tilegym-cutile-autotuning — Claude Code skill: tune CuTile kernel

What tilegym-cutile-autotuning does

The skill splits the job into stages: classify the kernel, pick candidate configs, call exhaustive_search from cuda.tile.tune, store the winning setup, and skip the search afterwards with a plain ct.launch. Templates cover elementwise ops, in-place kernels, matmul, persistent kernels, FMHA, FP8, grouped GEMM and varlen attention.

A task router points to the right section quickly, and the pitfalls list flags the mistakes that bite most often. Architecture-specific settings run across sm80–sm120, with tile sizes, occupancy and num_ctas treated as separate search dimensions. One trap matters more than the rest: rebuilding the kernel with replace_hints on every call forces a recompile, so both the config and the compiled kernel belong in the cache.

Simple occupancy-only kernels are handled by copying the quick pattern. The heavier cases need the reference files in references/. Tuning runs on a real GPU, so a CUDA setup with PyTorch has to be in place first. Test twice: once with tuning on, once with DISABLE_AUTOTUNE=1, because a broken in-place kernel shows up only in the second run.