1 paper
Jinjun Yi, Zhixin Zhao, Yitao Hu +7
LLM serving is increasingly dominated by decode attention, which is a memory-bound operation due to massive KV cache loading from global memory. Meanwhile, real-world workloads exh…