FastAPI Serving Endpoint AI Prompt

Build a production-ready FastAPI model serving endpoint for {{model_name}}.

1. Application structure:
   - Lifespan context manager for model loading at startup (not per-request)
   - Global model object stored in app state, not as a module-level global
   - Separate router for model endpoints

2. Request/response schemas (Pydantic v2):
   - Input schema: {{input_schema}} with field validators and example values
   - Response schema: prediction, confidence, model_version, latency_ms, request_id
   - Error response schema with error code and message

3. Inference endpoint POST /predict:
   - Input validation via Pydantic
   - Preprocessing: replicate exactly the training preprocessing pipeline
   - Inference with torch.no_grad() and model.eval()
   - Postprocessing: convert model output to human-readable format
   - Response with latency measurement

4. Health and readiness:
   - GET /health: returns 200 if service is up
   - GET /ready: returns 200 only if model is loaded and warm
   - GET /metrics: prediction count, p50/p95/p99 latency, error rate

5. Robustness:
   - Input size limits to prevent memory exhaustion
   - Timeout on inference (configurable)
   - Graceful error handling — never return a stack trace to the client

6. Concurrency:
   - For CPU models: use thread pool executor with asyncio.run_in_executor
   - For GPU models: serialize inference with asyncio.Lock or use a request queue

Return: complete FastAPI application code with Dockerfile.