CacheLens

CacheLens finds where a prompt layout breaks prefix caching on a self-hosted vLLM server. On Meta Muse Glimmer 30B (official muse-glimmer vLLM image, one Modal A100-80GB) a synthetic 18-request workload with a per-request timestamp placed before a ~5.8k-token shared prompt got a 1.0% engine prefix-cache hit rate. CacheLens' analyzer located the first divergence at character 28 from the captured prompts. Moving the timestamp after the shared prefix raised the hit rate to 93.6% (engine counters, single run) and reduced median completion wall time from 3.69s to 2.42s. Limits: one synthetic serial run, no cache reset between variants, no TTFT/cost/answer-equivalence claims. The pre-event Oct 26 SmolLM2 prototype is disclosed in the README; the repo code and evidence were created Nov 3.