Skip to content

Conversation

@LuFinch
Copy link
Contributor

@LuFinch LuFinch commented Sep 11, 2025

This is a draft PR to enable SYCL-TLA build in torch-xpu-ops so that we can test SYCL-TLA kernels' accuracy/performance in Pytorch when SDPA/GEMM kernels are ready.

After discussion with Eikan, we decided to put build logic in torch-xpu-ops while put kernels source code in Pytorch in-tree. Please put your SYCL-TLA kernel source code in Pytorch and set its path as part of ATen_XPU_SYCLTLA_SRCS in torch-xpu-ops/src/ATen/CMakeLists.txt.

Since SYCL-TLA has different compilation options compared with normal SYCL kernels in torch-xpu-ops, I make the logic in cmake/BuildFlags.cmake as a macro so that I can reuse the common compilation options.

Since there is not a determined plan of how to import sycl-tla repo, I git clone the main branch in cmake for debug convinence. We can pin commit after sycl-tla has first release tag

Depend on g++ upgrading to gcc13, otherwise the sycltla kernel won't build

@LuFinch LuFinch force-pushed the lfq/cutlass branch 6 times, most recently from aa10375 to e9800af Compare October 20, 2025 07:51
@LuFinch LuFinch changed the title [Cutlass] Enable Cutlass with host compiler [SYCL-TLA] Enable SYCL-TLA build with host compiler Oct 20, 2025
@LuFinch LuFinch marked this pull request as ready for review October 21, 2025 02:33

This comment was marked as abuse.

@LuFinch LuFinch requested a review from EikanWang October 21, 2025 02:34
@LuFinch LuFinch changed the title [SYCL-TLA] Enable SYCL-TLA build with host compiler [SYCL-TLA] Enable SYCL-TLA build Oct 21, 2025
@LuFinch
Copy link
Contributor Author

LuFinch commented Oct 22, 2025

@fengyuan14 @EikanWang Could you help review and give some comments?

@LuFinch
Copy link
Contributor Author

LuFinch commented Nov 13, 2025

@EikanWang @guangyey Can we merge this PR?

@guangyey
Copy link
Contributor

Of course, let's move forward. We can apply a minor fix if needed.

@guangyey guangyey enabled auto-merge November 13, 2025 05:44
@guangyey guangyey disabled auto-merge November 13, 2025 05:45
@guangyey guangyey enabled auto-merge November 13, 2025 05:45
@guangyey guangyey requested a review from Copilot November 13, 2025 05:48
@guangyey guangyey disabled auto-merge November 13, 2025 05:48
Copy link
Contributor

Copilot AI left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull Request Overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated 2 comments.

Comments suppressed due to low confidence (3)

cmake/BuildFlags.cmake:1

  • Removed line that defines __INTEL_LLVM_COMPILER_VERSION. This flag may be required by code that checks the compiler version. If this removal is intentional, ensure that no code depends on this definition.
# Setup building flags for SYCL device and host codes.

cmake/BuildFlags.cmake:1

  • This comment line was removed but the related code logic immediately following it remains unchanged. The comment should be preserved to document the FP64 conversion emulation logic for DG2/ATS-M targets.
# Setup building flags for SYCL device and host codes.

cmake/BuildFlags.cmake:1

  • Removed initialization of SYCL_flags variable. This variable is no longer used after renaming to SYCL_COMPILE_FLAGS, but verify that no other code depends on SYCL_flags being defined.
# Setup building flags for SYCL device and host codes.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

@guangyey guangyey enabled auto-merge November 13, 2025 05:49
@guangyey guangyey added this pull request to the merge queue Nov 13, 2025
Merged via the queue into main with commit 8384acf Nov 13, 2025
34 of 35 checks passed
@guangyey guangyey deleted the lfq/cutlass branch November 13, 2025 05:49
github-merge-queue bot pushed a commit that referenced this pull request Nov 26, 2025
This PR moves the sycltla kernels in
pytorch/pytorch#167056 into torch-xpu-ops.

This PR is based on #2030.
When the build PR merge, I will rebase this PR.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants