Physical model container, independent of task and neural architecture. AUTO chooses an available supported container; it never guesses a model architecture or silently converts a requested quantization.