ultralytics 8.4.102

v8.4.102 - 8x faster MuSGD optimizer step with batched Newton-Schulz (#25288)

on Python PyPI

🌟 Summary

🚀 Ultralytics 8.4.102 makes MuSGD training up to 8× faster while expanding YOLO26 edge deployment, improving export reliability, and strengthening dataset and GPU error handling.

📊 Key Changes

• ⚡ Faster MuSGD optimizer

  • Batches Newton–Schulz orthogonalization across compatible parameter matrices instead of processing each parameter separately.
  • Uses fused PyTorch foreach operations for momentum and SGD updates.
  • Reduces thousands of small CUDA kernel launches, making MuSGD much less launch-bound.
  • Preserves the existing single-tensor muon_update API while adding an optimized internal batched path.
  • Reported impact: up to 8× faster optimizer steps, particularly for models such as YOLO26n.

• 📱 YOLO26 semantic segmentation on Hailo

  • Adds direct Hailo HEF export and inference for YOLO26 semantic segmentation.
  • Hailo-8/8L devices return logits for host-side upsampling and class reduction.
  • Hailo-10/15 devices can compile multi-class upsampling and ArgMax directly into the accelerator graph, returning compact class maps.
  • Uses task-specific lightweight calibration datasets when no dataset is provided.

• 🛡️ More reliable GPU training recovery

  • Treats CUBLAS_STATUS_ALLOC_FAILED as a recoverable first-epoch memory error, allowing automatic batch-size reduction and retry behavior similar to out-of-memory errors.

• 📦 Improved classification dataset handling

  • Preserves global class IDs for sparse NDJSON train and validation splits.
  • Correctly maps all samples to class 0 when single_cls=True.
  • Uses the actual labels exposed by the dataset for class-count validation.

• 🔧 Safer model export and inference

  • RT-DETR ONNX export now validates that the resolved opset is compatible before export.
  • INT8 classification calibration respects the requested dataset split and falls back cleanly when a test split is unavailable.
  • Augmented inference now checks the model’s final detection head directly, improving compatibility with legacy RT-DETR checkpoints.

• ✅ Clearer validation errors

  • Segmentation now raises an actionable ValueError when mask_ratio is too large for the selected image size, instead of failing later inside OpenCV.

• 📚 Platform and deployment documentation updates

  • Documents optional immutable dataset versions for cloud training, linking trained models to the exact dataset snapshot used.
  • Updates Hailo accuracy-retention guidance for segmentation, pose, OBB, and classification.
  • Reflects the current Ultralytics Platform deployment coverage of 42 regions.

🎯 Purpose & Impact

• Faster training: MuSGD users should see substantially shorter optimizer steps and improved GPU utilization, especially on CUDA workloads with many parameters.
• Broader edge deployment: YOLO26 semantic segmentation can now be deployed directly to supported Hailo accelerators, including Raspberry Pi-oriented hardware.
• More dependable production workflows: Automatic recovery from additional CUDA allocation failures reduces avoidable training interruptions.
• Fewer dataset-related crashes: Correct handling of sparse class metadata and single-class training prevents invalid labels and CUDA assertion failures.
• More predictable exports: Earlier RT-DETR opset validation and improved calibration split handling provide clearer errors and more reliable INT8 export behavior.
• Better experiment reproducibility: Immutable dataset versions in Ultralytics Platform help ensure that models can be traced back to the precise data used for training.

What's Changed

• Document Save Dataset Version for cloud training by @glenn-jocher in #25281
• Fix GPU training failures from production logs by @glenn-jocher in #25284
• Document Hailo per-task INT8 accuracy retention by @JESUSROYETH in #25283
• Raise clean error when mask_ratio exceeds imgsz by @darshil929 in #25282
• Add Hailo semantic segmentation export and inference for YOLO26 by @JESUSROYETH in #25280
• Fix RT-DETR augmented prediction fallback by @glenn-jocher in #25289
• Validate classification calibration split by @glenn-jocher in #25292
• Relabel classification samples for single-class training by @glenn-jocher in #25290
• Validate RT-DETR ONNX opset before export by @glenn-jocher in #25291
• 8x faster MuSGD optimizer step with batched Newton-Schulz by @Y-T-G in #25288

New Contributors

• @darshil929 made their first contribution in #25282

Full Changelog: v8.4.101...v8.4.102