Voyager SDK

I am currently working with metis M2 card. Trying to run yolov8 model using inference.py file from the sdk. The throughput is very low. I set the AIPU core to 4 and verfied the same in the compile_config.json file. But there is another argument called multicore_mode("multicore_mode": "batch"). what exactly does the multi core mode do? and is there way to set it to other option than ‘batch’?

Hi, I was trying to get the voyager sdk v1.6.0 running on the arduino portenta board but I cannot get it to work. The board came with v1.4.0 preinstalled, which does work. I followed the guide for creating the docker container and used the install script with the no-driver flag to install v1.6.0, but every time I run axdevice I get the folowing error:axdevice[libaxldev_linux.c:1030] AXL_IOCTL_GET_RESOURCE_INFO failed: Inappropriate ioctl for device[AxeleraDevice.cpp:75] Device not found: metis-0:1:0WARNING: Failed to open device 0(metis-0:1:0)[libaxldev_linux.c:1030] AXL_IOCTL_GET_RESOURCE_INFO failed: Inappropriate ioctl for device[AxeleraDevice.cpp:75] Device not found: metis-0:1:0ERROR: Failed to open device metis-0:1:0ERROR: AXR_ERROR_CONNECTION_ERROR: Failed to open device metis-0:1:0I thought it might be due to the kernel driver version outside the container but the latest version I

I'm trying to deploy yolov8m-worldv2 on the Metis M.2, but it fails due to an error.Does Metis support yolov8m-worldv2?Is it because it's a VLM model that Metis can't deploy it?I would like to use a zero-shot model.Thank you for your support.

Hello, I am having issues reproducing the Axelera Model Zoo benchmarks on a Axelera Metis M.2 module.My issues seem to be related to pre/post processing, which is why I started with a classification model.I was hoping to use “fakevideo” as input, to bypass the pre-processing, but that seems to be fixed at 30fps.Another unresolved issue I have is that opencl is not working on my AMD PC.For ResNet-50 v1.5, here is what I am getting … (venv) voyager-sdk$ ./inference.py resnet50-imagenet data/coco --disable-opencl --no-display --show-stats========================================================================           Element                                         Time(𝜇s)   Effective FPS========================================================================axinplace-addstreamid0                                16        61,066.3vaapipostproc0                                     1,927           518.9videoco

I created a custom model based on yolo26n, 'yolo26n-aya'.The performance results for source code (bitmap (256x256) images and mp4 video) are shown below.(a) images    (venv) ubuntu@antelao-3588:~/voyager-sdk$ AXELERA_USE_CL_DOUBLE_BUFFER=0 ./inference.py yolo26n-aya ./data/aya100/images --show-stats --no-display    ========================================================================    Element                                         Time(??s)   Effective FPS    ========================================================================    axinplace-addstreamid0                               141         7,076.4    inference-task0:libtransform_resize_cl_0             371         2,692.7    inference-task0:libtransform_padding_0               844         1,184.0    inference-task0:inference                          2,961           337.7    inference-task0:Inference latency                208,874             n/a

SummaryVoyager SDK 1.6.0 quantizer rejects ONNX models that have a fan-out fork close tothe input (e.g., a tensor consumed by multiple parallel paths early in the graph)with the error:axelera.compiler.exceptions.QuantizationError: Input model cannot be quantized because the core graph has several inputs that are not currently supported by Quantizer. 

Hello.I'm expecting my Metis Compute Board to arrive on April 24th.The Axelera Metis Compute Board User Guide - Issue 2lists voyager-sdk version 1.4.0.Will the user guide be updated for version 1.6?Also,while there are instructions for "how to work with different video sources,"how can I perform four different YOLO26 models with a single USB camera:Object DetectionPose EstimationSegmentationOriented Bounding BoxesThank you in advance.

Hi everyone,I'm integrating a single-object tracking model with the Voyager SDK and hitting a wall with multi-input quantization.Onnx Model opset :   17 Inputs:- template: [1, 3, 112, 112]- online_template: [1, 3, 112, 112]- search: [1, 3, 224, 224]I have calibration data ready as .bin files for each input.The problem: compiler.quantize() doesn't accept multi-input calibration data.Things I've tried:- Yielding a dict: {'template': arr, 'online_template': arr, 'search': arr} → "Failed to get input shape from calibration dataset"- Yielding a list: [template, online_template, search] → "'list' object has no attribute 'shape'"- Yielding a tuple → same error- transform_fn returning a list → same errorI found the September 2025 post where it was mentioned multi-input isn't supported and the concat+identity conv workaround. However, my inputs have different spatial sizes (112x112 vs 224x224) which mak

Hey, we ordered the Aetina RK3588 setup in order to spend minimum time to get it running. When working the numbers are quite impressive, but sadly I encountered three separate issues while testing the same RK3588 setup.They might be unrelated, but I grouped them here for context.Current setup:Aetina RK3588, with the most recent SDK directly installed on it (no docker, in order to reduce failure points and save storage).Axdevice: Device 0: metis-0:1:0 1GiB m2 flver=1.4.0 bcver=7.0 clock=800MHz(0-3:800MHz) mvm=0-3:100%Issue 1: Gstreamer + display not workingWorking: pipe=gst --no-display, pipe=torch-aipu (with display), pipe=torch (with display)Not working: pipe=gst (with display) e.g.(venv) aetina@aetina:~/voyager-sdk$ ./inference.py yolov5s-v7-coco media/traffic1_1080p.mp42026-03-25 13:35:42.378368936 [W:onnxruntime:Default, device_discovery.cc:164 DiscoverDevicesForPlatform] GPU device discovery fai

I am using Metis dev board with Arduino x8. When running./inference_llm.py llama-3-2-1b-1024-4core-static --prompt "Give me a joke"it always deploys the model first. Even when I run this command one by one several times it is always deploying the model again, even thought it was just deployed.Is there a way to deploy the model once and then just inference it?

Hello,I looked through the model zoo provided in the github repo for voyager-sdk and could not find support for all the yolo11 pose estimation models. Specifically for yolo11m-pose and yolo11x-pose. Is support for these two models something that is being considered in the (near) future? Is it possible to deploy these two models with the repo being as is?

Hello, I would like to use the Metis AIPU to perform micro benchmarks, so very simple operations on which i would have the maximum control. Therefore i was wondering if there is a way to read and write for example in the cores’ L1 memory “manually”? For the moment i couldn’t find any of these, the closest being the DmaBufAllocator but i would guess it is not directly related to the core itself? Thank you very much 

Hey everyone,I'm trying to boot Ubuntu 22.04 from a SATA SSD on a Metis Compute Board (RK3588) running the Axelera BSP.Setup Board: Metis Compute Board (RK3588) BSP: Axelera Voyager Linux 1.2.1 (Mender-managed A/B boot) Goal: Boot Ubuntu 22.04 from SATA SSD (/dev/sda1) What I've Done Built a full Ubuntu 22.04 arm64 rootfs on the SATA SSD Copied BSP kernel modules (6.1.148-rockchip-standard) into the rootfs Set correct fstab using UUID Verified systemd and libc are present and valid Used mender_post_setup_commands to override mender_kernel_root to /dev/sda1 after Mender setup The ProblemMender overwrites mender_kernel_root every boot with the eMMC p

Hello, I would like to run Voxtral-Mini for Speech-To-Text (STT) and qwen3 tts 0.6B for Text-To-Speech. (TTS)Qwen3 tts would probably fit into a M.2 Metis AI board, but voxtral is a 4B model, therefore it would probably need a M.2 Metia AI Max.I have a few questions to start:I find a plenty of examples running Metis with images, but audio is almost not existent. Do you support it? If yes, can you point me to the specific documentation, please? If I use the 16GB of RAM of M.2 Metis AI Max, can I have two models loaded and running in parallel on the board? Voice applications are very sensitive to latency, being able to load two models and keeping them ready to work would keep latency low

Hello, I want to share that I think I found a bug in https://software.axelera.ai/ui/native/axelera-apt-source/axelera-runtime/ubuntu24/axelera-runtime_1.5.3-1_arm64.deb (and potentially other packages with the same version). pkg-config returns the wrong location for the headers of axruntime:(venv) ottavio@instance-20220410-1128:~/Projects/voyager-sdk$ pkg-config --cflags axruntime-I/opt/axelera/runtime-1.5.3-1/(venv) ottavio@instance-20220410-1128:~/Projects/voyager-sdk$but it shoud be (venv) ottavio@instance-20220410-1128:~/Projects/voyager-sdk$ pkg-config --cflags axruntime-I/opt/axelera/runtime-1.5.3-1/include(venv) ottavio@instance-20220410-1128:~/Projects/v

Hello,I am trying to understand how to add support for Metis to my embedded system, which is based on Buildroot.So far, this is what I have:I need to compile a kernel supporting DMA-BUF (done). I need to compile the kernel drivers (this is what I am working on now). I need to have the axelera_runtime package in my system, because the final application will depend on it.The last point is still unclear to me. I expected to find it in the Voyager SDK repository, but it is not there. The SDK depends on runtime files installed by the installation script, which downloads them from https://software.axelera.ai/ui/packages.Could you please clarify whether the source code for axelera_runtime is available? Or is it closed source?Thank you, Ottavio

Hi Team,I’m working with the Voyager SDK (using docker container) and trying to build the axstreamer and gstaxstreamer components. The source files examples reference:#include <axruntime/axruntime.hpp> (examples/axruntime/axruntime_example.cpp)But I can’t find axruntime.hpp anywhere in the SDK.I’m planning to run inference using a webcam as the input on my mac, with a pipeline that includes ffmpeg and mediamtx for streaming.Is axruntime supposed to be part of the SDK repo, a separate dependency, or is there a script I missed?Thanks!Val.

Hi,I just finished setting up everything and now I’m trying to run an example pipeline using.\inference.py yolov7-640x480-coco cars_clip.mp4Building the model seems to work fine but when the inference starts, it either crashes after some images withDetecting... : 2%|█▋ | 37/1802 [00:01<00:39, 44.29frames/s][ERROR][axeCommandQueueExecuteCommandListsAsync]: Wait kernel failed with return code -1.terminate called after throwing an instance of 'std::runtime_error' what(): axr_run_model failed with Error at zeCommandQueueExecuteCommandLists(cmdqueue, n_cmdlists, cmdlists, nullptr): cmdqueue_run_cmdlists: 319: Exit with error code: 0x7FFFFFFE : ZE_RESULT_ERROR_UNKNOWNAborted (core dumped)or instantly withINFO : Could not exec vainfo: Command '['vainfo'

For anyone building custom embedded Linux images with Yocto, there's now an official BSP layer available for Metis hardware.It’s pretty niche stuff, but if you're working on embedded systems where you need a stripped-down, customised Linux image (think industrial deployments, kiosks, dedicated edge boxes), this is for you.What's in it:The layer includes recipes for the Metis PCIe driver and the udev rules. Add it to your Yocto build, include the packages in your image, and your custom Linux will recognise and talk to Metis hardware out of the box.Currently supports Yocto 5.0 (Scarthgap), with branches for Kirkstone and Dunfell if you're on older versions.One important thing to note:The VoyagerSDK itself still doesn't run natively on Yocto yet. You'll need to run it inside a Docker container on your target system. The kernel driver runs on the host and handles the hardware communication, but inference and SDK

Hello, I try to use a YOLO model on 8 camera streams in parallel on the Metis PCIe card. 20-30 fps per camera stream is all I need. With the 548 fps end-to-end stated for YOLOv8s in https://axelera.ai/metis-aipu-benchmarks, it should be possible to reach ~68 fps per camera. As a small test I wrote a python script, which creates a inference stream with 8 videos as input:from axelera.app import config, display, inf_tracersfrom axelera.app.stream import create_inference_streamdef run(window, stream): for frame_result in stream: window.show(frame_result.image, frame_result.meta, frame_result.stream_id) fps = stream.get_all_metrics()['end_to_end_fp

Hello everyone,I am new to Axelera and working with a Metis device (4 GiB, PCIe, Voyager SDK). Device: metis-0:1:0 Memory: 4GiB Clock: 800MHz SDK: voyager-sdk  I am trying to deploy the InsightFace model for face recognition and would appreciate guidance on model conversion, supported formats, and deployment steps.Thank you in advance,Akash

Hello,I am facing an issue with a Metis Compute Board where any inference command (or any command using the accelerator) immediately fails with the error:ERROR: No devices found This happens regardless of the model or the command I run.Yesterday, I managed to temporarily recover by removing and reinstalling the Docker container with the Voyager SDK, after which inference started working again. I thought this was because I did not use the metis in a month so maybe it needed to be reinstalled.However, today the same error appears again, even though I did not change anything in my setup.I have already tried rebooting the board, but the problem persists. At the moment, no inference works at all, and reinstalling the container no longer

Hello,I’m currently working with the Metis Compute Board and the Voyager SDK, and I had a question regarding development outside of the SDK.Is it possible to run a script like a custom inference or a similar script directly on the board without going through the Voyager SDK pipeline ?I’m mainly asking to better understand the level of low-level access available on the Metis (runtime, APIs, or tools outside the SDK).Thanks in advance for your help.Best regards,

Hello,Following the question I had on running a GEMM only model, I have settled for a model that has Id Conv → Flatten → Gemm . I was wondering if the Flatten operation is just rearranging some pointer values or if it is doing a memory operation (like copying the tensor or something like that)?Thank you very much,