[OpenVINO GPU] OpenVINO EP shouldn't override the "ACCURACY" precision to "FP32"

[mingmingtasd]

Describe the issue

There is a regression caused by https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/providers/openvino/openvino_provider_factory.cc#L109 from this PR: #23553. Now users can't set "ACCURACY" precision for OV GPU device, it will be overridden as "FP32" which is not as expected. "ACCURACY" precision is really useful, see details: https://docs.openvino.ai/2025/openvino-workflow/running-inference/optimize-inference/precision-control.html#execution-mode, it's not equal to "FP32".

To reproduce

You can set "precision" as "ACCURACY" but it will behavior as "FP32", like https://onnxruntime.ai/docs/execution-providers/OpenVINO-ExecutionProvider.html#cc-api-20:

std::unordered_map<std::string, std::string> options;
options["device_type"] = "GPU";
options["precision"] = "ACCURACY";
session_options.AppendExecutionProvider("OpenVINO", options);

Urgency

Very urgent, it impacts my performance testing

Platform

Windows

OS Version

win11

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

30c6825

ONNX Runtime API

C++

Architecture

X64

Execution Provider

OpenVINO

Execution Provider Library Version

ov 2024.6

[mingmingtasd]

By the way, is that possible to expose the ov::hint::execution_mode directly and let it override the precision?

[huningxin]

For WebNN use case, we need OpenVINO EP to select execution precision on GPU device by respecting the model data types. Say if an operator in the model operates on FP32 tensor, execute it in FP32 precision, if an operator operates on FP16 tensor, execute it in FP16 precision.

Previously , we set OpenVINO EP precision option to ACCURACY: https://onnxruntime.ai/docs/execution-providers/OpenVINO-ExecutionProvider.html#summary-of-options

Internally, regarding to ACCURACY option, BasicBackend::PopulateConfigValue() sets OpenVINO ov::hint::inference_precision to ov::element::undefined and set execution_mode to ov::hint::execution_mode::ExecutionMode::ACCURACY that works for us.

https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/providers/openvino/backends/basic_backend.cc#L170

However, the recent commit: a6ea57b, changed changes this behavior and sets to highest precision on a device, for GPU, it is FP32:

https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/providers/openvino/openvino_provider_factory.cc#L109

This causes WebNN FP16 model execution is slower than before.