Set torch.cuda.manual_seed_all() for DDP

[glenn-jocher]

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced reproducibility in model training by ensuring consistent random seed settings.

📊 Key Changes

• Uncommented the setting of PYTHONHASHSEED to ensure a fixed hash seed.
• Reactivated CUDA random seed settings with torch.cuda.manual_seed and torch.cuda.manual_seed_all for single and multi-GPU setups.

🎯 Purpose & Impact

• 🎲 Consistency: These changes guarantee that experiments with the same seed are reproducible, which is critical for research and benchmarking.
• 🏗 Stability: Ensuring deterministic behavior improves the stability of models when trying to replicate results across different runs.
• 📈 Multi-GPU Support: By setting seeds for all GPUs, this ensures that multi-GPU training sessions behave more predictably.

[UnglvKitDe]

@glenn-jocher I looked at it. Right now I also have no idea what is missing. I will think about it.

[glenn-jocher]

@UnglvKitDe got it, thanks for looking into it! Let me know if you need any further assistance.