Starstruk Logo MarkStarstruk

v0.64.10 1.11.2

medium risk

roboflow/rf-detrSep 30, 2026 → Oct 4, 20264 days between releases

What's New

Hardens threaded inference, fixes data validation, and optimizes ONNX exports for Apple's CoreML.

Breaking Changes

  • Removed `rfdetr.datasets.yolo.REQUIRED_DATA_SUBDIRS` constant.
  • Refuses YOLO split directory resolving outside dataset root.
  • export(format='tflite', dynamic_batch=True) raises NotImplementedError.
  • [executorch] extra is capped at version 1.3.x.

New Features

  • Added `find_yolo_data_file` and `find_yolo_train_images` utilities.
  • ONNX exports optimized for Apple's CoreML provider.
  • Improved YOLO dataset validation and layout detection.

Improvements

  • Threads sharing a model no longer corrupt weights or disable cuDNN.
  • predict() now accepts flipped frames (negative strides) as documented.
  • ONNX exports run faster through Apple's CoreML.
  • YOLO data.yaml files load consistently with sorted class IDs.

Bug Fixes

  • Fixed threads corrupting model weights and disabling cuDNN.
  • Fixed predict() failing on flipped frames.
  • Fixed class_names root category shifting labels.
  • Fixed YOLO data.yaml loading inconsistencies.

Upgrade Notes

Remove inline imports of REQUIRED_DATA_SUBDIRS as it has been removed. Review the Migration guide for four behavior changes. Note that export(format='tflite', dynamic_batch=True) is no longer supported.

The Two Releases

v0.64.10

OpenHuman v0.64.10

Sep 30, 2026· 21 days ago

An upgrade focusing on performance, UX, and modularity, including faster agent loops, selectable memory engines, and improved integrations.

Full release notes
1.11.2

v1.11.2 — Safer Threaded Inference and Training Inputs

Oct 4, 2026· 18 days ago

Hardens threaded inference and fixes data validation issues, including thread safety for model weights and cuDNN.

Full release notes