Cannot install Microsoft.AI.Foundry.Local.WinML

[MikeYeager]

I have a Windows 11 ARM64 machine with VS 2026.

I created a console test application to try the .WinML version of FoundryLocal. At first, it told me it only supported TargetFramework net9.0-windows10.0.26100 (my target framework was net10.0). It kind of makes sense it would only work with a windows specific target. net10.0-windows10.0.26100 didn't work, but net9.0-windows10.0.26100 seemed to work after restarting VS.

Next it complained that my Platform was set to Any CPU, so I created an ARM64 profile.

When I try to install Microsoft.AI.Foundry.Local.WinML NuGet package, it eventually fails with a message "A NuGet restore loop has been detected in project ... Further restores have been aborted. The project might be in a bad state. Solution Explorer shows it tried to install Microsoft.ML.OnnxRuntime.Gpu.Linux, but failed.

I tried installing that package directly, which looked like it worked, but a build produced the following error: "Assets file 'D:\Dropbox\AI\FoundryLocalWinML\FoundryLocalWinML\obj\project.assets.json' doesn't have a target for 'net9.0-windows10.0.26100/win-arm64'. Ensure that restore has run and that you have included 'net9.0-windows10.0.26100' in the TargetFrameworks for your project. You may also need to include 'win-arm64' in your project's RuntimeIdentifiers."

Following that, I got a NUGET error on build: "The following PackageReference item(s) do not have a version specified: Microsoft.ML.OnnxRuntime.Gpu.Linux"

My .csproj file looks like this:

<Project Sdk="Microsoft.NET.Sdk">

  <PropertyGroup>
    <OutputType>Exe</OutputType>
    <TargetFramework>net9.0-windows10.0.26100</TargetFramework>
    <ImplicitUsings>enable</ImplicitUsings>
    <Nullable>enable</Nullable>
    <Platforms>ARM64</Platforms>
  </PropertyGroup>

  <ItemGroup>
    <PackageReference Include="Microsoft.AI.Foundry.Local.WinML" Version="1.0.0" />
    <PackageReference Include="Microsoft.Extensions.Hosting" Version="10.0.6" />
    <PackageReference Include="Microsoft.Extensions.Logging.Console" Version="10.0.6" />
  </ItemGroup>

  <ItemGroup>
    <PackageReference Update="Microsoft.ML.OnnxRuntime.Gpu.Linux" Version="1.24.4" />
  </ItemGroup>

</Project>

Microsoft.AI.Foundry.Local package works fine, it's only the Microsoft.AI.Foundry.Local.WinML version I have issues with. I don't see a sample that addresses using this package.

I found a related issue that suggested adding a NuGet package source https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/ORT/nuget/v3/index.json. and installing Microsoft.ML.OnnxRuntime.Foundry. That didn't fix the issue. What's the secret combination of things I need to use my NPU?

Thanks in advance,
Mike

[dondemcsak]

I have the exact same issue. I've been trying to figure it out for a couple days.

This is an upstream bug in [Microsoft.ML.OnnxRuntime.Foundry itself — it unconditionally declares OnnxRuntime.Gpu.Linux as a dependency for all TFMs (including netstandard), so it exists in the graph regardless of platform. The Gpu.Linux package then lists a win-arm64 native DLL in its .targets file that was never actually shipped.

This is a package bug that affects the Microsoft sample too — they just don't build on win-arm64.

The [Microsoft.ML.OnnxRuntime.Gpu.Linux.props] file injects a None item for win-arm64\native\onnxruntime.dll unconditionally (without an Exists() guard, unlike the x64 entries). That None item is what's failing since it resolves to the gpu.linux package folder which has no Windows ARM64 files.

[dondemcsak]

Here is the workaround. In your project file include:

  <!--
    Workaround for upstream bug in Microsoft.ML.OnnxRuntime.Gpu.Linux.props:
    the .props file (imported transitively via OnnxRuntime.Foundry) injects a
    None/CopyToOutputDirectory item for win-arm64/native/onnxruntime.dll without
    an Exists() guard. The path resolves into the gpu.linux package folder which
    ships no Windows binaries, so MSBuild fails with MSB3030 on win-arm64.
    Override the transitive reference with ExcludeAssets so the .props is never
    imported. NU1504 (duplicate PackageReference) is suppressed because NuGet
    sees both our direct reference and the transitive resolution.
  -->
  <PropertyGroup>
    <NoWarn>$(NoWarn);NU1504</NoWarn>
  </PropertyGroup>
  <ItemGroup>
    <PackageReference Include="Microsoft.ML.OnnxRuntime.Gpu.Linux" Version="1.23.2">
      <ExcludeAssets>build;buildTransitive</ExcludeAssets>
    </PackageReference>
  </ItemGroup>

[MikeYeager]

It worked! Thank you. I assume we should submit the bug to the Onnx team?

Not sure why I'm getting an unknown chip model name, but I'm on to the next hurdle.

Debugging shows the registration succeeding after the DownloadAndRegisterEpsAsync() call, but then the exception is thrown when GetCatalogAsync() is called.

I should mention that the app continues to run, but it list and runs only the cpu versions of the models. Apparently it's falling back when the provider fails.