anthony-maio

Triton Flash Attention V2 — Development skill for Claude Code

Development community

Implement FlashAttention v2 kernels in Triton with online softmax, causal masking, GQA head routing, multi-stream accumulators, and fused epilogues.

How to install Triton Flash Attention V2

Installs to ~/.claude/skills/anthony-maio-triton-skills-triton-flash-attention-v2/SKILL.md

Terminal

mkdir -p ~/.claude/skills/anthony-maio-triton-skills-triton-flash-attention-v2 && curl -fsSL https://raw.githubusercontent.com/anthony-maio/triton-skills/HEAD/skills/triton-flash-attention-v2.md -o ~/.claude/skills/anthony-maio-triton-skills-triton-flash-attention-v2/SKILL.md

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

What Triton Flash Attention V2 does

name: triton-flash-attention-v2 description: Implement FlashAttention v2 kernels in Triton with online softmax, causal masking, GQA head routing, multi-stream accumulators, and fused epilogues.

FlashAttention v2 kernels in Triton

**Targets:** Triton >= 2.1, SM70+/CDNA2+

Overview

FlashAttention v2 computes `O = softmax(QK^T / sqrt(d_k)) V` without materializing the N×N attention matrix. The kernel iterates over K/V blocks, maintains running softmax statistics `(m, l, acc)` in r

Alternatives in Development

Connect — Fancy stream processing made operationally mundane 8608 114 1 8.6k ★
Skills Directory — 462 skills across 35 categories Browse online: spawner.vibeship.co/skills 845 ★
StatsPAI — StatsPAI is the first Agent-native Python library for causal inference and applied econometrics — unified API 302 ★