ONNX Runtime: deploy a converted ONNX model behind a REST API (e.g. FastAPI) using an ONNX Runtime inference session

Steps

1. Install the runtime with `pip install onnxruntime` (CPU) or `pip install onnxruntime-gpu` (CUDA) — only one of the two packages should be installed in a given environment.
2. In your FastAPI app, create a single module-level `onnxruntime.InferenceSession("model.onnx")` at startup (not per request) so the loaded graph is reused across requests.
3. Discover the expected tensor names once via `session.get_inputs()[0].name` and `session.get_outputs()[0].name` so you can build a correctly-keyed input dict.
4. Implement a POST endpoint that parses the request body into a numpy array of the right dtype/shape and calls `session.run(output_names_or_None, {input_name: input_array})`, returning the resulting array(s) as JSON.
5. Select an execution provider explicitly when it matters, e.g. `onnxruntime.InferenceSession(path, providers=["CPUExecutionProvider"])`, since the provider determines which backend kernels run the graph.
6. Run the app with uvicorn and validate it via curl or the framework's interactive API docs before containerizing for deployment.