Splitwise: Efficient generative LLM inference using phase splitting

Its abstract states the two-phase structure of a request as a hardware fact: 'we find that there are two main phases during an LLM inference request: a compute-intensive prompt computation, and a mem[ory-intensive token generation]'. The paper then separates them onto different machines and notes that doing so 'requires state transfer from the machine running prompt computation over to the machine generating tokens', implemented over the cluster's fast backplane. This is the primary-source statement that prefill and decode are two different flop profiles, and the source for kv-transfer-as-state-transfer between phases. | SOURCE AUDIT 2026-10-03: the 'publisher' field held a paragraph of prose (authors + affiliations, 72 chars) instead of a publisher name; trimmed to 'arXiv' with the authorship detail moved here.