[ChenDAO1]

What this PR does / why we need it?

This PR adds a reproducible CPU-only package build and verification path for vllm-ascend. It validates that the package can be installed and built without Ascend NPU devices or the CANN toolkit, while preserving the normal NPU/runtime dependency contract.

Specifically, this PR:

• adds Dockerfile.cpu, based on openEuler 24.03 LTS, as a clean CPU-only build environment;
• adds tools/cpu_only_build.sh as the executable source of truth for dependency installation, package build, and verification;
• installs exact +cpu PyTorch wheels while resolving their generic dependencies from the main package index, then installs the complete pyproject.toml build-system requirements before building the package;
• builds with COMPILE_CUSTOM_KERNELS=0, an explicit SOC_VERSION, and TORCH_DEVICE_BACKEND_AUTOLOAD=0, so the verification path does not require CANN or NPU hardware;
• keeps arctic-inference==0.1.1 in requirements.txt and therefore in the normal vLLM Ascend runtime dependency contract;
• excludes only the exact arctic-inference==0.1.1 requirement from CPU-only verification because its source distribution is incompatible with the CPU-only build environment;
• filters Arctic Inference both when installing CPU-only dependencies and while setup.py generates editable-install package metadata, restoring the original requirements.txt through a shell trap on success or failure;
• verifies that no NPU device or CANN installation is visible, CPU PyTorch cannot access CUDA, Arctic Inference is absent only from the CPU-only environment, the package metadata is available, and pip check succeeds;
• adds a path-filtered GitHub Actions workflow on the linux-amd64-cpu-4 self-hosted runner;
• uses the runner's BuildKit path through docker/build-push-action@v7, with CPU-only verification executed as a Dockerfile build layer so the job does not require a local Docker daemon or docker run;
• pulls the openEuler base image through the self-hosted runner's reachable public SWR proxy and routes openEuler DNF, PyPI, CPU PyTorch, and Ascend package traffic through CI caches, while retaining Docker Hub and public-index defaults for local builds;
• updates installation and suffix speculative decoding documentation, plus dependency-contract regression tests.

arctic-inference was removed only from pyproject.toml build-system requirements because it is not required to execute the package build backend. Normal installations still receive the exact runtime dependency through requirements.txt and setup.py.

Does this PR introduce any user-facing change?

No user-facing API or NPU runtime behavior changes are introduced.

Normal vLLM Ascend installations continue to include arctic-inference==0.1.1, and suffix speculative decoding keeps its existing dependency. The Arctic Inference exclusion is local to the CPU-only CI verification path and does not affect normal NPU installations.

How was this patch tested?

The patch was validated with:

• bash -n tools/cpu_only_build.sh;

• git diff --check;

• actionlint for .github/workflows/cpu_only_build.yaml;

• python -m unittest tests.ut.test_dependency_documentation -v;

• CPU-only dependency, build, metadata, device/CANN absence, Arctic Inference absence, and pip check verification encoded in Dockerfile.cpu and tools/cpu_only_build.sh;

• the public SWR openEuler manifest and the DNF repository rewrite used by the self-hosted BuildKit path;

• GitHub Actions CPU-only BuildKit verification passed on the current PR head; the repository E2E selected tests were also triggered.

• vLLM version: v0.27.1

• vLLM main: vllm-project/vllm@58d3918

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request enhances the project's CI infrastructure by adding a dedicated CPU-only build verification path. This ensures that the package remains buildable and dependency-consistent even in environments lacking NPU hardware or the CANN toolkit. Additionally, it cleans up the dependency requirements by removing an optional plugin that was causing installation issues in restricted environments, providing clearer installation guidance for users.

Highlights

• CPU-only Build Verification: Introduced a new Dockerfile.cpu and a supporting shell script to enable reproducible, NPU-free build verification for the project.
• Dependency Management: Removed arctic-inference from the universal requirements list as it is an optional plugin, and updated documentation to reflect that it should be installed separately.
• CI/CD Pipeline: Added an executable regression layer to the CI pipeline to ensure the package can be built and verified in environments without CANN toolkit or NPU hardware.

New Features

🧠 You can now enable Memory (public preview) to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Ignored Files

• Ignored by pattern: .github/workflows/** (1)
  • .github/workflows/cpu_only_build.yaml

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize the Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counterproductive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

This pull request introduces a reproducible CPU-only build and verification environment by adding Dockerfile.cpu and tools/cpu_only_build.sh. It also removes arctic-inference from the core requirements, documenting it as an optional plugin instead. Feedback suggests updating the PR title and summary to align with the repository's style guide, and improving the device verification script to check for any /dev/davinci* device rather than just /dev/davinci0 to prevent false negatives.

[gemini-code-assist]

Here is the suggested PR Title and Summary according to the repository style guide:

Suggested PR Title:

[CI][Feature] Add reproducible CPU-only build verification

Suggested PR Summary:

### What this PR does / why we need it?

This PR introduces a reproducible CPU-only build verification path to ensure that `vllm-ascend` can be built and verified in environments without CANN or NPU device support. 

Specifically, it:
- Adds `Dockerfile.cpu` based on openEuler 24.03 LTS.
- Implements `tools/cpu_only_build.sh` to handle dependency installation, package building, and CPU-only verification.
- Removes `arctic-inference==0.1.1` from the universal requirements (`pyproject.toml` and `requirements.txt`) since its source build requires PyTorch 2.7, making it incompatible with the current CPU stack. It is now documented as an optional plugin.
- Updates documentation to reflect that `arctic-inference` is optional and should be installed separately.

### Does this PR introduce _any_ user-facing change?

No, this is a CI/build-system enhancement and documentation update. There are no user-facing API or behavioral changes to `vllm-ascend` itself, other than making `arctic-inference` an optional dependency.

### How was this patch tested?

The changes were validated via:
- CPU-only Docker image build verification.
- Dependency installation and package build with `COMPILE_CUSTOM_KERNELS=0`.
- Verification of CPU-only environment (no NPU/CANN) and successful `pip check`.

References

1. Adhere to the Pull Request Summary Style Guide for formatting the PR Title and Summary. ^(link)

[MrZ20]

Make sure the test case does not include it; otherwise, direct deletion is not recommended.

[ChenDAO1]

I will Keeps arctic-inference==0.1.1 in the normal vLLM Ascend dependency contract, while excluding it only from the CPU-only build verification path because it is incompatible with the CPU-only PyTorch build environment.

[MrZ20]

Please determine the scope in which this test is triggered. At minimum, I believe the following items should not be included in the trigger scope; otherwise, this test may become a per‑pr test, while its intended purpose is to perform validation only when specific content is modified.

      - 'vllm_ascend/**'
      - 'tests/ut/test_dependency_documentation.py'
      - '.github/workflows/scripts/test_config.yaml'
      - '.github/workflows/scripts/select_tests.py'
      - 'docs/source/installation.md'
      - 'docs/source/tutorials/features/suffix_speculative_decoding.md'
      - 'docs/source/user_guide/feature_guide/speculative_decoding.md'

[ChenDAO1]

Thanks for pointing this out. I agree that the current path filter is broader than necessary and may cause the CPU-only build verification to run for unrelated PRs. I'll narrow the trigger scope to dependency/build-related files. For vllm_ascend/**, I'll keep only the source files that are directly imported during package build, such as vllm_ascend/envs.py and vllm_ascend/device/hardware.py.

[ChenDAO1]

/rerun
[Bot]: rerun completed.

Failed:

• E2E: still in progress, retry /rerun after completion

[ChenDAO1]

/cancel
[Bot]: successfully force-cancelled the following runs:

https://github.com/vllm-project/vllm-ascend/actions/runs/32207228102

[ChenDAO1]

/rerun
[Bot]: rerun completed.

Failed:

• E2E: some jobs could not be re-run (still waiting on dependencies)

[zhangxinyuehfad]

Plz use self-host runner, don't use github runner.

[ChenDAO1]

already fixed this

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.