[CI][BUILD] update dockerfile and build script

[jikunshang]

…ild wheel script

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS ABOVE HAVE BEEN CONSIDERED.

Purpose

1. use config file to control which umd/igc to use when build docker,
2. add build wheel script

some usage

./build_script/build_wheel_docker.sh --output-dir /tmp/wheels --version 1.0.0
./build_script/build_wheel_docker.sh --no-cache --max-jobs 8
# build a wheel using specific igc/umd version
./build_script/build_wheel_docker.sh --gpu-profile igc-2.27.10-cr-26.01

default on should be 2.22 or 2.24, due to some known igc issue.

1. 2.11 not support g31
2. 2.27 build kernel hang
3. 2.28 some triton kernel failed in vllm-xpu ci.

Test Plan

Test Result

(Optional) Documentation Update

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing (anything written below this line will be removed by GitHub Actions)

[Copilot]

Pull request overview

This PR updates the XPU Docker build to install Intel GPU runtime packages via a selectable JSON “profile”, and adds a helper script to build the vllm-xpu-kernels wheel inside Docker for reproducibility.

Changes:

• Added a profile-driven GPU runtime installer script + JSON config of runtime package URLs.
• Updated Dockerfile.xpu to install the GPU runtime using the new profile mechanism (GPU_RUNTIME_PROFILE build-arg).
• Added a build_wheel_docker.sh script to build and export wheels from a Docker container.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 10 comments.

File	Description
build_script/install_gpu_runtime.sh	New installer that selects a profile from JSON and installs matching .deb packages.
build_script/gpu_runtime_packages.json	Defines available GPU runtime profiles and their package URLs.
build_script/build_wheel_docker.sh	New script to build wheels inside the Dockerfile.xpu-built image and copy artifacts to host.
Dockerfile.xpu	Switches GPU runtime install from hardcoded URLs to the new configurable installer/profile.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

WORKDIR=$(mktemp -d) is only cleaned up on the success path. With set -e, any failure (wget/dpkg/python) will leak the temp directory. Add a trap to remove ${WORKDIR} on EXIT/ERR after creation.

[Copilot]

The copied config file name is /tmp/gpu_runtime_packages.json, but the RUN line passes /tmp/gpu_runtime_packages.json (missing gpu_runtime_). As written, install_gpu_runtime.sh will fail with “Config file not found”. Fix the path (or rename the copied file) so it matches the actual filename.

Suggested change

	COPY build_script/gpu_runtime_packages.json build_script/install_gpu_runtime.sh /tmp/
	RUN bash /tmp/install_gpu_runtime.sh /tmp/gpu_runtime_packages.json ${GPU_RUNTIME_PROFILE}
	COPY build_script/gpu_runtime_packages.json /tmp/gpu_packages.json
	COPY build_script/install_gpu_runtime.sh /tmp/
	RUN bash /tmp/install_gpu_runtime.sh /tmp/gpu_packages.json ${GPU_RUNTIME_PROFILE}

[Copilot]

The python snippets interpolate ${CONFIG_FILE} / ${PROFILE} directly into Python source code inside a double-quoted python3 -c "..." string. If PROFILE (or the file path) contains quotes/newlines, this can break parsing and can allow code injection during the build. Prefer passing values via environment variables / argv (e.g., python3 -c '...' -- "$CONFIG_FILE" "$PROFILE") or using a here-doc with no shell interpolation.

Suggested change

	PROFILE=$(python3 -c "
	import json, sys
	cfg = json.load(open('${CONFIG_FILE}'))
	print(cfg['default'])
	")
	fi
	echo "==> Installing GPU runtime packages from profile: ${PROFILE}"
	# Extract the list of URLs for the selected profile
	URLS=$(python3 -c "
	import json, sys
	cfg = json.load(open('${CONFIG_FILE}'))
	profile = cfg['profiles'].get('${PROFILE}')
	if profile is None:
	print(f'ERROR: Profile \"${PROFILE}\" not found. Available: {list(cfg[\"profiles\"].keys())}', file=sys.stderr)
	sys.exit(1)
	print('\n'.join(profile['packages']))
	")
	PROFILE=$(python3 - "$CONFIG_FILE" << 'PY'
	import json
	import sys
	config_path = sys.argv[1]
	with open(config_path, 'r') as f:
	cfg = json.load(f)
	print(cfg['default'])
	PY
	)
	fi
	echo "==> Installing GPU runtime packages from profile: ${PROFILE}"
	# Extract the list of URLs for the selected profile
	URLS=$(python3 - "$CONFIG_FILE" "$PROFILE" << 'PY'
	import json
	import sys
	config_path = sys.argv[1]
	profile_name = sys.argv[2]
	with open(config_path, 'r') as f:
	cfg = json.load(f)
	profiles = cfg.get('profiles', {})
	profile = profiles.get(profile_name)
	if profile is None:
	available = list(profiles.keys())
	print(f'ERROR: Profile "{profile_name}" not found. Available: {available}', file=sys.stderr)
	sys.exit(1)
	packages = profile.get('packages', [])
	print('\n'.join(packages))
	PY
	)

[Copilot]

Installing the downloaded .debs via dpkg -i can leave the image in a broken state if dependencies are missing (dpkg doesn’t resolve deps). Prefer apt-get install -y ./*.deb (after apt-get update) or follow dpkg -i with apt-get -f install -y to reliably resolve dependencies.

Suggested change

	dpkg -i "${WORKDIR}"/*.deb
	apt-get update && apt-get install -y "${WORKDIR}"/*.deb

[Copilot]

One of the package URLs has a trailing space. Because install_gpu_runtime.sh uses IFS= with read -r, that whitespace will be preserved and passed to wget, causing the download to fail. Please remove the trailing space (or trim URLs in the installer).

Suggested change

	"https://github.com/intel/intel-graphics-compiler/releases/download/v2.24.8/intel-igc-core-2_2.24.8+20344_amd64.deb ",
	"https://github.com/intel/intel-graphics-compiler/releases/download/v2.24.8/intel-igc-core-2_2.24.8+20344_amd64.deb",

[Copilot]

The profile description for igc-2.27.10-cr-26.01 states a Compute Runtime version 25.53.36486.8, but the package URLs in this profile are for 26.01.36711.4. Update the description (or URLs) so the metadata matches what will actually be installed.

[Copilot]

The in-container git reinitialization path runs git commit, which will fail if user.name / user.email aren’t configured in the container (common in CI/base images). Configure them locally before committing (or use git -c user.name=... -c user.email=... commit ...) to make this script robust.

Suggested change

	git commit -q -m "build" --allow-empty
	git tag -a v0.0.0 -m "placeholder" 2>/dev/null || true
	}
	else
	git init -q
	git add -A
	git commit -q -m "build" --allow-empty
	git -c user.name="vllm-build-bot" -c user.email="vllm-build-bot@example.com" commit -q -m "build" --allow-empty
	git tag -a v0.0.0 -m "placeholder" 2>/dev/null || true
	}
	else
	git init -q
	git add -A
	git -c user.name="vllm-build-bot" -c user.email="vllm-build-bot@example.com" commit -q -m "build" --allow-empty

[Copilot]

Argument parsing assumes every option that takes a value is followed by $2. If a user passes e.g. --output-dir as the last argument, the script will error under set -u with a less helpful message. Add an explicit check for missing option values and print a clear usage error.

Suggested change

	# ─── Parse arguments ───────────────────────────────────────────────────────
	while [[ $# -gt 0 ]]; do
	case "$1" in
	--output-dir)
	OUTPUT_DIR="$2"; shift 2 ;;
	--image-name)
	IMAGE_NAME="$2"; shift 2 ;;
	--gpu-profile)
	GPU_RUNTIME_PROFILE="$2"; shift 2 ;;
	--version)
	VLLM_VERSION_OVERRIDE="$2"; shift 2 ;;
	--max-jobs)
	MAX_JOBS="$2"; shift 2 ;;
	--no-cache)
	DOCKER_NO_CACHE="--no-cache"; shift ;;
	-h|--help)
	sed -n '2,/^$/p' "$0" | sed 's/^# \?//'; exit 0 ;;
	*)
	echo "ERROR: Unknown option: $1" >&2; exit 1 ;;
	usage() {
	sed -n '2,/^$/p' "$0" | sed 's/^# \?//'
	}
	# ─── Parse arguments ───────────────────────────────────────────────────────
	while [[ $# -gt 0 ]]; do
	case "$1" in
	--output-dir)
	if [[ $# -lt 2 ]]; then
	echo "ERROR: Missing value for option '$1'." >&2
	usage
	exit 1
	fi
	OUTPUT_DIR="$2"; shift 2 ;;
	--image-name)
	if [[ $# -lt 2 ]]; then
	echo "ERROR: Missing value for option '$1'." >&2
	usage
	exit 1
	fi
	IMAGE_NAME="$2"; shift 2 ;;
	--gpu-profile)
	if [[ $# -lt 2 ]]; then
	echo "ERROR: Missing value for option '$1'." >&2
	usage
	exit 1
	fi
	GPU_RUNTIME_PROFILE="$2"; shift 2 ;;
	--version)
	if [[ $# -lt 2 ]]; then
	echo "ERROR: Missing value for option '$1'." >&2
	usage
	exit 1
	fi
	VLLM_VERSION_OVERRIDE="$2"; shift 2 ;;
	--max-jobs)
	if [[ $# -lt 2 ]]; then
	echo "ERROR: Missing value for option '$1'." >&2
	usage
	exit 1
	fi
	MAX_JOBS="$2"; shift 2 ;;
	--no-cache)
	DOCKER_NO_CACHE="--no-cache"; shift ;;
	-h|--help)
	usage
	exit 0 ;;
	*)
	echo "ERROR: Unknown option: $1" >&2
	usage
	exit 1 ;;

[Copilot]

The usage comment doesn’t match the script’s argument order. The script treats $1 as CONFIG_FILE and $2 as PROFILE, but the header suggests calling it with only a profile name. Either update the usage/docs or add argument detection (e.g., if $1 is an existing file, treat it as CONFIG_FILE; otherwise treat it as PROFILE).

Suggested change

	# RUN bash /tmp/install_gpu_runtime.sh [PROFILE_NAME]
	#
	# If PROFILE_NAME is omitted or "default", the default profile from the JSON is used.
	set -euo pipefail
	CONFIG_FILE="${1:-/tmp/gpu_runtime_packages.json}"
	PROFILE="${2:-default}"
	# # Use default config file (/tmp/gpu_runtime_packages.json) and optional profile:
	# RUN bash /tmp/install_gpu_runtime.sh [PROFILE_NAME]
	# # Or specify an explicit config file and optional profile:
	# RUN bash /tmp/install_gpu_runtime.sh [CONFIG_FILE] [PROFILE_NAME]
	#
	# If PROFILE_NAME is omitted or "default", the default profile from the JSON is used.
	set -euo pipefail
	CONFIG_FILE_DEFAULT="/tmp/gpu_runtime_packages.json"
	if [ $# -ge 1 ] && [ -f "$1" ]; then
	# First argument is a config file path; optional second argument is the profile name.
	CONFIG_FILE="$1"
	PROFILE="${2:-default}"
	elif [ $# -ge 1 ]; then
	# First argument is treated as the profile name; use the default config file path.
	CONFIG_FILE="${CONFIG_FILE_DEFAULT}"
	PROFILE="$1"
	else
	# No arguments: use default config file and default profile.
	CONFIG_FILE="${CONFIG_FILE_DEFAULT}"
	PROFILE="default"
	fi

[Copilot]

This script invokes python3, but Dockerfile.xpu only installs python3.12 (no explicit python3 package). To avoid a build-time failure if /usr/bin/python3 isn’t present in the base image, either call python3.12 here or ensure python3 is installed in the image before running this script.