Themis.Inference.MIGraphX 2.0.1

Themis.Inference.MIGraphX

Themis.Inference.MIGraphX adds optional AMD MIGraphX plugin registration and GPU selection for Themis ONNX inference. It reuses Themis' existing ONNX model bundles and typed decision contracts; it does not add a model format or output mapping.

The package does not include ROCm, MIGraphX, or native provider binaries. Supply an absolute path to the MIGraphX plugin library when registering the provider:

var pluginPath = Environment.GetEnvironmentVariable("THEMIS_MIGRAPHX_EP_LIBRARY")
    ?? throw new InvalidOperationException("Set THEMIS_MIGRAPHX_EP_LIBRARY to the MIGraphX plugin path.");

services.AddThemis(options => options.Model = "laya-production");
services.AddThemisMIGraphX(pluginPath);

The default configuration requires MIGraphXExecutionProvider on a GPU and fails instead of silently using CPU. Pass an InferenceRuntimeConfiguration with PreferAccelerator and no RequiredExecutionProviderName to allow the normal runtime-default fallback when MIGraphX is unavailable. ONNX Runtime keeps registered plugin libraries alive for its environment; the plugin is not unloaded while the process is running.

MIGraphX is a Linux provider. The plugin library must implement ONNX Runtime's plugin execution-provider ABI and be built against the same ONNX Runtime version as the managed runtime used by Themis (currently 1.31.0). Its ROCm and MIGraphX shared-library dependencies must be available to the Linux loader. This is separate from Windows ML's MIGraphX distribution. This package does not bundle those native assets; a compatible, provenance- reviewed native build is required. The published InferenceEngine.ROCm.Runtime.linux-x64 package uses ONNX Runtime 1.24.1 and is not compatible with Themis' 1.31.0 managed runtime, so this integration does not reference it.

Build or obtain the plugin from a trusted source, review its source, build configuration, license, and native dependencies, then publish the consumer with linux-x64. A registered provider or successful inference call alone does not prove that model nodes ran on the AMD GPU. The repository's MIGraphX hardware test checks typed answers and ONNX Runtime profiling when a compatible GPU/plugin pair is available; this repository has not validated that hardware path.

References:

• ONNX Runtime plugin execution-provider usage
• AMD ONNX Runtime plugin execution providers
• ONNX Runtime MIGraphX requirements