vllm-xpu

MEDIUM

maintainer cantosun99 0 votes scanned 2026-10-18 11:24:11.986802

View on AUR

Why flagged

The package pipes bootstrap.pypa.io/get-pip.py directly to a Python interpreter at build time (unverified remote code execution), then installs packages from https://wheels.vllm.ai/xpu/ — a non-official, non-whitelisted wheel index that could serve arbitrary prebuilt binaries; while wheels.vllm.ai appears to be the vLLM project's own distribution host and bootstrap.pypa.io is the canonical pip bootstrap, the combination of piping a remote script to a shell and pulling prebuilt wheels from an extra index with no checksum verification represents a meaningful supply-chain risk.

Triggered rules

Medium External download from an untrusted host, not in source=() external_download_not_in_source

curl/wget fetches a URL on a non-allowlisted host that is not part of source=(), so it is not checksum-verified by makepkg.

PKGBUILD:56 curl -sS https://bootstrap.pypa.io/get-pip.py | "$pkgdir/opt/vllm-xpu/bin/python"

Medium Package install from an untrusted index/registry host install_untrusted_index

An install command is pointed at a custom package index / registry on a non-allowlisted host (--index-url / --registry / --extra-index-url), a classic dependency-confusion vector.

PKGBUILD:74 $_pip install -v --extra-index-url https://wheels.vllm.ai/xpu/ -r "$_req_file"
PKGBUILD:78 $_pip install -v --extra-index-url https://wheels.vllm.ai/xpu/ .

Low Few votes, recently uploaded zero_votes_recent

Uploaded within the last 14 days with 2 or fewer community votes — little peer review so far.

Medium AI review llm_review

An AI model (anthropic/claude-sonnet-4.6) reviewed this and agrees it is MEDIUM (confidence 72%): The package pipes bootstrap.pypa.io/get-pip.py directly to a Python interpreter at build time (unverified remote code execution), then installs packages from https://wheels.vllm.ai/xpu/ — a non-official, non-whitelisted wheel index that could serve arbitrary prebuilt binaries; while wheels.vllm.ai appears to be the vLLM project's own distribution host and bootstrap.pypa.io is the canonical pip bootstrap, the combination of piping a remote script to a shell and pulling prebuilt wheels from an extra index with no checksum verification represents a meaningful supply-chain risk.

PKGBUILD

3 offending line(s) highlighted

      1# Maintainer: cantosun99 <privat at cantosun dot de>
    
      2pkgname=vllm-xpu
    
      3pkgver=0.32.0
    
      4pkgrel=1
    
      5pkgdesc="vLLM optimized for Intel Arc GPUs"
    
      6arch=('x86_64')
    
      7url="https://github.com/vllm-project/vllm"
    
      8license=('MIT')
    
      9# python311 is required because vLLM and Triton XPU explicitly reject Python 3.14
    
      10depends=('python311' 'intel-compute-runtime' 'level-zero-loader' 'intel-gmmlib' 'intel-deep-learning-essentials')
    
      11makedepends=('python311' 'cmake' 'ninja' 'gcc' 'curl')
    
      12provides=('vllm')
    
      13conflicts=('vllm')
    
      14source=("vllm-${pkgver}.tar.gz::https://github.com/vllm-project/vllm/archive/refs/tags/v${pkgver}.tar.gz")
    
      15sha256sums=('bf592c132f27bbb5f65cff0bbd2ee873eae722b79c8286934fd5b5cf88b43d61')
    
      16install=$pkgname.install
    
      17
    
      18build() {
    
      19  cd "vllm-${pkgver}"
    
      20
    
      21  # Source Intel OneAPI environment for the SYCL compiler
    
      22  if [ -f /opt/intel/oneapi/setvars.sh ]; then
    
      23    source /opt/intel/oneapi/setvars.sh > /dev/null 2>&1 || true
    
      24  else
    
      25    echo "ERROR: /opt/intel/oneapi/setvars.sh not found."
    
      26    echo "Please ensure intel-deep-learning-essentials is fully installed."
    
      27    exit 1
    
      28  fi
    
      29
    
      30  # Set XPU target and limit RAM usage (MAX_JOBS)
    
      31  export VLLM_TARGET_DEVICE=xpu
    
      32  export MAX_JOBS=${MAX_JOBS:-$(nproc)}
    
      33
    
      34  # Fix for setuptools-scm failing on GitHub release tarballs (missing .git folder)
    
      35  export SETUPTOOLS_SCM_PRETEND_VERSION=$pkgver
    
      36}
    
      37
    
      38package() {
    
      39  cd "vllm-${pkgver}"
    
      40
    
      41  # Re-source OneAPI for the package step
    
      42  if [ -f /opt/intel/oneapi/setvars.sh ]; then
    
      43    source /opt/intel/oneapi/setvars.sh > /dev/null 2>&1 || true
    
      44  fi
    
      45
    
      46  export VLLM_TARGET_DEVICE=xpu
    
      47  export MAX_JOBS=${MAX_JOBS:-$(nproc)}
    
      48  export SETUPTOOLS_SCM_PRETEND_VERSION=$pkgver
    
      49
    
      50  # Create an isolated virtual environment directly in the install prefix.
    
      51  # We use --without-pip just in case Arch's python311 split package removed ensurepip.
    
      52  # This perfectly bypasses PEP 668 (EXTERNALLY-MANAGED) restrictions!
    
      53  /usr/bin/python3.11 -m venv --without-pip "$pkgdir/opt/vllm-xpu"
    
      54
    
      55  # Download and install pip directly into the isolated venv
    
      56  curl -sS https://bootstrap.pypa.io/get-pip.py | "$pkgdir/opt/vllm-xpu/bin/python"
    
      57
    
      58  local _pip="$pkgdir/opt/vllm-xpu/bin/python -m pip"
    
      59
    
      60  # Upgrade pip and wheel
    
      61  $_pip install --upgrade pip wheel setuptools
    
      62
    
      63  # Find the correct XPU requirements file (vLLM changed its folder structure recently)
    
      64  local _req_file=""
    
      65  if [ -f requirements-xpu.txt ]; then
    
      66    _req_file="requirements-xpu.txt"
    
      67  elif [ -f requirements/xpu.txt ]; then
    
      68    _req_file="requirements/xpu.txt"
    
      69  fi
    
      70
    
      71  # Install XPU specific requirements.
    
      72  # We force the extra index URL to guarantee pip finds triton==3.7.2+xpu!
    
      73  if [ -n "$_req_file" ]; then
    
      74    $_pip install -v --extra-index-url https://wheels.vllm.ai/xpu/ -r "$_req_file"
    
      75  fi
    
      76
    
      77  # Build and install vLLM
    
      78  $_pip install -v --extra-index-url https://wheels.vllm.ai/xpu/ .
    
      79
    
      80  # Fix shebang paths that were hardcoded to the build directory.
    
      81  # During package creation, pip writes shebangs pointing to the build tree
    
      82  # (e.g., /home/user/.../pkg/vllm-xpu/opt/vllm-xpu/bin/python), which no longer
    
      83  # exists after installation. Rewrite them to point to the final install location.
    
      84  local _build_python="$pkgdir/opt/vllm-xpu/bin/python"
    
      85  local _final_python="/opt/vllm-xpu/bin/python"
    
      86  find "$pkgdir/opt/vllm-xpu/bin" -type f -print0 | \
    
      87    xargs -0 sed -i "s|#!${_build_python}|#!${_final_python}|g"
    
      88
    
      89  # Create wrapper scripts in /usr/bin
    
      90  install -d "$pkgdir/usr/bin"
    
      91
    
      92  # Wrapper for OpenAI API server (legacy entry point alias)
    
      93  cat <<'EOF' > "$pkgdir/usr/bin/vllm-xpu"
    
      94#!/bin/bash
    
      95[ -f /opt/intel/oneapi/setvars.sh ] && source /opt/intel/oneapi/setvars.sh > /dev/null 2>&1 || true
    
      96exec /opt/vllm-xpu/bin/vllm serve "$@"
    
      97EOF
    
      98  chmod +x "$pkgdir/usr/bin/vllm-xpu"
    
      99
    
      100  # Wrapper for vllm CLI (e.g. vllm serve, vllm bench)
    
      101  cat <<'EOF' > "$pkgdir/usr/bin/vllm-serve-xpu"
    
      102#!/bin/bash
    
      103[ -f /opt/intel/oneapi/setvars.sh ] && source /opt/intel/oneapi/setvars.sh > /dev/null 2>&1 || true
    
      104exec /opt/vllm-xpu/bin/vllm "$@"
    
      105EOF
    
      106  chmod +x "$pkgdir/usr/bin/vllm-serve-xpu"
    
      107}
    
      108

Scan history

Scanned at (UTC)	Severity	Rules
2026-10-18 11:24:11	Medium	4
2026-10-18 11:22:05	Medium	4