skills.rest
yo-steven
yo-stevenCOMMUNITY
Explore all skills by this author →
GitHub

converting-cutile-to-triton

Convert cuTile @ct.kernel GPU kernels to Triton @triton.jit implementations.

Updated May 23, 2026
Search Tags:#attention#triton#performance validation#autotuning#tma#cutile#gpu kernel conversion
One-click install
npx skills add https://github.com/yo-steven/skills-exploration-20260522 --skill converting-cutile-to-triton-yo-steven
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: converting-cutile-to-triton
Source: https://github.com/yo-steven/skills-exploration-20260522/tree/main/skills/TileGym/converting-cutile-to-triton
Command: npx skills add https://github.com/yo-steven/skills-exploration-20260522 --skill converting-cutile-to-triton-yo-steven

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

Converting cuTile GPU kernels (@ct.kernel) into Triton (@triton.jit) safely and efficiently is error-prone, especially when correctness breaks or performance regresses due to incorrect TMA usage, dtype handling, or layout-specific logic.

Core Features & Use Cases

  • Deterministic conversion workflow with gates: enforces an analyze → convert → validate → test → benchmark sequence with explicit stop conditions before perf sign-off.
  • cuTile → Triton API translation guidance: maps common constructs (ct.load/store, ct.mma, ct.Constant, indexing, and ct.launch) to Triton equivalents, including host-side launch syntax.
  • Mandatory performance-critical TMA conversion: prevents 5–20× regressions by requiring TMA tensor descriptors for any 2D+ block-shaped tile loads/stores.
  • Advanced handling for attention/GQA/Gemma-style kernels: applies a mandatory optimization strategy checklist (not just “TMA first”), including dual-kernel layout flags like transpose/transposes_v patterns.

Quick Start

Convert your cuTile kernel file to a Triton @triton.jit implementation by following the phase-gated workflow (analyze → convert → validate → test → benchmark) in translations/workflow.md, making sure every 2D+ tile load/store uses tl.make_tensor_descriptor TMA.

Frequently Asked Questions about converting-cutile-to-triton

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I convert cuTile GPU kernels to Triton without performance regression?▼

To convert cuTile to Triton without performance regression, follow a phase-gated workflow: analyze, convert, validate, test, and benchmark. Use mandatory TMA tensor descriptors for 2D+ tile loads to avoid major slowdowns.

Why does my cuTile to Triton kernel conversion cause a 5x to 20x slowdown?▼

Converting cuTile to Triton causes massive slowdowns when 2D+ block-shaped tile loads bypass TMA. Using tl.make_tensor_descriptor TMA for these accesses prevents the 5–20× performance regressions seen in naive ports.

How do I map cuTile APIs like ct.load and ct.mma to Triton equivalents?▼

Mapping cuTile APIs to Triton requires translating ct.load/store, ct.mma, ct.Constant, and ct.launch into Triton equivalents. The conversion workflow provides guidance for standard GEMM/BMM and attention-style kernel ports.

Can I use this cuTile to Triton conversion workflow for attention and GQA kernels?▼

This cuTile to Triton conversion workflow supports attention and GQA-style kernels. It applies an optimization checklist including dual-kernel layout flags like transpose and transposes_v patterns to preserve correctness.

What is the required workflow for porting cuTile kernels to Triton?▼

The required cuTile to Triton porting workflow is a deterministic sequence: analyze, convert, validate, test, and benchmark. It enforces explicit stop conditions and validation gates before declaring the kernel conversion complete.

Related Skills

Explore More Software Engineering Skills→
NVIDIANVIDIA
796

converting-cutile-to-triton

Translate cuTile GPU kernels to Triton with dual-layout and META-driven autotuning.

Official
Advanced
AscendAscend
31

triton-operator-code-gen

Generate Triton kernels from Ascend NPU operator designs with smoke tests.

Official
Advanced
ShekswessShekswess
3

triton

Author, debug, and optimize GPU kernels with Triton in Python.

Community
Advanced
LeightonWangLeightonWang

triton-skills

Write, debug, and optimize Triton and Gluon GPU kernels.

Community
Advanced