LightStudio.Onnx 1.30.0.3

LightStudio.Onnx

ONNX Runtime 1.30.0 with the complete upstream .NET managed API, embedded Dawn/WebGPU on Linux, and CoreML on macOS. Add only LightStudio.Onnx; do not also reference Microsoft.ML.OnnxRuntime, .Managed, or .EP.WebGpu.

On Linux, enable WebGPU explicitly:

using Microsoft.ML.OnnxRuntime;

using var options = new SessionOptions();
options.AppendExecutionProvider("WebGPU", new Dictionary<string, string>
{
    ["preferredLayout"] = "NHWC"
});
using var session = new InferenceSession("model.onnx", options);

The standard namespaces, InferenceSession, DenseTensor<T>, OrtValue, I/O binding, metadata, and profiling APIs are preserved. Accelerator selection is explicit; an unconfigured session uses CPU. Unsupported operators may fall back to CPU. Use options.AddSessionConfigEntry("session.disable_cpu_ep_fallback", "1") when CPU fallback must be rejected. WebGPU graph capture is opt-in and requires static shapes. Native training, CUDA, and ROCm are not enabled.

Platform Providers

For macOS, use the generic CoreML API with the unchanged upstream desktop assemblies. The legacy AppendExecutionProvider_CoreML helper is conditional in upstream managed builds.

options.AppendExecutionProvider("CoreML", new Dictionary<string, string>
{
    ["ModelFormat"] = "MLProgram",
    ["MLComputeUnits"] = "ALL"
});

CoreML falls back to CPU for unsupported operators by default. Acceleration depends on device capabilities and operator support; enabling CoreML does not guarantee GPU or NPU execution.

Platforms

Dawn is embedded in the Linux runtime library. The macOS runtime includes CoreML integration using macOS system APIs. The OS, graphics driver, Vulkan loader (Linux WebGPU), and macOS system C++ runtime remain system prerequisites. Telemetry is disabled. Standard framework dependencies System.Memory and System.Numerics.Tensors remain normal NuGet dependencies.

Managed assets cover .NET Standard 2.0 and .NET 8+. Mobile platforms, musl RIDs, and static macOS linking are not included.

Browser Decision

browser-wasm is intentionally omitted, including both thread flavors. ORT's browser WebGPU implementation relies on asynchronous JavaScript and Asyncify/JSPI for adapter creation, device creation, and buffer readback. Its supported ORT Web API is asynchronous JavaScript, not the synchronous managed InferenceSession API. Shipping archives alone would not provide a supported threaded .NET bridge. This does not mean that WebGPU or ORT Web generally lack browser threading.

Sources And Licensing

Native sources are pinned to ONNX Runtime commit f2c39fe2f838cf35ce7da92824f5a5e3ee6e88a7. Dawn and other dependencies use its pinned dependency manifest. The official 1.30.0 managed assemblies are included unchanged, with no Microsoft ONNX Runtime package dependency in the final NuGet. Upstream and dependency notices are included under licenses/.

Build a native RID with bash scripts/build-onnx.sh <rid>, then pack after all four native outputs are present:

dotnet pack package/onnx/LightStudio.Onnx.csproj -c Release -o artifacts/packages

For local testing only, a deliberately partial prerelease can be built with -p:OnnxRuntimeIdentifiers=linux-x64 -p:OnnxAllowPartialPackage=true -p:PackageVersion=1.30.0-local.1. Such a package is not the all-platform release.

Use the pinned Linux Dockerfile for release builds; compiling directly on a newer distribution can raise the glibc requirement. The GitHub Actions workflow builds all four RIDs on onnx-v* tags or manual dispatch, validates native dependencies and .NET consumers, and uploads the complete package. Only manual dispatch with publish=true pushes to NuGet.org using NUGET_API_KEY.