The framing that the bottleneck was memory management and not the matrix multiply is the part people miss, and the jump from roughly 20 to 38 percent KV utilization up to 96 is what makes it concrete. The OS paging analogy holds well because the win is bounding internal fragmentation to about one block per request rather than making attention itself faster. Since you are teasing SGLang next, the interesting contrast is prefix and radix caching, which attacks a different waste, the repeated prompt prefixes across requests rather than the per-request KV layout.
The framing that the bottleneck was memory management and not the matrix multiply is the part people miss, and the jump from roughly 20 to 38 percent KV utilization up to 96 is what makes it concrete. The OS paging analogy holds well because the win is bounding internal fragmentation to about one block per request rather than making attention itself faster. Since you are teasing SGLang next, the interesting contrast is prefix and radix caching, which attacks a different waste, the repeated prompt prefixes across requests rather than the per-request KV layout.