In this section, you’ll explore how to run a complete inference test using the Scenario Runner from Arm’s ML SDK for Vulkan. You’ll also learn what’s provided on Arm’s Hugging Face page, including downloadable binaries and assets that demonstrate the ML extensions for Vulkan in action.
The SDK provides a collection of tools and runtime components that help you integrate neural networks into Vulkan-based applications. While the ML extensions for Vulkan (VK_ARM_data_graph and VK_ARM_tensors) define the runtime interface, the SDK provides a practical workflow for converting, packaging, and deploying ML models in real-time applications such as games.
| Component | Description | Usage Context | GitHub link |
|---|---|---|---|
| Model Converter | Converts TOSA IR into SPIR-V graphs and packages them into .vgf files for runtime execution. | Used in asset pipelines for model deployment | https://github.com/arm/ai-ml-sdk-model-converter |
| VGF Library | Lightweight runtime decoder for .vgf files containing graphs, constants, and shaders. | Integrate into game engine to load/use graphs | https://github.com/arm/ai-ml-sdk-vgf-library |
| Scenario Runner | Executes ML workloads declaratively using JSON-based scenario descriptions. | Ideal for rapid prototyping and validation | https://github.com/arm/ai-ml-sdk-scenario-runner |
| Emulation Layer | Vulkan layer that emulates data graph and tensor extensions using compute shaders. | For testing on devices without native ML extensions for Vulkan support | https://github.com/arm/ai-ml-emulation-layer-for-vulkan |
Visit the NSS model page on Hugging Face
The landing page contains a minimal example - a scenario - to run NSS with an actual frame. It contains a Windows-compatible Scenario Runner binary, the VGF model, and a single frame of input and expected output data. This allows you to run an end-to-end flow, and the landing page provides resources to explore the VGF model in more detail.
In the following section, you’ll explore how to debug and inspect the workloads in this Learning Path using RenderDoc.