Scenario A — snapshot-mode KV compression on vLLM: Δppl +29% / top-1 74% (improves top-1 by 15 pp, Δppl only by 6 pp)#17

FluffyAIcode merged 31 commits into

mainFluffyAIcode/LLM-KV--Cache-compress:mainfrom

AgentMemory/v1-3-ppl-snapshot-mode-vllm-102eFluffyAIcode/LLM-KV--Cache-compress:AgentMemory/v1-3-ppl-snapshot-mode-vllm-102eCopy head branch name to clipboard