Package Details: python-vllm-cuda 0.20.1-1

Git Clone URL: https://aur.archlinux.org/python-vllm-cuda.git (read-only, click to copy)
Package Base: python-vllm-cuda
Description: high-throughput and memory-efficient inference and serving engine for LLMs
Upstream URL: https://github.com/vllm-project/vllm
Licenses: Apache-2.0
Conflicts: python-vllm
Provides: python-vllm
Submitter: envolution
Maintainer: ncihnegn
Last Packager: ncihnegn
Votes: 3
Popularity: 1.01
First Submitted: 2024-12-01 16:12 (UTC)
Last Updated: 2026-05-04 07:10 (UTC)

Pinned Comments

Latest Comments

arzeth commented on 2026-07-04 16:46 (UTC)

I figured out how to update to 0.24.0.
Step 0/2: change pkgver; add 'python-cutlass' 'cutlass' 'cutlass-headers' 'python-cbor2' 'python-flashinfer' to depends=(...). Note: I can't test whether AUR package cutlass conflicts with non-AUR cutlass-headers because VLLM's qutlass doesn't support pre-Blackwell (so maybe *cutlass* packages should be in optdepends instead?).

Step 1/2: new prepare() {}:

prepare() {
  cd $_pkgname

  _gitRepo=$(awk '/FetchContent_Declare\(/,/\)/{if(/GIT_REPOSITORY/){print $2}}' cmake/external_projects/vllm_flash_attn.cmake)
  _gitTag=$( awk '/FetchContent_Declare\(/,/\)/{if(/GIT_TAG/){print $2}}' cmake/external_projects/vllm_flash_attn.cmake)
  mkdir .deps || rm -rf .deps/vllm-flash-attn-src
  git clone --recurse-submodules "$_gitRepo" .deps/vllm-flash-attn-src
  cd .deps/vllm-flash-attn-src
  git checkout --recurse-submodules "$_gitTag"
  patch -Np1 -i ${srcdir}/121.patch
}

Step 2/2: add

export VLLM_FLASH_ATTN_SRC_DIR="${srcdir}/${_pkgname}/.deps/vllm-flash-attn-src"

into build() {}

P.S. Since this is python-vllm-cuda, maybe change python-torchvision to python-torchvision-cuda, python-pytorch to python-pytorch-opt-cuda for those who don't have any *torch* packages installed yet? Or does pacman (like yay) asks which variant to choose anyway? Though, a user could be using --noconfirm and thus get the inferior variants.

orkut commented on 2026-07-04 13:20 (UTC)

Thanks for this beautiful repo:)

Can you update it to the latest version please?

ganelon commented on 2026-05-10 02:07 (UTC)

@Strykar It got past that part when I manually cloned it rather than using yay. However, it is not yet done building... we'll see

Strykar commented on 2026-04-30 23:50 (UTC)

Build fails at:

/home/strykar/.cache/aurutils/sync/python-vllm-cuda/PKGBUILD: line 94: cd: .deps/vllm-flash-attn-src: No such file or directory
==> ERROR: A failure occurred in prepare().
    Aborting...
FAILED: [code=4] python-vllm-cuda.stamp

Is this even required anymore?

envolution commented on 2025-02-12 21:30 (UTC)

@Sherlock-Holo thanks for your report, it's been added to makedepends

Sherlock-Holo commented on 2025-02-11 09:28 (UTC) (edited on 2025-02-11 09:30 (UTC) by Sherlock-Holo)

when build this package, it says

Traceback (most recent call last):
  File "/home/sherlock/.cache/yay/python-vllm-cuda/src/vllm/setup.py", line 18, in <module>
    from setuptools_scm import get_version
ModuleNotFoundError: No module named 'setuptools_scm'

if add the miss makedepends python-setuptools-scm, it will fail with

Traceback (most recent call last):
  File "/home/sherlock/.cache/yay/python-vllm-cuda/src/vllm/setup.py", line 633, in <module>
    version=get_vllm_version(),
            ~~~~~~~~~~~~~~~~^^
  File "/home/sherlock/.cache/yay/python-vllm-cuda/src/vllm/setup.py", line 527, in get_vllm_version
    raise RuntimeError("Unknown runtime environment")
RuntimeError: Unknown runtime environment

envolution commented on 2024-12-28 04:47 (UTC) (edited on 2024-12-28 04:51 (UTC) by envolution)

Not working currently due to lack of python 3.13 support in vllm-flash-attention. Try python-vllm-bin or the cpu version python-vllm