mesh-llm skippy-spec-bench
Use this skill for target/draft speculative compatibility work.
What It Checks
- Target and draft tokenization agreement.
- Baseline target decode versus draft-verified decode.
- Draft acceptance/rejection behavior.
- Batched verification and checkpoint/restore behavior.
- Recurrent-state implications for rollback.
Repo Notes
The old source repo used a standalone llama-spec-bench crate. It may not be
present in this mesh checkout yet, so verify available packages before running
commands:
cargo metadata --no-deps --format-version 1 | jq -r '.packages[].name' | sort
If the spec bench is imported, keep it as a diagnostics/benchmark tool. Do not make normal mesh serving depend on it.