# llama.cpp CUDA server — serves the Gemma-4 E4B GGUF + mmproj with vision.
# Self-contained: no torch / transformers / bitsandbytes version hell.
FROM ghcr.io/ggml-org/llama.cpp:server-cuda

ENV HF_HOME=/tmp/hf

# llama-server is the image entrypoint; these are its args.
# -hf auto-downloads the quant + the repo's mmproj (vision) file.
# Serves an OpenAI-compatible /v1/chat/completions (+ /health) on 7860.
CMD ["--host", "0.0.0.0", "--port", "7860", \
     "-hf", "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic-GGUF:Q5_K_M", \
     "-ngl", "99", "-c", "4096", "--jinja"]