vllm - 💡(How to fix) Fix [Doc]: cu129 nightly channel is stale — latest wheel pre-dates PyPI stable, breaks documented install

📚 The doc issue

Description

The cu129 nightly channel's latest wheel is 0.20.2rc1.dev233+g5497ffbf7.cu129, while PyPI now ships 0.20.2 final (CUDA 13). The cu129 pre-release sorts below the PyPI stable per PEP 440, so the documented install command resolves to PyPI:

uv pip install -U vllm --pre
--extra-index-url https://wheels.vllm.ai/nightly/cu129
--extra-index-url https://download.pytorch.org/whl/cu129
--index-strategy unsafe-best-match

→ installs vllm 0.20.2; every compiled .so links libcudart.so.13 (confirmed via readelf -d); CUDA-12 envs crash with ImportError: libcudart.so.13.

Under healthy release flow, cu129 nightlies should already be ahead of stable (e.g. 0.20.3rc1.devN+cu129) and would win the resolution — the command would do the right thing. The artifact catalog at wheels.vllm.ai/nightly/cu129/ appears stalled.

Suggest a potential alternative/fix

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.