Compare configurations. Keep the context.
Filter inference examples, inspect throughput and latency, and compare serving settings. Explore imported optimization runs and separately labeled synthetic examples. Model, draft model, seed, and workload stay attached to every recorded trial.
Loading the bundled dataset…
Inference metrics only.
This explorer is about output-token throughput, E2E latency, and TTFT. It excludes quality scores, fine-tuning outcomes, and adaptation lift.
Recorded benchmarks use p95 E2E latency and p95 TTFT; synthetic examples use mean latency. Incomplete metadata is disclosed without removing runs. Match model, draft model, seed, workload, and known hardware before interpreting comparisons. No general engine ranking is inferred.
Inspect the measurement contract