skills.rest
NVIDIA
NVIDIAOFFICIAL
Explore all skills by this author →
WebsiteGitHub

tilegym-converting-cutile-triton-to-cutile-rs

Convert Triton-TileIR and cuTile-Python GPU kernels to cutile-rs Rust via a multi-agent pipeline.

807|85|Updated Nov 13, 2025
Search Tags:#Rust#GPU kernel#Triton#Multi-Agent#cutile-rs#cuda tile#kernel conversion
One-click install
npx skills add https://github.com/NVIDIA/TileGym --skill tilegym-converting-cutile-triton-to-cutile-rs
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: tilegym-converting-cutile-triton-to-cutile-rs
Source: https://github.com/NVIDIA/TileGym/tree/main/skills/tilegym-converting-cutile-triton-to-cutile-rs
Command: npx skills add https://github.com/NVIDIA/TileGym --skill tilegym-converting-cutile-triton-to-cutile-rs

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) components.

What problem does it solve?

Porting tile-based GPU kernels from Triton-TileIR or cuTile-Python to cutile-rs (Rust) requires coordinating IR dumping, device kernel authoring, FFI host wiring, correctness testing, and performance benchmarking, which is error-prone when done ad hoc.

Core Features & Use Cases

  • Bounded multi-agent pipeline: Orchestrates Agents A through F with a strict routing table, spawn caps, and verdict-based handoffs covering IR analysis, kernel conversion, host FFI, correctness, and performance stages.
  • Deterministic validation scripts: Ships preflight, IR validation, per-agent validators, IR diffing, and a final 17-file aggregate kernel gate.
  • Worked examples and references: Includes softmax (raw-pointer launch) and bmm (tiled-output launch) templates plus coding rules, op mapping, and IR diff checklists.
  • Use Case: A developer adding a cutile-rs backend to a TileGym operator runs the pipeline to produce kernel.rs, ffi.rs, a Python wrapper, backend registration, and CUPTI-backed correctness and performance reports.

Quick Start

Ask the agent to convert a TileGym kernel such as softmax to a cutile-rs backend and let the skill run its preflight and agent pipeline automatically.

Frequently Asked Questions about tilegym-converting-cutile-triton-to-cutile-rs

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

FAQPage Schema
How do I convert a Triton or cuTile-Python kernel to cutile-rs Rust?▼

Invoke this skill with the target kernel name; it runs preflight.sh, then routes through Agent A (IR dump and analysis), Agent B (device kernel.rs), Agent D (ffi.rs, wrapper, correctness), and Agent E (performance), with Agent C diagnosing failures.

What does the cutile-rs conversion pipeline produce?▼

It produces kernel.rs device code, an ffi.rs C-ABI launcher compiled into libcutile_kernels.so, a cffi-based Python wrapper registered as the cutile-rs backend, plus correctness and CUPTI performance reports under the kernel output directory.

Does the skill support multi-variant kernels with different tile configurations?▼

Yes. Agent A records structural variants in analysis.json with per-variant reference IR, and Agent B writes one cutile entry function per structural variant, including const-generic or literal-specialized entries for baked constants.

What happens when correctness or performance validation fails?▼

Agent D classifies failures as host or kernel, and Agent C diagnoses IR diffs with an owner tag. The routing table respawns the owning agent within hard spawn caps, and unrecoverable cases end in PIPELINE_COMPLETE or STOP.

What are the limitations of the cutile-rs conversion skill?▼

It is forward-only (no gradient support), targets the pinned cutile 0.2.0 crates.io dependencies, and stops on environment blockers such as missing GPU, toolchain, or tileiras compiler rather than attempting repairs.

Related Skills

Explore More Software Engineering Skills→
NVIDIANVIDIA
3.2k

tilegym-converting-cutile-to-triton

Converts cuTile GPU kernels to Triton with TMA optimization and debugging guidance.

Official
Advanced
SignmanalSignmanal
1

tensorrt-llm

Optimize NVIDIA GPU inference for large language models with quantization and parallelism.

Community
Advanced
drunkcodingdrunkcoding
4

benchmark-kernel

Benchmark FlashInfer kernel performance across backends using CUPTI timing.

Community
Intermediate
yo-stevenyo-steven

adding-cutile-kernel

Adds a cuTile GPU kernel operator to TileGm with full integration.

Community
Advanced