Ultralytics v8.4.126 Is Here
Ultralytics v8.4.126 delivers safer, thread-safe checkpoint loading, up to 40% faster restricted model loading, and a cleaner RLE prior calculation—all while preserving backward compatibility.
Highlights
Reliable restricted checkpoint loading across concurrent threads.
Approximately 40% less restricted-loading overhead in affected workflows.
Improved secure-loading compatibility for PyTorch 2.6 and newer.
Simplified RLE loss calculations with better mixed-precision behavior.
Continued compatibility with older checkpoints and unrestricted workflows.
New Features
Thread-safe restricted checkpoint loading
Restricted PyTorch checkpoint loading is now safe in concurrent environments. The shared allow-list remains registered for the process lifetime, preventing intermittent errors when multiple threads load models simultaneously.
A new regression test validates 32 concurrent restricted loads, helping protect multi-threaded inference services, Platform CPU workers, and other production workloads.
Improvements
Faster restricted model loading
Instead of rebuilding a large allow-list for every load, Ultralytics YOLO now registers only the checkpoint globals that each file references. This reduces restricted-loading overhead by approximately 40% in affected cases.
Improved secure-loading compatibility
Restricted loading uses PyTorch functionality available in version 2.6 and newer. Older PyTorch versions continue to fall back to standard loading behavior, while normal unrestricted loading remains unchanged.
Official YOLO26 checkpoints—including YOLO26x—are now protected from failures caused by concurrent loading.
Simplified RLE prior calculation
The runtime multivariate distribution object has been replaced with a simpler closed-form implementation. This change:
- Reduces unnecessary internal overhead.
- Improves numerical behavior during mixed-precision training.
- Retains existing checkpoint buffers so models saved before v8.4.126 can resume correctly.
Bug Fixes
- Fixed a race condition affecting concurrent restricted checkpoint loads.
- Prevented one thread from removing shared safe globals while another load was still running.
- Preserved expected behavior for existing unrestricted-loading workflows and older saved checkpoints.
Contributors
This release includes PR #25885: Make restricted checkpoint loading thread-safe and 40% faster, simplify RLE prior by @pderrenger.
Try v8.4.126
Upgrade to the latest release:
pip install -U ultralytics
Review the Ultralytics v8.4.126 release or explore the full changelog from v8.4.125 to v8.4.126.
Give the release a try and share your feedback, performance results, or any issues you encounter with the Ultralytics community!