Skip to content

Tilegym cutile autotuning vs Tilegym improve cutile kernel perf

Tilegym cutile autotuning is an open-source agent skill for developer tools. Tilegym improve cutile kernel perf is an open-source agent skill for developer tools.

  • Tilegym cutile autotuning is best for add autotuning to a new cutile elementwise or reduction kernel.
  • Tilegym improve cutile kernel perf is best for speed up a slow cutile attention or matmul kernel in tilegym.
What it is
Tilegym cutile autotuning is an open-source agent skill for developer tools.
Tilegym improve cutile kernel perf is an open-source agent skill for developer tools.
Best for
Add autotuning to a new CuTile elementwise or reduction kernel
Speed up a slow cuTile attention or matmul kernel in TileGym
Kind
Skill
Skill
Pricing
Open source
Open source
Platforms
Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, Goose
Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, Goose
GitHub stars
4K
4K
Last commit
Nov 2026
Nov 2026
License
Apache-2.0
Apache-2.0
Open source
Good for
Add autotuning to a new CuTile elementwise or reduction kernel; Fix wrong results in an in-place CuTile kernel after the first run; Reduce a slow autotune step that takes several minutes to compile
Speed up a slow cuTile attention or matmul kernel in TileGym; Tune autotune configs for a cuTile kernel on a Blackwell GPU; Benchmark and refine a cuTile kernel over several iterations

What Tilegym cutile autotuning does

  • Kernel type classification

    Helps decide which search dimensions a kernel needs, from occupancy-only to full tile searches.

  • Tune-once caching pattern

    Provides a template that caches the best config and compiled kernel per input shape and launches directly.

  • Per-architecture search spaces

    Covers how to design configs for different GPU architectures and keep the final search space small.

  • Common pitfall fixes

    Lists recurring autotune problems such as in-place corruption, slow compilation and recompilation on hot paths, with solutions.

  • Worked before and after examples

    Includes RMSNorm, GEMM and RoPE examples showing fixed launches converted to autotuned ones.

  • Validation and fallback guidance

    Describes A/B benchmarking against a fixed config and a DISABLE_AUTOTUNE path for testing.

What Tilegym improve cutile kernel perf does

  • Setup and baseline capture

    Creates a branch, locates the cuTile kernel and launch wrapper, and records baseline latency.

  • Bottleneck classification

    Sorts kernels into memory-bound, balanced or compute-bound using arithmetic intensity to set priorities.

  • One-change experiment loop

    Applies a single optimization per iteration, verifies correctness, and re-benchmarks before deciding.

  • Tunable parameter catalog

    Reference covering tile sizes, occupancy, TMA, persistent scheduling, num_ctas and latency hints.

  • IR dump analysis

    Guides inspecting compiled cuTile IR to compare variants and diagnose errors.

  • Progress tracking table

    Logs each iteration's latency, correctness result and keep or revert status in a Markdown table.

Tilegym cutile autotuning vs Tilegym improve cutile kernel perf: questions

Is Tilegym cutile autotuning or Tilegym improve cutile kernel perf cheaper?

Both are open source. Compare current plans on each official site.

What is the difference between Tilegym cutile autotuning and Tilegym improve cutile kernel perf?

Tilegym cutile autotuning is an open-source agent skill for developer tools. Tilegym improve cutile kernel perf is an open-source agent skill for developer tools.

Keep comparing