🌟 Summary 🚀 Ultralytics v8.4.87 delivers a cleaner, safer GPU device-selection system, plus stability and performance fixes for training, inference, tracking, exports, and dataset checks. 📊 Key Changes Clean-sheet CUDA device selection 🧭 Added parse_device() to normalize device inputs such as cuda:0, 0,1, lists/tuples, torch.device, and -1 idle-GPU auto-selection. select_device() no longer mutates CUDA_VISIBLE_DEVICES, making device selection predictable across repeated calls and long-running Python processes. Explicit single-GPU requests now use torch.cuda.set_device() instead of environment-variable remapping. Trainer, DDP setup, validation, autobatch, and distributed barriers now consistently use resolved CUDA device indices. Added documentation for ultralytics.utils.torch_utils.parse_device. Stronger GPU training tests 🧪 Added a cold-process nonzero-GPU training test to better match real CLI and Ultralytics Platform training behavior. Verifies that training on GPUs like device=1 or higher works correctly from a fresh process without relying on previous CUDA initialization. Fixed DataLoader worker cleanup at training shutdown 🧹 Added a close() method to InfiniteDataLoader. Training now explicitly shuts down persistent train and validation workers before Python exits. Helps prevent end-of-run DataLoader worker ... killed by signal: Terminated errors after results are already saved. Improved inference warmup for standard NMS ⚡ AutoBackend.warmup() now preloads torchvision for non-end-to-end models. This helps later non-max suppression calls use faster torchvision NMS when appropriate, reducing first-inference latency after warmup. Corrected dataset file-speed reporting 💾 Fixed an inverted condition in check_file_speeds(). Slow storage, such as network-mounted datasets, should now trigger the intended warning instead of being incorrectly reported as "Fast image access ✅". Tracking ReID device alignment 🎯 Trackers now pass the predictor device into ReID encoders. ReID models are initialized and run on the same device as prediction where applicable, improving consistency for tracking workflows. Export reliability improvements 📦 TensorFlow SavedModel export now distinguishes CUDA vs non-CUDA export paths more carefully. CPU exports hide TensorFlow GPUs where possible to avoid unnecessary GPU memory use. ONNX Runtime and Paddle dependency checks now better handle interchangeable CPU/GPU package variants to avoid unnecessary or conflicting installs. Paddle export now uses the actual export device to decide whether GPU Paddle is needed. 🎯 Purpose & Impact More reliable GPU behavior 🚀 Users should see fewer surprises when training, validating, predicting, or exporting repeatedly in the same Python session. This is especially important for notebooks, services, CI, distributed training, and production systems where changing CUDA_VISIBLE_DEVICES mid-process can cause hard-to-debug issues. Better support for nonzero GPU training 🖥️ Training on GPUs beyond CUDA:0 is now more robust, including cold-start CLI usage common in production and Ultralytics Platform environments. Cleaner shutdowns after training ✅ Persistent DataLoader workers are now cleaned up explicitly, reducing noisy shutdown crashes and improving confidence that completed runs exit cleanly. Lower latency after warmup ⚡ Standard detection workflows can benefit from smoother post-warmup inference performance by ensuring faster NMS paths are ready when needed. More accurate dataset diagnostics 📊 Users with slow disks or network storage will receive correct warnings, helping them identify dataset I/O bottlenecks that can slow training. More consistent tracking and export workflows 🔄 ReID tracking components now better follow the selected prediction device. Export paths are less likely to allocate unwanted GPU memory or install conflicting runtime packages. What's Changed Add cold-process nonzero-device GPU train test by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25019 Fix inverted read-speed condition in dataset file speed check by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/25025 Fix leaked dataloader workers at end of training (atexit killed by signal: Terminated crash) by @Bovey0809 in https://github.com/ultralytics/ultralytics/pull/25024 Preload torchvision during warmup for non-end2end NMS path by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/25023 Clean-sheet device selection: stop mutating CUDA_VISIBLE_DEVICES by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/25021 Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.86...v8.4.87