SPLASH is a serving system for large language models that switches the parallel layout of attention while requests are running, improving serving throughput by 1.3-1.73x compared to fixed-layout deployments.