xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-09 02:34:56 +08:00
Code Issues Packages Projects Releases Wiki Activity

Revert #29787 and #29690 (#29815)

Browse Source
This commit is contained in:
Kevin H. Luu 2025-12-01 13:42:03 -08:00 committed by GitHub
parent eaf81485ed
commit 1336a1ea24
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
9 changed files with 185 additions and 580 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

46
.buildkite/generate_index.py Normal file
View File

@ -0,0 +1,46 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import argparse
import os
template = """<!DOCTYPE html>
<html>
<body>
<h1>Links for vLLM</h1/>
<a href="../{x86_wheel_html_escaped}">{x86_wheel}</a><br/>
<a href="../{arm_wheel_html_escaped}">{arm_wheel}</a><br/>
</body>
</html>
"""
parser = argparse.ArgumentParser()
parser.add_argument("--wheel", help="The wheel path.", required=True)
args = parser.parse_args()
filename = os.path.basename(args.wheel)
with open("index.html", "w") as f:
print(f"Generated index.html for {args.wheel}")
# sync the abi tag with .buildkite/scripts/upload-wheels.sh
if "x86_64" in filename:
x86_wheel = filename
arm_wheel = filename.replace("x86_64", "aarch64").replace(
"manylinux1", "manylinux2014"
)
elif "aarch64" in filename:
x86_wheel = filename.replace("aarch64", "x86_64").replace(
"manylinux2014", "manylinux1"
)
arm_wheel = filename
else:
raise ValueError(f"Unsupported wheel: {filename}")
# cloudfront requires escaping the '+' character
f.write(
template.format(
x86_wheel=x86_wheel,
x86_wheel_html_escaped=x86_wheel.replace("+", "%2B"),
arm_wheel=arm_wheel,
arm_wheel_html_escaped=arm_wheel.replace("+", "%2B"),
)
)

16
.buildkite/release-pipeline.yaml
View File

@ -8,7 +8,7 @@ steps:
commands:
# #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
# https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
- "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
- "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg VLLM_MAIN_CUDA_VERSION=12.9 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
- "mkdir artifacts"
- "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
- "bash .buildkite/scripts/upload-wheels.sh"
@ -30,6 +30,19 @@ steps:
DOCKER_BUILDKIT: "1"
# x86 + CUDA builds
- label: "Build wheel - CUDA 12.8"
depends_on: ~
id: build-wheel-cuda-12-8
agents:
queue: cpu_queue_postmerge
commands:
- "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
- "mkdir artifacts"
- "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
- "bash .buildkite/scripts/upload-wheels.sh"
env:
DOCKER_BUILDKIT: "1"
- label: "Build wheel - CUDA 12.9"
depends_on: ~
id: build-wheel-cuda-12-9
@ -96,6 +109,7 @@ steps:
- label: "Annotate release workflow"
depends_on:
- create-multi-arch-manifest
- build-wheel-cuda-12-8
id: annotate-release-workflow
agents:
queue: cpu_queue_postmerge

369
.buildkite/scripts/generate-nightly-index.py
View File

@ -1,369 +0,0 @@
#!/usr/bin/env python3
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# do not complain about line length (for docstring)
# ruff: noqa: E501
import argparse
import json
import re
import sys
from dataclasses import asdict, dataclass
from pathlib import Path
from typing import Any
from urllib.parse import quote
if not sys.version_info >= (3, 10):
raise RuntimeError("This script requires Python 3.10 or higher.")
INDEX_HTML_TEMPLATE = """<!DOCTYPE html>
<html>
<meta name="pypi:repository-version" content="1.0">
<body>
{items}
</body>
</html>
"""
@dataclass
class WheelFileInfo:
package_name: str
version: str
build_tag: str | None
python_tag: str
abi_tag: str
platform_tag: str
variant: str | None
filename: str
def parse_from_filename(file: str) -> WheelFileInfo:
"""
Parse wheel file name to extract metadata.
The format of wheel names:
{package_name}-{version}(-{build_tag})?-{python_tag}-{abi_tag}-{platform_tag}.whl
All versions could contain a variant like '+cu129' or '.cpu' or `.rocm` (or not).
Example:
vllm-0.11.0-cp38-abi3-manylinux1_x86_64.whl
vllm-0.10.2rc2+cu129-cp38-abi3-manylinux2014_aarch64.whl
vllm-0.11.1rc8.dev14+gaa384b3c0-cp38-abi3-manylinux2014_aarch64.whl
vllm-0.11.1rc8.dev14+gaa384b3c0.cu130-cp38-abi3-manylinux1_x86_64.whl
"""
wheel_file_re = re.compile(
r"^(?P<package_name>.+)-(?P<version>[^-]+?)(-(?P<build_tag>[^-]+))?-(?P<python_tag>[^-]+)-(?P<abi_tag>[^-]+)-(?P<platform_tag>[^-]+)\.whl$"
)
match = wheel_file_re.match(file)
if not match:
raise ValueError(f"Invalid wheel file name: {file}")
package_name = match.group("package_name")
version = match.group("version")
build_tag = match.group("build_tag")
python_tag = match.group("python_tag")
abi_tag = match.group("abi_tag")
platform_tag = match.group("platform_tag")
# extract variant from version
variant = None
if "dev" in version:
ver_after_dev = version.split("dev")[-1]
if "." in ver_after_dev:
variant = ver_after_dev.split(".")[-1]
version = version.removesuffix("." + variant)
else:
if "+" in version:
version, variant = version.split("+")
return WheelFileInfo(
package_name=package_name,
version=version,
build_tag=build_tag,
python_tag=python_tag,
abi_tag=abi_tag,
platform_tag=platform_tag,
variant=variant,
filename=file,
)
def generate_project_list(subdir_names: list[str]) -> str:
"""
Generate project list HTML content linking to each project & variant sub-directory.
"""
href_tags = []
for name in sorted(subdir_names):
name = name.strip("/").strip(".")
href_tags.append(f' <a href="{name}/">{name}/</a><br/>')
return INDEX_HTML_TEMPLATE.format(items="\n".join(href_tags))
def generate_package_index_and_metadata(
wheel_files: list[WheelFileInfo], wheel_base_dir: Path, index_base_dir: Path
) -> tuple[str, str]:
"""
Generate package index HTML content for a specific package, linking to actual wheel files.
"""
href_tags = []
metadata = []
for file in sorted(wheel_files, key=lambda x: x.filename):
relative_path = (
wheel_base_dir.relative_to(index_base_dir, walk_up=True) / file.filename
)
# handle with '+' in URL, and avoid double-encoding '/' and already-encoded '%2B'
# NOTE: this is AWS S3 specific behavior!
file_path_quoted = quote(relative_path.as_posix(), safe=":%/")
href_tags.append(f' <a href="{file_path_quoted}">{file.filename}</a><br/>')
file_meta = asdict(file)
file_meta["path"] = file_path_quoted
metadata.append(file_meta)
index_str = INDEX_HTML_TEMPLATE.format(items="\n".join(href_tags))
metadata_str = json.dumps(metadata, indent=2)
return index_str, metadata_str
def generate_index_and_metadata(
whl_files: list[str],
wheel_base_dir: Path,
index_base_dir: Path,
default_variant: str | None = None,
alias_to_default: str | None = None,
):
"""
Generate index for all wheel files.
Args:
whl_files (list[str]): List of wheel files (must be directly under `wheel_base_dir`).
wheel_base_dir (Path): Base directory for wheel files.
index_base_dir (Path): Base directory to store index files.
default_variant (str | None): The default variant name, if any.
alias_to_default (str | None): Alias variant name for the default variant, if any.
First, parse all wheel files to extract metadata.
We need to collect all wheel files for each variant, and generate an index for it (in a sub-directory).
The index for the default variant (if any) is generated in the root index directory.
If `default_variant` is provided, all wheels must have variant suffixes, and the default variant index
is purely a copy of the corresponding variant index, with only the links adjusted.
Otherwise, all wheels without variant suffixes are treated as the default variant.
If `alias_to_default` is provided, an additional alias sub-directory is created, it has the same content
as the default variant index, but the links are adjusted accordingly.
Index directory structure:
index_base_dir/ (hosted at wheels.vllm.ai/{nightly,$commit,$version}/)
index.html # project list, linking to "vllm/" and other packages, and all variant sub-directories
vllm/
index.html # package index, pointing to actual files in wheel_base_dir (relative path)
metadata.json # machine-readable metadata for all wheels in this package
cpu/ # cpu variant sub-directory
index.html
vllm/
index.html
metadata.json
cu129/ # cu129 is actually the alias to default variant
index.html
vllm/
index.html
metadata.json
cu130/ # cu130 variant sub-directory
index.html
vllm/
index.html
metadata.json
...
metadata.json stores a dump of all wheel files' metadata in a machine-readable format:
[
{
"package_name": "vllm",
"version": "0.10.2rc2",
"build_tag": null,
"python_tag": "cp38",
"abi_tag": "abi3",
"platform_tag": "manylinux2014_aarch64",
"variant": "cu129",
"filename": "vllm-0.10.2rc2+cu129-cp38-abi3-manylinux2014_aarch64.whl",
"path": "../vllm-0.10.2rc2%2Bcu129-cp38-abi3-manylinux2014_aarch64.whl" # to be concatenated with the directory URL and URL-encoded
},
...
]
"""
parsed_files = [parse_from_filename(f) for f in whl_files]
if not parsed_files:
print("No wheel files found, skipping index generation.")
return
# Group by variant
variant_to_files: dict[str, list[WheelFileInfo]] = {}
for file in parsed_files:
variant = file.variant or "default"
if variant not in variant_to_files:
variant_to_files[variant] = []
variant_to_files[variant].append(file)
print(f"Found variants: {list(variant_to_files.keys())}")
# sanity check for default variant
if default_variant:
if "default" in variant_to_files:
raise ValueError(
"All wheel files must have variant suffixes when `default_variant` is specified."
)
if default_variant not in variant_to_files:
raise ValueError(
f"Default variant '{default_variant}' not found among wheel files."
)
if alias_to_default:
if "default" not in variant_to_files:
# e.g. only some wheels are uploaded to S3 currently
print(
"[WARN] Alias to default variant specified, but no default variant found."
)
elif alias_to_default in variant_to_files:
raise ValueError(
f"Alias variant name '{alias_to_default}' already exists among wheel files."
)
else:
variant_to_files[alias_to_default] = variant_to_files["default"].copy()
print(f"Alias variant '{alias_to_default}' created for default variant.")
# Generate index for each variant
subdir_names = set()
for variant, files in variant_to_files.items():
if variant == "default":
variant_dir = index_base_dir
else:
variant_dir = index_base_dir / variant
subdir_names.add(variant)
variant_dir.mkdir(parents=True, exist_ok=True)
# gather all package names in this variant
packages = set(f.package_name for f in files)
if variant == "default":
# these packages should also appear in the "project list"
# generate after all variants are processed
subdir_names = subdir_names.union(packages)
else:
# generate project list for this variant directly
project_list_str = generate_project_list(sorted(packages))
with open(variant_dir / "index.html", "w") as f:
f.write(project_list_str)
for package in packages:
# filter files belonging to this package only
package_files = [f for f in files if f.package_name == package]
package_dir = variant_dir / package
package_dir.mkdir(parents=True, exist_ok=True)
index_str, metadata_str = generate_package_index_and_metadata(
package_files, wheel_base_dir, package_dir
)
with open(package_dir / "index.html", "w") as f:
f.write(index_str)
with open(package_dir / "metadata.json", "w") as f:
f.write(metadata_str)
# Generate top-level project list index
project_list_str = generate_project_list(sorted(subdir_names))
with open(index_base_dir / "index.html", "w") as f:
f.write(project_list_str)
if __name__ == "__main__":
"""
Arguments:
--version <version> : version string for the current build (e.g., commit hash)
--current-objects <path_to_json> : path to JSON file containing current S3 objects listing in this version directory
--output-dir <output_directory> : directory to store generated index files
--alias-to-default <alias_variant_name> : (optional) alias variant name for the default variant
"""
parser = argparse.ArgumentParser(
description="Process nightly build wheel files to generate indices."
)
parser.add_argument(
"--version",
type=str,
required=True,
help="Version string for the current build (e.g., commit hash)",
)
parser.add_argument(
"--current-objects",
type=str,
required=True,
help="Path to JSON file containing current S3 objects listing in this version directory",
)
parser.add_argument(
"--output-dir",
type=str,
required=True,
help="Directory to store generated index files",
)
parser.add_argument(
"--alias-to-default",
type=str,
default=None,
help="Alias variant name for the default variant",
)
args = parser.parse_args()
version = args.version
if "/" in version or "\\" in version:
raise ValueError("Version string must not contain slashes.")
current_objects_path = Path(args.current_objects)
output_dir = Path(args.output_dir)
if not output_dir.exists():
output_dir.mkdir(parents=True, exist_ok=True)
# Read current objects JSON
with open(current_objects_path) as f:
current_objects: dict[str, list[dict[str, Any]]] = json.load(f)
# current_objects looks like from list_objects_v2 S3 API:
"""
"Contents": [
{
"Key": "e2f56c309d2a28899c68975a7e104502d56deb8f/vllm-0.11.2.dev363+ge2f56c309-cp38-abi3-manylinux1_x86_64.whl",
"LastModified": "2025-11-28T14:00:32+00:00",
"ETag": "\"37a38339c7cdb61ca737021b968075df-52\"",
"ChecksumAlgorithm": [
"CRC64NVME"
],
"ChecksumType": "FULL_OBJECT",
"Size": 435649349,
"StorageClass": "STANDARD"
},
...
]
"""
# Extract wheel file keys
wheel_files = []
for item in current_objects.get("Contents", []):
key: str = item["Key"]
if key.endswith(".whl"):
wheel_files.append(key.split("/")[-1]) # only the filename is used
print(f"Found {len(wheel_files)} wheel files for version {version}: {wheel_files}")
# Generate index and metadata, assuming wheels and indices are stored as:
# s3://vllm-wheels/{version}/<wheel files>
# s3://vllm-wheels/<anything>/<index files>
wheel_base_dir = Path(output_dir).parent / version
index_base_dir = Path(output_dir)
generate_index_and_metadata(
whl_files=wheel_files,
wheel_base_dir=wheel_base_dir,
index_base_dir=index_base_dir,
default_variant=None,
alias_to_default=args.alias_to_default,
)
print(f"Successfully generated index and metadata in {output_dir}")

119
.buildkite/scripts/upload-wheels.sh
View File

@ -2,28 +2,6 @@
set -ex
# ======== part 0: setup ========
BUCKET="vllm-wheels"
INDICES_OUTPUT_DIR="indices"
DEFAULT_VARIANT_ALIAS="cu129" # align with vLLM_MAIN_CUDA_VERSION in vllm/envs.py
PYTHON=${PYTHON_PROG:=python3} # try to read from env var, otherwise use python3
SUBPATH=$BUILDKITE_COMMIT
S3_COMMIT_PREFIX="s3://$BUCKET/$SUBPATH/"
# detect if python3.10+ is available
has_new_python=$($PYTHON -c "print(1 if __import__('sys').version_info >= (3,10) else 0)")
if [[ "$has_new_python" -eq 0 ]]; then
# use new python from docker
docker pull python:3-slim
PYTHON="docker run --rm -v $(pwd):/app -w /app python:3-slim python3"
fi
echo "Using python interpreter: $PYTHON"
echo "Python version: $($PYTHON --version)"
# ========= part 1: collect, rename & upload the wheel ==========
# Assume wheels are in artifacts/dist/*.whl
wheel_files=(artifacts/dist/*.whl)
@ -32,69 +10,74 @@ if [[ ${#wheel_files[@]} -ne 1 ]]; then
echo "Error: Expected exactly one wheel file in artifacts/dist/, but found ${#wheel_files[@]}"
exit 1
fi
# Get the single wheel file
wheel="${wheel_files[0]}"
# current build image uses ubuntu 20.04, which corresponds to manylinux_2_31
# refer to https://github.com/mayeut/pep600_compliance?tab=readme-ov-file#acceptable-distros-to-build-wheels
manylinux_version="manylinux_2_31"
# Detect architecture and rename 'linux' to appropriate manylinux version
arch=$(uname -m)
if [[ $arch == "x86_64" ]]; then
manylinux_version="manylinux1"
elif [[ $arch == "aarch64" ]]; then
manylinux_version="manylinux2014"
else
echo "Warning: Unknown architecture $arch, using manylinux1 as default"
manylinux_version="manylinux1"
fi
# Rename 'linux' to the appropriate manylinux version in the wheel filename
if [[ "$wheel" != *"linux"* ]]; then
echo "Error: Wheel filename does not contain 'linux': $wheel"
exit 1
fi
new_wheel="${wheel/linux/$manylinux_version}"
mv -- "$wheel" "$new_wheel"
wheel="$new_wheel"
echo "Renamed wheel to: $wheel"
# Extract the version from the wheel
version=$(unzip -p "$wheel" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
echo "Version in wheel: $version"
pure_version="${version%%+*}"
echo "Pure version (without variant): $pure_version"
echo "Version: $version"
# copy wheel to its own bucket
aws s3 cp "$wheel" "$S3_COMMIT_PREFIX"
normal_wheel="$wheel" # Save the original wheel filename
# ========= part 2: generate and upload indices ==========
# generate indices for all existing wheels in the commit directory
# this script might be run multiple times if there are multiple variants being built
# so we need to guarantee there is little chance for "TOCTOU" issues
# i.e., one process is generating indices while another is uploading a new wheel
# so we need to ensure no time-consuming operations happen below
# If the version contains "dev", rename it to v1.0.0.dev for consistency
if [[ $version == *dev* ]]; then
suffix="${version##*.}"
if [[ $suffix == cu* ]]; then
new_version="1.0.0.dev+${suffix}"
else
new_version="1.0.0.dev"
fi
new_wheel="${wheel/$version/$new_version}"
# use cp to keep both files in the artifacts directory
cp -- "$wheel" "$new_wheel"
wheel="$new_wheel"
version="$new_version"
fi
# list all wheels in the commit directory
echo "Existing wheels on S3:"
aws s3 ls "$S3_COMMIT_PREFIX"
obj_json="objects.json"
aws s3api list-objects-v2 --bucket "$BUCKET" --prefix "$SUBPATH/" --delimiter / --output json > "$obj_json"
mkdir -p "$INDICES_OUTPUT_DIR"
# Upload the wheel to S3
python3 .buildkite/generate_index.py --wheel "$normal_wheel"
# call script to generate indicies for all existing wheels
# this indices have relative paths that could work as long as it is next to the wheel directory in s3
# i.e., the wheels are always in s3://vllm-wheels/<commit>/
# and indices can be placed in /<commit>/, or /nightly/, or /<version>/
if [[ ! -z "$DEFAULT_VARIANT_ALIAS" ]]; then
alias_arg="--alias-to-default $DEFAULT_VARIANT_ALIAS"
# generate index for this commit
aws s3 cp "$wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
aws s3 cp "$normal_wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
if [[ $normal_wheel == *"cu129"* ]]; then
# only upload index.html for cu129 wheels (default wheels) as it
# is available on both x86 and arm64
aws s3 cp index.html "s3://vllm-wheels/$BUILDKITE_COMMIT/vllm/index.html"
aws s3 cp "s3://vllm-wheels/nightly/index.html" "s3://vllm-wheels/$BUILDKITE_COMMIT/index.html"
else
alias_arg=""
echo "Skipping index files for non-cu129 wheels"
fi
$PYTHON .buildkite/scripts/generate-nightly-index.py --version "$SUBPATH" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" $alias_arg
# generate index for nightly
aws s3 cp "$wheel" "s3://vllm-wheels/nightly/"
aws s3 cp "$normal_wheel" "s3://vllm-wheels/nightly/"
# copy indices to /<commit>/ unconditionally
echo "Uploading indices to $S3_COMMIT_PREFIX"
aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "$S3_COMMIT_PREFIX"
# copy to /nightly/ only if it is on the main branch and not a PR
if [[ "$BUILDKITE_BRANCH" == "main" && "$BUILDKITE_PULL_REQUEST" == "false" ]]; then
echo "Uploading indices to overwrite /nightly/"
aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/nightly/"
if [[ $normal_wheel == *"cu129"* ]]; then
# only upload index.html for cu129 wheels (default wheels) as it
# is available on both x86 and arm64
aws s3 cp index.html "s3://vllm-wheels/nightly/vllm/index.html"
else
echo "Skipping index files for non-cu129 wheels"
fi
# copy to /<pure_version>/ only if it does not have "dev" in the version
if [[ "$version" != *"dev"* ]]; then
echo "Uploading indices to overwrite /$pure_version/"
aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/$pure_version/"
fi
aws s3 cp "$wheel" "s3://vllm-wheels/$version/"
aws s3 cp index.html "s3://vllm-wheels/$version/vllm/index.html"

15
docs/getting_started/installation/cpu.md
View File

@ -46,23 +46,10 @@ vLLM is a Python library that supports the following CPU variants. Select your C
### Pre-built wheels
Please refer to the instructions for [pre-built wheels on GPU](./gpu.md#pre-built-wheels).
When specifying the index URL, please make sure to use the `cpu` variant subdirectory.
For example, the nightly build index is: `https://wheels.vllm.ai/nightly/cpu/`.
Currently, there are no pre-built CPU wheels.
### Build wheel from source
#### Set up using Python-only build (without compilation) {#python-only-build}
Please refer to the instructions for [Python-only build on GPU](./gpu.md#python-only-build), and replace the build commands with:
```bash
VLLM_USE_PRECOMPILED=1 VLLM_PRECOMPILED_WHEEL_VARIANT=cpu VLLM_TARGET_DEVICE=cpu uv pip install --editable .
```
#### Full build (with compilation) {#full-build}
=== "Intel/AMD x86"
--8<-- "docs/getting_started/installation/cpu.x86.inc.md:build-wheel-from-source"

73
docs/getting_started/installation/gpu.cuda.inc.md
View File

@ -26,50 +26,43 @@ uv pip install vllm --torch-backend=auto
??? console "pip"
```bash
# Install vLLM with CUDA 12.9.
pip install vllm --extra-index-url https://download.pytorch.org/whl/cu129
# Install vLLM with CUDA 12.8.
pip install vllm --extra-index-url https://download.pytorch.org/whl/cu128
```
We recommend leveraging `uv` to [automatically select the appropriate PyTorch index at runtime](https://docs.astral.sh/uv/guides/integration/pytorch/#automatic-backend-selection) by inspecting the installed CUDA driver version via `--torch-backend=auto` (or `UV_TORCH_BACKEND=auto`). To select a specific backend (e.g., `cu128`), set `--torch-backend=cu128` (or `UV_TORCH_BACKEND=cu128`). If this doesn't work, try running `uv self update` to update `uv` first.
We recommend leveraging `uv` to [automatically select the appropriate PyTorch index at runtime](https://docs.astral.sh/uv/guides/integration/pytorch/#automatic-backend-selection) by inspecting the installed CUDA driver version via `--torch-backend=auto` (or `UV_TORCH_BACKEND=auto`). To select a specific backend (e.g., `cu126`), set `--torch-backend=cu126` (or `UV_TORCH_BACKEND=cu126`). If this doesn't work, try running `uv self update` to update `uv` first.
!!! note
NVIDIA Blackwell GPUs (B200, GB200) require a minimum of CUDA 12.8, so make sure you are installing PyTorch wheels with at least that version. PyTorch itself offers a [dedicated interface](https://pytorch.org/get-started/locally/) to determine the appropriate pip command to run for a given target configuration.
As of now, vLLM's binaries are compiled with CUDA 12.9 and public PyTorch release versions by default. We also provide vLLM binaries compiled with CUDA 12.8, 13.0, and public PyTorch release versions:
As of now, vLLM's binaries are compiled with CUDA 12.8 and public PyTorch release versions by default. We also provide vLLM binaries compiled with CUDA 12.6, 11.8, and public PyTorch release versions:
```bash
# Install vLLM with a specific CUDA version (e.g., 13.0).
# Install vLLM with a specific CUDA version (e.g., 11.8 or 12.6).
export VLLM_VERSION=$(curl -s https://api.github.com/repos/vllm-project/vllm/releases/latest | jq -r .tag_name | sed 's/^v//')
export CUDA_VERSION=130 # or other
uv pip install https://github.com/vllm-project/vllm/releases/download/v${VLLM_VERSION}/vllm-${VLLM_VERSION}+cu${CUDA_VERSION}-cp38-abi3-manylinux_2_31_x86_64.whl --extra-index-url https://download.pytorch.org/whl/cu${CUDA_VERSION}
export CUDA_VERSION=118 # or 126
uv pip install https://github.com/vllm-project/vllm/releases/download/v${VLLM_VERSION}/vllm-${VLLM_VERSION}+cu${CUDA_VERSION}-cp38-abi3-manylinux1_x86_64.whl --extra-index-url https://download.pytorch.org/whl/cu${CUDA_VERSION}
```
#### Install the latest code
LLM inference is a fast-evolving field, and the latest code may contain bug fixes, performance improvements, and new features that are not released yet. To allow users to try the latest code without waiting for the next release, vLLM provides wheels for every commit since `v0.5.3` on <https://wheels.vllm.ai/nightly>. There are multiple indices that could be used:
* `https://wheels.vllm.ai/nightly`: the default variant (CUDA with version specified in `VLLM_MAIN_CUDA_VERSION`) built with the last commit on the `main` branch. Currently it is CUDA 12.9.
* `https://wheels.vllm.ai/nightly/<variant>`: all other variants. Now this includes `cu130`, and `cpu`. The default variant (`cu129`) also has a subdirectory to keep consistency.
To install from nightly index, run:
LLM inference is a fast-evolving field, and the latest code may contain bug fixes, performance improvements, and new features that are not released yet. To allow users to try the latest code without waiting for the next release, vLLM provides wheels for Linux running on an x86 platform with CUDA 12 for every commit since `v0.5.3`.
```bash
uv pip install -U vllm \
--torch-backend=auto \
--extra-index-url https://wheels.vllm.ai/nightly # add variant subdirectory here if needed
--extra-index-url https://wheels.vllm.ai/nightly
```
!!! warning "`pip` caveat"
Using `pip` to install from nightly indices is _not supported_, because `pip` combines packages from `--extra-index-url` and the default index, choosing only the latest version, which makes it difficult to install a development version prior to the released version. In contrast, `uv` gives the extra index [higher priority than the default index](https://docs.astral.sh/uv/pip/compatibility/#packages-that-exist-on-multiple-indexes).
If you insist on using `pip`, you have to specify the full URL of the wheel file (which can be obtained from the web page).
??? console "pip"
```bash
pip install -U https://wheels.vllm.ai/nightly/vllm-0.11.2.dev399%2Bg3c7461c18-cp38-abi3-manylinux_2_31_x86_64.whl # current nightly build (the filename will change!)
pip install -U https://wheels.vllm.ai/${VLLM_COMMIT}/vllm-0.11.2.dev399%2Bg3c7461c18-cp38-abi3-manylinux_2_31_x86_64.whl # from specific commit
pip install -U vllm \
--pre \
--extra-index-url https://wheels.vllm.ai/nightly
```
`--pre` is required for `pip` to consider pre-released versions.
##### Install specific revisions
If you want to access the wheels for previous commits (e.g. to bisect the behavior change, performance regression), you can specify the commit hash in the URL:
@ -78,13 +71,33 @@ If you want to access the wheels for previous commits (e.g. to bisect the behavi
export VLLM_COMMIT=72d9c316d3f6ede485146fe5aabd4e61dbc59069 # use full commit hash from the main branch
uv pip install vllm \
--torch-backend=auto \
--extra-index-url https://wheels.vllm.ai/${VLLM_COMMIT} # add variant subdirectory here if needed
--extra-index-url https://wheels.vllm.ai/${VLLM_COMMIT}
```
The `uv` approach works for vLLM `v0.6.6` and later and offers an easy-to-remember command. A unique feature of `uv` is that packages in `--extra-index-url` have [higher priority than the default index](https://docs.astral.sh/uv/pip/compatibility/#packages-that-exist-on-multiple-indexes). If the latest public release is `v0.6.6.post1`, `uv`'s behavior allows installing a commit before `v0.6.6.post1` by specifying the `--extra-index-url`. In contrast, `pip` combines packages from `--extra-index-url` and the default index, choosing only the latest version, which makes it difficult to install a development version prior to the released version.
??? note "pip"
If you want to access the wheels for previous commits (e.g. to bisect the behavior change,
performance regression), due to the limitation of `pip`, you have to specify the full URL of the
wheel file by embedding the commit hash in the URL:
```bash
export VLLM_COMMIT=33f460b17a54acb3b6cc0b03f4a17876cff5eafd # use full commit hash from the main branch
pip install https://wheels.vllm.ai/${VLLM_COMMIT}/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl
```
Note that the wheels are built with Python 3.8 ABI (see [PEP
425](https://peps.python.org/pep-0425/) for more details about ABI), so **they are compatible
with Python 3.8 and later**. The version string in the wheel file name (`1.0.0.dev`) is just a
placeholder to have a unified URL for the wheels, the actual versions of wheels are contained in
the wheel metadata (the wheels listed in the extra index url have correct versions). Although we
don't support Python 3.8 any more (because PyTorch 2.5 dropped support for Python 3.8), the
wheels are still built with Python 3.8 ABI to keep the same wheel name as before.
# --8<-- [end:pre-built-wheels]
# --8<-- [start:build-wheel-from-source]
#### Set up using Python-only build (without compilation) {#python-only-build}
#### Set up using Python-only build (without compilation)
If you only need to change Python code, you can build and install vLLM without compilation. Using `uv pip`'s [`--editable` flag](https://docs.astral.sh/uv/pip/packages/#editable-packages), changes you make to the code will be reflected when you run vLLM:
@ -108,24 +121,18 @@ This command will do the following:
In case you see an error about wheel not found when running the above command, it might be because the commit you based on in the main branch was just merged and the wheel is being built. In this case, you can wait for around an hour to try again, or manually assign the previous commit in the installation using the `VLLM_PRECOMPILED_WHEEL_LOCATION` environment variable.
```bash
export VLLM_PRECOMPILED_WHEEL_COMIMT=$(git rev-parse HEAD~1) # or earlier commit on main
export VLLM_USE_PRECOMPILED=1
export VLLM_COMMIT=72d9c316d3f6ede485146fe5aabd4e61dbc59069 # use full commit hash from the main branch
export VLLM_PRECOMPILED_WHEEL_LOCATION=https://wheels.vllm.ai/${VLLM_COMMIT}/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl
uv pip install --editable .
```
There are more environment variables to control the behavior of Python-only build:
* `VLLM_PRECOMPILED_WHEEL_LOCATION`: specify the exact wheel URL or local file path of a pre-compiled wheel to use. All other logic to find the wheel will be skipped.
* `VLLM_PRECOMPILED_WHEEL_COMMIT`: override the commit hash to download the pre-compiled wheel. It can be `nightly` to use the last **already built** commit on the main branch.
* `VLLM_PRECOMPILED_WHEEL_VARIANT`: specify the variant subdirectory to use on the nightly index, e.g., `cu129`, `cpu`. If not specified, the CUDA variant with `VLLM_MAIN_CUDA_VERSION` will be tried, then fallback to the default variant on the remote index.
You can find more information about vLLM's wheels in [Install the latest code](#install-the-latest-code).
!!! note
There is a possibility that your source code may have a different commit ID compared to the latest vLLM wheel, which could potentially lead to unknown errors.
It is recommended to use the same commit ID for the source code as the vLLM wheel you have installed. Please refer to [Install the latest code](#install-the-latest-code) for instructions on how to install a specified wheel.
#### Full build (with compilation) {#full-build}
#### Full build (with compilation)
If you want to modify C++ or CUDA code, you'll need to build vLLM from source. This can take several minutes:

2
docs/getting_started/installation/gpu.md
View File

@ -52,7 +52,7 @@ vLLM is a Python library that supports the following GPU variants. Select your G
--8<-- "docs/getting_started/installation/gpu.xpu.inc.md:set-up-using-python"
### Pre-built wheels {#pre-built-wheels}
### Pre-built wheels
=== "NVIDIA CUDA"

118
setup.py
View File

@ -310,6 +310,9 @@ class cmake_build_ext(build_ext):
class precompiled_build_ext(build_ext):
"""Disables extension building when using precompiled binaries."""
def run(self) -> None:
assert _is_cuda(), "VLLM_USE_PRECOMPILED is only supported for CUDA builds"
def build_extensions(self) -> None:
print("Skipping build_ext: using precompiled extensions.")
return
@ -319,17 +322,14 @@ class precompiled_wheel_utils:
"""Extracts libraries and other files from an existing wheel."""
@staticmethod
def extract_precompiled_and_patch_package(
wheel_url_or_path: str, download_filename: str | None
) -> dict:
def extract_precompiled_and_patch_package(wheel_url_or_path: str) -> dict:
import tempfile
import zipfile
temp_dir = None
try:
if not os.path.isfile(wheel_url_or_path):
# use provided filename first, then derive from URL
wheel_filename = download_filename or wheel_url_or_path.split("/")[-1]
wheel_filename = wheel_url_or_path.split("/")[-1]
temp_dir = tempfile.mkdtemp(prefix="vllm-wheels")
wheel_path = os.path.join(temp_dir, wheel_filename)
print(f"Downloading wheel from {wheel_url_or_path} to {wheel_path}")
@ -648,102 +648,38 @@ package_data = {
]
}
def _fetch_metadata_for_variant(
commit: str, variant: str | None
) -> tuple[list[dict], str]:
variant_dir = f"{variant}/" if variant is not None else ""
repo_url = f"https://wheels.vllm.ai/{commit}/{variant_dir}vllm/"
meta_url = repo_url + "metadata.json"
logger.info("Trying to fetch metadata from {}", meta_url)
from urllib.request import urlopen
with urlopen(meta_url) as resp:
# urlopen raises HTTPError on unexpected status code
wheels = json.loads(resp.read().decode("utf-8"))
return wheels, repo_url
# If using precompiled, extract and patch package_data (in advance of setup)
if envs.VLLM_USE_PRECOMPILED:
# Attempts:
# 1. user-specified wheel location (can be either local or remote, via
# VLLM_PRECOMPILED_WHEEL_LOCATION)
# 2. user-specified variant from nightly repo (current main commit via
# VLLM_PRECOMPILED_WHEEL_VARIANT)
# 3. the variant corresponding to VLLM_MAIN_CUDA_VERSION from nightly repo
# 4. the default variant from nightly repo (current main commit)
assert _is_cuda(), "VLLM_USE_PRECOMPILED is only supported for CUDA builds"
wheel_location = os.getenv("VLLM_PRECOMPILED_WHEEL_LOCATION", None)
if wheel_location is not None:
wheel_url = wheel_location
download_filename = None
logger.info("Using user-specified precompiled wheel location: %s", wheel_url)
else:
import platform
arch = platform.machine()
# try to fetch the wheel metadata from the nightly wheel repo
main_variant = envs.VLLM_MAIN_CUDA_VERSION.replace(".", "")
variant = os.getenv("VLLM_PRECOMPILED_WHEEL_VARIANT", main_variant)
commit = os.getenv(
"VLLM_PRECOMPILED_WHEEL_COMMIT",
precompiled_wheel_utils.get_base_commit_in_main_branch(),
)
logger.info(
"Using precompiled wheel commit %s with variant %s", commit, variant
)
try_default = False
wheels, repo_url, download_filename = None, None, None
try:
wheels, repo_url = _fetch_metadata_for_variant(commit, variant)
except Exception:
logger.warning(
"Failed to fetch precompiled wheel metadata for variant %s",
variant,
exc_info=True,
)
try_default = True # try outside handler to keep the stacktrace simple
if try_default:
logger.info("Trying the default variant")
wheels, repo_url = _fetch_metadata_for_variant(commit, None)
# if this also fails, then we have nothing more to try / cache
assert wheels is not None and repo_url is not None, (
"Failed to fetch precompiled wheel metadata"
)
# The metadata.json has the following format:
# see .buildkite/scripts/generate-nightly-index.py for details
"""[{
"package_name": "vllm",
"version": "0.11.2.dev278+gdbc3d9991",
"build_tag": null,
"python_tag": "cp38",
"abi_tag": "abi3",
"platform_tag": "manylinux1_x86_64",
"variant": null,
"filename": "vllm-0.11.2.dev278+gdbc3d9991-cp38-abi3-manylinux1_x86_64.whl",
"path": "../vllm-0.11.2.dev278%2Bgdbc3d9991-cp38-abi3-manylinux1_x86_64.whl"
},
...]"""
for wheel in wheels:
# TODO: maybe check more compatibility later? (python_tag, abi_tag, etc)
if wheel.get("package_name") == "vllm" and arch in wheel.get(
"platform_tag", ""
):
logger.info("Found precompiled wheel metadata: %s", wheel)
if "path" not in wheel:
raise ValueError(f"Wheel metadata missing path: {wheel}")
wheel_url = repo_url + wheel["path"]
download_filename = wheel.get("filename")
logger.info("Using precompiled wheel URL: %s", wheel_url)
break
if arch == "x86_64":
wheel_tag = "manylinux1_x86_64"
elif arch == "aarch64":
wheel_tag = "manylinux2014_aarch64"
else:
raise ValueError(
f"No precompiled vllm wheel found for architecture {arch} "
f"from repo {repo_url}. All available wheels: {wheels}"
)
patch = precompiled_wheel_utils.extract_precompiled_and_patch_package(
wheel_url, download_filename
)
raise ValueError(f"Unsupported architecture: {arch}")
base_commit = precompiled_wheel_utils.get_base_commit_in_main_branch()
wheel_url = f"https://wheels.vllm.ai/{base_commit}/vllm-1.0.0.dev-cp38-abi3-{wheel_tag}.whl"
nightly_wheel_url = (
f"https://wheels.vllm.ai/nightly/vllm-1.0.0.dev-cp38-abi3-{wheel_tag}.whl"
)
from urllib.request import urlopen
try:
with urlopen(wheel_url) as resp:
if resp.status != 200:
wheel_url = nightly_wheel_url
except Exception as e:
print(f"[warn] Falling back to nightly wheel: {e}")
wheel_url = nightly_wheel_url
patch = precompiled_wheel_utils.extract_precompiled_and_patch_package(wheel_url)
for pkg, files in patch.items():
package_data.setdefault(pkg, []).extend(files)

7
vllm/envs.py
View File

@ -74,7 +74,7 @@ if TYPE_CHECKING:
VLLM_MEDIA_CONNECTOR: str = "http"
VLLM_MM_INPUT_CACHE_GIB: int = 4
VLLM_TARGET_DEVICE: str = "cuda"
VLLM_MAIN_CUDA_VERSION: str = "12.9"
VLLM_MAIN_CUDA_VERSION: str = "12.8"
MAX_JOBS: str | None = None
NVCC_THREADS: str | None = None
VLLM_USE_PRECOMPILED: bool = False
@ -445,9 +445,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
# Target device of vLLM, supporting [cuda (by default),
# rocm, cpu]
"VLLM_TARGET_DEVICE": lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),
# Main CUDA version of vLLM. This follows PyTorch but can be overridden.
# Main CUDA version of vLLM, supporting [12.6, 12.8, 12.9],
# 12.8 is the default. This follows PyTorch but can be overridden.
"VLLM_MAIN_CUDA_VERSION": lambda: os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower()
or "12.9",
or "12.8",
# Maximum number of compilation jobs to run in parallel.
# By default this is the number of CPUs
"MAX_JOBS": lambda: os.getenv("MAX_JOBS", None),