What problem does it solve? Optimizing GPU kernels in the Flash Linear Attention (FLA) library over multiple iterations risks silently breaking correctness, gaming benchmarks, or losing track of what was tried. This Skill imposes a disciplined loop — frozen test contract, three-phase plan, one-change-per-iteration protocol, and evidence logging — so speedups on fla/ops/** kernels are real, reproducible, and promotable. ## Core Features & Use Cases - Frozen correctness gate: Treats the op's pytest file and naive reference as immutable, banning tolerance loosening, shape dropping, and one-sided numeric-flag relaxation, with NaN-poisoned memory catching partial writes. - Structured iteration protocol: Enforces one change per iteration followed by gate run, benchmark, log entry, and commit, with stall handling after three non-improving rounds and explicit no-go criteria. - Evidence and trap catalogs: Provides references/TRAPS.md (TF32 reference inflation, autotune staleness, clock drift, int64 overflow) and log templates for iteration rows, kept-candidate headers, and shape-dispatch justification. - Use Case: While making chunk_gla faster in Triton, you write a task contract, profile with NCU, try three candidate directions, log each result in OPT_LOG.md, and promote only the candidate with a full green gate and a profiler-explained win. ## Quick Start Ask the agent to optimize the FLA operator chunk_gla for a 1.2x speedup on the target shapes using this optimization loop, starting with a task contract draft and the frozen verify gate.
