ultralytics 8.4.211

v8.4.211 - Encode images up to 9% faster with compiled SAM 3 (#27158)

on Python PyPI

🌟 Summary

v8.4.211 speeds up compiled SAM image encoding and expands model deployment options, with additional gains and fixes across training, OCR, and exports.

📊 Key Changes

• 🚀 Faster compiled SAM 3 and SAM 3.1: The image encoders run 7–9% faster in the tested torch.compile setup. The optimization is specific to compiled execution; eager-mode results remain unchanged.
• 📦 Static INT8 export for ExecuTorch: Models can now be exported in a smaller INT8 format with calibration data. Detection, segmentation, pose, and oriented-box heads remain in FP32 to help preserve prediction quality.
• 📝 Faster OCR on NVIDIA GPUs: OCR can use a cached FP16 TensorRT engine, which the release reports is about 2.5× faster than ONNX Runtime on dense pages after its initial engine build. The OCR guide also clarifies which writing systems the default recognizer supports.
• ⚡ Faster deterministic training: Semantic segmentation and depth workloads use a faster resizing path when deterministic mode is enabled; measured speedups reached 3.9× for one semantic segmentation setup.
• ⚖️ Class weighting for classification: The cls_pw training option now accounts for class imbalance in classification datasets, as well as detection.
• 🖥️ Faster CPU pooling: SPPF uses an equivalent pooling approach on CPU, reducing measured block latency by 22–45% without changing GPU or exported-model behavior.
• 🛠️ Export and input fixes: TensorRT INT8 keeps additional sensitive head layers in floating point; legacy CoreML INT8 exports are fixed; and fixes address Tiger-Pose keypoint flips and classification on float NumPy inputs.
• 📚 Platform and documentation updates: Guides now cover items such as PDF prediction uploads, crowded prediction results, stopped-deployment metrics, dataset cloning limits, and PP-OCRv6 language coverage.

🎯 Purpose & Impact

• Compiled SAM users can process images faster without changing the eager-mode path.
• ExecuTorch users gain a smaller deployment option, while keeping important prediction heads at higher precision.
• OCR users on NVIDIA GPUs can reduce recognition time on dense pages, while users of unsupported scripts are less likely to mistake detections for accurate text recognition.
• Training and CPU inference improvements can reduce run time for relevant workloads; actual gains depend on the device, model, and settings.
• Export and input fixes improve reliability when preparing models for TensorRT, CoreML, and other deployment workflows.

What's Changed

• Fix broken RT-DETR TensorRT 10 INT8 engines and keep all head output layers in float like QAT by @Y-T-G in #26220
• Speed up SPPF on CPU with separable max pooling by @glenn-jocher in #26082
• Run the PP-OCRv6 recognizer as a cached FP16 TensorRT engine by @glenn-jocher in #27143
• Platform docs: Clone Images is the visitor's route; owners copy and paste by @glenn-jocher in #27144
• Platform docs: Predict tab PDFs and dense results, clone model limit, stopped deployment metrics, Gemini 4 Carbon by @pderrenger in #27145
• docs: include language and character support for PP-OCRv6 by @onuralpszr in #27155
• Let CPU CI runners save and restore zstd-compressed Actions caches by @fcakyon in #27156
• Add static INT8 quantization to ExecuTorch export by @raimbekovm in #27150
• Apply cls_pw class weights to classification training by @raimbekovm in #27148
• Speed up semantic and depth training up to 3.9x under deterministic mode by @raimbekovm in #27154
• Show which languages YOLO OCR can read (Cyrillic is not supported yet) by @glenn-jocher in #27157
• Fix Tiger-Pose flips and classification predict on float NumPy arrays by @sohailshk in #27146
• Fix legacy CoreML INT8 weight quantization by @amanharshx in #27131
• Encode images up to 9% faster with compiled SAM 3 by @glenn-jocher in #27158

Full Changelog: v8.4.210...v8.4.211