causal-coverage

The paper explores CoWindow Attention (CoWA) and MassAlloc Attention (MALA) as optimized attention mechanisms for AI models, achieving substantial speedups in training and inference while reducing computational costs and maintaining performance.