ultralytics 8.3.189

v8.3.189 - `ultralytics 8.3.189` 3x faster inplace `model.fuse()` (#21844)

on Python PyPI

🌟 Summary

ultralytics 8.3.189 delivers a 3x faster in-place model.fuse() for leaner, quicker inference, plus better NVIDIA Jetson handling with version-aware detection and more robust autobackend behavior. 🚀

📊 Key Changes

• In-place layer fusion (primary change) ⚡

  • Conv/BatchNorm and Deconv/BatchNorm fusion now happens in-place, updating existing layers directly with gradients disabled.
  • Smarter bias handling (including transposed conv): correctly registers or updates fused bias without creating new layer objects.
  • Result: faster fuse(), lower memory use, and cleaner models. See PR “3x faster inplace model.fuse()” by @glenn-jocher.

• Jetson-aware autobackend improvements ✅

  • On Jetson with JetPack 5: move the model to device before fuse() to avoid runtime issues; other devices fuse first, then move to device.
  • attempt_load_weights now performs fuse().eval() before a single, final .to(device) for consistency.
  • is_jetson is imported and cached for faster repeated checks. See PR “Autobackend model.fuse() order of operations speedup” by @Laughing-q.

• Version-aware Jetson detection 🔍

  • is_jetson(jetpack=4|5|6) now checks JetPack versions via /etc/nv_tegra_release with an internal L4T mapping and returns False on failure.
  • Backward compatible when called without arguments; results cached with lru_cache. See PR “Add Jetpack4, 5 and 6 version checks” by @glenn-jocher and @Laughing-q.

• Version bump

  • version → 8.3.189.

🎯 Purpose & Impact

• Faster inference setups and smaller memory footprint 🚀

  • In-place fuse() eliminates extra allocations and object creation, accelerating model preparation and inference.

• More reliable Jetson deployments, especially on JetPack 5 🧩

  • Adjusted fuse/device order fixes fusion-related errors on edge devices, improving stability in production pipelines.

• Easier, precise hardware targeting 🛠️

  • Version-aware is_jetson() enables conditional logic by JetPack version (e.g., selecting models, dependencies, or optimizations) with a simple call.

• Backward compatible and safer defaults ✅

  • Existing usage of model.fuse() and is_jetson() continues to work, now faster and more robust.

Quick examples:

• Faster in-place fusion

  from ultralytics import YOLO
  
  model = YOLO('yolov8n.pt').model
  model = model.fuse()  # now runs ~3x faster and in-place

• Jetson version checks

  from ultralytics.utils import is_jetson
  
  if is_jetson():        # Any Jetson
      ...
  if is_jetson(5):       # Specifically JetPack 5
      ...

Happy building and deploying! ✨

What's Changed

• Add Jetpack4, 5 and 6 version checks by @glenn-jocher in #21854
• Autobackend model.fuse() order of operations speedup by @Laughing-q in #21853
• ultralytics 8.3.189 3x faster inplace model.fuse() by @glenn-jocher in #21844

Full Changelog: v8.3.188...v8.3.189