tune-golden

What problem does it solve?

Manually maintaining and updating per-GPU matmul golden configuration YAMLs is error-prone, time-consuming, and often leaves stale configs that underperform compared to newly tuned greedy kernel picks. This skill eliminates that manual overhead by automating full golden dataset re-tuning, live A/B validation, and auditable update reporting.

Core Features & Use Cases

• Full Golden Dataset Re-tuning: Automatically re-tunes all matmul shapes in a target GPU's golden config set using the emmy CLI's built-in tune workflow, training the learned prior across all shapes in one pass.
• Live A/B Validation: Benchmarks the deployable greedy kernel pick against each recorded golden config in the same run to identify genuine performance wins, accounting for benchmark noise with reproducible re-runs.
• Hand-Safe YAML Updates: Edits the hand-maintained golden YAML files while preserving their required flow-style knob formatting and key order, avoiding destructive auto-serialization.
• Detailed Findings Reports: Generates structured reports analyzing prior shortfalls, knob mismatches, and workflow friction, with actionable recommendations for improving the tuning pipeline.
• Use Case: ML compiler teams maintaining emmy's GPU kernel performance can use this skill to regularly refresh golden configs, validate that deployed greedy picks outperform recorded goldens, and track systematic gaps in the search prior's performance.

Quick Start

Use the tune-golden skill to re-tune all matmul golden configs for your target GPU, validate greedy pick performance against recorded goldens, and generate a findings report of updates and performance gaps.