In the professional context (apart of individual apps distributed by small creators / indiehackers) usually models are run using standardized runtimes in native

In the professional context (apart of individual apps distributed by small creators / indiehackers) usually models are run using standardized runtimes in native code (C++ usually), using runtimes TensorRT (for Nvidia Devices), onnxruntime (agnostic), etc.