at-dispatch-v2

AT_DISPATCH_V2 is the modern PyTorch dispatch macro format that replaces legacy AT_DISPATCH_ALL_TYPES_AND* and AT_DISPATCH_FLOATING_TYPES* macros. Convert by reordering arguments to (scalar_type, name, AT_WRAP(lambda)), adding #include <ATen/Dispatch_v2.h>, wrapping lambdas with AT_WRAP(), and expanding type groups with AT_EXPAND().

ATen kernel files under aten/src/ATen/native/ that use dispatch macros require conversion, including ATen kernels, CUDA kernels, and native operators. The scope covers all legacy AT_DISPATCH variants in these directories.

AT_DISPATCH_V2 modernizes PyTorch's kernel dispatch system with clearer argument ordering, explicit type handling via AT_EXPAND(), and reduced macro complexity. It improves code maintainability and reduces errors during large-scale codebase modernization.

The conversion requires argument reordering to the AT_DISPATCH_V2 format: scalar_type, name, AT_WRAP(lambda). This automation handles the reordering and lambda wrapping, eliminating manual refactoring errors across multiple kernel files.

Yes, AT_DISPATCH_V2 conversion preserves original behavior across multi-line lambdas. The AT_WRAP() wrapper handles complex lambda bodies while maintaining dispatch logic, supporting kernels with extensive computation spans.

Automated macro conversion across all ATen native files ensures consistency and reduces refactoring time compared to manual updates. This approach minimizes bug introduction risk when modernizing numerous kernel files simultaneously.