Nuget package Microsoft.ML.OnnxRuntime.Gpu version >= 1.17.0 not working

[EinarBjorn]

Describe the issue

I've been developing ONNX-based applications for Windows in C++ using the Microsoft.ML.OnnxRuntime.Gpu NuGet package. However, it only works up to version 1.16.3.

In the version 1.17.0 changelog, I noticed the entry: “Split ONNX Runtime GPU NuGet package into two packages,” which likely refers to the Microsoft.ML.OnnxRuntime.Gpu.Windows and Microsoft.ML.OnnxRuntime.Gpu.Linux packages. Unfortunately, I can’t find any documentation on how to migrate from version 1.16.3 to 1.17.0 or newer.

When I install only Microsoft.ML.OnnxRuntime.Gpu, the project fails to build because the header files are missing from the NuGet package. When I install only Microsoft.ML.OnnxRuntime.Gpu.Windows, the project again fails to build because the header files aren’t found—though I can see them manually under \buildTransitive\native\include.

Installing both packages yields the same result as installing just Microsoft.ML.OnnxRuntime.Gpu.Windows.

In order to build, I must install Microsoft.ML.OnnxRuntime.Gpu, but doing so means I lose the CUDA provider—even if I also install the other two packages.

To reproduce

update Microsoft.ML.OnnxRuntime.Gpu from 1.16.3 to 1.20.1 for example

Urgency

No response

Platform

Windows

OS Version

Window 10, 22H2

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.170 and newer

ONNX Runtime API

C++

Architecture

X64

Execution Provider

CUDA

Execution Provider Library Version

No response