anthony-maio

Triton Persistent Warp Matmul

Name: Triton Persistent Warp Matmul
Author: anthony-maio

AI community

SOURCE https://github.com/anthony-maio/triton-skills/blob/main/skills/triton-persistent-warp-matmul.md

Description

--- name: triton-persistent-warp-matmul description: Teach an AI agent to implement persistent, warp-specialized matmul kernels in Triton using TMA and producer/consumer warps. --- # Persistent & Warp-Specialized Matmul Kernels in Triton > **Targets:** Triton >= 3.0; TMA/warp specialization requires SM90+ (Hopper) Overview This skill teaches how to implement a persistent GEMM in Triton where fewer thread blocks than output tiles are launched and each block iterates over multiple tiles. It cov

Installation

Installs to ~/.claude/skills/anthony-maio-triton-skills-triton-persistent-warp-matmul/SKILL.md

Terminal

mkdir -p ~/.claude/skills/anthony-maio-triton-skills-triton-persistent-warp-matmul && curl -fsSL https://raw.githubusercontent.com/anthony-maio/triton-skills/HEAD/skills/triton-persistent-warp-matmul.md -o ~/.claude/skills/anthony-maio-triton-skills-triton-persistent-warp-matmul/SKILL.md

Restart Claude Code, or start a new session, for it to be picked up.