WebGPU Plugin EP — NuGet Packaging

This directory contains the C# NuGet package project and test app for the WebGPU plugin Execution Provider.

Directory Structure

csharp/
├── pack_nuget.py                                   # Helper script to build the NuGet package
├── Microsoft.ML.OnnxRuntime.EP.WebGpu/
│   ├── Microsoft.ML.OnnxRuntime.EP.WebGpu.csproj   # NuGet package project (netstandard2.0)
│   ├── WebGpuEp.cs                                 # Helper class for native library resolution
│   └── README.md                                   # Package readme (shipped inside .nupkg)
└── test/
    └── WebGpuEpNuGetTest/
        ├── WebGpuEpNuGetTest.csproj                # Test console app (net8.0)
        ├── Program.cs                              # Registers EP, runs inference, validates output
        ├── mul.onnx                                # Test model (element-wise multiply)
        └── generate_mul_model.py                   # Script to regenerate mul.onnx

Prerequisites

• .NET SDK 8.0 or later
• A built WebGPU plugin EP shared library

Building the NuGet Package

Use pack_nuget.py to stage native binaries and run dotnet pack. The script copies everything into a staging directory before building — the source tree is never modified. By default, an auto-cleaned temporary directory is used; pass --staging-dir to use an explicit one (required when running with --build-only or --pack-only).

At least one binary directory (or --artifacts-dir with matching subdirectories) must be provided. Platforms without a binary directory are skipped. Run python pack_nuget.py --help for the full list of options and their defaults.

Pack with a local build (single platform)

cd plugin-ep-webgpu/csharp

python pack_nuget.py --version 0.1.0-dev `
  --binary-dir-win-x64 <path-to-win-x64-binaries>

Pack multiple platforms

Each --binary-dir-* points at the directory containing that platform's already-built native binaries. In practice the four binaries are produced on different machines and combined in CI; locally you'd typically only set the one(s) you have available.

python pack_nuget.py --version 0.1.0-dev `
  --binary-dir-win-x64 <path-to-win-x64-binaries> `
  --binary-dir-win-arm64 <path-to-win-arm64-binaries> `
  --binary-dir-linux-x64 <path-to-linux-x64-binaries> `
  --binary-dir-macos-arm64 <path-to-macos-arm64-binaries>

Versioning

The package version is supplied to pack_nuget.py via --version. In the packaging pipeline, the release or pre-release version is derived from plugin-ep-webgpu/VERSION_NUMBER.

Testing the Package

The test app registers the WebGPU EP, creates a session, runs a simple Mul model, and validates the output.

# Point the test project's nuget.config at the pack output
$localFeed = (Resolve-Path nuget_output).Path
@"
<?xml version="1.0" encoding="utf-8"?>
<configuration>
  <packageSources>
    <clear />
    <add key="nuget.org" value="https://api.nuget.org/v3/index.json" />
    <add key="local" value="$localFeed" />
  </packageSources>
</configuration>
"@ | Set-Content test/WebGpuEpNuGetTest/nuget.config

# Build and run
dotnet run --project test/WebGpuEpNuGetTest/WebGpuEpNuGetTest.csproj --configuration Release

A successful run prints PASSED: All outputs match expected values. and exits with code 0.

Regenerating the Test Model

python test/WebGpuEpNuGetTest/generate_mul_model.py

Requires the onnx Python package.

CI Pipeline

The NuGet packaging is integrated into the WebGPU plugin pipeline:

• Pipeline: tools/ci_build/github/azure-pipelines/plugin-webgpu-pipeline.yml
• Packaging stage: tools/ci_build/github/azure-pipelines/stages/plugin-webgpu-nuget-packaging-stage.yml

The CI stage downloads build artifacts from all enabled platform stages, invokes pack_nuget.py, ESRP-signs the package, and runs the test app on a GPU agent.

Native Binaries Per Platform

On Windows, dxil.dll and dxcompiler.dll are the DirectX Shader Compiler binaries downloaded from the DXC GitHub releases. The CI pipeline handles this automatically.