ultralytics 8.3.224

v8.3.224 - `ultralytics 8.3.224` Accelerate `crop_mask` on GPU with `is_cuda` check (#22575)

on Python PyPI

🌟 Summary

Ultralytics 8.3.224 boosts segmentation performance and stability on GPUs, adds CPU-only ExecuTorch export/benchmarks for YOLO11, fixes multi‑GPU evaluation and ONNX device selection, and improves logging + docs for a smoother developer experience. 🚀

📊 Key Changes

• GPU-accelerated crop_mask with device-safe logic (PR #22575) ⚡
  • Ensures boxes live on the same device as masks and avoids slow Python loops on CUDA.
  • Switches to loop-only when n < 50 and on CPU; vectorized paths on GPU.
• Correct multi-GPU COCO evaluation jdict gather (PR #22541) 🧩
  • Aggregates predictions across ranks using dist.gather_object and cleans up worker memory.
• ExecuTorch export and benchmark support (CPU-only) for YOLO11 (PR #22552) 🧠
  • Adds format guards (no YOLOWorldv2, no E2E, no Pose) and benchmark table entry.
• Respect selected CUDA device in ONNX Runtime (PR #22546) 🎯
  • Passes device_id to CUDAExecutionProvider to avoid accidental GPU 0 usage.
• More reliable W&B logging (PR #22563) 📈
  • Commits metrics each epoch and aligns log order for real-time, clean dashboards.
• Rust ONNXRuntime example fix and dependency updates (PR #22557) 🦀
  • Uses download-binaries for ONNX Runtime to reduce setup friction and version conflicts.
• Massive docstring/style cleanup across codebase (PRs #22554, #22565) 📚
  • Standardizes one-line summaries, arg sections, and readability with no logic changes.
• Export formats updated: ExecuTorch now listed CPU=True, GPU=False in exporter matrix 🧪

🎯 Purpose & Impact

• Faster, safer segmentation on GPU 💨
  • Avoids device mismatch errors and slow loops; users should see snappier, more stable mask processing.
• Accurate distributed validation ✅
  • Multi-GPU validation now correctly aggregates predictions for COCO-style evaluation and JSON exports.
• Broader deployment options with ExecuTorch 🧩
  • Enables CPU-only export and benchmarking for YOLO11—handy for mobile/edge experimentation, with clear guardrails.
• Correct GPU selection for ONNX inference 🔧
  • Ensures the specified GPU (e.g., cuda:1) is honored, improving reliability in multi-GPU environments.
• Better experiment tracking in W&B 📊
  • Immediate per-epoch updates and aligned steps yield clearer, real-time dashboards with minimal overhead.
• Easier Rust example usage 🛠️
  • Reduces dependency issues; quicker “just run it” developer path.
• Cleaner docs and IDE help ✨
  • Consistent docstrings improve readability, navigation, and autocompletion across modules.

Quick tips:

• Export to ExecuTorch (CPU-only) for YOLO11:

  yolo export model=yolo11n.pt format=executorch device=cpu

• Ensure ONNX uses the intended GPU automatically when available—no code changes needed.

What's Changed

• Fix gather jdict for COCO evaluation by @Laughing-q in #22541
• feat: ✨ Add ExecuTorch benchmark and update export settings and add format validation in benchmarks by @onuralpszr in #22552
• Fix docstrings by @glenn-jocher in #22554
• Commit Weights & Biases logs after each epoch to force dashboard update by @Y-T-G in #22563
• Fixing rust example (fixes #22556) by @andrenatal in #22557
• Update Google-style docstrings by @glenn-jocher in #22565
• Specify device ID for ONNX inference on CUDA by @Y-T-G in #22546
• ultralytics 8.3.224 Accelerate crop_mask on GPU with is_cuda check by @Laughing-q in #22575

New Contributors

• @andrenatal made their first contribution in #22557

Full Changelog: v8.3.223...v8.3.224