A docker image environment for LLM development on GPUs (Nvidia), useful with gcube or vastai.
Include torch, transformers, unsloth, flash-attn, axolotl.
Use port 8000 for jupyter notebook.
docker pull ghcr.io/deveworld/gpu-dev:cuda-12.6
docker pull ghcr.io/deveworld/gpu-dev:cuda-12.8
Need >= 210GB RAM
git clone https://github.com/deveworld/gpu-dev
sudo docker build -t ghcr.io/deveworld/gpu-dev:latest -t ghcr.io/deveworld/gpu-dev:cuda-12.6 . -f 12.6/Dockerfile
sudo docker build -t ghcr.io/deveworld/gpu-dev:cuda-12.8 . -f 12.8/Dockerfile
sudo docker login --username {username} --password-stdin
sudo docker push ghcr.io/deveworld/gpu-dev --all-tags