Ultralytics v8.4.102: Up to 8× Faster MuSGD and Expanded YOLO26 Hailo Support
Ultralytics v8.4.102 delivers up to 8× faster MuSGD optimizer steps, adds YOLO26 semantic segmentation export and inference for Hailo accelerators, and improves GPU recovery, dataset handling, validation, and export reliability.
Explore the Ultralytics v8.4.102 release for the complete release details.
New Features
Up to 8× Faster MuSGD
MuSGD now batches Newton–Schulz orthogonalization across compatible parameter matrices and uses fused PyTorch foreach operations for momentum and SGD updates.
This reduces thousands of small CUDA kernel launches and can make optimizer steps up to 8× faster, particularly for models such as YOLO26n. The existing single-tensor muon_update API remains unchanged.
This improvement was introduced in PR #25288 by @Y-T-G.
YOLO26 Segmentation on Hailo
YOLO26 semantic segmentation now supports direct Hailo HEF export and inference:
- Hailo-8 and Hailo-8L: Return logits for host-side upsampling and class reduction.
- Hailo-10 and Hailo-15: Can compile multi-class upsampling and ArgMax into the accelerator graph.
- Automatic calibration: Uses task-specific lightweight calibration datasets when none are provided.
The implementation arrived in PR #25280 by @JESUSROYETH.
Reliability and Bug Fixes
More Reliable GPU Training Recovery
CUBLAS_STATUS_ALLOC_FAILED is now treated as a recoverable first-epoch memory error. Training can automatically reduce the batch size and retry, similar to existing out-of-memory recovery behavior.
See PR #25284 by @glenn-jocher.
Improved Classification Dataset Handling
Sparse NDJSON training and validation splits now preserve global class IDs correctly. When single_cls=True, every sample is consistently mapped to class 0.
- Single-class relabeling was fixed in PR #25290 by @glenn-jocher.
- INT8 calibration now respects the requested classification split and falls back cleanly when a test split is unavailable, as implemented in PR #25292 by @glenn-jocher.
Clearer Segmentation Validation Errors
Segmentation now raises an actionable ValueError when mask_ratio is too large for the selected imgsz, preventing a less informative downstream OpenCV failure.
This fix comes from PR #25282 by first-time contributor @darshil929. Welcome to the YOLO community! 
Export and Inference Improvements
RT-DETR Export Validation
RT-DETR ONNX export now checks whether the resolved opset is compatible before beginning export, producing earlier and clearer feedback.
See PR #25291 by @glenn-jocher.
RT-DETR Augmented Inference Compatibility
Augmented inference now inspects the model’s final detection head directly, improving compatibility with legacy RT-DETR checkpoints.
This update was added in PR #25289 by @glenn-jocher.
Platform and Deployment Documentation
Cloud training documentation now covers optional immutable dataset versions, helping connect trained models to the exact dataset snapshot used for better reproducibility.
- Dataset-version documentation was added in PR #25281 by @glenn-jocher.
- Updated Hailo INT8 accuracy-retention guidance for segmentation, pose, OBB, and classification appears in PR #25283 by @JESUSROYETH.
The documentation also reflects deployment coverage across 42 regions in the Ultralytics Platform, where users can annotate datasets, train models, export deployments, and monitor results.
Get Started
Upgrade to the latest release:
pip install -U ultralytics
Then try your existing training, export, or inference workflow with v8.4.102. MuSGD users and teams deploying YOLO26 segmentation to Hailo hardware should see the most immediate benefits.
Review every change in the v8.4.101 to v8.4.102 full changelog, and please share your results, questions, or issues with the Ultralytics community. 