpu_threads_per_process 12 lora_train_command_line.py. If you see a message saying that libcudart.so wasn't found, figure out where it is and set an environment variable called whatever you want with the directory where it is. Here's what I did on Arch:
⎗

✓
1
FUCK_BITSANDBYTES=/opt/cuda/targets/x86_64-linux/lib/ accelerate launch --num_cpu_threads_per_process 12 lora_train_command_line.py