Agent skill · software engineering · nvidia
tilegym-cutile-autotuning
Use when adding, modifying, optimizing, or debugging CuTile autotuning code. Trigger signals: `exhaustive_search` / `replace_hints` / `hints_fn` / `cuda.tile.tune` in code, `autotune` in filenames, or correctness/performance issues in autotuned CuTile kernels. Covers: tune-once/cache/launch pattern, per-architecture configs (sm80–sm120), parameter space design (tile sizes, occupancy, num_ctas), and 7 common pitfalls with solutions.
Why this skill is useful
Adds executable autotuning scripts and specific patterns for optimizing CuTile kernels that the AI wouldn't generate on its own.
What it needs
Requires cuda, torch installed locally. About 9k tokens when loaded. Last updated 2026-08-07. 2,816 stars on the source repository.
What this skill does
CuTile Autotuning Add autotuning to CuTile kernels using the exhaustivesearch API with tune-once/cache/direct-launch pattern. Instructions Follow the decision tree to classify the kernel, design a search space, implement the tune-once/cache/launch pattern, and validate performance. 1. Classify — use the Decision Tree to determine search dimensions (occupancy-only vs full tile search) 2. Design search space — select the matching template from references/kernel-type-templates.md; prune to ≤ 30 configs in the final code via arch filters (directed exploration probes may temporarily exceed this — see Design Philosophy) 3. Implement — add exhaustivesearch + cache + ct.launch following the Step-by-Step Workflow; handle in-place writes with split-buffer if needed 4. Test — run correctness with autotune enabled and with DISABLEAUTOTUNE=1 5. Validate — A/B benchmark against fixed best-known config; see references/search-strategies.md 6. Shrink — prune dead-weight configs that never win, targeting ≤ 8 configs per architecture to minimize compilation cost (Step 10) Task Router — Jump to What You Need What are you trying to do? Go to --- --- Add autotune to a new kernel (most common) Quick Reference below → Workflow: Adding Autotune → references/kernel-type-templates.md (pick by kernel type: T1=elementwise, T2=in-place, T3=matmul, T4=persistent, T5=FMHA, T6=FP8, T7=grouped GEMM, T8=varlen attention, T9=dual-GEMM fusion) Debug: data corruption / wrong results after first run Pitfall #1 (In-Place Kernel) Debug: autotune taking 5+ minutes Pitfall #2 (Compilation Timeout) Debug: search space generator returning zero configs Pitfall #5 first; also check arch filters, size guards, and numctas constraints Optimize an existing autotune config Workflow: Optimizing an Existing Config Quick Reference — Occupancy-Only Autotune (Tune-Once/Cache/Launch) Most CuTile kernels (elementwise, reduction, LayerNorm) need only occupancy tuning. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills nvidia/tilegym-cutile-autotuning