LinML.Catalog 0.5.1

LinML.Catalog

Execution provider catalog for ONNX Runtime on Linux.

LinML.Catalog is the Linux counterpart to the Windows ML (WinML) ExecutionProviderCatalog. It automates vendoring, SHA-256 integrity verification, dynamic registration, and session attachment for external plugin execution providers on Linux.

Supported Execution Providers

LinML.Catalog provides unified discovery, readiness probing, and session attachment across both built-in and external plugin execution providers:

Built-In Providers (ProviderSource.BuiltIn)

• CPU (CPUExecutionProvider): Baseline execution provider, always ready.
• CUDA (CUDAExecutionProvider): Probed via host runtime (e.g. Microsoft.ML.OnnxRuntime.Gpu) and verified before session append.
• Classic TensorRT (TensorrtExecutionProvider): Probed via host runtime and verified before session append.

Plugin Providers (ProviderSource.Plugin)

• OpenVINO Plugin EP (onnxruntime-ep-openvino 1.6.1): Pinned PyPI wheel automated vendoring, SHA-256 verification, and registration.
• TensorRT-RTX Plugin EP (onnxruntime-ep-nv-tensorrt-rtx 0.4.0): Pinned PyPI wheel automated vendoring, SHA-256 verification, SONAME symlink management, and registration (on WSL2 with GPU passthrough the catalog installs and registers the LinML-hosted patched build instead, default on; opt out with LINML_TRT_RTX_WSL2=0; TRT_RTX_CUSTOM_EP overrides both). In containers the install root is $HOME/.local/linml-deps unless Create(root) is used; mount a volume there to keep installs across runs.
• MIGraphX Plugin EP (onnxruntime-ep-amdgpu 2611.1): Pinned standalone plugin release archive vendoring, SHA-256 verification, and registration.

Installation

dotnet add package LinML.Catalog

Prerequisite: LinML.Catalog references Microsoft.ML.OnnxRuntime.Managed to avoid native library conflicts. The consuming host application must supply a native ONNX Runtime package (such as Microsoft.ML.OnnxRuntime.Gpu or Microsoft.ML.OnnxRuntime).

Usage

using LinML.Catalog;
using LinML.Contracts;
using Microsoft.ML.OnnxRuntime;

// 1. Get default catalog instance
IExecutionProviderCatalog catalog = ExecutionProviderCatalog.GetDefault();

// 2. Discover OpenVINO plugin
IExecutionProvider openVino = catalog.Find(ProviderId.OpenVino)!;

// 3. Ensure plugin binaries are ready (downloads & verifies SHA-256 checksum)
ProviderEnsureResult ensured = await openVino.EnsureReadyAsync(new EnsureOptions(AllowDownloads: true));

// 4. Configure SessionOptions with verified plugin EP
using var sessionOptions = new SessionOptions();
if (ensured.IsReady)
{
    ProviderAppendResult result = sessionOptions.AppendFromCatalog(openVino, [OrtHardwareDeviceType.CPU]);
}

// 5. Create ONNX Runtime session
using var session = new InferenceSession("model.onnx", sessionOptions);

Design Principles

• No Conflicting Binaries: References only Microsoft.ML.OnnxRuntime.Managed. The host application supplies the native ONNX Runtime (e.g. Microsoft.ML.OnnxRuntime.Gpu), avoiding duplicate libonnxruntime.so collisions.
• Integrity Verified: Plugin binaries are verified against strict cryptographic SHA-256 digests.
• Shared Registry: Shares the PluginLibraryRegistry with the LinML CLI.

Documentation & Repository

For more details, CLI tools, and architecture documentation, visit the LinML GitHub Repository.