ultralytics 8.4.105

v8.4.105 - Fuse distillation checkpoints to the student model (#25415)

on Python PyPI

🌟 Summary

v8.4.105 improves reliability for distillation checkpoints, expands YOLO26 deployment options, and enhances CoreML, mobile, Hailo, QNN, Platform, and documentation workflows.

📊 Key Changes

• 🛠️ Fixed distillation checkpoint crashes by adding fuse() delegation to DistillationModel. Checkpoints saved during or after interrupted training can now be used with yolo export, predict, and val without failing because the training-only wrapper lacks a fusion method. By @glenn-jocher.
• 🚀 Added YOLO26 depth deployment support for Hailo and Qualcomm QNN, enabling depth estimation on Hailo-8L and Snapdragon devices. Depth outputs are decoded into metric depth maps on the host.
• 📱 Standardized mobile model input sizes:
  • Classification: 224
  • Detection, segmentation, semantic segmentation, depth, pose, and OBB: 640
• 🍎 Improved CoreML dynamic-shape exports for YOLO26 segmentation and depth models by replacing unsupported dynamic-size upsampling with fixed scale-based operations.
• ⚙️ Improved AutoBatch with multi-scale training by aligning memory-profiling sizes to valid model strides, preventing silent fallback or shape errors.
• 🧠 Added opt-in CUDA channels_last support for training and native PyTorch prediction, providing reported throughput improvements while preserving accuracy. Standard NCHW behavior remains unchanged by default.
• 🔧 Improved model fusion and quantization reliability:
  • Fixed convolution fusion for channels_last weights.
  • Corrected ConvTranspose2d + BatchNorm fusion along the proper weight axis.
  • Preserved DFL operations in floating point during OpenVINO INT8 quantization.
  • Fixed CUDA calibration failures caused by inference tensors.
• 📦 Expanded export compatibility and documentation with more accurate opset and simplify options for ONNX-backed formats such as MNN, RKNN, QNN, DEEPX, Edge TPU, and TensorFlow exports.
• 🧪 Improved testing and performance internals:
  • Added targeted depth-loss fuzz combinations.
  • Removed redundant WorldDetect feature-map cloning.
  • Vectorized tracker Kalman-filter operations.
  • Reduced heatmap memory usage and removed unnecessary validator device transfers.
• 🧰 Improved training and dataset workflows:
  • Fixed RMSprop incorrectly selecting MuSGD.
  • Fixed single-class depth training.
  • Added class-name aliases for better fine-tuning weight transfer.
  • Consolidated dataset label-cache logic across detection, depth, and semantic datasets.
• 🌐 Expanded Ultralytics Platform capabilities and documentation:
  • Added clearer YOLO26 depth API responses and configurable 8-, 12-, or 16-bit depth maps.
  • Improved annotation with automatic saving and YOLO-assisted predictions.
  • Added activity export, updated billing and team behavior, and clarified deployment and API limits.
  • Updated Platform documentation and examples to reflect current behavior.

🎯 Purpose & Impact

• ✅ More checkpoints remain usable: Interrupted or mid-training distillation checkpoints can now flow through validation, prediction, and export normally.
• 🚀 Broader edge deployment: YOLO26 depth models can run on additional accelerators, including Hailo hardware and Snapdragon NPUs, making on-device depth estimation more practical.
• 📈 Better performance potential: CUDA channels_last, tracker vectorization, lighter heatmap storage, and optimized loss calculations can improve throughput or reduce memory and CPU overhead.
• 🧩 More dependable exports: CoreML dynamic exports, model fusion, calibration, and quantization are less likely to crash or produce incorrect results.
• 📱 Simpler mobile integration: Consistent input-size guidance and refreshed CoreML, LiteRT, and QNN documentation reduce deployment ambiguity.
• 🧪 Safer future changes: Expanded fuzz coverage and cleaner shared dataset infrastructure improve regression detection and maintainability.
• ℹ️ User action: Existing users should update with pip install -U ultralytics to receive the checkpoint, export, and deployment fixes.

What's Changed

• Fix CoreML dynamic=True segment export upsample mapping error by @glenn-jocher in #25347
• Document YOLO26-Depth task in Platform docs and API reference by @glenn-jocher in #25349
• Fix CoreML dynamic=True depth export bilinear upsample mapping error by @glenn-jocher in #25350
• Stride-align AutoBatch probe size to true multi-scale maximum by @glenn-jocher in #25352
• Document depth predict bits parameter and model-resolution maps by @glenn-jocher in #25354
• Advertise Platform after export failures by @glenn-jocher in #25353
• Update fuzz coverage for depth task by @glenn-jocher in #25357
• fix: align ONNX-backed export arguments by @amanharshx in #25356
• Keep the DFL float during OpenVINO INT8 quantization by @synml in #25306
• Delete duplicate Python setup from CI by @glenn-jocher in #25366
• Fix conv/bn fusion crash on channels_last weights by @raimbekovm in #25363
• Store the Heatmap accumulator single-channel by @JESUSROYETH in #25342
• Return the validator correct matrix with torch.from_numpy instead of a device round-trip by @JESUSROYETH in #25344
• Drop the redundant per-forward feature-map clone in WorldDetect.forward by @JESUSROYETH in #25345
• Vectorize the per-frame multi_predict and multi_gmc track loops by @JESUSROYETH in #25343
• Add gated channels_last memory format for CUDA train and predict by @raimbekovm in #25362
• Refactor dataset label caching into shared template with per-class hooks by @Laughing-q in #25359
• Fix fuse_deconv_and_bn scaling the wrong weight axis for ConvTranspose2d by @raimbekovm in #25364
• Fix Model.calibrate() inference-tensor crash on CUDA by @JESUSROYETH in #25351
• Compute BboxLoss weights and pose keypoint strides on matched anchors only by @JESUSROYETH in #25346
• Add Hailo export and inference for the YOLO26 depth task by @JESUSROYETH in #25348
• Collapse conv/bn fusion to broadcast form and delete input-side channels_last conversions by @glenn-jocher in #25367
• Fix depth training with single-class mode by @glenn-jocher in #25370
• Build Docker runner before export image by @glenn-jocher in #25371
• Point README banner link to YOLO Vision 2026 registration by @raimbekovm in #25377
• Add https://youtu.be/7d1XWgh1pmk to docs by @RizwanMunawar in #25374
• Fix optimizer=RMSprop silently building MuSGD by @raimbekovm in #25376
• Point remaining banner links to YOLO Vision 2026 registration by @raimbekovm in #25381
• Fix reCamera links and refine core contribution principles by @onuralpszr in #25380
• docs: clarify YOLOE export workflow for C++ ONNX Runtime inference by @23z156-wq in #25373
• Ruff 0.16.0 compatibility changes by @glenn-jocher in #25390
• Threshold SAM3 semantic masks at the model logit threshold by @Y-T-G in #25386
• Add CPU ONNX and T4 TensorRT10 speed columns to depth perf table by @Bovey0809 in #25382
• Preserve falsey activation arguments by @glenn-jocher in #25391
• Resolve MuSGD only through the explicit optimizer path by @raimbekovm in #25383
• Gather end2end top-k results with a zero-copy expand instead of a materialised repeat by @JESUSROYETH in #25369
• Add cross-dataset class-row weight transfer for RT-DETR by @artest08 in #24423
• Add stable Platform screenshot markers by @glenn-jocher in #25388
• Add T4 TensorRT speed comparison against Depth Anything V2 to depth docs by @raimbekovm in #25399
• Fix shared export table rendering by @glenn-jocher in #25401
• Standardize mobile exports at 224/640 by @glenn-jocher in #25387
• Fix yolo26-seg semantic target labelling empty images as class 0 when overlap_mask=False by @JESUSROYETH in #25392
• Build TAL get_box_metrics gather indices on device by @raimbekovm in #25397
• fix(coreml): record NMS effective precision by @amanharshx in #25389
• Vectorize GMC point filter loops with boolean masks by @raimbekovm in #25398
• Fix gmc docstring examples to run as written by @raimbekovm in #25402
• Make GMC outlier filter reject deviations on both sides of the mean by @raimbekovm in #25404
• Pad all multispectral channels in RandomPerspective warps by @JESUSROYETH in #25411
• Fix TRACKTRACK.update IndexError on a single torch-backed detection by @JESUSROYETH in #25394
• Fix intel:gpu device alias resolution on multi-GPU OpenVINO machines by @JESUSROYETH in #25393
• Bake the semantic argmax class map into the OpenVINO export graph by @JESUSROYETH in #25395
• fix: derive export argument validation from format table by @amanharshx in #25368
• Document class-alias weight transfer for fine-tuning by @artest08 in #25409
• Add colorization options for depth maps in documentation by @onuralpszr in #25400
• Validate pose, kobj, rle, angle and erasing at the cfg layer by @usegitdivyansh in #25384
• Fix NCNN inference dropping all but the first image of a batch by @JESUSROYETH in #25412
• Skip the redundant float32 input copy in the OpenVINO and Paddle backends by @JESUSROYETH in #25413
• Remove redundancy from AGENTS.md Core Principles by @glenn-jocher in #25414
• Fuse distillation checkpoints to the student model by @glenn-jocher in #25415

New Contributors

• @23z156-wq made their first contribution in #25373
• @usegitdivyansh made their first contribution in #25384

Full Changelog: v8.4.104...v8.4.105