Few CMake improvements

[Silv3S]

• align torch-xpu-ops version to match target PyTorch version
• simplify install_xpu_headers macro and add CONFIGURE_DEPENDS to track these headers

[Copilot]

Pull Request Overview

This PR removes the BUILD_SPLIT_KERNEL_LIB option and its associated logic for splitting kernels into multiple libraries, which was only needed for older compiler versions. It also updates the project version to match PyTorch 2.10.0, fixes a comment typo, and improves the header installation macro.

Key changes:

• Removes split kernel library build paths for both Windows and Linux
• Updates project version from 2.3.0 to 2.10.0
• Simplifies the install_xpu_headers macro with CONFIGURE_DEPENDS

Reviewed Changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/BuildOnWindows.cmake	Removes 200+ lines of split kernel library logic for Windows builds
src/BuildOnLinux.cmake	Removes 87 lines of split kernel library logic for Linux builds
src/ATen/CMakeLists.txt	Simplifies header installation macro and moves mkl file glob inside conditional
cmake/Modules/FindSYCL.cmake	Corrects comment from "FindCUDA.cmake" to "FindSYCL.cmake"
cmake/Modules/FindONEMKL.cmake	Refactors MKL library list construction and improves warning message
CMakeLists.txt	Updates project version to 2.10.0 and removes BUILD_SPLIT_KERNEL_LIB variable

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[PawelSwider2000]

LGTM, please confirm that we could remove BUILD_SPLIT_KERNEL_LIB

[guangyey]

Overall LGTM. The only thing we need to discuss is the removal of BUILD_SPLIT_KERNEL_LIB. As I understand it, this flag was originally used to

1. Work around the binary size when too many AOT targets were enabled, however, this is no longer necessary since the compiler now supports device code suppression feature.
2. Facilitate debugging - some developers might still use this flag to debug. It is useful for reducing incremental build time, especially when debugging ATen XPU ops code.

Personally, I haven't used BUILD_SPLIT_KERNEL_LIB recently, so I can't decide whether it should be removed. Let's ask for opinions from others.
I suggest you split this PR into two separate ones. I can help land the other changes, and we can keep the BUILD_SPLIT_KERNEL_LIB removal PR for further discussion.

[guangyey]

LGTM. Thanks~

[Silv3S]

Thanks for review. I agree that it will be better to split this PR and continue discussion about BUILD_SPLIT_KERNEL_LIB removal. I assumed it's not necessary since new compilers don't have such limitations and also noticed that this flag can't be used for debug builds without customizing CMake.

if(CMAKE_BUILD_TYPE MATCHES "(Debug|RelWithDebInfo)")
  set(BUILD_SEPARATE_OPS TRUE)
endif()

For debug builds BUILD_SEPARATE_OPS takes precedence, making BUILD_SPLIT_KERNEL_LIB block unreachable

if(BUILD_SEPARATE_OPS)
  ...
elseif(BUILD_SPLIT_KERNEL_LIB OR __INTEL_LLVM_COMPILER LESS 20250004 OR ICX_DATE LESS 20241205)
  ...
else()
  ...
endif()

Continued in #2208

[Silv3S]

It fails on CI, but as these changes affect only build files it's unlikely that this PR introduced functional regressions. @guangyey can this be somehow marked as not relevant or should I re-run CI until all tests are passing?

[guangyey]

Land this due to the low risk.