Ultralytics v8.4.171: AMD GPU Support and More Reliable Workflows
Ultralytics YOLO v8.4.171 expands hardware support with AMD ROCm and MIGraphX, bringing AMD GPU acceleration to training and ONNX inference. This release also improves dataset caching, fixes RT-DETR training behavior, and streamlines model export.
The focus is broader hardware compatibility and smoother workflows—no new model architecture is introduced.
New Features: AMD GPU Support
PR #24137 by @itikhono adds AMD support across training and deployment:
- Train and predict with ROCm: Run native PyTorch models on supported AMD GPUs.
- Accelerate ONNX inference with MIGraphX: Execution-provider selection is automatic on supported systems, with CPU fallback when MIGraphX is unavailable.
- Start repeat inference sessions faster: Compiled MIGraphX models are cached to avoid recompilation on subsequent loads. The first run still requires compilation.
- Simplify deployment: A ROCm-based
latest-amdDocker image, dedicated setup guidance, and AMD GPU CI testing support the new workflow.
MIGraphX support targets Linux x86_64 with Python 3.11 or newer. Published Radeon 8060S benchmarks show matching accuracy and faster ONNX/MIGraphX inference than PyTorch ROCm in the tested configurations.
Bug Fixes and Workflow Improvements
- Dataset caching and task behavior: PR #26461 by @glenn-jocher addresses regressions identified during third-party PR audits, including image-cache reading across cache modes, archive timestamp preservation to avoid unnecessary label rescans, pose keypoint scaling in LiteRT/RKNN, and mask and analytics behavior.
- More reliable RT-DETR training: PR #26465 by @Nikhi00718 corrects contrastive denoising group indices, fixing denoising query matching during training.
- Lower temporary storage usage during export: PR #26466 by @amanharshx removes an unnecessary temporary weights copy from Core ML NMS export.
Documentation and Testing
- Apple M4 performance guidance: PR #26459 by @onuralpszr adds Core ML and Core AI latency benchmarks to the Core AI documentation.
- Cleaner test logs: PR #26462 by @UltralyticsAssistant suppresses expected, benign tracer warnings in pytest, matching the package’s warning handling.
Community Contributions
Thanks to all contributors and the YOLO community for helping improve this release! A special welcome to @itikhono, whose first contribution brings the new AMD support.
Try v8.4.171
Upgrade to this release:
pip install --upgrade ultralytics==8.4.171
Explore the v8.4.171 release notes for release details, or review the full changelog to see all changes.
Give it a try and share your feedback below! If you’re testing AMD acceleration, include your GPU, operating system, Python version, and workload to help the community compare results.