Automates the migration of legacy PyTorch AT_DISPATCH macros to the modern AT_DISPATCH_V2 API.
The at-dispatch-v2 skill streamlines the process of upgrading PyTorch C++ codebases by converting legacy dispatch macros to the newer V2 format. It handles complex transformations including argument reordering, wrapping lambdas with AT_WRAP to prevent macro expansion issues, and mapping traditional dispatch patterns to extensible type groups using AT_EXPAND. This is essential for developers maintaining ATen kernels, CUDA implementations, or native operators who need to leverage the more composable and flexible dispatch system introduced in modern PyTorch versions.
主な機能
01Intelligent lambda wrapping with AT_WRAP for complex expressions
02Support for multi-type dispatch patterns (AND2, AND3, etc.)
03Automatic header injection for ATen/Dispatch_v2.h
04Automated argument reordering for V2 API compliance
05Mapping legacy macros to modern AT_EXPAND type groups
0681 GitHub stars
ユースケース
01Refactoring CUDA kernels using multi-type dispatch macros
02Porting legacy ATen kernel files to the modern PyTorch dispatch API
03Standardizing custom operator implementations for modern PyTorch releases
