gpu-optimization.

Why FlashAttention Breaks the Profiler (And Why That's Good)

FlashAttention shows low GPU occupancy yet outperforms all other attention backends. Here's what the profiler isn't telling you about modern kernel design.