Are header files exposed in torch ?

[YangKai0616]

🚀 The feature, motivation and pitch

Have you considered exposing the library's header files in torch? This would allow developers to directly include xpu sycl header files. For example, #include <ATen/native/xpu/sycl/Loops.h>. Just like how it's done with cuda: #include <ATen/native/cuda/Loops.cuh>.

In some integrated environments, I am unable to manually download the current repo. When I need to develop a kernel implemented with SYCL, I have to locate the implementations of all header files and extract them individually as new header files......

Alternatives

No response

Additional context

No response

[CuiYifeng]

@YangKai0616 Thank you for submitting this issue. Let me rephrase your question:

• Based on "unable to manually download the current repo", your requirement is to be able to directly find XPU SYCL header files in the repo pytorch instead of in torch-xpu-ops, right?

If not, XPU SYCL header files have been exposed (but not exist) in PyTorch. For example, #include <ATen/native/xpu/sycl/FillKernel.h> and call at::native::xpu::fill_kernel. Please let us know if you need more help.

[YangKai0616]

@CuiYifeng , thanks for the clarification! Let me confirm:

• Based on "unable to manually download the current repo", your requirement is to be able to directly find XPU SYCL header files in the repo pytorch instead of in torch-xpu-ops, right?

Yes, that’s correct. When I install PyTorch via pip, and then try to write custom kernels for the XPU backend, I cannot directly include headers like: #include <ATen/native/xpu/sycl/Loops.h> (https://github.com/intel/torch-xpu-ops/blob/24fab67b6ecf7620d0cf776047a3056c5b518bab/src/ATen/native/xpu/sycl/Loops.h). The compiler will report:fatal error: 'ATen/native/xpu/sycl/Loops.h' file not found.

However, when writing similar kernels for the CUDA backend, I can directly include headers such as: #include <ATen/native/cuda/Loops.cuh> (https://github.com/pytorch/pytorch/blob/f34744d2a5eae350682cd234178ade441057e1d6/aten/src/ATen/native/cuda/Loops.cuh) and it works without issues.

So my question is: do you have any plans to make XPU SYCL headers (like Loops.h) available in the pip-installed PyTorch package, similar to how CUDA headers are exposed?

Thanks!

[CuiYifeng]

@YangKai0616 Thanks for the confirmation.
Yes, we are considering gradually contributing torch-xpu-ops code (including XPU SYCL headers) to torch.

[EikanWang]

@guangyey , I remembered we have a dedicated task to do it and have implemented it, right?

[dvrogozh]

@EikanWang, we had a task to install XPU specific Aten ops codegen header files such as:

$ find torch/include/ATen/ops -name "*xpu*"
torch/include/ATen/ops/baddbmm_xpu_dispatch.h
torch/include/ATen/ops/bmm_xpu_dispatch.h
torch/include/ATen/ops/mm_xpu_dispatch.h
...

That task is completed with these commits in torch-xpu-ops and pytorch repos:

• pytorch/pytorch@57ab39f
• 51d33c6

The change is targeting to first appear in Pytorch 2.9 release.

However, the request from @YangKai0616 is to expose different set of files. They are not being installed as far as I see:

# In the source tree:
$ find . -name Loops.h
./torch/include/ATen/native/cpu/Loops.h
./aten/src/ATen/native/cpu/Loops.h
./third_party/torch-xpu-ops/src/ATen/native/xpu/sycl/Loops.h

# In the installed tree:
$ find . -name Loops.h
./lib/python3.12/site-packages/torch/include/ATen/native/cpu/Loops.h

As you can see only CPU version is getting installed for the XPU enabled build. I am not sure if there is a task in a todo list already to handle these headers as well (@guangyey?).

Just like how it's done with cuda: #include <ATen/native/cuda/Loops.cuh>

I wonder if there are any pytorch level tests getting a try to these headers for CUDA. We could then extend for XPU and verify the fix once it will be ready.

[guangyey]

We expose some xpu internal functions as TORCH_XPU_API, however, these header files are never installed to PyTorch. It doesn't make sense. I will install these header files in PyTorch.