ultralytics 8.3.241

v8.3.241 - `ultralytics 8.3.241` ORT CoreML execution provider when `device="mps"` (#22984)

on Python PyPI

🌟 Summary

v8.3.241 boosts ONNX Runtime performance on macOS by preferring CoreML on Apple hardware, stabilizes ExecuTorch exports, and expands hardware/image support for Rockchip and 2‑channel data — plus a round of documentation and tooling polish. 🍏🚀

📊 Key Changes

• Mac ONNX Runtime: Prefer CoreML on Apple hardware (@glenn-jocher)

  • On macOS with device="mps", Ultralytics now prefers CoreMLExecutionProvider over CPU for ONNX Runtime.
  • New provider selection logic: CUDA > CoreML (MPS) > CPU, with clearer logging of which provider is actually used.
  • IO binding is now only used when it’s supported (CUDA), avoiding issues with CoreML.

• ExecuTorch export reliability improvements (@glenn-jocher, @onuralpszr)

  • export_executorch() now pins numpy<=2.3.5 to avoid known coremltools failures with newer NumPy.
  • Removed a redundant import torch in the ExecuTorch export path for a cleaner exporter.

• New hardware support: Rockchip RV1126B (@venjye, @glenn-jocher)

  • Added "rv1126b" to supported Rockchip chips in code.
  • Updated the Rockchip RKNN docs to explicitly mention RV1126B as supported.

• Better image handling: 2‑channel images in Annotator (@kenanking)

  • ultralytics.utils.plotting.Annotator now correctly handles 2‑channel numpy images by converting them to 3‑channel before drawing and saving.

• TF/SavedModel & TensorRT backend cleanups (@glenn-jocher)

  • TensorFlow SavedModel loading simplified to always use tf.saved_model.load for this path.
  • TensorRT engine parsing refactored for clearer handling of dynamic shapes and data types across TRT <10 and TRT ≥10.

• SAM 3 docs: correct save usage (@dingjie-ai)

  • SAM3SemanticPredictor examples now set save=True through the overrides dict at initialization, not per call.

• Sony IMX500 & solutions docs polish (@glenn-jocher)

  • Cleaned and re-indented the YOLOSegment example for the Sony IMX500 integration docs.
  • Improved wording and types in various docs macros (auto-annotation, solutions args, tracking/visualization args, validation args).

• Docs site & authorship updates (@glenn-jocher, @RizwanMunawar, @onuralpszr)

  • Updated Streamlit live inference guide with a new YouTube demo video.
  • Bumped the docs chat widget to ultralytics/llm@v0.1.8 for a better in-docs chat experience.
  • Updated mkdocs-ultralytics-plugin dev dependency to >=0.2.4.
  • Fixed and enriched GitHub authors mapping (add @dingjie-ai, @jshakes, and improve @0xSynapse avatar URL).

🎯 Purpose & Impact

• Faster inference on macOS with ONNX Runtime 🍏⚡

  • Users running ONNX models on Mac with device="mps" now automatically benefit from Apple Neural Engine / GPU acceleration via CoreML when available.
  • Real-world speedup example: ~20.3 ms → ~9.0 ms per inference vs CPU provider, nearly a 2× improvement.
  • CUDA behavior and fallbacks remain unchanged; only macOS/MPS selection is improved and more robust.

• More reliable ExecuTorch exports 📦🧩

  • The NumPy pin greatly reduces “mysterious” ExecuTorch/CoreML export failures caused by incompatible NumPy versions.
  • Users targeting ExecuTorch (e.g., mobile/embedded) should see more stable export runs without needing to manually juggle NumPy versions.

• Broader deployment options on edge devices 🧱🚀

  • Adding RV1126B ensures users can export Ultralytics models directly for that chip without custom patches.
  • Updated docs make it clearer which Rockchip devices are expected to work out of the box.

• Smoother visualization for non-standard data 🌈🖼️

  • 2‑channel image workflows (e.g., spectral or custom sensor data) can now be annotated and saved without errors, aligning with existing 1‑channel, 3‑channel, and multispectral support.

• Cleaner and more predictable backends 🧠🔧

  • Refined TensorRT and TF SavedModel handling makes backend behavior more consistent across versions and shapes.
  • IO binding is now used only where it’s supported (CUDA), avoiding subtle bugs with CoreML.

• Clearer documentation & better tooling experience 📚✨

  • SAM 3 examples now match real behavior, reducing confusion around where save=True should be configured.
  • IMX500 example, solution/validation/tracking argument docs, and auto-annotation docs are easier to read and more accurate.
  • The docs chat widget and authors list improvements enhance the support and community experience without affecting core training/inference code.

Overall, v8.3.241 is a backend and compatibility–focused release: faster ONNX Runtime on macOS, more dependable exports, better edge-device coverage, and smoother visualization for specialized image formats — with documentation and tooling kept in sync. 🎛️🚀

What's Changed

• Update chat.min.js to version 0.1.7 by @glenn-jocher in #22987
• Update chat.min.js to version 0.1.8 by @glenn-jocher in #22988
• Docs: clarify save argument usage for SAM3SemanticPredictor by @dingjie-ai in #22989
• docs: 📝 remove broken GitHub author entry for 0xSynapse by @onuralpszr in #22992
• Fix Sony syntax by @glenn-jocher in #22994
• Add https://youtu.be/1H9ktpHUUB8 to docs by @RizwanMunawar in #22993
• Add support for handling 2-channel images in Annotator class by @kenanking in #22999
• Add support for rv1126b Rockchip processor by @venjye in #22991
• fix: 🐞 remove redundant import of torch from exporter.py by @onuralpszr in #23007
• Fix executorch bug by @glenn-jocher in #23006
• Update rockchip-rknn.md with RV1126B by @glenn-jocher in #23003
• Fix macros by @glenn-jocher in #23010
• ultralytics 8.3.241 ORT CoreML execution provider when device="mps" by @glenn-jocher in #22984

New Contributors

• @venjye made their first contribution in #22991
• @kenanking made their first contribution in #22999
• @dingjie-ai made their first contribution in #22989

Full Changelog: v8.3.240...v8.3.241