Overview
Run Neural Graphics Workloads Natively on Vulkan
The ML SDK for Vulkan is a set of libraries and tools that enable developers to run machine learning workloads directly through the Vulkan API. Built on Vulkan ML extensions such as VK_ARM_data_graph and VK_ARM_tensors, it provides a portable, hardware-accelerated way to execute neural networks alongside graphics and compute pipelines.
Designed for graphics engines with neural rendering capabilities, the SDK offers practical workflows for model conversion, inference graph packaging, and efficient execution on Vulkan-capable hardware.
Explore SDK
Start Building With the SDK
| Component | Features | Developer Benefits | GitHub Link |
| ML SDK for Vulkan | Provides the tooling required to build the infrastructure to run neural networks. | Achieve real-time neural graphics via the ML for Vulkan extensions. | Access Repo |
| Model Converter | Converts TOSA IR into SPIR-V graphs and packages them into .vgf files for runtime execution. | Convert trained models into Vulkan-ready assets for deployment. | Access Repo |
| VGF Library | Lightweight runtime encoder and decoder for .vgf files containing graphs, constants, and shaders. | Load and execute packaged ML graphs at runtime in engines and apps. | Access Repo |
| Scenario Runner | Executes ML workloads declaratively using JSON-based scenario descriptions. | Rapidly prototype and validate ML workloads. | Access Repo |
| Emulation Layer | Vulkan layer that emulates data graph and tensor extensions using compute shaders. | Develop and test Vulkan ML workflows without native ML hardware support. | Access Repo |
Resources
Access Docs, Learning Paths, and Open Source Models
Releases
