Ultralytics v8.4.154 Released
Ultralytics v8.4.154 improves CoreML export and inference reliability, restores accurate RT-DETR OpenVINO INT8 deployment, accelerates YOLO26 training, reduces SAM3 memory usage, and strengthens dataset validation and Platform workflows.
Explore the v8.4.154 release for the complete release details.
Highlights
Reliable dynamic CoreML export and static multi-image inference
Faster Ultralytics YOLO training with reduced GPU overhead
Accurate RT-DETR OpenVINO INT8 export
Memory-bounded SAM3 semantic mask processing
Safer classification, segmentation, and pose validation
Platform training callbacks powered by the Platform SDK
New Features and Improvements
Faster YOLO26 Training
Training now avoids unnecessary memory initialization, activation copies, host-device synchronization, and repeated EMA state reconstruction. Fused Adam and AdamW optimizers are also enabled where supported.
In one measured environment, a YOLO26x COCO training step on a B200 GPU improved from 228.9 ms to 183.5 ms, representing approximately a 1.25× speedup. Results may vary by hardware and workload.
This work was contributed by @TarzanZhao in PR #26180.
Lower-Memory SAM3 Mask Processing
Large SAM3 semantic masks are now upscaled in bounded chunks instead of being processed all at once. This avoids multi-gigabyte temporary allocations while preserving mask results.
This improvement was contributed by @BSchilperoort in PR #26200.
Platform SDK Training Callbacks
Platform training callbacks now use the generated Platform SDK rather than duplicated HTTP and retry logic. The update adds controlled POST retries while preserving authentication, cancellation, checkpoint signing, payload sanitization, and quiet console-error behavior.
This integration requires ultralytics-platform>=0.1.45 and was implemented by @JaviChulvi in PR #26190.
Bug Fixes
CoreML Dynamic Export and Batch Inference
Detection, segmentation, pose, and OBB models can now export with dynamic=True without triggering the coremltools arange conversion error.
Static CoreML models also correctly process batches containing multiple images, with proper output stacking for raw predictions, embedded NMS, segmentation, and classification models.
The fix and v8.4.154 release update were delivered by @raimbekovm in PR #26199.
RT-DETR OpenVINO INT8 Accuracy
RT-DETR OpenVINO INT8 export now keeps the decoder in floating point while applying NNCF transformer-aware quantization.
Reported RT-DETR-L accuracy improved from approximately 0.0002 to 0.6513 mAP50-95, with nearly unchanged CPU inference speed.
The fix was contributed by @onuralpszr in PR #26191.
Correct Classification Class Mapping
Classification validation and training splits are now aligned with the model’s class names rather than relying on each directory’s local alphabetical ordering. Classes absent from the model are skipped with a warning, preventing silently incorrect metrics and model-selection results.
This correction was contributed by @cainiao33 in PR #26192.
Clearer Pose Validation Errors
Standalone pose validation now reports an actionable error when kpt_shape is missing, including cases involving stale label caches.
The validation improvement was contributed by @cainiao33 in PR #26193.
Dataset and Documentation Updates
Dataset checks and documentation now cover:
- Rejection of box-only labels and mismatched polygon-to-box counts in segmentation datasets
- Small-, medium-, and large-object mAP reporting when detection validation uses
save_json=True - Resume behavior, where the checkpoint dataset is restored unless an explicit
data=override is supplied - CoreML restrictions involving dynamic inputs, NMS, classification, RT-DETR, and batch sizes
These updates were contributed by @raimbekovm in PR #26198.
Platform documentation also adds guidance for semantic PNG mask imports, similar-image search, generated image variations, moving models between projects, remembered training settings, verified dataset uploads, API rate limits, dataset slugs, and upload integrity checks.
These workflow updates were contributed by @raimbekovm in PR #26195.
New Contributors
A warm welcome and thank you to our first-time contributors:
- @TarzanZhao, who contributed the training performance improvements in PR #26180
- @BSchilperoort, who contributed memory-bounded SAM3 mask processing in PR #26200
Upgrade Today
Install or upgrade to the latest release with:
pip install -U ultralytics
Please try v8.4.154 and share your results, benchmarks, and feedback with the Ultralytics YOLO community. You can review every change in the v8.4.153 to v8.4.154 full changelog. 