What problem does it solve? Tuning Triton kernels on AMD Instinct GPUs fails silently in three common ways: AMD-specific knobs (waves_per_eu, matrix_instr_nonkdim, kpack) get rejected as keyword arguments and silently dropped, a wrong autotune key serves configs tuned for different shapes (measured 3.75x slowdowns with no error), and single-pass timing reports wins that are smaller than the measurement noise. This Skill provides the correct method for each. ## Core Features & Use Cases - AMD knob placement: Shows that matrix_instr_nonkdim, kpack, and waves_per_eu must go in the first positional dict of triton.Config, verified via all_kwargs(), with measured gains (e.g. +7.7% TFLOPS for nonkdim=16 on gfx942). - Autotune key correctness: Explains that key= must name every dimension that changes the best config, with measured decode-shape penalties when M is omitted. - Search space pruning and regime splitting: Provides an analytical viability predicate (LDS limits, thread utilization) and shows 34-44% gains from tuning decode and general regimes separately. - Noise-aware measurement: Requires independent repeats with median and spread, refusing to claim gains smaller than the noise floor. - Use Case: You are optimizing a Triton GEMM for LLM decode (M=1) on an MI300X. Use this Skill to build a decode-specific config space with AMD knobs, set key=['M','N','K'], clear the Triton cache, and confirm the winner beats torch.mm beyond the spread. ## Quick Start Ask the AI to tune your Triton kernel on AMD Instinct by building a pruned autotune space with the AMD knobs in the positional config dict, a correct key, and noise-aware re-measurement of the winner.
