Can it handle complex dispatch patterns like ALL_TYPES_AND_COMPLEX_AND2?

Absolutely. It maps these complex legacy patterns to multiple AT_EXPAND calls combined with individual trailing type arguments for a precise conversion.

Does this skill handle the necessary header includes?

Yes, the skill is designed to identify missing Dispatch_v2.h includes and add them near existing ATen/Dispatch.h headers while preserving backward compatibility.

Is the AT_WRAP macro mandatory in the new format?

Yes, AT_WRAP is required around the lambda body to ensure the C++ preprocessor correctly handles internal commas, which are common in template arguments and multi-argument function calls.

Why should I use AT_DISPATCH_V2 instead of legacy macros?

AT_DISPATCH_V2 removes macro arity limits (like AND2, AND3), allows for composable type sets using AT_EXPAND, and provides a more extensible architecture for adding new data types.

PyTorch AT_DISPATCH V2 Converter

Name: PyTorch AT_DISPATCH V2 Converter
Author: Activer007

byActiver007

0•

开发者工具

Refactors legacy PyTorch AT_DISPATCH macros into the modern AT_DISPATCH_V2 format for improved extensibility and clearer type grouping.

This skill automates the transition from legacy PyTorch AT_DISPATCH macros to the newer V2 API found in aten/src/ATen/Dispatch_v2.h. It intelligently handles complex transformations, including reordering arguments to place scalar types and names first, wrapping lambdas with AT_WRAP to prevent C++ preprocessor comma parsing issues, and expanding type groups with AT_EXPAND. It is a critical tool for core PyTorch developers and contributors maintaining ATen kernels, CUDA implementations, or native operators who need to adopt the more flexible and composable dispatch patterns of the V2 API.

主要功能

010 GitHub stars

02Support for individual type trailing arguments (e.g., kHalf, kBFloat16)

03Automatic argument reordering for V2 API compliance

04Automated header management for ATen/Dispatch_v2.h

05Intelligent mapping of legacy macros to AT_EXPAND type groups

06Mandatory AT_WRAP implementation for complex lambda bodies

使用场景

01Porting legacy ATen kernels to the modern Dispatch V2 API

02Refactoring CUDA kernels to support additional data types without macro arity limits

03Modernizing native operator implementations in the PyTorch core library

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add activer007/ordinary-claude-skills at-dispatch-v2

For use in Claude.ai and ChatGPT

Download Skill