v1.11.0 1.11.1

medium risk

0xNyk/council-of-high-intelligenceSep 30, 2026 → Sep 30, 20260 days between releases

What's New

Fixes critical LoRA loading, TensorRT inference, and Crowd mAP bugs while renaming `clip_gradients`.

Breaking Changes

  • `RFDETRModelModule.clip_gradients` renamed to `configure_gradient_clipping`
  • `TRTInference` raises `ValueError` for invalid inputs
  • `TensorRTExporter.build_engine` raises `ValueError` on `dynamic_batch=True`
  • `RFDETR.from_checkpoint()` raises `ValueError` instead of `KeyError`

New Features

  • Added `decode_pil_image` / `decode_pil_image_bytes` decoders
  • Added `CrowdRegion` / `crowd_regions_from_coco` for crowd annotations
  • Added `resize_masks_nearest` for mask resizing
  • Added `is_installed(name)` utility
  • Added `Exporter.check_dependencies` classmethod

Improvements

  • Renamed `RFDETRModelModule.clip_gradients` to `configure_gradient_clipping`

Bug Fixes

  • `backbone_lora=True` checkpoints now load correctly instead of silently using random initialization
  • `TRTInference` now refuses inputs that would have returned wrong detections
  • Val/test mAP calculations now match pycocotools on COCO crowd regions
  • `RFDETR.from_checkpoint()` now raises `ValueError` instead of silently failing with `KeyError`

Upgrade Notes

Most users require no action as this release exposes previously silent errors. Update code to use `configure_gradient_clipping` instead of `clip_gradients`. Verify TensorRT inputs and `dynamic_batch` settings to avoid new `ValueError` exceptions.

The Two Releases

v1.11.0

v1.11.0

Sep 30, 2026· 22 days ago

This release focuses on inbox enhancements, including GIF rendering and autoplay, while generalizing the Threads channel to all users. It introduces the AI Agent feature, comment automation filters, and improves WhatsApp contact linking when numbers change.

Full release notes
1.11.1

v1.11.1: Correctness Fixes for LoRA, TensorRT & Crowd mAP

Sep 30, 2026· 22 days ago

A correctness patch that fixes silent bugs in LoRA loading, TensorRT inference, and Crowd mAP calculations, and renames a gradient clipping method.

Full release notes