Fix Windows CUDA 12.9 SM120 compilation

[Copilot]

Description

• CUDA header workaround
  • Shadow CUDA 12.9’s generated clusterlaunchcontrol.h for MSVC builds.
  • Replace invalid 32-bit long2 PTX operands with upstream-correct longlong2.
• Scope
  • Applies only to MSVC with CUDA 12.9.
  • Preserves native SM120 compilation and becomes a no-op for corrected headers.

Motivation and Context

CUDA 12.9’s clusterlaunchcontrol.h uses Windows’ 32-bit long for PTX operands requiring 64 bits. This causes widespread .cu compilation failures when targeting SM120 on Windows, while Linux builds succeed.

• Fixes [Build] Build of 1.26.0 fails on Windows with CUDA 12.9 and sm_120 #29535

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

Pull request overview

Fixes Windows CUDA 12.9 compilation for SM120 by shadowing the faulty toolkit header with a generated corrected copy.

Changes:

• Adds a CMake helper replacing invalid long2 PTX operand casts with longlong2.
• Applies the workaround only for MSVC and CUDA 12.9.x.
• Removes stale shadow headers when the installed toolkit no longer needs correction.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
cmake/external/cuda_configuration.cmake	Adds the header-patching helper.
cmake/CMakeLists.txt	Activates and prioritizes the CUDA 12.9 workaround on MSVC.

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.