Handling Asynchronous FastAPI Workers for Heavy AI Inference Endpoints

When wrapping heavy machine learning models in a FastAPI backend, mixing synchronous blocking code inside async route handlers will instantly deadlock your event loop.

Always ensure that CPU-bound operations (like image pre-processing, matrix multiplications, or vector searches) are offloaded using run_in_threadpool or handled by dedicated background worker queues like Celery or RQ.

How do you structure your dependency injection for model weights so they are loaded once into GPU memory on startup rather than re-instantiated on every request?