Can I define custom rules if the automatic differentiation fails?

Absolutely. This skill includes patterns for using EnzymeRules to define custom forward and reverse passes for specific functions.

Does this skill support GPU differentiation?

Yes, the skill provides specific implementation patterns for CUDA.jl, using autodiff_deferred to compute gradients of GPU kernels.

How do I handle mutable arrays during differentiation?

You should use the Duplicated(x, dx) annotation, where x is your original array and dx is a 'shadow' array that accumulates the gradients.

What makes Enzyme.jl different from other Julia AD tools?

Unlike high-level source-to-source tools, Enzyme operates at the LLVM IR level. This allows it to differentiate through a wider range of code with significantly less overhead and better performance.

Enzyme.jl Automatic Differentiation

Name: Enzyme.jl Automatic Differentiation
Author: plurigrid

byplurigrid

•

データサイエンスとML

Provides high-performance LLVM-level automatic differentiation for Julia code on CPU and GPU architectures.

This skill integrates Enzyme.jl's powerful LLVM-level automatic differentiation capabilities into Claude, allowing for precise gradient, Jacobian, and Hessian computations within Julia workflows. It supports both forward and reverse modes, handles complex data structures via specialized type annotations, and offers seamless integration with CUDA.jl for GPU kernel differentiation. By operating at the LLVM IR level, it generates highly optimized derivative code that bypasses the overhead of traditional source-to-source AD tools, making it ideal for performance-critical scientific computing and machine learning optimization.

主な機能

01CUDA.jl integration for differentiating high-performance GPU kernels

02Support for Forward, Reverse, and Split differentiation modes

032 GitHub stars

04Custom derivative rule definition through the EnzymeRules system

05Comprehensive type annotations including Const, Active, and Duplicated

06LLVM-level differentiation for near-native performance

ユースケース

01Differentiating GPU kernels for physics-informed neural networks (PINNs)

02Optimizing custom machine learning loss functions with efficient gradients

03Performing sensitivity analysis on complex scientific simulations and physical models

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add plurigrid/asi enzyme-autodiff

For use in Claude.ai and ChatGPT

Download Skill