anthony-maio

Triton Fused Epilogue Kernels — Development skill for Claude Code

Development community

Fuse epilogue ops (normalization, gating, residual, activation, dropout) into Triton attention/matmul kernels to eliminate HBM round-trips.

How to install Triton Fused Epilogue Kernels

Installs to ~/.claude/skills/anthony-maio-triton-skills-triton-fused-epilogue-kernels/SKILL.md

Terminal
mkdir -p ~/.claude/skills/anthony-maio-triton-skills-triton-fused-epilogue-kernels && curl -fsSL https://raw.githubusercontent.com/anthony-maio/triton-skills/HEAD/skills/triton-fused-epilogue-kernels.md -o ~/.claude/skills/anthony-maio-triton-skills-triton-fused-epilogue-kernels/SKILL.md

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

What Triton Fused Epilogue Kernels does


name: triton-fused-epilogue-kernels description: Fuse epilogue ops (normalization, gating, residual, activation, dropout) into Triton attention/matmul kernels to eliminate HBM round-trips.

Fused Epilogue Kernels in Triton

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

Overview

Fusing epilogue work directly into attention or GEMM kernels avoids extra HBM writes/reads and kernel launches. Perform all final math in-register immediately before the final `tl.store`. Use `tl.constexpr` bo

Alternatives in Development

  • Evox Short — CEO → EVOX: The Loop + Opus 4.6 Activation 104 ★
  • Arnold — General-purpose planning and execution harness for LLMs — structured phases, critique, gating, and review 97 ★
  • Innovate — Draft N novel sparse-attention submissions in one shot — algorithm-innovation focus, all must compile, no benc 68 ★

Full documentation available on GitHub

View Source Repository