🌟 Summary
Ultralytics 8.4.183 expands hardware deployment and improves reliability across training, prediction, data handling, and evaluation. 🚀
📊 Key Changes
- Axelera Metis support for every YOLO26 task: With Voyager SDK 1.9.0, detection, instance and semantic segmentation, depth estimation, pose, OBB, and classification models can all export, predict, and validate through
yolo. The NMS-free detection path is used by default. SDK 1.8 exports remain compatible; the updated setup requiresmetis-dkms1.7.0 or newer and supports PyTorch versions below 2.15. - Higher Axelera throughput: In the documented COCO128 benchmark, SDK 1.9 reached 109 images per second at batch 32, compared with 40 images per second on SDK 1.8. Larger prediction batches can improve throughput; batch size 1 remains useful when per-image latency matters.
- Intel acceleration for LiteRT models: LiteRT exports can now run on Intel CPUs and NPUs, including support for YOLO26 task types such as detection, segmentation, depth, pose, OBB, and classification.
- More reliable training and depth calibration: Fixes address warmup and gradient accumulation edge cases, MPS memory growth during training, and depth calibration replacing an already-good calibration with a worse one. Interrupted runs can also resume under the same Ultralytics Platform model, and W&B resuming works on Windows.
- More accurate and robust data workflows: Fixes improve ImageNet-subset classifier validation, accept UTF-8 labels with a byte-order mark, preserve non-ASCII grounding annotations, convert COCO RLE masks for segmentation, and align 16-bit NumPy images with image-file prediction behavior.
- Prediction, export, and visualization fixes: Negative FastSAM points now reliably exclude masks regardless of prompt order. Other fixes cover very thin images, rectangular INT8 calibration, grayscale LiteRT models, semantic overlay colors, depth-map saving, and several additional edge cases.
- Platform and documentation updates: Platform documentation now includes Gemma 4 and Nemotron as hosted annotation models and explains new visibility controls and workflows. Benchmark instructions were also clarified for NMS-free evaluation.
🎯 Purpose & Impact
- Deploy YOLO26 on more hardware and tasks: Axelera users can use one integrated workflow across all seven YOLO26 tasks, while Intel LiteRT users gain CPU and NPU inference options.
- Get better throughput when processing batches: Axelera SDK 1.9 can substantially increase prediction throughput for larger batches, though actual results depend on hardware and workload.
- Trust training, validation, and predictions more: The fixes reduce avoidable failures and make reported results better reflect the supplied data and model—for example, correcting classifier label alignment changes ImageNette validation accuracy from 9.45% to 70.37% in the reported test.
- Keep existing workflows working: Many changes are targeted bug fixes or compatibility improvements; notably, Axelera models exported with SDK 1.8 do not need to be re-exported for SDK 1.9.
What's Changed
- Fix R/B channel swap of semantic mask overlays in train/val mosaics by @cainiao33 in #26755
- Scale 16-bit NumPy prediction inputs to 8 bits by @Nikhi00718 in #26741
- Restore pre-existing temporary module aliases by @Vaishnavi220506 in #26756
- Respect negative FastSAM points regardless of prompt order by @Nikhi00718 in #26759
- Align ImageNet subset folders with classifier class IDs by @Nikhi00718 in #26760
- Fix INT8 calibration crash when exporting with rect=True by @Y-T-G in #26750
- Fix LiteRT layout detection for grayscale exports by @aswanth-07 in #26747
- Clamp warmup accumulate ramp to the capped nominal accumulate by @cainiao33 in #26753
- Bucket target padding on MPS to stop per-shape driver memory growth by @vtarasv in #26744
- Keep one-row, one-column and 1x1 depth maps 2D in save_depth_png by @cainiao33 in #26754
- Read GroundingDataset JSON annotations as UTF-8 by @MohammadHijjawi97 in #26745
- Keep UTF-8 BOM labels during dataset scanning by @mottopanikeiku in #26761
- Add hosted Gemma 4 and Nemotron to the Platform annotation docs by @glenn-jocher in #26765
- Bump eviden-actions/clean-self-hosted-runner from v1.4.41 to v1.4.43 in /.github/workflows by @UltralyticsAssistant in #26776
- Add Intel CPU and NPU inference support to LiteRT backend by @onuralpszr in #26779
- Require ultralytics-thop>=2.2.3 for functional linear and attention counts by @glenn-jocher in #26785
- Update Axelera SDK to 1.9.0 and support every YOLO26 task by @a-tangfan in #26782
- Preserve rectangular classification INT8 calibration shapes by @Nikhi00718 in #26783
- Keep letterbox resize dimensions positive for thin images by @aswanth-07 in #26771
- Continue the Platform model when local training resumes by @raimbekovm in #26689
- Document Platform default visibility, Annotate similar, the TensorRT JetPack 6 nms=True error and DOTA submission flags by @raimbekovm in #26775
- Resume distillation runs under restricted checkpoint loading by @raimbekovm in #26777
- Decode COCO RLE masks in convert_coco by @Vaishnavi220506 in #26774
- Resume W&B runs on Windows without the latest-run symlink by @raimbekovm in #26772
- Convert NDJSON validation data before YOLOE prompt setup by @raimbekovm in #26770
- Compare URL schemes and archive extensions case-insensitively by @raimbekovm in #26781
- Keep a depth model's existing calibration in calibrate() when it scores best by @Y-T-G in #26778
- Fix warmup endpoint, Conv2 fusion, BOM source lists, image_open file errors, semantic mask names and float64 OBB corners by @cainiao33 in #26766
- ultralytics 8.4.183 Update Axelera SDK to 1.9.0 and support every YOLO26 task by @glenn-jocher in #26788
New Contributors
- @vtarasv made their first contribution in #26744
- @mottopanikeiku made their first contribution in #26761
Full Changelog: v8.4.182...v8.4.183