$S^3$: Increasing GPU Utilization during Generative Inference for Higher Throughput | Read Paper on Bytez