I might not be fully getting what we want to achieve by resolving this issue.

Looking at the driver.py:find_sycl() code after #3135 it seems that it's already generic enough to not depend on icpx (if icpx is not installed the sycl headers are deduced from either ONEAPI_PATH or intel-sycl-rt package).

My understanding judging by this comment is that resolving this issue implies placing the sycl headers in third_party/intel/backend/includes, by either placing them manually (like AMD) or by downloading them from somewhere in setup.py (like Nvidia). But what's the point of that if Triton XPU requires us to have intel-sycl-rt installed, so we always can deduce the sycl headers with this existing logic?

cc @anmyachev (I have seen you've reviewed the original PR, so I guess you may have some thoughts on that)

SYCL has several use scenarios in Triton XPU, see #3135 (comment). The second scenario is using g++ but relying on icpx.

This issue is to align with others' practice of preparing runtime headers libs in advance and eliminate the hidden dependence on icpx, just like gcc compile cuda program using cudart without the participation of nvcc. https://github.com/intel/intel-xpu-backend-for-triton/blob/main/third_party/nvidia/backend/driver.py#L15-L47

Shipping the SYCL headers as part of Triton is not ideal because they might go out of sync with the SYCL-RT version. I think Triton should continue to use the headers shipped with the SYCL-RT.

@vlad-penkin WDYT ?