0.33.4 1.10.0
roboflow/rf-detrAug 24, 2026 → Sep 4, 202611 days between releases
What's New
Significant performance gains with default value changes affecting batch size and validation behavior.
Breaking Changes
- `grad_accum_steps` default changed to 1, reducing effective batch size.
- Validation now evaluates one model per epoch instead of two.
- Dataset builders require a complete pipeline-option namespace.
- `log_per_class_metrics` default changed to False.
New Features
- `TrainConfig.pack_targets` for packed target transport across workers.
- `TrainConfig.eval_batch_size` to decouple validation batch size.
- `TrainConfig.best_model_metric` to rank checkpoints by mAR.
- Training progress bar restoration with live GPU memory and LR metrics.
- `deploy_to_roboflow()` version auto-resolution and `ROBOFLOW_HOME` alias support.
Improvements
- Training time reduced by 44% via optimizer parameter group merging.
- Inference speed increased by up to 46% via uint8 widening on accelerator.
- DataLoader throughput increased by up to 2.26x using packed targets.
- New GPU batched linear-assignment solver for the matcher on CUDA.
- Optimizer parameter groups merged by hyperparameter (465 to 28).
Bug Fixes
- Dataset builders now validate their config instead of silently falling back to wrong defaults.
Upgrade Notes
Upgrading requires attention to `grad_accum_steps` defaulting to 1, which reduces effective batch size. Validation metrics now reflect the EMA model only. Ensure dataset builders have a complete pipeline-option namespace.
The Two Releases
0.33.4
Aug 24, 2026· 2 months ago
This release updates the playground UI by applying the Geist font style to cards to improve visual consistency.
Full release notesv1.10.0: Faster Training & Inference
Sep 4, 2026· 2 months ago
A performance-focused release reducing training time by 44% (55 minutes to 31 minutes) and improving inference speed, accompanied by several default value changes.
Full release notes