Jafner.net/docker-llm-amd

What we have so far

1. Ollama loads and serves a few models via API.
   • Ollama itself doesn't have a UI. CLI and API only.
   • The API can be accessed at https://api.ollama.jafner.net.
   • Ollama running as configured supports ROCm (GPU acceleration).
   • Configured models are described here, and
   • Run Ollama with: HSA_OVERRIDE_GFX_VERSION=11.0.0 OLLAMA_HOST=192.168.1.135:11434 OLLAMA_ORIGINS="app://obsidian.md*" OLLAMA_MAX_LOADED_MODELS=0 ollama serve
2. Open-webui provides a pretty web interface for interacting with Ollama.
   • The web UI can be accessed at https://ollama.jafner.net.
   • The web UI is protected by Traefik's lan-only rule, as well as its own authentication layer.
   • Run open-webui with: cd ~/Projects/LLMs/open-webui && docker compose up -d && docker compose logs -f
     • Then open the page and log in.
     • Connect the frontend to the ollama instance by opening the settings (top-right), clicking "Connections", and setting "Ollama Base URL" to "https://api.ollama.jafner.net". Hit refresh and begin using.
3. SillyTavern provides a powerful interface for building and using characters.
   • Run SillyTavern with: cd ~/Projects/LLMs/SillyTavern && ./start.sh
4. Oobabooga provides a more powerful web UI than open-webui, but it's less pretty.
   • Run Oobabooga with: cd ~/Projects/LLMs/text-generation-webui && ./start_linux.sh
   • Requires the following environment variables be set in one_click.py (right after import statements):

os.environ["ROCM_PATH"] = '/opt/rocm'
os.environ["HSA_OVERRIDE_GFX_VERSION"] = '11.0.0'
os.environ["HCC_AMDGPU_TARGET"] = 'gfx1100'
os.environ["PATH"] = '/opt/rocm/bin:$PATH'
os.environ["LD_LIBRARY_PATH"] = '/opt/rocm/lib:$LD_LIBRARY_PATH'
os.environ["CUDA_VISIBLE_DEVICES"] = '0'
os.environ["HCC_SERIALIZE_KERNEL"] = '0x3'
os.environ["HCC_SERIALIZE_KERNEL"]='0x3'
os.environ["HCC_SERIALIZE_COPY"]='0x3'
os.environ["HIP_TRACE_API"]='0x2' 
os.environ["HF_TOKEN"]='<my-huggingface-token>'

- Requires the following environment variable be set in `start_linux.sh` for access to non-public model downloads:

# config
HF_TOKEN="<my-huggingface-token>"

That's where we're at.

Set Up Models Directory

1. Navigate to the source directory with all models: cd "~/Nextcloud/Large Language Models/GGUF/"
2. Symlink each file into the docker project's models directory: for model in ./*; do ln $(realpath $model) $(realpath ~/Git/docker-llm-amd/models/$model); done
   • Note that the symlinks must be hardlinks or they will not be passed properly into containers.
3. Launch ollama: docker compose up -d ollama
4. Create models defined by the modelfiles: docker compose exec -dit ollama /modelfiles/.loadmodels.sh

Roadmap

• Set up StableDiffusion-web-UI.

• Get characters in SillyTavern behaving as expected.

  • Repetition issues.
  • Obsession with certain parts of prompt.
  • Refusals.

• Set up something for character voices.

  • Coqui TTS - Docker install.
  • TTS Generation Web UI.

• Set up Extras for SillyTavern.

Notes

• So many of these projects use Python with its various version and dependencies and shit.
  • Always use a Docker container or virtual environment.
  • It's like a condom.