Enable Windows on ARM CUDA target

[mfuntowicz]

Summary

This PR enables building the CUDA backend for Windows on Arm by adding a target architecture option to the Windows builder, plus a small CMake fix required by newer CUDA toolkits.
Previously, builder.ps1 always derived the build architecture from the host process architecture, making it impossible to cross-compile ARM64 binaries from an x64 machine. This PR introduces an explicit -TargetArch parameter so the CUDA backend can now be built natively on ARM64 machines or cross-compiled from x64 hosts.

Changes

• New -TargetArch parameter (x64 | ARM64), defaulting to the host architecture — preserving existing behavior.
• Initialize-VSEnvironment selects the correct vcvarsall.bat argument for the host/target combination (arm64, x64_arm64, arm64_x64), enabling cross-compilation.
• Get-CMakeConfigureArgs passes the target architecture to the Visual Studio generator (-A ARM64 / -A x64) instead of hardcoding the detected host platform.
• Informative message when targeting Windows ARM64 with CUDA (requires CUDA 13.0+ for Windows ARM64 support), and a warning when ARM64 is requested with a non-CUDA backend (only CUDA officially supports Windows ARM64).
  kernel-builder/src/pyproject/templates/torch/preamble.cmake
• When building with MSVC, add -Xcompiler=/Zc:preprocessor to the nvcc flags. CCCL shipped with CUDA 13+ hard-errors when nvcc uses MSVC's traditional (non-conforming) preprocessor as the host preprocessor, which is required for the ARM64 CUDA build to succeed.

Usage

Native ARM64 build on an ARM64 machine

.\builder.ps1 -SourceFolder ./examples/relu -Backend cuda -TargetArch ARM64 -ArchList "12.1" -Build

Cross-compile from an x64 host

.\builder.ps1 -SourceFolder ./examples/relu -Backend cuda -TargetArch ARM64 -ArchList "12.1" -Build

Notes

• CUDA support for Windows ARM64 requires CUDA toolkit 13.0 or newer.
• x64-only workflows are unaffected: without -TargetArch, the script behaves exactly as before.

[mfuntowicz]

Example output:

(kernels) PS C:\Users\momo-\workspace\kernels> .\nix-builder\scripts\windows\builder.ps1 .\examples\kernels\relu\ -TargetArch ARM64 -Backend CUDA -Build -Force
[*] Found kernel-builder at: C:\Users\momo-\workspace\kernels\nix-builder\scripts\windows\..\..\..\target\debug\kernel-builder.exe
[*] Generating CUDA backend...
[*] Executing: C:\Users\momo-\workspace\kernels\nix-builder\scripts\windows\..\..\..\target\debug\kernel-builder.exe create-pyproject C:\Users\momo-\workspace\kernels\examples\kernels\relu\ --force
[+] Generation completed successfully!
[*] Targeting Windows ARM64 with CUDA backend - requires a CUDA toolkit with Windows ARM64 support (CUDA 13.0+)
[*] Building project with CMake...
[*] Configuration: Release
[*] Target architecture: ARM64
[*] Initializing Visual Studio environment...
[*] Visual Studio environment already initialized
[*] Configuring CMake project...
[*] Target architecture: ARM64, using Visual Studio platform: ARM64
[*] Using Python from environment: C:\Users\momo-\workspace\kernels\.venv/Scripts\python.exe
[*] Detected PyTorch version: 2.15.0.dev20260831+cu134
CMake Warning (dev) at CMakeLists.txt:17 (project):
  project() called with 'LANGUAGES' as first argument.  The first parameter
  should be the project name, not a keyword argument.  See the
  cmake-commands(7) manual for correct usage of the project() command.
This warning is for project developers.  Use -Wno-dev to suppress it.

-- Selecting Windows SDK version 10.0.26100.0 to target Windows 10.0.28000.
-- FetchContent base directory: C:/Users/momo-/workspace/kernels/examples/kernels/relu/build/_deps
-- Found Python3: C:\Users\momo-\workspace\kernels\.venv/Scripts\python.exe (found version "3.13.15") found components: Development Development.SABIModule Interpreter Development.Module Development.Embed
-- Using backend: cuda, GPU language: CUDA
-- PyTorch: CUDA detected: 13.4
-- PyTorch: CUDA nvcc is: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v13.4/bin/nvcc.exe
-- PyTorch: CUDA toolkit directory: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v13.4
-- PyTorch: Header version is: 13.4
-- C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v13.4/lib/x64/nvrtc.lib shorthash is 629d40cf
-- USE_CUDNN is set to 0. Compiling without cuDNN support
-- USE_CUSPARSELT is set to 0. Compiling without cuSPARSELt support
-- USE_CUDSS is set to 0. Compiling without cuDSS support
-- USE_CUFILE is set to 0. Compiling without cuFile support
-- Autodetected CUDA architecture(s):  12.1(2.0)
-- Added CUDA NVCC flags for: -gencode;arch=compute_20,code=sm_121
CMake Warning at C:/Users/momo-/workspace/kernels/.venv/Lib/site-packages/torch/share/cmake/Torch/TorchConfig.cmake:22 (message):
  library kineto not found.
Call Stack (most recent call first):
  C:/Users/momo-/workspace/kernels/.venv/Lib/site-packages/torch/share/cmake/Torch/TorchConfig.cmake:119 (append_torchlib_if_found)
  CMakeLists.txt:55 (find_package)


-- CUDA base archs used for intersection with kernel archs: 7.5;8.0;8.6;8.7;8.9;9.0;10.0;11.0;12.0;12.1
-- Default CUDA kernel architectures: 7.5;8.0;8.6;8.7;8.9;9.0;10.0;11.0;12.0;12.1+PTX
-- Generated build name: torch215-cu134-aarch64-windows
-- CUDA kernel: relu, capabilities: 7.5;8.0;8.6;8.7;8.9;9.0;10.0;11.0;12.0;12.1+PTX
-- Using PACKAGE_NAME: relu
-- GPU archs: 10.0;11.0;12.0;12.1+PTX;7.5;8.0;8.6;8.7;8.9;9.0
-- Added install rules for _relu_cuda_54964b2_dirty -> torch215-cu134-aarch64-windows
-- Added install rules for _relu_cuda_54964b2_dirty -> build/torch215-cu134-aarch64-windows
-- Configuring done (10.5s)
-- Generating done (0.0s)
CMake Warning:
  Manually-specified variables were not used by the project:

    Python_EXECUTABLE


-- Build files have been written to: C:/Users/momo-/workspace/kernels/examples/kernels/relu/build
[*] Building project...
MSBuild version 18.9.1+a81b43525 for .NET Framework

  torch_binding.cpp
[...]
[+] Build completed successfully!