detect_ort_provider() only tested for CUDAExecutionProvider inside
#ifdef SAE_ORT_WITH_TRT_EP, so any build that did not also opt into the
TensorRT execution provider could never select CUDA and fell straight
through to the CPU. The two are independent: the TRT EP needs the
headers and profile plumbing and is rightly an opt-in, CUDA is a plain
ORT provider and is not.
The failure is silent rather than loud, which is why it survived --
inference runs on the CPU and every answer is still correct, just far
slower. Measured on the VR-012 study: 0.32 s/crop against 0.0021 s/crop
once a GPU backend is actually used, with the card sitting at 212 MiB
and 0% utilisation throughout.
Only the TensorrtExecutionProvider line stays inside the guard.
dtourolle