performance-optimization

---

name: performance-optimization description: Optimize training performance for Kaggle competitions. Use PROACTIVELY when training is slow, GPU memory issues occur, or user asks about performance tuning. Keywords: 遅い, 最適化, OOM, CUDA, batch_size, num_workers, 高速化, パフォーマンス

When to Use (PROACTIVE)

This skill should be activated automatically when:

• User reports slow training
• CUDA Out of Memory (OOM) errors occur
• User asks about batch_size or num_workers tuning
• User mentions: "遅い", "最適化", "OOM", "高速化"
• Training speed is suboptimal

What This Skill Does

Provides GPU-specific optimization strategies:

• Batch size and num_workers recommendations by GPU
• Mixed Precision Training implementation
• Gradient Accumulation patterns
• Memory optimization techniques
• Parallel processing patterns

How to Use

Refer to gpu-tuning-guide.md for:

• GPU-specific hyperparameter recommendations
• Mixed precision training code
• Gradient accumulation implementation
• Memory profiling techniques
• Common performance bottlenecks and solutions