Fix: lazy model load, GGUF integrity check, re-download corrupt files

Fix: lazy model load on first request, add error handling

Fix: use background process with signal trap instead of exec

Fix: replace llama_cpp.server with custom FastAPI server

Fix: replace llama_cpp.server with custom FastAPI server (no mlock, no health check loop)

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Fix: --api_key (underscore) not --api-key (dash) for llama_cpp.server

Fix: always set api-key (default: not-set) so health check auth works

Fix: correct CLI flags for llama_cpp.server 0.3.32 (use_mlock, no cont_batching)

Fix: add uvicorn+fastapi deps (llama_cpp.server requires them)

Fix: correct GGUF magic bytes check (47475546 not 46474755)

Fix: use hf_hub_download instead of python -m huggingface_hub

Fix: install git (required by HF Spaces build system)

Pre-built CPU wheel: pip install in 30s instead of 10min source compile

CPU perf: OpenBLAS build, auto-threads, mlock, batched inference

Pin all deps: python:3.10-slim-bookworm, exact pip versions in requirements.txt

Fix: add build-essential cmake for llama-cpp-python source build

Switch from vLLM to llama-cpp-python (CPU, free HF tier, GGUF Q3_K_M)

Pin to v0.7.3 tag (immutable, cacheable on HF)

Switch to latest tag for faster HF cache hit

Production vLLM deploy: SHA-pinned Docker, pre-flight checks, ai.js client, usage guide

Create README.md

Create Dockerfile

initial commit