Ultralytics YOLO

🌟 Summary

ultralytics 8.3.199 boosts startup speed with lazy model loading, refines export/runtime stability, and modernizes GPU Docker docs—delivering faster imports, smoother deployments, and clearer tooling. ⚡🐳

📊 Key Changes

• Lazy model loading for faster imports (primary)
  • Models like YOLO, SAM, RTDETR, NAS, YOLOE, FastSAM, YOLOWorld are now loaded on first access via __getattr__, preserving the public API. About ~3% faster import ultralytics.
    See PR: 3% Faster Imports with Lazy Loading (#21985) by @RizwanMunawar
• More consistent export outputs
  • Quantized export NMS wrapper now returns unpackable tensors (boxes, scores, labels, n_valid) for non-keypoint tasks; keypoint outputs unchanged.
    See PR: Fix imx object detection export outputs (#22045) by @Laughing-q
• Smarter TensorRT installation on Linux
  • Auto-installs the CUDA-matching TensorRT wheel (e.g., tensorrt-cu12) and blocks known-bad versions for more reliable exports.
    See PR: Specify CUDA version during TensorRT installation (#22060) by @Y-T-G
• Safer torch.compile defaults
  • attempt_compile() now warns on mode="max-autotune" and uses max-autotune-no-cudagraphs instead; docs updated accordingly.
    See PR: Add warning and default to no-cudagraphs (#22040) by @Y-T-G
• Clearer hyperparameter tuning plots
  • New default filters out zero-fitness points in plot_tune_results(..., exclude_zero_fitness_points=True) for cleaner visuals.
    See PR: Exclude zero-fitness points in Tuner plots (#22047) by @glenn-jocher
• Robustness fix for custom model parsing
  • Prevents undefined scale errors in parse_model() when scales isn't provided.
    See PR: Fix undefined variable in parse_model (#22054) by @Y-T-G
• GPU test coverage re-enabled
  • ONNX export with NMS for OBB re-enabled; CUDA export tests and GPU benchmarks now run when GPUs are available.
    See PR: Re-enable TensorRT export in GPU tests (#22062) by @Laughing-q
• Docker docs modernized for NVIDIA Container Toolkit
  • Replaces deprecated NVIDIA Docker approach with NVIDIA Container Toolkit; adds distro-specific install steps and standardizes --runtime=nvidia.
    See PRs: Docker Quickstart update (#21994), Standardize GPU Docker commands (#22052) by @onuralpszr
• CI reliability and maintenance
  • GPU runner updated to gpu-latest; Slack alerts now target specific failed jobs; runner image version parameterized.
    See PRs: Update GPU runner label (#22051) by @glenn-jocher, Refine Slack notifications (#22012) by @lakshanthad, Parametrize runner version in Dockerfile (#22049) by @glenn-jocher
• New reference docs for lazy imports
  • Adds a reference page explaining lazy imports in ultralytics/__init__.py.

🎯 Purpose & Impact

• Faster startup and the same API 🏎️
  • Importing Ultralytics is quicker with zero code changes. You can still do:
    • from ultralytics import YOLO
    • ultralytics.YOLO("yolo11n.pt")
• More reliable deployment pipelines 🧰
  • Standardized NMS export outputs simplify integration with ONNX/TensorRT and downstream code.
  • Correct TensorRT package selection reduces install/export friction on Linux.
• Safer compilation defaults 🛡️
  • torch.compile now prefers max-autotune-no-cudagraphs, avoiding CUDA Graphs issues while keeping performance benefits.
• Cleaner experiment insights 📈
  • Tuning plots focus on meaningful runs by default, making it easier to spot what works.
• Improved docs and GPU usability 🧪
  • NVIDIA Container Toolkit guidance and consistent --runtime=nvidia examples make GPU containers more predictable across distros.
• Better CI signal and stability 🔔
  • Targeted Slack alerts and updated runners improve reliability without affecting user-facing features.

Helpful snippets:

• Import remains the same:

  from ultralytics import YOLO
  model = YOLO("yolo11n.pt")
  

• Tuner plots with zero-fitness points visible (previous behavior):

  from ultralytics.utils.plotting import plot_tune_results
  plot_tune_results("tune_results.csv", exclude_zero_fitness_points=False)
  

• GPU Docker run examples:

  sudo docker run -it --ipc=host --runtime=nvidia --gpus all ultralytics/ultralytics:latest
  

What's Changed

• docs: 📝 Update Docker Quickstart Guide to include NVIDIA Container Toolkit by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/21994
• Fix imx object detection export outputs by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22045
• Fix Slack notifications on scheduled CI failure by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/22012
• Exclude zero-fitness points from Tuner plots by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22047
• Add warning when using mode="max-autotune" with compile by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22040
• Parameterize runner version in Dockerfile-runner by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22049
• Update ci.yml for A100 GPU DDP runners by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22051
• Re-enable TensorRT export in GPU tests by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22062
• Fix undefined variable error in parse_model() by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22054
• docs: 📝 Update Docker commands to use NVIDIA runtime for GPU support by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/22052
• Specify CUDA version during TensorRT installation by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22060
• ultralytics 8.3.199 3% Faster Ultralytics Imports with Lazy Model Loading by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/21985

Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.198...v8.3.199