New Release: Ultralytics v8.4.173

Discussion
, ,
1

:rocket: Ultralytics v8.4.173: Native AMD Xilinx Export and Leaner Docker Images

Ultralytics YOLO v8.4.173 adds native export for AMD Versal AI Edge Series Gen 2 NPUs, bringing a new deployment path to supported AMD edge hardware. This release also reduces Docker image sizes, expands PyTorch dependency support, and improves documentation.

:new_button: New Features: AMD Xilinx Export

The native AMD Xilinx export in PR #26519, contributed by @glenn-jocher, introduces:

  • A new xilinx export format: Export through AMD Quark to a quantized ONNX model and Vitis AI configuration. Accepted aliases include vitis, vitisai, and versal.
  • Mixed-precision deployment: Most of the model uses INT8 quantization, while the head remains in floating point for the Vitis AI compiler to run in BF16.
  • Validation before deployment: Validate the exported model on a CPU, then use AMD’s Vitis AI tools to compile it into a .rai model for the NPU.
  • Targeted hardware support: Native export supports Versal AI Edge Gen 2 devices, including the VEK385.

:pushpin: This is not a universal Xilinx export. Older Zynq, Kria, and first-generation Versal devices still require separate, device-specific AMD workflows.

The AMD Xilinx deployment guide added in PR #26517, also contributed by @glenn-jocher, covers export, deployment, hardware compatibility, and which operations may run on the NPU versus the CPU.

INT8 quantization can reduce deployment footprint, while a BF16 head helps preserve accuracy. Actual accuracy and speed depend on calibration data, compilation, and the target device.

:spouting_whale: Improvements: Smaller Images and Broader Dependency Support

  • Leaner Docker images: The Docker cleanup in PR #26514 by @glenn-jocher removes unused OpenCV GUI libraries, gnupg, and other unnecessary system packages. Reported root-filesystem savings range from about 184 MB to 305 MB, depending on the image. Users who need GUI display can install the GUI OpenCV build explicitly.
  • Broader PyTorch support: The dependency update in PR #26463 by @dependabot[bot] expands the updated torch requirement from >=2.12.0,<2.13.0 to >=2.12.0,<2.15.0.

:wrench: Documentation Fixes

Three updates from @glenn-jocher improve documentation usability:

:speech_balloon: Try It and Share Your Feedback

Install this release with:

pip install --upgrade ultralytics==8.4.173

Explore the v8.4.173 release notes for release details, or review the full changelog to see every change.

Testing on supported AMD hardware or using the slimmer Docker images? Share your results, questions, and feedback below. Thanks to the Ultralytics team and YOLO community for helping improve every release! :tada: