Security article

AI - Stable diffussion with dreambooth

AI - Stable diffussion with dreambooth: CUDA 11.8 on Xubuntu (GCC 12) — My DreamBooth Prep Notes • Knowledge • cuda, dreambooth

2023-05-192 tags

CUDA 11.8 on Xubuntu (GCC 12) — My DreamBooth Prep Notes

I’ve spent a lot of time getting a local Linux box ready to train image models of me. Cloud (Drive + rented GPUs) works, but I wanted it running on my own metal. The catch: tons of moving pieces that must play nicely together — GCC, CUDA toolkit, kernel/NVIDIA drivers, etc. The “latest and greatest” combo didn’t behave for me, so I pinned a known-good stack: CUDA 11.8 on Xubuntu 22.04 with GCC 12.

1) NVIDIA driver (clean slate → 525)

On a fresh Xubuntu 22.04 install I wiped any old NVIDIA bits and installed nvidia-driver-525:

sudo apt update    # optional but recommended
sudo apt upgrade   # optional but recommended
sudo apt install nvidia-driver-525

Reboot after installing the driver.

2) Install CUDA 11.8 (runfile)

Grab the CUDA 11.8 runfile installer and run it:

wget https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda_11.8.0_520.61.05_linux.run
sudo sh cuda_11.8.0_520.61.05_linux.run

Quick sanity check:

nvidia-smi

You should see your GPUs and a driver line (e.g., Driver Version: 525.xx, CUDA Version: 12.0).

3) Add CUDA to your environment

Append to ~/.bashrc:

export PATH="/usr/local/cuda/bin:$PATH"
export LD_LIBRARY_PATH="/usr/local/cuda/lib64:$LD_LIBRARY_PATH"

Reload the shell:

source ~/.bashrc

sudo bash -c "echo /usr/local/cuda/lib64 > /etc/ld.so.conf"
sudo ldconfig

Confirm it’s visible:

ldconfig -p | grep cuda

4) Verify your CUDA version

cat /usr/local/cuda/version.json | grep version -B 2
# Expect something like:
# "cuda" : { "name" : "CUDA SDK", "version" : "11.8.20220929" },

5) Build CUDA Samples (optional but reassuring)

Samples moved to GitHub:

git clone https://github.com/NVIDIA/cuda-samples.git
cd cuda-samples/
git checkout v11.8

Install a required dependency:

sudo apt install libfreeimage-dev

Compile using all cores and tail the log:

make -j"$(nproc)" > compile.log 2>&1 &
tail -f compile.log

You’ll get warnings for older GPUs — ignore them. When the log ends with Finished building CUDA samples, try a couple examples:

# CUBLAS matrix multiply
cd Samples/4_CUDA_Libraries/matrixMulCUBLAS
./matrixMulCUBLAS

# (Optional) P2P bandwidth/latency if you have multiple GPUs
cd ~/cuda-samples/Samples/5_Domain_Specific/p2pBandwidthLatencyTest
./p2pBandwidthLatencyTest

NVLink matrix (if applicable):

nvidia-smi nvlink -s

Clean up if you like:

rm -rf ~/cuda_11.8.0_520.61.05_linux.run ~/cuda-samples

6) Python env + Diffusers / DreamBooth deps

You can use Anaconda or plain venv. If you prefer Anaconda (example version — grab the latest from their site):

wget https://repo.anaconda.com/archive/Anaconda3-2022.05-Linux-x86_64.sh
chmod +x ./Anaconda3-2022.05-Linux-x86_64.sh
./Anaconda3-2022.05-Linux-x86_64.sh

Clone the DreamBooth-ready Diffusers fork and install pieces:

# Fork with DreamBooth examples
git clone https://github.com/ShivamShrirao/diffusers
cd diffusers

# CUDA-enabled PyTorch (11.7 wheels play well with a 11.8 toolkit install)
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117

# DreamBooth example requirements
cd examples/dreambooth
pip install -r requirements.txt

# Back to repo root to install diffusers in editable/standard mode
cd ../..
pip install .

# Extra performance libs
pip install -U --pre triton
pip install ninja bitsandbytes

# xFormers (specific commit used)
pip install git+https://github.com/facebookresearch/xformers@51dd119#egg=xformers

Configure accelerate (I set all device indices to 0 and chose FP16):

accelerate config

Authenticate with Hugging Face:

huggingface-cli login

Convert a Diffusers model to an original Stable Diffusion checkpoint (example):

python convert_diffusers_to_original_stable_diffusion.py   --model_path /home/MODEL-TO-CONVERT   --checkpoint_path /OUTPUT_PATH

That’s all. This combo (Xubuntu 22.04 + driver 525 + CUDA 11.8 + GCC 12) was the first setup that behaved consistently for me and didn’t fight my DreamBooth runs.

← Back to overview