2 ms·Local 1M Context Inference at 15 tokens/s and ~100% "Needle In a Haystack"1 points by quxinxin 2y ago