[Build] Can NPU enablement be optional?

[mc-nv]

Describe the issue

I see with this change compilation flag is persistent.
https://github.com/microsoft/onnxruntime/pull/22027/files

Is it possible to make it optional?

Urgency

No response

Target platform

Linux

Build script

N/A

Error / output

Compilation is not failing, but the models generation and environment configuration requires specific/undocumented changes.

Visual Studio Version

No response

GCC / Compiler Version

No response

[mc-nv]

cc: @snnn @metang

[github-actions]

This issue has been automatically marked as stale due to inactivity and will be closed in 30 days if no further activity occurs. If further support is needed, please provide an update and/or more details.

[yinggeh]

Hi @snnn @metang. Please advise.

[jywu-msft]

@mc-nv can you provide details on the build options and what the resulting error(s) are?

[mc-nv]

Environment details:

BUILDTOOLS_VERSION:17.12.35506.116 
CMAKE_VERSION:3.30.5 
CUDA_VERSION:12.6.3 
CUDNN_VERSION:9.6.0.74 
PYTHON_VERSION:3.12.3 
TENSORRT_VERSION:10.7.0.23 
VCPGK_VERSION:2024.03.19

Build details:

git clone --recursive -b rel-1.20.1 https://github.com/microsoft/onnxruntime.git

onnxruntime/tools/ci_build/build.py `
   --cmake_generator "Visual Studio 17 2022" `
   --config Release `
   --cmake_extra_defines "CMAKE_CUDA_ARCHITECTURES=75;80;86;90;100;120" `
   --skip_submodule_sync `
   --parallel `
   --build_shared_lib `
   --compile_no_warning_as_error `
   --skip_tests `
   --update `
   --build_dir /workspace/build `
   --use_cuda `
   --cuda_home ${env:CUDA_PATH} `
   --cudnn_home ${env:CUDA_PATH} `
   --use_tensorrt --tensorrt_home "/tensorrt" ; 

[jywu-msft]

I was told this was fixed in ORT 1.21
+@preetha-intel

[mc-nv]

Let me check it, thank you.

[mc-nv]

Seems there was no changes:
https://github.com/microsoft/onnxruntime/blob/main/cmake/onnxruntime_providers_openvino.cmake#L20-L22

 if(OpenVINO_VERSION VERSION_GREATER_EQUAL 2024.4)
    add_definitions(-DUSE_OVEP_NPU_MEMORY=1)
  endif()

Value remain default.

[jywu-msft]

@preetha-intel fyi

[preetha-intel]

The check is handled internally at runtime. NPU enablement is optional

[mc-nv]

Hi @preetha-intel
I have submitted review to handle the configuration issue we observe.

#24557

[preetha-intel]

Hi @mc-nv The check for NPU is handled interenally here. The flag is basically set to ensure that from the specified version of OV the NPU allocators are available.
Can you share the specific/undocumented changes that are required so that we can check and get back