Add NPU condition for "cmake/onnxruntime_providers_openvino.cmake"

[mc-nv]

Description

Make NPU flag optional in cmake/onnxruntime_providers_openvino.cmake

Motivation and Context

Motivation for this issue is to make NPU flag optional
#22985

[snnn]

In https://github.com/microsoft/onnxruntime/blob/main/cmake/onnxruntime_providers.cmake we have

if (onnxruntime_USE_OPENVINO)
  include(onnxruntime_providers_openvino.cmake)
endif()

So how does it affect you?

[jywu-msft]

@preetha-intel @sfatimar does this make sense?

[sfatimar]

onnxruntime_USE_OPENVINO_NPU is only enabled when we build ORT with OpenVINO and default device type as NPU which is not always the case. Even when we build with CPU we should be able to use NPU and NPU Device memory as a runtime option

[jywu-msft]

@mc-nv, I think what @sfatimar is saying is that -DUSE_OVEP_NPU_MEMORY=1 should always be set, whether or not the system has an NPU. onnxruntime_USE_OPENVINO_NPU just builds with NPU as default device type, but we want to allow NPU for even CPU default device type.
The issue you previously encountered should have been fixed. you shouldn't be seeing any runtime errors on systems that don't have an NPU. Are you seeing differently?

[mc-nv]

Thank you for the explanation.
Let me move the PR in draft till I finish my internal testing.

[preetha-intel]

Hi @mc-nv The NPU allocator will be used only when NPU HW is selected. Refer here. The flag is basically set to ensure that from the specified version of OV the NPU allocators are available.

[jywu-msft]

In https://github.com/microsoft/onnxruntime/blob/main/cmake/onnxruntime_providers.cmake we have

if (onnxruntime_USE_OPENVINO)
  include(onnxruntime_providers_openvino.cmake)
endif()

So how does it affect you?

I think they enable multiple EP's for triton inference server onnxruntime backend including OpenVINO EP.
https://github.com/triton-inference-server/onnxruntime_backend/blob/f1fec69b8668a59d923719a4da6560113a5f0584/CMakeLists.txt#L92