Update OpenVINO NPU dynamic external-initializer expectation

[GopalakrishnanN]

Description

Updates the OpenVINO external-initializer dynamic-shape test to account for NPU behavior.

OVEP_ExtInit_DynamicEmbed_Tests.ModelWithDynamicShapeEmbedsWeights previously expected session construction to throw for every device when a dynamic-shape model with >2GB external weights is loaded from memory. That still holds for CPU/GPU, where dynamic backend initialization happens at session construction and the weights are embedded into the model proto, hitting protobuf's 2GB limit.

For NPU, backend creation for this dynamic-shape case is deferred at session initialization. As a result, the external weights are not embedded into the model proto at this point and session construction succeeds. The test now keeps the existing EXPECT_THROW for CPU/GPU and uses EXPECT_NO_THROW for NPU.

Testing

Installed OpenVINO 2026.2.1 into the repo .venv and built an OpenVINO NPU provider-test binary:

.venv\Scripts\python.exe tools\ci_build\build.py `
  --config Release `
  --build_dir build\Windows_OpenVINO_NPU_2026 `
  --parallel `
  --build `
  --use_openvino NPU `
  --build_shared_lib `
  --cmake_generator "Visual Studio 18 2026" `
  --cmake_extra_defines CMAKE_PREFIX_PATH="C:\Users\gnallasamy\Codes\onnxruntime\.venv\Lib\site-packages\openvino\cmake" `
  --target onnxruntime_provider_test

Build result: onnxruntime_provider_test.exe built successfully with onnxruntime_providers_openvino.dll.

Ran all CPU/GPU/NPU parameters for the updated test:

.\onnxruntime_provider_test.exe --gtest_filter="OVEP_DynamicEmbed_Tests/OVEP_ExtInit_DynamicEmbed_Tests.ModelWithDynamicShapeEmbedsWeights/*"

Result:

[==========] 3 tests from 1 test suite ran.
[  PASSED  ] 3 tests.

Also reran the self-contained NPU smoke test:

.\onnxruntime_provider_test.exe --gtest_filter="OVEP_Tests/OVEP_ExtInit_Tests.ModelFromExtInit/2"

Result:

[  PASSED  ] 1 test.

[Copilot]

Pull request overview

Updates the OpenVINO EP external-initializer dynamic-shape unit test expectations to reflect that NPU handles dynamic-shape backend creation differently than CPU/GPU during session construction.

Changes:

• Split the session-construction expectation by device: keep EXPECT_THROW for CPU/GPU and use EXPECT_NO_THROW for NPU.
• Add an explanatory assertion message describing the NPU-specific behavior.

[github-actions]

You can commit the suggested changes from lintrunner.