check batch_size % utilized_device_count

[glenn-jocher]

Bug fix to check batch_size divisibility of utilized CUDA device count vs total system CUDA device count. Partially addresses issue #2491.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved CUDA device selection logic in the YOLOv5 codebase.

📊 Key Changes

• Modified the CUDA device selection to support specifying devices via a comma-separated string.
• Ensured the batch size is divisible by the number of selected GPUs, rather than the total available GPUs.
• Updated the display of GPU properties to reflect the new device selection method.

🎯 Purpose & Impact

• Allows users more flexibility in choosing specific GPUs for training, especially on systems with multiple GPUs. 🏗️
• Guarantees the batch size is appropriately set for the number of GPUs in use, which can enhance performance. 🚀
• Enhances clarity for users when multiple GPUs are selected by displaying the respective device properties. 🖥️