ultralytics 8.3.204

v8.3.204 - `ultralytics 8.3.204` Scope `batch_size` check from `select_device()` to Trainer (#22254)

on Python PyPI

🌟 Summary

YOLO 8.3.204 sharpens training ergonomics, stabilizes exports, broadens device support, and refreshes docs—making multi-GPU projects, ONNX workflows, and edge deployments smoother than ever. 🚀

📊 Key Changes

• 🔄 Trainer batch handling upgrade: The primary update from @Laughing-q moves multi-GPU batch checks out of select_device() and into the Trainer, requiring explicit batch sizes when using more than one GPU and delivering clearer error guidance.
• 🧠 YOLOE segmentation workflow: New YOLOESegTrainerFromScratch, cleaner loss initialization, and shared preprocessing streamline from-scratch and fine-tune training paths.
• 📏 Export & device refinements: CUDA-aware ONNX opset selection (fewer runtime errors), TensorFlow/ONNX NMS guards, image-size propagation during export, plus safer non_blocking transfers on MPS.
• 📦 Lean inference weights: strip_optimizer() now removes AMP scaler state, trimming checkpoint clutter.
• 🌐 Streamlit edge support: RKNN/OpenVINO detection works reliably in the live inference app, benefiting Rockchip deployments.
• 📚 Docs & metadata refresh: New YOLO26 preview page, updated YOLO11 FLOPs/parameters, clarified YAML guide, and accurate contributor credits.

🎯 Purpose & Impact

• ✅ More predictable multi-GPU runs: Users must set a real batch size, preventing silent defaults and helping balance workloads across GPUs.
• 🛡️ Stabler model exports: GPU ONNX exports throw fewer errors, NMS requirements are explicit, and exported engines now honor your chosen imgsz, reducing mismatches in production.
• ⚙️ Broader device compatibility: Adjusted tensor transfers eliminate MPS corruption, while RKNN support speeds up Rockchip deployments via Streamlit.
• 🧪 Enhanced YOLOE training flexibility: Dedicated trainers and cleaned-up loss logic make experimenting with YOLOE segmentation (YOLO11-Seg, etc.) easier and more torch.compile-friendly.
• 📊 Cleaner release assets: Stripped checkpoints load faster for inference without lingering AMP baggage.
• 📖 Up-to-date guidance: YOLO26 sneak peek and refreshed YOLO11 metrics help teams plan upgrades with accurate performance data.

What's Changed

• Fix --slow CUDA tests by @glenn-jocher in #22193
• Verify passed dataset is a directory and not a file for classification training by @Y-T-G in #22192
• Expand PyTorch version text matrix by @glenn-jocher in #22152
• Fix AttributeError error for validation with visualize by @Y-T-G in #22194
• Fix --slow flag in ci.yml pip install command by @Laughing-q in #22196
• Remove uv version 0.8.19 restraint by @Laughing-q in #22197
• Disable non_blocking device transfers for MPS to prevent output corruption by @Y-T-G in #22181
• Enable Python Fault Handler in CI by @Y-T-G in #22203
• Remove the compromise of criterion initialization for torch.compile by @Laughing-q in #22205
• YOLOE: Cleanup duplicate preprocess_batch by @Laughing-q in #22206
• Strip scaler from model checkpoints by @Y-T-G in #22207
• Docs: Add new page for YOLO26 by @Laughing-q in #22213
• Add RKNN model support in streamlit live-inference solution by @RizwanMunawar in #22212
• YOLOE: Use passed imgsz for visual prompt extraction by @Y-T-G in #22238
• Fix RKNN model support in streamlit live-inference solution by @lakshanthad in #22230
• Docs: Update YOLO11 profiling data in README by @lmycross in #22258
• Add Mengyu to mkdocs_github_authors.yaml by @glenn-jocher in #22263
• Fix Docs YAML guide missing docstrings by @glenn-jocher in #22264
• ultralytics 8.3.204 Scope batch_size check from select_device() to Trainer by @Laughing-q in #22254

New Contributors

• @lmycross made their first contribution in #22258

Full Changelog: v8.3.203...v8.3.204