tune-model

What problem does it solve?

This Skill eliminates the tedious, error-prone manual work of profiling compiled LLM kernels, benchmarking against PyTorch eager and torch.compile, and diagnosing root causes of performance gaps. Without it, developers spend hours sifting through NCU profiles, tune databases, and emitted CUDA to pinpoint why kernels underperform, and struggle to produce reproducible, actionable findings reports.

Core Features & Use Cases

• Automated Clean Autotuning: Runs reproducible, cache-free autotuning passes for single layers or full LLM models, with support for dynamic sequence length shapes to produce deployable masked-tile kernels.
• Comprehensive Benchmarking: Generates end-to-end and per-kernel benchmark tables comparing emmy performance against PyTorch eager and torch.compile, plus optional serving A/B benchmarks against vanilla vLLM for servable embedding models.
• Root-Cause Triage: Automatically classifies underperforming kernels into four failure categories (search shortfall, tier/optimization lockout, codegen quality, bench failures) with supporting evidence from NCU metrics, tune-DB rows, and emitted CUDA code.
• Structured Findings Reports: Produces standardized, reproducible reports saved to the plans/ directory, including repro commands, suggested fixes, and workflow retrospective notes for continuous improvement of the emmy compiler.
• Use Case Example: A developer optimizing the Qwen/Qwen3-Embedding-0.6B model can run a single workflow to identify that a specific attention kernel is slow due to a tensor core eligibility gate, get a pinned repro command, and receive a suggested knob adjustment to unlock tensor core performance.

Quick Start

Use the tune-model skill to autotune the first layer of the Qwen/Qwen3-Embedding-0.6B model, benchmark its kernels against PyTorch eager and torch.compile, and generate a structured performance findings report saved to the plans/ directory.