Multi-stage Docker build: Stage 1 compiles llama-cpp-python once, Stage 2 reuses compiled wheels - NO TIMEOUT! Build time 8-12 minutes first time, then cached.

Multi-stage Docker build: Stage 1 compiles llama-cpp-python to wheel, Stage 2 installs pre-built wheel - NO TIMEOUT! Pre-download fast-chat model at build time.

Switch to Ollama for zero-compilation deployment - pre-downloads models at startup

Revert to llama-cpp-python with storage optimization - only load fast-chat at startup

Fix: use mistral type for qwen models (not supported directly) + fallback to llama

Fix: only load fast-chat at startup (350MB) - skip other large models to save storage

Fix: map model types correctly for ctransformers (qwen, phi, llama, mistral)

Switch to ctransformers (pre-built, no compilation!) - faster HF Spaces deploy

Add FastAPI backend with Docker for HuggingFace Spaces