New Release: Ultralytics v8.3.130

[glenn-jocher]

Ultralytics v8.3.130 Release: Faster Startups, Smarter Monitoring, and Improved Export Reliability

Summary

We’re excited to announce Ultralytics v8.3.130, a release focused on accelerating your model workflows, making training progress easier to track, and boosting reliability and security. This update brings significant speed enhancements to model initialization, improved guidance for training metric monitoring, stronger ONNX export testing, and workflow security improvements. Whether you’re deploying models or developing new solutions, this release delivers a smoother, more robust experience.

Release URL: Ultralytics v8.3.130 Release Notes

---

New Features & Improvements

Faster Model Initialization

• Optimized Layer Fusion: Model startup is now quicker thanks to an enhanced model.fuse() process, which performs layer fusion on the CPU before transferring to the GPU. This results in noticeably faster loading and lower memory pressure, especially if you’re working on systems with limited GPU resources.
  • Contributed by @dianyo — see the details in PR #20466.

Better Training Monitoring

• Improved Callback Documentation: Developers can now more easily access and print key training metrics after each checkpoint using the on_model_save callback. The documentation has been expanded with a clear Python example, making it easier for users at all levels to monitor, debug, and optimize their models.
  • Contributed by @RizwanMunawar — see PR #20531.

Enhanced ONNX Export Testing & Workflow Security

• Expanded ONNX Export CI: ONNX export testing has been greatly expanded, with CUDA support included to catch more potential issues and ensure reliable model conversion across varied environments.

  • Implemented by @glenn-jocher — in PR #20448.

• Stronger Workflow Permissions: GitHub workflow permissions have been tightened for safer and more secure automated formatting and labeling processes, keeping the codebase secure for all contributors.

  • Contributed by @glenn-jocher — in PR #20556.

• License Compliance: All relevant files now include proper license headers to ensure open-source compliance and code reuse clarity.

---

What’s Changed

• Add on_model_save callback with Python example to Callbacks docs by @RizwanMunawar
• Potential fix for workflow permissions in format.yml by @glenn-jocher
• Add ONNX CUDA continuous integration in ONNX CI PR by @glenn-jocher
• Faster model.fuse() operations in PR #20466 by @dianyo

Full changelog

---

New Contributors

A special welcome and thanks to @dianyo for making their first contribution with the model speedup PR!

---

Get Involved & Share Feedback

We invite you to update to v8.3.130 and experience the improvements firsthand. Let us know how these enhancements impact your workflow, and as always, your feedback is invaluable in guiding future development.

Thank you to the amazing YOLO community and the Ultralytics team for your continued passion, ideas, and contributions. We look forward to your insights and suggestions—let’s keep building together!

Happy experimenting and deploying!

— The Ultralytics Team