[release] Add nightly wheel release index

[khluu]

Summary

Add a nightly wheel release pipeline for vllm-omni. Wheels are hosted on the existing vllm-wheels S3 bucket under the omni/ prefix, served via wheels.vllm.ai.

How it works

1. Build — Installs uv with Python 3.12, runs python3 -m build to produce a pure Python wheel
2. Upload — Uploads the wheel to s3://vllm-wheels/omni/<commit>/
3. Index — Generates a PEP 503 compliant index and uploads it to:
   • s3://vllm-wheels/omni/<commit>/ (always)
   • s3://vllm-wheels/omni/nightly/ (when NIGHTLY=1 env var is set)
   • s3://vllm-wheels/omni/<version>/ (for non-dev release versions only)

The nightly index at omni/nightly/ uses relative paths pointing back to the latest commit's wheel directory.

Installation

uv pip install vllm-omni --extra-index-url https://wheels.vllm.ai/omni/nightly --torch-backend=auto

Files added

File	Description
.buildkite/nightly-release-pipeline.yaml	Two-step pipeline: build wheel, then generate index
.buildkite/scripts/upload-nightly-wheels.sh	Uploads wheel from dist/ to S3 under omni/<commit>/
.buildkite/scripts/generate-and-upload-nightly-index.sh	Lists wheels from S3, generates index, uploads to nightly/commit/version paths
.buildkite/scripts/generate-nightly-index.py	PEP 503 index generator (simplified from vLLM's script)

S3 layout

s3://vllm-wheels/
├── omni/
│   ├── <commit>/
│   │   ├── vllm_omni-*.whl
│   │   ├── index.html
│   │   └── vllm-omni/
│   │       ├── index.html
│   │       └── metadata.json
│   └── nightly/          # symlink-like copy of latest commit's index
│       ├── index.html
│       └── vllm-omni/
│           ├── index.html
│           └── metadata.json
├── nightly/              # existing vLLM nightly (unaffected)
└── ...

🤖 Generated with Claude Code

[khluu]

You can try it out with uv pip install vllm-omni --extra-index-url https://wheels.vllm.ai/omni/nightly --torch-backend=auto

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 9c66d3671e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Parse +npu/+xpu release suffixes as variants

Update parse_from_filename to recognize all release device suffixes the project can emit. The current check only treats rocm, cu, and cpu suffixes as variants, so release wheels like ...+npu or ...+xpu (produced by setup.py) are misclassified as the default variant. If multiple device wheels are uploaded under one commit/version directory, the generated default index can point at NPU/XPU artifacts instead of isolating them under variant subdirectories, which leads to incorrect wheel resolution from the generic index.

Useful? React with 👍 / 👎.

[khluu]

Sample run: https://buildkite.com/vllm/omni-release/builds/9

[Gaohan123]

LGTM. Let's check if it can solve the pipeline failure https://buildkite.com/vllm/omni-release/builds/21/steps/canvas

[david6666666]

Could you add a documentation explaining how users can use this feature? Thank you very much.