tuning-core

What problem does it solve? GPU kernel tuning efforts frequently report speedups that never materialize in production because tuned configs fail to engage, measurements are noisy or wrong, or correctness is never gated. This Skill provides the language-agnostic discipline that prevents those failures. ## Core Features & Use Cases - Six-step tuning loop: Scope, Baseline, Search, Gate, Deploy, Verify — applicable to Triton, HIP, CK, hipBLASLt, aiter, and live vLLM/sglang servers. - Measurement rigor: Rules for synchronization, warmup, noise floors, interleaved A/B timing, and graph-captured benchmarking on gfx942/gfx950. - Correctness gating: Relative error metrics (err_ratio < 0.05) with per-dtype baselines and FP8 dialect (FNUZ vs OCP) checks. - Use Case: Before tuning a decode GEMM in a live sglang server, follow this loop to pin GPUs, measure the restart-level noise floor, prune the config space by LDS limits, gate on relative error, deploy the tuned CSV, and verify engagement via rocprofv3 kernel traces. ## Quick Start Read this skill first, then follow its six-step loop to tune my GEMM kernel and verify the tuned config actually engages in the live serving path.