v1.6: Release notes
The Voyager SDK makes it easy to build high-performance inferencing applications with Axelera AI Metis devices. The sections below provide links to code examples, tutorials and reference documentation.
This is a production-ready release of Voyager SDK. Software components and features that are in development are marked with one of the following maturity labels:
- Experimental: May change or be removed without notice; no support guarantees.
- Alpha: Usable but incomplete; breaking changes possible.
- Beta: Feature-complete but not fully stable; committed to developing this further in future releases.
| Document | Description |
|---|---|
Installation guide - original install.sh |
Explains how to set up the Voyager SDK repository and toolchain on your development system using install.sh |
| Python pip installation | Explains how to set up the Voyager SDK repository and toolchain on your development system using new standalone Python wheels |
| Quick start guide | Explains how to deploy and run your first model |
| Windows setup guide | Explains how to install Voyager SDK and run a model in Windows 11 |
| AxDevice manual | AxDevice is a tool that lists all Metis boards connected to your system and can configure their settings |
| Board firmware update guide | Explains how to update your board firmware (for customers with older boards who have received instructions) |
| Document | Description |
|---|---|
| Model zoo | Lists all models supported by this release of the Voyager SDK |
Deployment manual (deploy.py) |
Explains all options provided by the command-line deployment tool |
| Custom weights tutorial | Explains how to deploy a model using your own weights |
| Custom model tutorial | Explains how to deploy a custom model |
| Document | Description |
|---|---|
| Benchmarking guide | Explains how to measure end-to-end performance and accuracy |
Inferencing manual (inference.py) |
Explains all options provided by command-line interencing tool |
| Application integration tutorial (high level) | Explains how to integrate a YAML pipeline within your application |
| Application integration tutorial (low level) | Explains how to integrate an AxInferenceNet model within your application |
The Voyager SDK allows you to develop inferencing pipelines and end-user applications at different levels of abstraction.
| API | Description |
|---|---|
| InferenceStream (high level) | Library for directly reading pipeline image and inference metadata from within your application |
| AxInferenceNet (middle level) | C/C++ API reference for integrating model inferencing and pipeline construction directly within an application |
| AxRuntime (low level) | Python and C/C++ APIs for manually constructing, configuring and executing pipelines |
| GStreamer | Plugins for integrating Metis inferencing within a GStreamer pipeline |
The InferenceStream library is the easiest to use and enables most users to achieve the highest performance. The lower-level APIs enable expert users to integrate Metis within existing video streaming frameworks.
The Voyager SDK makes it easy to construct pipelines that combine multiple models in different ways. A number of end-to-end reference pipelines are provided, which you can use as templates for your own projects.
| Directory | Description |
|---|---|
/ax_models/reference/parallel |
Multiple pipelines running in parallel |
/ax_models/reference/cascade |
Cascaded pipelines in which the output of one model is input to a secondary model |
/ax_models/reference/cascade/with_tracker |
Cascaded pipelines in which the output of the first model is tracked prior to being input to a secondary model |
/ax_models/reference/image_preprocess |
Pipelines in which the camera input is first preprocessed prior to being used for inferencing |
A new Python-native API for building, running, and packaging ML inference pipelines. The entire pipeline, from model loading through post-processing and tracking, can be expressed as a composable Python expression.
- Composable operators:
op.seq()for sequential,op.par()for parallel,op.foreach()for cascade (per-object) processing. - Data routing:
op.select(i)to extract from tuples,op.pack()/op.unpack()for explicit tuple conversion. - 30+ operators across preprocessing, inference, postprocessing, filtering, tracking, and result types.
- Model loading:
op.load('model.axm')- hardware inference on AIPU.op.onnx_model('model.onnx')- CPU inference, no AIPU required.op.load('pipeline.axe')- portable pipeline package (new .axe format).
- Tracker integration:
op.tracker(algo='bytetrack')- supports ByteTrack, OC-SORT, SORT, TrackTrack. Full lifecycle states viareturn_all_states=True(new, tracked, lost, removed). - Pipeline optimizer: Automatic SIMD-accelerated fusion of operator chains (e.g., NchwToNhwc + Quant + Pad into single QuantizeTransposePad).
- Typed result objects:
DetectedObject,PoseObject,SegmentedObject,TrackedObject,Classificationwith protocol-based interfaces and.draw()visualization.
This section provides links to additional documentation available in the Voyager SDK repository.
| Document | Description |
|---|---|
| Advanced deployment tutorials | Advanced deployment options [experimental] |
| AxRunmodel manual | AxRunModel is a tool that can run deployed models on Metis hardware using different features available in the AxRuntime API (such as DMA buffers, double buffering and multiple cores) |
| Compiler CLI | Compiler Command Line Interface [beta] |
| Compiler API | Python Compiler API [experimental] |
| ONNX operator support | List of ONNX operators supported by the Axelera AI compiler |
| Thermal and Power Management Guide | Document detailing the thermal behavior and power management for Metis and instructions to make changes |
| SLM/LLM inference tutorial | Explains how to run Language Models on Metis devices [experimental] |
For blog posts, projects and technical support please visit Axelera AI Community.
For technical documents and guides please visit Customer Portal.