New Release: Ultralytics v8.3.126

Discussion
,
1

:rocket: Ultralytics v8.3.126 Release: Smarter GPU Auto-Selection & Enhanced Usability!

Hi everyone!

We’re excited to announce Ultralytics v8.3.126—a release focused on smarter hardware utilization, improved developer experience, and more user-friendly documentation. Here’s what’s new and why you should give it a try!

:glowing_star: Summary

This update introduces auto-selection of the most idle GPUs for training, removing the guesswork from resource allocation in multi-GPU environments. Alongside this, you’ll find improved documentation, logging, and compatibility, helping both new and advanced users streamline their workflows.

:tada: New Features

Idle GPU Auto-Selection

  • Effortless GPU Choice:
    Simply provide device=-1 (or device=[-1, -1]) and let Ultralytics automatically select the least busy GPU(s) for your training jobs.
  • Smart Monitoring:
    The new GPUInfo utility tracks GPU usage, memory, temperature, and power to make the best choice for your workloads.
  • Seamless Integration:
    CUDA-based tests and training now leverage this functionality, making setup easier than ever.

:books: Documentation Enhancements

  • Updated Training Docs:
    Discover detailed instructions and examples for using the idle GPU selection feature in the training documentation.
  • New Reference Pages:
    Explore the new autodevice utility reference for an in-depth look at how device selection works.
  • Clearer Device Options:
    Device selection methods are now clearly explained in configuration and argument reference guides.

:wrench: Improvements & Bug Fixes

  • Sorted Logging for Readability:
    Function argument logs are now alphabetically sorted for easier reading (PR by @ambitious-octopus).
  • Progress Bar Compatibility:
    Improved compatibility for custom progress bars and device type hints (PR by @glenn-jocher).
  • TQDM Warning Fix:
    Addressed warnings related to collections.Iterable and tqdm usage (PR by @glenn-jocher).
  • CUDA Idle Device Assignment:
    The new auto-assignment logic improves training setup on shared or multi-user GPU machines (PR by @glenn-jocher).

:bullseye: Purpose & Impact

  • Smarter & More Efficient Training:
    Automatically identifies and uses available GPUs—ideal for shared teams and multi-job servers.
  • Simplified Setup:
    Avoid manual GPU checks and selection—just set device=-1 and go!
  • Improved Readability & Usability:
    Cleaner logs and expanded documentation help everyone make the most of these features.

:vertical_traffic_light: Ready to Try?

Get started with the v8.3.126 release and let us know how the new GPU auto-selection works for you in your projects! Your feedback is invaluable in helping us further enhance the Ultralytics ecosystem.

Explore the full changelog for all the details:

Thanks to our amazing contributors and the YOLO community for making these improvements possible!
We look forward to your thoughts, bug reports, and suggestions as you try out the latest release.

Happy training!
— The Ultralytics Team