v3.0.1 1.11.1
data-privacy-stack/presidioSep 28, 2026 → Sep 30, 20262 days between releases
What's New
RF-DETR 1.11.1 is a correctness patch fixing silent failures and renaming the `clip_gradients` API method.
Breaking Changes
- `RFDETRModelModule.clip_gradients` renamed to `configure_gradient_clipping`.
- `TRTInference` now raises `ValueError` for invalid inputs instead of returning wrong results.
- `TensorRTExporter.build_engine` now raises `ValueError` for `dynamic_batch=True` ONNX without specific config.
- `RFDETR.from_checkpoint()` now raises `ValueError` instead of `KeyError: 'args'` for invalid checkpoints.
New Features
- Added `rfdetr.export._benchmark` helpers for latency and memory measurement.
- Added `decode_pil_image` and `decode_pil_image_bytes` for direct PIL decoding.
- Added `CrowdRegion` and `crowd_regions_from_coco` utilities for evaluation.
- Added `rfdetr.utilities.package.is_installed()` for dependency checking.
- Added `Exporter.check_dependencies` as an overridable classmethod hook.
Improvements
- LoRA backbone checkpoints now load correctly instead of silently failing.
- TensorRT inference now raises explicit errors for invalid inputs rather than returning wrong results.
- Crowd-region mAP evaluation now matches pycocotools standards.
Bug Fixes
- Fixed silent failure when loading LoRA backbone checkpoints.
- Fixed TensorRT inference returning wrong detections for invalid inputs.
- Fixed crowd-region mAP overcounting false positives on COCO datasets.
Upgrade Notes
Update code to use `configure_gradient_clipping` instead of `clip_gradients`. Update error handling in `RFDETR.from_checkpoint()` and TensorRT inference to catch `ValueError` instead of `KeyError`.
The Two Releases
v3.0.1
Sep 28, 2026· 23 days ago
Makes stable update reminders part of normal diagram delivery, ensuring users are notified of new verified releases without automatic installation.
Full release notesv1.11.1: Correctness Fixes for LoRA, TensorRT & Crowd mAP
Sep 30, 2026· 22 days ago
This release serves as a correctness patch for RF-DETR, fixing silent failures in LoRA checkpoint loading, TensorRT inference input validation, and COCO crowd region evaluation metrics.
Full release notes