fix: override ENTRYPOINT to run start.sh as shell script

refactor: switch to official llama.cpp server Docker image

fix: restore runtime download, sync all config changes

fix: install llama-cpp-python from pre-built CPU wheel to prevent OOMKilled

fix: upgrade pydantic to V2 and add sed to apt deps

fix: strip CRLF from start.sh via sed in Dockerfile and tr -d in script

fix: use PyPI directly, llama-cpp-python==0.2.76, no custom index

fix: pinned deps in single requirements.txt, /simple index url, simplified start.sh

fix: split llama-cpp-python wheel install from server deps to avoid source build

fix: use --index-url to force pre-built CPU wheel, avoid source build

fix: use pre-built CPU wheel, remove cmake/build-essential and requirements.txt

feat: stable final version with pinned deps and MODEL_URL env var

fix: correct CMD syntax with backslash continuation; use curl for model download

fix: inline CMD to avoid BOM/CRLF; use hf instead of huggingface-cli

fix: use bash to run start.sh to avoid exec format error

fix: use --index-url for pre-built llama-cpp-python CPU wheel

fix: use pre-built llama-cpp-python CPU wheel to avoid compile timeout

fix: switch to llama-cpp-python server (ghcr image not found)

Initial deploy: llama-server with guardrails_typhoon25_q4_k_m