pypi ultralytics 8.4.216
v8.4.216 - `ultralytics 8.4.216` ST Edge AI export for STM32 devices (#27291)

one hour ago

🌟 Summary

Ultralytics 8.4.216 adds export support for STM32 edge devices and improves training, inference reliability, and deployment guidance across the ecosystem.

📊 Key Changes

  • STM32 export with ST Edge AI: The new format="stedgeai" export creates ONNX models for ST Edge AI Core and STM32 devices, including STM32N6. It supports float and INT8 export, with normalized detection and keypoint coordinates to support whole-graph quantization. This is the release’s main feature and also enables Ultralytics Platform to offer the format.
  • Better training behavior on AMD GPUs: ROCm training now keeps the NCHW layout by default, avoiding a layout that was substantially slower on some Radeon hardware. AMD GPU metrics are also reported during ROCm training.
  • More reliable model use and exports: Fixes include support for YOLOE prompts and TorchScript with PyTorch 2.16, correct CPU handling for device="cpu:0", and restored top-1 accuracy for INT8 LiteRT and NXP classification exports.
  • Fixes for validation, masks, and video: Validation now keeps separate per-image metrics for same-named files in different folders and handles mixed relative and absolute paths. Segmentation masks remain correct when results are converted to another dtype, and saved video duration is preserved when the effective frame rate is below 1 FPS.
  • More robust image and solution handling: Local HEIC and AVIF images are supported in LLM requests; several image-reading paths now respect photo orientation. Heatmaps and tracking zones also handle changes in frame size more reliably.
  • Expanded practical guidance: The preprocessing guide now includes a runnable dataset audit and advice for addressing class imbalance. The model-configuration guide explains how to start a YOLO26 P2 model from standard pretrained weights, with the trade-off that the extra small-object output increases computation.
  • Updated deployment documentation and examples: Several integration guides—including Core ML, LiteRT, NCNN, RKNN, Ascend, and DEEPX—were reorganized with clearer export requirements, output details, and limitations. Python examples were consolidated into shared ONNX Runtime and OpenCV DNN examples covering multiple tasks.

🎯 Purpose & Impact

  • Bring YOLO to smaller, low-power hardware: STM32 users gain a direct export path for compiling models with ST’s tools, including an INT8 option suited to constrained devices.
  • Avoid slow or incorrect workflows: ROCm layout changes improve training performance on affected GPUs, while the bug fixes reduce errors and preserve outputs across validation, inference, and export.
  • Make deployment easier to plan: Clearer integration guides help users choose settings, understand exported outputs, and spot hardware or version requirements before deployment.
  • Help users catch dataset issues earlier: The audit makes it easier to find class imbalance, validation-only classes, and duplicate images before spending time on training.

What's Changed

  • Platform docs: rotate shortcut and download after rotation by @pderrenger in #27223
  • Read HEIC and AVIF image paths in LLM requests as YOLO predict does by @MohammadHijjawi97 in #27221
  • Restructure the Rockchip RKNN integration page to the integration template by @raimbekovm in #27226
  • Restructure the Tencent NCNN integration page to the integration template by @raimbekovm in #27229
  • Restructure the LiteRT integration page to the integration template by @raimbekovm in #27224
  • Keep per-image metrics for images that share a filename across subfolders by @raimbekovm in #27222
  • Restore top-1 accuracy of INT8 LiteRT and NXP classification exports by @raimbekovm in #27225
  • Add a dataset audit and class imbalance section to the preprocessing guide by @pderrenger in #27238
  • Embed live Platform predict panel on YOLO model pages by @Y-T-G in #27230
  • Preserve saved video duration below 1 FPS by @AhmetMert01 in #27234
  • Fix RT-DETR CoreML 8-bit export with non-finite anchor constants by @raimbekovm in #27245
  • Keep validation and training working when an image list mixes relative and absolute paths by @Vaishnavi220506 in #27243
  • Skip a full-frame copy when predicting on upright PIL images by @aswanth-07 in #27239
  • Keep segmentation masks correct when converting results to another dtype by @Nikhi00718 in #27232
  • Read photos upright in similarity search and multispectral conversion by @cainiao33 in #27235
  • Keep each call's prompts when one FastSAM model serves concurrent requests by @Nikhi00718 in #27233
  • Run every YOLO task from one ONNX Runtime and one OpenCV DNN Python example by @onuralpszr in #27242
  • Limit NCNN CPU threads to the CPUs available to the process on Linux by @raimbekovm in #27244
  • Restructure the Apple Core ML integration page to the integration template by @raimbekovm in #27240
  • Keep COCO conversion labels inside the labels directory for absolute and .. image paths by @Lornatang in #27248
  • Add ST Edge AI export for STM32 devices by @glenn-jocher in #27211
  • Date the Core ML compute-unit crash comment to macOS 26.5 by @glenn-jocher in #27249
  • Keep NCHW for AMD ROCm training by @laodouya in #27257
  • Load distillation teachers before model construction by @Laughing-q in #27259
  • Fix docs author link for WangQvQ by @raimbekovm in #27260
  • Keep parking zones drawn on phone photos aligned with ParkingManagement frames by @Yannis-A-D in #27251
  • Restructure the Huawei Ascend integration page and document CANN 9.2.1 setup by @raimbekovm in #27261
  • Keep YOLOE text prompts and TorchScript export and inference working on PyTorch 2.16 by @klin2024 in #27266
  • Cut tiled segmentation memory by keeping masks inside their boxes by @raimbekovm in #27265
  • Describe the iOS SDK confidence for nms=True Core ML models by version by @raimbekovm in #27256
  • Docs: train YOLO26 P2 from pretrained weights, current auto-annotation models, pose keypoint multi-select by @pderrenger in #27268
  • Reorganize integrations docs navigation and naming by @raimbekovm in #27164
  • Fit parking zone previews within both canvas dimensions by @Vaishnavi220506 in #27273
  • Read exported class names without loading the inference backend by @Lornatang in #27269
  • Recover sparse optical flow after frame size changes by @AhmetMert01 in #27270
  • Split DOTA train/val images above the Pillow pixel limit by @raimbekovm in #27267
  • Keep Comet ML and Hailo search terms in docs titles and descriptions by @glenn-jocher in #27276
  • Platform docs: image source records, .pt-only model uploads, chat attachments, current VLM leaders by @pderrenger in #27282
  • Stop the false "Skipping labels" warning when validating single_cls models by @MohammadHijjawi97 in #27280
  • Restructure the Qualcomm QNN integration page by @raimbekovm in #27284
  • Restructure the AMD Xilinx integration page around the export workflow by @raimbekovm in #27286
  • Show every epoch-0 training batch image on Platform for fast runs by @laodouya in #27279
  • Keep Heatmap and TrackZone running when the frame size changes by @Vaishnavi220506 in #27283
  • Run on CPU with device="cpu:0" instead of raising an invalid CUDA error by @MohammadHijjawi97 in #27275
  • Show AMD GPU metrics on Platform during ROCm training by @laodouya in #27281
  • Fix nightly broken-links check: absolute Platform predict URLs and ONNX Runtime links by @glenn-jocher in #27290
  • Restructure the DEEPX integration page by @raimbekovm in #27289
  • Score only the selected classes in COCO and LVIS faster-coco-eval mAP by @aswanth-07 in #27288
  • ultralytics 8.4.216 ST Edge AI export for STM32 devices by @glenn-jocher in #27291

New Contributors

Full Changelog: v8.4.215...v8.4.216

Don't miss a new ultralytics release

NewReleases is sending notifications on new releases.