This document is the current public evidence boundary for Adapter Cache Tradeoffs.
E001 is preregistered but has not been canonically run on SGLang. The repository currently supports local smoke verification, dry-run budget checks, historical evidence review, and OpenAI-compatible serving integrations.
simulated_cached_prefix_tokensandsimulated_cached_prompt_tokensare benchmark-side estimates. They are not server prefix-cache counters.server_prefix_cache_*fields are server-level counters when the runtime exposes them. They are not per-request adapter-cache measurements.- The mock backend is deterministic systems scaffolding. It is not model quality evidence.
- Whitespace tokenization approximates prefix blocks and memory footprint. It does not match exact runtime tokenization or byte-accurate KV allocation.
- Historical vLLM results remain useful prior evidence, but some predate the
current run lifecycle contract requiring
status.json. - New public claims should come from current-contract runs with raw request records, resolved configs, manifests, summaries, and deterministic analysis.
- The E001 SGLang run, Vidur policy screening, and vLLM replication are pending.
- Quality scores from fixture-style evals are bounded to those fixtures and should not be described as standard public benchmark performance.
- This repository does not vendor SGLang, Vidur, or vLLM.
- It does not claim production serving recommendations across all models, adapter ranks, context lengths, GPUs, or traffic mixes.
- Activated-LoRA-style cache behavior is simulator-side unless a serving runtime exposes matching kernel/cache-key behavior.
- Cost and power frontiers are not measured here.