kernel-tileir-optimization

Automate TileIR-based autotune config classification and tuning for Triton kernels on Blackwell GPUs.

2.8k|332|Updated Feb 25, 2026
One-click install
npx skills add https://github.com/NVIDIA/skills --skill kernel-tileir-optimization
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: kernel-tileir-optimization
Source: https://github.com/NVIDIA/skills/tree/main/skills/TensorRT-LLM/kernel-tileir-optimization
Command: npx skills add https://github.com/NVIDIA/skills --skill kernel-tileir-optimization

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Optimizes existing Triton kernels for NVIDIA's TileIR backend on Blackwell GPUs by automatically generating and applying TileIR-specific autotune configurations, enabling higher performance through TMA descriptor usage, 2CTA configurations, and occupancy tuning. This skill coordinates the end-to-end workflow from kernel classification to optimization and benchmarking, reducing manual tuning effort and ensuring TileIR compatibility.

Core Features & Use Cases

  • TileIR-aware autotuning for dot-related kernels (GEMM, attention) using TMA descriptors and 2CTA strategies.
  • Norm-like and reduction kernel support with high-occupancy configurations and varied num_warps.
  • Element-wise kernels optimization with extended block sizes and occupancy ranges, plus robust validation against PTX baseline.
  • End-to-end workflow: classify kernel, apply TileIR optimizations, validate, and benchmark TileIR vs PTX backends on Blackwell hardware.

Quick Start

Run the classification and TileIR optimization workflow on a sample Triton kernel to generate optimized configs.

Frequently Asked Questions about kernel-tileir-optimization

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

FAQPage Schema
How do I optimize Triton kernels for Blackwell GPUs?

You can optimize Triton kernels for Blackwell GPUs by automating TileIR-based autotune configuration generation, applying TMA descriptors and 2CTA strategies to classify and tune dot-related, element-wise, and reduction kernels.

What is TileIR autotuning and how does it work with Triton?

TileIR autotuning automatically generates and applies backend-specific configurations to Triton kernels, enabling TMA descriptor usage and occupancy tuning to achieve higher performance on Blackwell hardware compared to standard PTX backends.

Do I need a Blackwell GPU to run TileIR kernel optimization?

Yes, TileIR kernel optimization requires a Blackwell (sm_100+) GPU, the nvtriton package installation, and the ENABLE_TILE=1 environment flag to execute the classification, transformation, validation, and benchmarking stages.

Can I benchmark TileIR performance against PTX baselines for Triton kernels?

Yes, the workflow validates optimized kernels and benchmarks TileIR performance against PTX baselines across dot-related, norm-like, element-wise, and reduction kernel types to ensure compatibility and measure performance gains.

What types of Triton kernels are supported for TileIR optimization?

TileIR optimization supports dot-related kernels like GEMM and attention using TMA descriptors and 2CTA strategies, as well as norm-like, reduction, and element-wise kernels with high-occupancy configurations and varied num_warps.

Why is my Triton autotuning not improving performance on Blackwell GPUs?

Standard Triton autotuning may lack TileIR-specific configurations like TMA descriptor usage and 2CTA strategies; applying automated TileIR optimization classifies and tunes kernel types to improve Blackwell GPU performance.