Compute/Throughput Baseline & Regression Gate

What problem does it solve?

This Skill provides a repeatable, machine-readable gate to detect regressions in training throughput and quality, enabling CI to block merges when performance regresses beyond defined tolerances.

Core Features & Use Cases

• End-to-end gating: environment capture, deterministic micro-benchmarks, MFU estimation, and baseline comparisons keyed by machine_profile.
• CI workflow generation: templates for CI to fetch baselines, run gates, and update baselines on main.
• Use cases include measuring tokens_per_sec_p50, step_time_p50, memory, perplexity, and probe accuracy to protect production model training pipelines.

Quick Start

Run the full perf gate locally with python assets/run_template.py --bench --quality --compare to reproduce CI results.