Ultralytics v8.4.216: STM32 export, faster ROCm training, and more
Ultralytics v8.4.216 brings YOLO to STM32 devices with ST Edge AI export. This release also avoids a training slowdown on some AMD Radeon GPUs, fixes issues across validation and inference, and makes it easier to audit datasets and plan deployments.
New: ST Edge AI export for STM32
Use format="stedgeai" to export ONNX models for ST Edge AI Core and STM32 devices, including the STM32N6. The export supports floating-point and INT8 models, with normalized detection and keypoint coordinates to support whole-graph quantization. @glenn-jocher’s ST Edge AI export PR #27211 adds the export path and enables Ultralytics Platform to offer the format.
Training, inference, and export improvements
- AMD ROCm training: @laodouya’s NCHW update #27257 keeps NCHW as the default layout to avoid a slowdown on some Radeon GPUs. Their AMD GPU metrics update #27281 also brings ROCm GPU metrics to Ultralytics Platform.
- YOLOE and CPU compatibility: @klin2024’s first contribution, PR #27266, fixes YOLOE text prompts and TorchScript export and inference on PyTorch 2.16. @MohammadHijjawi97’s CPU device fix #27275 makes
device="cpu:0"run on CPU instead of raising a CUDA error. - Export accuracy: @raimbekovm’s INT8 export fix #27225 restores top-1 accuracy for LiteRT and NXP classification exports.
- Validation, masks, and video: @raimbekovm’s per-image metrics fix #27222 keeps results separate for same-named files in different folders. @Vaishnavi220506’s path fix #27243 lets training and validation handle lists that mix relative and absolute paths. @Nikhi00718’s mask fix #27232 preserves segmentation masks during dtype conversion, and @AhmetMert01’s video fix #27234 preserves saved video duration below 1 FPS.
- Images and video solutions: @MohammadHijjawi97’s image input update #27221 adds local HEIC and AVIF support to LLM requests. @cainiao33’s photo orientation fix #27235 reads photos upright in similarity search and multispectral conversion, and @Vaishnavi220506’s frame-size fix #27283 keeps Heatmap and TrackZone running when input dimensions change.
Better guidance for datasets and deployment
- Catch dataset issues before training. @pderrenger’s dataset audit guide update #27238 adds a runnable audit for class imbalance, validation-only classes, and duplicate images, plus practical advice on addressing imbalance.
- Start YOLO26 P2 from pretrained weights. @pderrenger’s YOLO26 P2 guide update #27268 explains how to start from standard pretrained weights and what to budget for: the extra small-object output increases computation.
- Plan exports with clearer requirements. @raimbekovm’s Core ML guide #27240, LiteRT guide #27224, NCNN guide #27229, RKNN guide #27226, Ascend guide #27261, and DEEPX guide #27289 clarify export requirements, outputs, and limitations. @onuralpszr’s Python examples update #27242 consolidates task-specific demos into one ONNX Runtime example and one OpenCV DNN example.
🔍 More changes and contributor PRs
- Training and evaluation: @Laughing-q’s distillation fix #27259 loads teachers before model construction; @MohammadHijjawi97’s validation fix #27280 removes a false warning for
single_clsmodels; and @aswanth-07’s evaluation fix #27288 scores only selected classes for COCO and LVIS mAP with faster-coco-eval. - Inference and exports: @Nikhi00718’s FastSAM fix #27233 keeps prompts separate across concurrent requests, and @aswanth-07’s PIL optimization #27239 avoids a full-frame copy for upright images. @raimbekovm’s RT-DETR export fix #27245 addresses non-finite anchor constants in Core ML 8-bit export, and their NCNN threading fix #27244 respects the CPUs available to the process on Linux. @Lornatang’s class-name update #27269 reads exported class names without loading the inference backend.
- Datasets and segmentation: @Lornatang’s COCO conversion fix #27248 keeps generated labels inside the labels directory. @raimbekovm’s DOTA update #27267 handles images above Pillow’s pixel limit, and their tiled segmentation optimization #27265 cuts memory use by keeping masks inside their boxes.
- Parking and tracking: @Yannis-A-D’s parking zone fix #27251 aligns zones drawn on phone photos with ParkingManagement frames. @Vaishnavi220506’s preview fix #27273 fits parking previews within both canvas dimensions, and @AhmetMert01’s optical flow fix #27270 recovers tracking after frame-size changes.
- Ultralytics Platform: @Y-T-G’s live predict panel #27230 brings interactive predictions to YOLO model pages, and @laodouya’s training preview fix #27279 shows every epoch-0 training batch image on fast runs. @pderrenger’s rotation docs #27223 cover rotation shortcuts and downloading rotated images, and their Platform docs update #27282 covers image sources,
.pt-only model uploads, chat attachments, and current VLM leaders. - Integration docs and maintenance: @raimbekovm’s navigation update #27164 reorganizes the integrations docs. Their Qualcomm QNN guide #27284, AMD Xilinx guide #27286, iOS SDK confidence notes #27256, and author-link correction #27260 round out the update. @glenn-jocher’s macOS 26.5 note #27249 dates a Core ML compute-unit crash comment, their search metadata update #27276 preserves Comet ML and Hailo search terms, and their nightly link-check fix #27290 corrects Platform predict and ONNX Runtime links. Their release PR #27291 finalizes v8.4.216.
Give it a try
pip install --upgrade ultralytics==8.4.216
The v8.4.216 release notes summarize the update, and the full changelog from v8.4.215 to v8.4.216 shows every change.
Trying STM32 export or training on ROCm? Tell us which board or GPU you’re using, what worked, and where you got stuck. Thanks to everyone who contributed, and welcome to @klin2024!