Anyone else tried out KV cache blending?

Idea is simple-ish in abstract: instead of running normal prefill over all of a given prompt, split it into parts - generates caches for part A and part B in isolation, concatenate the result, feed it into decode like normal. I honestly thought it'd totally fail. But I've been trying it out on Ling3

中文全文翻译正在补齐，可先查看下方英文原文。

Idea is simple-ish in abstract: instead of running normal prefill over all of a given prompt, split it into parts - generates caches for part A and part B in isolation, concatenate the result, feed it into decode like normal. I honestly thought it'd totally fail. But I've been trying it out on Ling3