đ Summary
v8.4.195 adds early Python 3.15 compatibility checks, fixes AutoBatch profiling on older PyTorch versions, and improves documentation accuracy across hundreds of pages.
đ Key Changes
- đ Python 3.15 CI support: Adds a Linux/Python 3.15 test job and the Python 3.15 package classifier. Since ONNX Runtime and OpenVINO do not yet provide Python 3.15 wheels, the job installs a reduced set of extras and skips the affected export tests for now.
- đ ī¸ AutoBatch compatibility fix: Adjusts when FLOPs are counted during profiling to avoid gradient-related failures on older PyTorch versions, including failures observed on NVIDIA Jetson.
- đ Documentation refresh: Reviews and clarifies documentation across 301 Markdown files, correcting details and improving links to related guides, models, datasets, and platform resources.
đ¯ Purpose & Impact
- â Gives maintainers an early way to catch compatibility issues as Python 3.15 support becomes available, while clearly limiting tests that depend on packages without compatible wheels.
- đ Makes AutoBatch profiling more reliable for users running older PyTorch versions and helps restore affected Jetson tests.
- đ Makes documentation easier to trust and navigate, with clearer dataset guidance and more accurate examples.
- âšī¸ No major model or architecture changes are included in this release.
What's Changed
- Improve docs accuracy and clarity, and link related guides, models and datasets by @glenn-jocher in #26976
- Fix AutoBatch profiling on older torch by counting FLOPs after the timed pass by @glenn-jocher in #26977
ultralytics 8.4.195Add Python 3.15 CI and classifier by @glenn-jocher in #26975
Full Changelog: v8.4.194...v8.4.195