Files
zeekayandHanzo Dev b7e883d13b feat(engine): the throughput line reports draft acceptance
Speculative acceptance was counted (speculative::stats, bumped every verify
round by the driver) and read only by an example and a test. A served engine
never reported it, so a benchmark against one could only infer acceptance from
tokens per second — which also moves with prefill, batching and the target's
own speed.

The interval line now carries each window's share when a draft is attached:

    ..., Drafts 3.00 of 7.00 accepted per verify (42.9%), 4.00 tokens per verify, ...

"tokens per verify" adds the one token every verify step emits beyond its
accepted drafts — the target's correction, or the bonus when every draft held —
which is the figure sglang reports as its accept length, so two engines serving
one draft compare directly. The counters are cumulative; the line reports the
window, reading them every interval whether or not it logs, and a reset between
two readings is taken as the start of the window rather than wrapped.

With no draft the segment is empty and the line is exactly what it was.

Co-authored-by: Hanzo Dev <[email protected]>
2026-09-19 03:21:15 -07:00
..

hanzo-engine

Core crate of hanzo including the models and associated executors.

Documentation: https://docs.rs/hanzo-engine/