Performance Guide

Benchmark code must protect measured regions from avoidable overhead.

Core practices:

• Prepare models, inputs, and backend state outside measured loops.
• Use monotonic high-resolution clocks.
• Separate warmup runs from measured runs.
• Collect heavy profiler data only when requested.
• Generate reports after measurement completes.
• Preserve raw observations so metric algorithms can evolve without rerunning workloads.