tilelang

Generates GPU kernels for CUDA, HIP, Metal, and CPU via Python.

4|Updated May 6, 2026

Search Tags:#gpu #metal #kernel #tensorflow #code-generation #cuda #hip

Ready to run

npx skills add https://github.com/jstzwj/ai-infra-plugins --skill tilelang

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: tilelang
Source: https://github.com/jstzwj/ai-infra-plugins/tree/main/plugins/tilelang/skills/tilelang
Command: npx skills add https://github.com/jstzwj/ai-infra-plugins --skill tilelang

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires tvm, torch, numpy, and includes scripts (resource) and references (resource) and assets (resource) components.

What problem does it solve?

TileLang simplifies the development of high-performance GPU kernels by providing a framework that automatically dispatches to hardware-optimized instructions, reducing complexity for engineers.

Core Features & Use Cases

Kernel Automation: Generates efficient code for CUDA, HIP, Metal, and CPU architectures with minimal manual tuning.
Deep API Access: Offers primitives for memory management, data movement, tensor core operations, and control flow.
Use Case: Develops optimized GEMM kernels and advanced tensor operations for AI model training and inference, streamlining workflows and maximizing hardware utilization.

Quick Start

Write a TileLang kernel to perform matrix multiplication, compile, and execute it directly within your Python environment.

Frequently Asked Questions about tilelang

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

FAQPage Schema

How do I automate GPU kernel generation for CUDA and Metal backends?▼

Automate GPU kernel generation by writing a Python interface that dispatches hardware-optimized instructions to CUDA, HIP, Metal, or CPU backends, minimizing manual tuning for deep learning and scientific computation scenarios.

What is the best way to develop high-performance GEMM kernels for AI inference?▼

Develop high-performance GEMM kernels by leveraging deep API primitives for memory management and tensor core operations, generating efficient code that maximizes hardware utilization for AI model training and inference.

Do I need TensorFlow to compile and execute GPU kernels in a Python environment?▼

You do not need TensorFlow; compilation and execution require a Python environment along with dependencies including TVM, Torch, and NumPy to facilitate the kernel creation workflow.

Can I access low-level memory management and data movement primitives for HIP architectures?▼

You can access low-level primitives for memory management, data movement, and tensor core operations, enabling fine-grained control over hardware-optimized instructions for HIP architectures.

Why does manual GPU kernel tuning take so much effort compared to automated code dispatching?▼

Manual GPU kernel tuning requires extensive low-level hardware knowledge, whereas automated code dispatching abstracts these complexities, automatically generating optimized instructions for CUDA, HIP, Metal, and CPU architectures.