Voyager SDK

Hello, I’m new to voyager SDK and everything related to AXELERA and i need help deploying a custom onnx model that has 2 inputs and 20+ outputs, all i want for now is to manage to run an inference with my own dataset/image and to save to disk the raw outputs of the model. The input is a pair of 2 files. I have data in yuv format and it’s saved in input_y.npy and input_uv.npy. Do you have any idea how can i do this or if it’s possible from the voyager 1.3 or do i need to add a split layer so i need just 1 input and only then i can deploy it?  Or to upgrade the voyager version to 1.4 

Does anyone have any suggestions on how to implement SAHI inference with Yolo via voyager-sdk to run on metis m2?

Our development team requires specialized expertise for Axelera AI platform implementation and edge AI acceleration optimization. We're seeking an experienced consultant with 5-7weeks availability to support us with

Hello everyone,I have been trying to compile the PRL Tracker model (see link below) for about a month using both Hailo and Vitis AI, but so far I have not been able to fully meet the requirements.I am now considering using the Metis Accelerator, but I am not sure whether I would run into the same issues.Has anyone here successfully deployed this model on the Metis Accelerator, or could you confirm if it is supported?Repository: https://github.com/vision4robotics/PRL-Track/tree/main Any guidance or shared experience would be greatly appreciated.Thanks in advance! 

Hello everyone, I am trying to use the Metis PCIe (4GB) in combination with the firefly ITX-3588J.Issue: I can not load the Metis kernel module (driver) into the Linux kernelSetup: I have the Metis PCIe with 4GB of ram  the Firefly ITX-3588J with 16GB of ram with a RockChip. OS: Ubuntu 22.04.4 LTS. I have a Gen3 Pcie with 4 lanes, details below```        LnkCap:    Port #0, Speed 8GT/s, Width x4, ASPM L0s L1, Exit Latency L0s <4us, L1 <16us            ClockPM- Surprise- LLActRep- BwNot- ASPMOptComp+``` So far:I installed the ubuntu OS as described here: https://support.axelera.ai/hc/en-us/articles/25556437653138-System-Imaging-Guide-Firefly-RK3588 I installed the latest 1.3.3 voyager sdk release.  i can see the metis board from lspci``` firefly@firefly:~$ lspci | grep accelerators01:00.0 Processing accelerators: Axelera AI Metis AIPU (rev 02)```however, i do not detect any drivers for the board, even after installing the sdk. Both the following commands do not return anything:```

Hi, I'm working on an RPI5 and I have the voyager SDK inside a container with ubuntu 22 and it works. I wanted to try running an inference on a stream using GStreamer. First, I tried to access the stream and I managed with the following command:GST_PLUGIN_PATH=/voyager-sdk/operators/lib AXELERA_DEVICE_DIR=../opt/axelera/device-1.2.5-1/omega LD_LIBRARY_PATH=/voyager-sdk/operators/lib:/opt/axelera/runtime-1.2.5-1/lib:$LD_LIBRARY_PATH gst-launch-1.0 rtspsrc location='rtsp://…..' latency=200 ! rtph264depay ! avdec_h264 ! videoconvert ! autovideosink Now I wanted to try the axinferencenet plugin but I'm having problems. I've tested various commands and formats, such as: GST_PLUGIN_PATH=/voyager-sdk/operators/lib AXELERA_DEVICE_DIR=../opt/axelera/device-1.2.5-1/omega LD_LIBRARY_PATH=/voyager-sdk/operators/lib:/opt/axelera/runtime-1.2.5-1/lib:$LD_LIBRARY_PATH gst-launch-1.0 rtspsrc location='rtsp://...' latency=200 ! rtph264depay ! avdec_h264 ! videoconvert ! video/x-raw,format=BGRA !  axtran

Hi everyone,I'm currently developing a web application on top of the Axelera PCIe card and Voyager SDK, and I’ve hit a roadblock.Use Case Overview: I want to relay this annotated stream back to my web frontend for live viewing. I’ve tried: Using GStreamer pipelines for HLS (e.g., with hlssink) Embedding video using <video> tag on the frontend pointing to HLS or .mp4 System Details: Hardware: Axelera Metis PCIe Card Software: Voyager SDK (latest) YOLOv8n-based custom model Python 3.10 + GStreamer on Ubuntu 22.04

Hi,Recently I tried running some ROI tracking algorithms (the ones included with opencv, such as CSRT, MOSSE, etc) on an embedded ARM board and found that the tracking performance and latency is not as good as the application needs. One reason is that all these algorithms run on the CPU, which explains why they are not the fastest. Searching around, I found the VOT (Visual object tracking) challenge related to visual object tracking and it seems that most of the state-of-art trackers now use AI/ML. I guess that’s a good thing since in theory it should be easier to run these on dedicated AI accelerator hardware? Anyway, I want to try running a visual tracking algorithm on the Metis hardware to see how fast and well it can work. However, there are so many algorithms and I reckon the Metis may not support all the operators used in those networks? So before I embark on this task, i wanted to ask the Axelera experts for some guidance on what to go for and what to avoid in order to find a tr

Hi everyone,I’m new to working with the Axelera system. I see that there is a CenterCrop style preprocessing function, but in my use case I need more control. Specifically, I’d like to crop an image (for RTSP streams or training data) by specifying the x-offset, y-offset, width, and height of the region of interest.Is there a built-in preprocessing function in the Axelera SDK/YAML configs for this kind of custom crop? Or is the recommended way to handle this via GStreamer (e.g., videocrop/videobox) before passing frames to axinferencenet?Any help would be greatly appreciated.Thank you! 🙏

Hello.I installed the SDK on Ubuntu 22.04, butit stopped with the following error:[172/194] Install libgstreamer-plugins-bad1.0-dev[172/194] E: Error, pkgProblemResolver::Resolve stopped. This is probably due to a non-modifiable package.ERROR: Failed to install libgstreamer-plugins-bad1.0-devI would appreciate any advice.

Hello,I’m deploying a model onnx (opset 17) and got a large bunch of warnings of unsupported operations. One of them is the unsupported Transpose:WARNING : Configuration of node '/backbone/3/0/Transpose' may not be supported. 'Transpose' parameters:WARNING : perm: [0, 3, 1, 2]WARNING : data: Metadata(shape=(1, 48, 64, 160), is_constant=False)WARNING : transposed: Metadata(shape=(1, 160, 48, 64), is_constant=False)WARNING : Unsatisfied constraint: perm == [0, 1, 2, 3]I guess the rule should have been perm != [0, 1, 2, 3] because here it is requesting a non no-op operation perm = [0, 3, 1, 2].I think this is also causing later the errorERROR : RuntimeError: The expanded size of the tensor (192) must match the existing size (160) at non-singleton dimension 1. Target sizes: [1, 192, 256, 160]. Tensor sizes: [1, 160, 1, 1]as it looks that indeed the transpose has not been applied.Could someone shade some light on this behavior?Thanks in advance 

Hello everyone,I'm new to the Axelera AI system and am currently working with a Metis PCIe card with 4 GB memory.I want to understand how much of this memory is being used by the neural network model I am deploying. In a way, I'm looking for a tool or command that provides a report similar to what nvidia-smi does for NVIDIA GPUs, but for the Metis core.From my research, I understand that the Metis core's DRAM is a fixed, dedicated resource and that the memory usage is determined at the time of compilation, not at runtime. I've heard that the Voyager SDK's compiler logs might contain this information.My main question is: Where in the compiler output or log files can I find the exact memory footprint and percentage utilization for my model? Is there a specific argument or command I need to use to generate this report?Any guidance on this would be greatly appreciated.Thank you!

The release notes for the v1.4 SDK state;“Compiled models are expected to run on Intel Core-i 12th and 13th generations (x86), AMD Ryzen (x86) and ARM64 host CPUs”.It would be good to support older generation Intel CPUs, so that servers from 5 or more years ago can be repurposed as AI edge accelerators.

Hi,I tried the Yolo11-coco-onnx models from Model Zoo and got terrible results. Can anyone give me some advice? ./inference.py yolo11n-coco-onnx dataset --no-display --show-device-fpsINFO    : Model:      yolo11n-coco-onnx                                         INFO    : Dataset:    CocoDataset-COCO2017INFO    : Date:       2025-08-21 12:33:18.028455INFO    : Inference Time: 67178.89msINFO    : Evaluation Time: 36222.56msINFO    : Evaluation Metrics:INFO    : =========================INFO    : | mAP_box       | 0.03% |INFO    : | mAP50_box     | 0.05% |INFO    : | precision_box | 0.08% |INFO    : | recall_box    | 1.70% |INFO    : =========================INFO    : Key Metric (mAP_box): 0.03%INFO    : Core Temp : 49.0°C                                                    INFO    : Metis : 804.0fpsINFO    : CPU % : 31.4%INFO    : End-to-end : 74.7fps./inference.py yolo11s-coco-onnx dataset --no-display --show-device-fpsINFO    : Model:      yolo11s-coco-onnx                            

 SDK v1.3:14:32:45 [INFO] Dump used CLI arguments to: /repo/build/cli_args.json 14:32:45 [INFO] Dump used compiler configuration to: /repo/build/conf.json 14:32:45 [INFO] Input model has static input shape(s): ((1, 1, 2, 2),). Use it for quantization. 14:32:45 [INFO] Data layout of the input model: NCHW 14:32:45 [INFO] Using dataset of size 100 for calibration. 14:32:45 [INFO] In case of compilation failures, turn on 'save_error_artifact' and share the archive with Axelera AI. 14:32:45 [INFO] Quantizing '' using QToolsV2. 14:32:46 [INFO] ONNX model validation can be turned off by setting 'validate_operators' to 'False'. 14:32:46 [INFO] Checking ONNX model compatibility with the constraints of opset 17. Calibrating... ✨✨✨✨✨✨✨✨✨✨✨✨✨✨✨✨✨✨✨✨ | 100% | 390.54it/s | 100it | 14:32:46 [INFO] Exporting '' using GraphExporterV2. 14:32:47 [INFO] Quantization finished. 14:32:47 [INFO] Quantization took: 1.57 seconds. 14:32:47 [INFO] Export quantized model manifest to JSON file: /repo/build/quantize

I’m doing some testing with the PCI-Express based AI card. After some initial hiccups due to motherboard compatibility which required me to swap place for GPU and AI-card and then exposing the card into a ubuntu-22.04 docker container it’s working fine. I’m impressed by the performance of the card compared to using --pipe torch and having a CUDA based backend for the benchmark problems identifying traffic. I want to do inference with not only custom weights, but also custom nets and see how it performs. My wish is to have a sample that uses numpy f32 arrays as input and output and does inference on the card given a model from a ONNX file. To keep it simple, I made a PyTorch net [3, 64][ Relu ][64, 64][ Relu ][64, 1] and trained iton a mathematical function f(x1, x2, x3) = y.This toy net was exported to ONNX. From there I wanted to test inference on the AI card before moving to more complex things.I could not see that my wanted usecase was compatible with `deploy.py` and the required it

Hi,My goal is to try out a few tutorials with the latest SDK release. Looks like a python package is not found, any clues what I could try?System: Axelere M.2 Card with Aetina Eval SystemVoyager SDK v1.3.1 - a fresh install of the SDK this timeFirst installed the latest driver https://software.axelera.ai/artifactory/axelera-apt-source/metis-dkms/metis-dkms_1.0.2_all.debthen I ran the installer with./install.sh --all --media --user <email addr> --token <token>I got a few suspicious warnings and hit ‘y’ a number of times. Then install of 186 packages completed.[185/186] Install gstreamer1.0-rtspbuilding operatorsrefreshing pcie and firmware0000:01:00.0 : DeviceDevice 0: metis-0:1:0 1GiB m2 flver=1.2.0-rc2 bcver=1.0 clock=800MHz(0-3:800MHz) mvm=0-3:100%Installation complete, but with unresolved issues (see above) firefly@aetina:~/Documents/g2-testing/voyager-sdk$ source venv/bin/activate I proceeded to try inference.py with one of the included videos as input(venv) firefly@aet

Hi,I was wondering if there exist an astral uv (pip alternative) build for the voyager-sdk?This would speed up setup of the sdk and allow easier merging with our development environment.Thanks in advance!

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 get an error when I want to deploy the WASB model for sports ball detection and tracking, converted to onnx with opset 15:ERROR   : Traceback (most recent call last):ERROR   :   File "<frozen compiler.top_level>", line 584, in quantizeERROR   :   File "<frozen qtools_tvm_interface.graph_exporter_v2.graph_exporter>", line 254, in exportERROR   :   File "<frozen qtools_tvm_interface.graph_exporter_v2.graph_exporter>", line 209, in _convert_operatorsERROR   :   File "<frozen qtools_tvm_interface.graph_exporter_v2.replacement_functions>", line 242, in closureERROR   : RuntimeError: Multiple consecutive arithmetic operations by runtime ops found in the graph. This is not allowed presently.ERROR   : ERROR   : The above exception was the direct cause of the following exception:ERROR   : ERROR   : Traceback (most recent call last):ERROR   :   File "/home/aetina/Desktop/voyager-sdk/axelera/app/compile.py", line 456, in compileERROR   :     the_manifest = top_level.c

Hi,When trying to deploy yolo11 with custom weights, I noticed a significant performance drop due to the model conversion to onnx. Can I ask what method and parameters were used to convert the axelera yolo11s-coco-onnx model?Thank you!

Hi ​Axelera  team,I’m running parallel and cascading pipelines together with inference.py—the pipelines initialize and generally work. The issue is input routing: during inference I’m unable to reliably select/bind the correct video (file/RTSP/USB) to each network. For example, launching two independent detectors plus one cascade with multiple media sources (e.g., ./inference.py modelA modelB media/a.mp4 media/b.mp4 --pipe=gst --verbose) often results in the streams attaching to the wrong model or failing to open, while one stream runs as expected.Is there a supported way to explicitly map inputs to models/pipelines when using inference.py? For instance: CLI flags to pin inputs by index/name (e.g., --input[0]=..., --input[1]=...) YAML fields that define per-network input URIs A recommended PipeManager pattern for deterministic source‑to‑network routing when combining parallel and cascading graphs   the command i wrote:./inference.py parellel5_yv11pose_scpmyv8_bcfyv8_yv11_ppeyv8 med

We’re excited to announce that Axelera is now fully supported in the DeGirum Cloud Compiler, our browser-based tool that lets you compile YOLO models in just a few clicks. No toolchains, no local setup, no guesswork. Just upload your PyTorch checkpoint and receive an Axelera-optimized binary, often in under 10 minutes.Note: A free DeGirum AI Hub account is required to access the compiler and test your models directly in the browser.✅ What It Supports YOLOv8 and YOLO11 All model sizes: n, s, m, l, x Tasks: Object Detection Classification Semantic Segmentation Oriented Bounding Box Detection Keypoint Estimation Custom input resolutions during compilation Integrated C++ postprocessing via DeGirum PySDK Instant in-browser testing — no download required Fast compilation — typically under 10 minutes 🎯 Who This Is ForThis tool is built for developers and researchers who want to: Skip manual toolchain setup Quickly compile PyTorch YOLO models to run on Ax

Hi Axelera Team, I'm exploring the Axelera Metis AI device for edge deployment and had a quick question: Is it possible to run our own custom Python scripts or shell commands on the device Or is the device strictly limited to running AI model inference pipelines through your SDK, without the ability to execute external scripts or commands? Looking forward to your guidance. Thanks in advance!