ultralytics 8.4.179

v8.4.179 - ultralytics 8.4.179 Cap rknn-toolkit2 below 2.5 and smoke RKNN at imgsz=64 (#26727)

on Python PyPI

🌟 Summary

Ultralytics 8.4.179 focuses on export reliability and inference correctness, led by a compatibility fix for RKNN on Rockchip devices. It also improves SAM, validation, benchmarking, and training stability.

📊 Key Changes

• RKNN export compatibility: Caps rknn-toolkit2 at versions below 2.5, because version 2.5.0 can fail to build YOLOv8 and YOLO11 models across Rockchip targets, and YOLO26 on RK3588. The RKNN smoke test now uses imgsz=64 to better catch this failure.
• SAM improvements: Segment-everything prompts now target the image area rather than letterbox padding, avoiding empty masks and improving results on narrow images. Mask resizing is also handled centrally in the prompt encoder.
• More reliable export and training: JetPack 6 with TensorRT 10.3 now returns an actionable error for unsupported INT8 exports with embedded NMS. YOLO26 pose loss is more numerically stable, and torch.compile falls back gracefully on CUDA systems without working Triton.
• More accurate evaluation: RT-DETR validation now filters out unrequested classes before applying the detection limit. YOLOE and YOLO-World benchmarks use the dataset’s class vocabulary for a fairer comparison between PyTorch and exported models.
• Performance and robustness fixes: TorchScript warmup now happens before timed CPU inference, ZIP archives with a root ./ entry no longer crash extraction, and transient GitLab mirror failures get one retry.
• Safeguards and documentation: An unsafe TorchScript graph rewrite was removed to preserve correct model outputs. Documentation also clarifies ONNX INT8 quantization and updates guidance for Roboflow imports and Platform features.

🎯 Purpose & Impact

• Rockchip users avoid a known-broken RKNN dependency version, while the stronger smoke test is more likely to catch similar export issues before release.
• SAM users should see fewer padding-related empty masks and more reliable prompts across different image shapes.
• Model developers get clearer failure messages and more dependable validation, benchmarking, and training behavior.
• No new model architecture is introduced in this release; the main impact is improved compatibility, correctness, and usability.

What's Changed

• Fix cohort regressions at their owners and remove test growth by @glenn-jocher in #26698
• Handle ZIP root directory entries without IndexError by @Nikhila-Dhevi-R in #26712
• Raise actionable diagnostic for TensorRT 10.3 INT8 nms=True exports on JetPack 6 by @JohnnyWilson16 in #26711
• Sample SAM segment-everything points on the image, not the letterbox padding by @raimbekovm in #26708
• Document the selectable Roboflow import preview, chart zoom reset and Platform 409 rules by @raimbekovm in #26706
• Document what ONNX INT8 export quantizes by @AtomicAyush in #26705
• Benchmark exported YOLOE and YOLO-World models with the dataset vocabulary by @raimbekovm in #26704
• Filter RT-DETR validation predictions by requested classes by @Nikhi00718 in #26702
• Skip torch.compile on CUDA when Triton is unavailable by @sayaksatpathi in #26701
• Warm up TorchScript on CPU before timed inference by @raimbekovm in #26700
• Add https://youtu.be/hPXsZ_fVz6M to docs by @RizwanMunawar in #26709
• Prevent sigma underflow from disabling YOLO26 pose RLE loss by @Nikhi00718 in #26713
• Retry failed GitLab mirrors once before failing by @glenn-jocher in #26714
• ultralytics 8.4.179 Cap rknn-toolkit2 below 2.5 and smoke RKNN at imgsz=64 by @glenn-jocher in #26727

New Contributors

• @JohnnyWilson16 made their first contribution in #26711
• @AtomicAyush made their first contribution in #26705
• @Nikhila-Dhevi-R made their first contribution in #26712
• @sayaksatpathi made their first contribution in #26701

Full Changelog: v8.4.178...v8.4.179