Real fix: preload nvidia-*-cu12 pip packages' shared libraries via ctypes before llama_cpp import (libcudart.so.12 not found on ZeroGPU worker, confirmed via real runtime logs)

Real fix: restore demo.launch() as entrypoint (required for ZeroGPU detection), expose /generate via Gradio's native api_name mechanism instead of a custom FastAPI route

Fix /generate route being shadowed by Gradio's own catch-all -- mount Gradio at /ui instead of adding routes post-hoc to demo.app

Switch to Gradio SDK -- ZeroGPU only works with Gradio SDK, not Docker/FastAPI

chatPDB API: FastAPI + llama-cpp-python ZeroGPU backend, Q4_K_M GGUF