ultralytics 8.3.208

v8.3.208 - `ultralytics 8.3.208` Expand FP16 export support (#22316)

on Python PyPI

🌟 Summary

Expands safe FP16 (half-precision) export support across ONNX and TorchScript, improves NCNN export reliability by switching to the ONNX pipeline, and delivers faster RT-DETR inference via anchor caching — plus clearer docs and smoother dependency handling. 🚀

📊 Key Changes

• FP16 export behavior (PR #22316)

  • Extends FP16 handling to both ONNX and TorchScript exports.
  • Enforces GPU-only rule for FP16; if half=True on CPU, it now warns and automatically uses half=False.
  • Expands tests to cover ONNX FP16 combinations for higher reliability.
  • Docs updated to clearly state FP16 constraints (not compatible with INT8 or CPU-only export). ⚠️

• NCNN export pipeline update (PR #22315)

  • NCNN export now uses ONNX → PNNX instead of TorchScript → PNNX for better compatibility and stability.
  • Automatically triggers ONNX when format='ncnn'. TorchScript is only produced if explicitly requested. 🔄

• RT-DETR performance optimization (PR #22318)

  • Caches decoder anchors and masks; regenerates only if shapes change or dynamic=True.
  • Reduces redundant computation to speed up training/inference. ⚡

• Smarter dependency checks with interchangeable packages (PR #22321)

  • check_requirements() supports alternatives like ("onnxruntime", "onnxruntime-gpu"), making installs more flexible. ✅
  • Internal usage updated (e.g., ONNX benchmarking) to accept either CPU or GPU variants.

• Validation docs clarity (PR #22319)

  • Clarifies that validation uses model.names (the model’s class set), not the dataset YAML classes. 📝

• Better examples for check_requirements usage (PR #22317)

  • Shows passing custom pip args (e.g., PyTorch CPU index) and updates version constraints. 🧰

• CI and maintenance improvements

  • Use missing_ok=True when cleaning INT8 TensorRT cache files in tests to avoid edge-case errors (PR #22320).
  • Bump astral-sh/setup-uv from v6 to v7 for GitHub Actions (PR #22327).
  • Tidy Dependabot config and labels; remove deprecated reviewer field (PR #22322). 🧹

🎯 Purpose & Impact

• Safer, clearer FP16 exports on GPU only

  • Prevents confusing or invalid CPU FP16 exports by auto-correcting and warning.
  • Users get predictable, faster FP16 ONNX/TorchScript exports on supported hardware. 💨

• More reliable NCNN exports

  • ONNX-based NCNN conversion reduces TorchScript-related issues and simplifies the export path.
  • Improves portability and success rate across platforms. 📦

• Faster RT-DETR workflows

  • Anchor/mask caching trims unnecessary computation, accelerating common inference loops. ⚙️

• Smoother installs and environment setup

  • Interchangeable package support (e.g., onnxruntime vs onnxruntime-gpu) reduces friction across CPU/GPU setups.
  • Updated examples help users configure installs for their hardware quickly. 🧑‍💻

• Reduced test flakes and cleaner maintenance

  • More robust tests and streamlined CI/dependency settings lower maintenance noise without user-facing changes. ✅

Quick examples:

• Export ONNX in FP16 on GPU:

from ultralytics import YOLO

YOLO("yolo11n.pt").export(format="onnx", half=True, device=0)  # ✅ GPU FP16

• Attempting FP16 export on CPU will auto-disable FP16 and warn:

YOLO("yolo11n.pt").export(format="onnx", half=True, device="cpu")  # ℹ️ Will warn and use half=False

• Check requirements with interchangeable packages:

from ultralytics.utils.checks import check_requirements

check_requirements([("onnxruntime", "onnxruntime-gpu"), "numpy"])

What's Changed

• Show check_requirements pip install extras example by @glenn-jocher in #22317
• Add missing_ok=True to TensorRT INT8 cache cleanup test by @glenn-jocher in #22320
• Fix validation data argument and class names description by @jb297686 in #22319
• Drop deprecated dependabot fields by @Borda in #22322
• Interchangeable packages feature for check_requirements() by @glenn-jocher in #22321
• Cache anchors in RTDETR head to avoid repeated initialization by @Y-T-G in #22318
• Bump astral-sh/setup-uv from 6 to 7 in /.github/workflows by @dependabot[bot] in #22327
• Create NCNN from ONNX model instead of TorchScript by @Y-T-G in #22315
• ultralytics 8.3.208 Expand FP16 export support by @glenn-jocher in #22316

New Contributors

• @jb297686 made their first contribution in #22319
• @Borda made their first contribution in #22322

Full Changelog: v8.3.207...v8.3.208