# dynamo-bench
Benchmarks and trace-export entrypoints for Dynamo. Hosts:
- `multiturn_bench` — concurrent multi-turn chat benchmark against an
OpenAI-compatible endpoint, with optional speculative prefill.
- `offline_replay_bench` — Rust-native replay loop using the mocker's perf
model, for profiling replay overhead.
- `kv_router/{mooncake,active_sequences}_bench` — kv-router microbenchmarks.
- `claude_trace_export` — converts local Claude sessions into canonical Dynamo
request traces for direct replay.
## Guardrails
- Dynamo request traces replay directly through `--trace-format dynamo`; do not
add an intermediate Mooncake file or converter.
- Benchmarks here are CI-checked via clippy (`--all-targets -- -D warnings`)
and the dedicated `mooncake_trace` test under
`--features mocker-kvbm-offload`. Keep both green.
- This is a benchmark crate, not a library — keep public surface area
minimal and treat the binaries as the primary product.