LLM systems, built from scratch and measured, not asserted.
A from-scratch KV cache for GPT-2 cuts the math per token by 1000× at long context. It makes each token only 6× faster. This post measures the gap and follows it to its source.
