ultralytics/ultralytics v8.3.220

v8.3.220 - `ultralytics 8.3.220` Add ExecuTorch export (.pte) with XNNPACK (#22244)

on GitHub

🌟 Summary

ExecuTorch export lands! You can now export YOLO11/YOLO26 models to .pte with XNNPACK acceleration for fast, native on-device inference on mobile and edge devices. Plus, a stability fix for CLIP embedding extraction in VisualAI. 📱⚡🧠

📊 Key Changes

• ExecuTorch export (primary update)
  • New export target: format=executorch → outputs a .pte model inside a <name>_executorch_model/ directory
  • Integrated into export_formats() and the exporter flow
  • New export_executorch() method using torch.export, XnnpackPartitioner, and to_edge_transform_and_lower
  • Saves model metadata alongside the .pte file for easy integration
  • Backend/runtime support added in AutoBackend for .pte loading and inference dispatch
  • Docs: New ExecuTorch integration guide and export tables updated
  • Tests: Added CLI and matrix export tests (skipped on Windows)
• VisualAI stability improvement
  • Safely detaches CLIP embeddings before converting to NumPy to avoid autograd errors and reduce memory overhead

Minimal usage:

• CLI: yolo export model=yolo11n.pt format=executorch
• Python:

  from ultralytics import YOLO
  YOLO("yolo11n.pt").export(format="executorch")

Helpful links:

• Learn more in the ExecuTorch integration docs
  • ExecuTorch overview and usage: Deploy YOLO11 with ExecuTorch
  • ExecuTorch project: Meta’s ExecuTorch

🎯 Purpose & Impact

• Mobile and edge deployment made easy 🚀
  • One-line export to .pte for running models on iOS, Android, and embedded Linux devices via ExecuTorch
  • Uses XNNPACK by default for optimized CPU performance on mobile
  • Early benchmarks on devices like Raspberry Pi indicate notable speedups (around 2x in example tests), with accuracy parity
• Streamlined developer workflow ✅
  • Feels like any other Ultralytics export target; consistent UX and metadata packaging
  • Clear docs and examples for integrating on mobile and embedded platforms
• Compatibility notes 🧩
  • Requires Python 3.10+, PyTorch ≥ 2.9, executorch and flatbuffers installed
  • Export writes both .pte and metadata.yaml; inference via ExecuTorch runtime (not standard Python YOLO() loading)
• Improved stability for VisualAI 🔒
  • Detaching CLIP embeddings prevents gradient-related errors and reduces memory use during similarity search

Overall, this release expands deployment options for YOLO11/YOLO26—especially for mobile and edge users—while improving robustness in VisualAI workflows. 🎉

What's Changed

• Fix VisualAI clip feature extraction numpy() grad error by @RizwanMunawar in #22463
• ultralytics 8.3.220 Add ExecuTorch export (.pte) with XNNPACK by @ambitious-octopus in #22244

Full Changelog: v8.3.219...v8.3.220