[CI/CD] build wheel in manylinux container

[jikunshang]

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS ABOVE HAVE BEEN CONSIDERED.

Purpose

use pytorch build wheel docker pytorch/manylinux2_28-builder:xpu-main
https://github.com/pytorch/pytorch/blob/main/.ci/manywheel/build_xpu.sh
igc in this docker image is 2.20.5

Test Plan

CI job pass
https://github.com/vllm-project/vllm-xpu-kernels/actions/runs/22673444106

Test Result

(Optional) Documentation Update

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing (anything written below this line will be removed by GitHub Actions)

[Copilot]

Pull request overview

Updates the per-commit wheel build workflow to run inside PyTorch’s manylinux2_28 XPU builder container, aiming to produce wheels in an environment closer to PyTorch’s official manywheel build setup.

Changes:

• Switch the wheel build job container image to pytorch/manylinux2_28-builder:xpu-main.
• Source Intel oneAPI environment before building, and install build requirements with pip.
• Post-process the produced wheel filename to use a manylinux_2_28_x86_64 platform tag.

---

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

[Copilot]

The workflow uses a moving container tag (pytorch/manylinux2_28-builder:xpu-main). This makes builds non-reproducible and can break unexpectedly when the image changes. Consider pinning the image by digest (or a versioned tag) so the wheel build environment is stable over time.

Suggested change

	image: pytorch/manylinux2_28-builder:xpu-main
	image: pytorch/manylinux2_28-builder@sha256:1234567890abcdef1234567890abcdef1234567890abcdef1234567890abcdef

[Copilot]

LD_LIBRARY_PATH is exported after setup.py bdist_wheel runs, so it won’t affect compilation/linking during the wheel build. If this path is required for building (e.g., to find Torch shared libs), export it before invoking python3 setup.py ...; otherwise this line is dead code and should be removed to avoid confusion.

Suggested change

	python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38
	export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/conda/lib/python3.11/site-packages/torch/lib/
	export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/conda/lib/python3.11/site-packages/torch/lib/
	python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38

[Copilot]

Renaming the wheel filename from linux_x86_64 to manylinux_2_28_x86_64 can mislabel the wheel without actually ensuring manylinux compliance (e.g., external shared library dependencies). Instead of mv, use a proper manylinux repair step (typically auditwheel repair, or build with the correct --plat-name and verify policy compliance) so the tag accurately reflects the wheel’s ABI/packaging.

Suggested change

	python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38
	export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/conda/lib/python3.11/site-packages/torch/lib/
	for f in dist/*-linux_x86_64.whl; do mv "$f" "${f/linux_x86_64/manylinux_2_28_x86_64}"; done
	pip install auditwheel
	python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38
	export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/conda/lib/python3.11/site-packages/torch/lib/
	auditwheel repair --plat manylinux_2_28_x86_64 -w dist dist/*-linux_x86_64.whl