@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 1336a1ea24

Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 12, 06:00 UTC

Revert #29787 and #29690 (#29815)

Kevin H. Luu committed Dec 1, 2025 at 13:42 UTC 1336a1ea244fa8bfd7e72751cabbdb5b68a0c11a

9 files changed +185 -580

.buildkite/generate_index.py new

+46

@@ -0,0 +1,46 @@

1 +# SPDX-License-Identifier: Apache-2.0

2 +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 +

4 +import argparse

5 +import os

6 +

7 +template = """<!DOCTYPE html>

8 +<html>

9 + <body>

10 + <h1>Links for vLLM</h1/>

11 + <a href="../{x86_wheel_html_escaped}">{x86_wheel}</a><br/>

12 + <a href="../{arm_wheel_html_escaped}">{arm_wheel}</a><br/>

13 + </body>

14 +</html>

15 +"""

16 +

17 +parser = argparse.ArgumentParser()

18 +parser.add_argument("--wheel", help="The wheel path.", required=True)

19 +args = parser.parse_args()

20 +

21 +filename = os.path.basename(args.wheel)

22 +

23 +with open("index.html", "w") as f:

24 + print(f"Generated index.html for {args.wheel}")

25 + # sync the abi tag with .buildkite/scripts/upload-wheels.sh

26 + if "x86_64" in filename:

27 + x86_wheel = filename

28 + arm_wheel = filename.replace("x86_64", "aarch64").replace(

29 + "manylinux1", "manylinux2014"

30 + )

31 + elif "aarch64" in filename:

32 + x86_wheel = filename.replace("aarch64", "x86_64").replace(

33 + "manylinux2014", "manylinux1"

34 + )

35 + arm_wheel = filename

36 + else:

37 + raise ValueError(f"Unsupported wheel: {filename}")

38 + # cloudfront requires escaping the '+' character

39 + f.write(

40 + template.format(

41 + x86_wheel=x86_wheel,

42 + x86_wheel_html_escaped=x86_wheel.replace("+", "%2B"),

43 + arm_wheel=arm_wheel,

44 + arm_wheel_html_escaped=arm_wheel.replace("+", "%2B"),

45 + )

46 + )

.buildkite/release-pipeline.yaml

+15 -1

@@ -8,7 +8,7 @@ steps:

8 commands:

9 # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:

10 # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7

11 - - "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 ."

11 + - "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 ."

12 - "mkdir artifacts"

13 - "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'"

14 - "bash .buildkite/scripts/upload-wheels.sh"

@@ -30,6 +30,19 @@ steps:

30 DOCKER_BUILDKIT: "1"

32 # x86 + CUDA builds

33 + - label: "Build wheel - CUDA 12.8"

34 + depends_on: ~

35 + id: build-wheel-cuda-12-8

36 + agents:

37 + queue: cpu_queue_postmerge

38 + commands:

39 + - "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 ."

40 + - "mkdir artifacts"

41 + - "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'"

42 + - "bash .buildkite/scripts/upload-wheels.sh"

43 + env:

44 + DOCKER_BUILDKIT: "1"

45 +

46 - label: "Build wheel - CUDA 12.9"

47 depends_on: ~

48 id: build-wheel-cuda-12-9

@@ -96,6 +109,7 @@ steps:

109 - label: "Annotate release workflow"

110 depends_on:

111 - create-multi-arch-manifest

112 + - build-wheel-cuda-12-8

113 id: annotate-release-workflow

114 agents:

115 queue: cpu_queue_postmerge

.buildkite/scripts/generate-nightly-index.py deleted

-369

@@ -1,369 +0,0 @@

1 -#!/usr/bin/env python3

2 -# SPDX-License-Identifier: Apache-2.0

3 -# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

4 -

5 -# do not complain about line length (for docstring)

6 -# ruff: noqa: E501

7 -

8 -import argparse

9 -import json

10 -import re

11 -import sys

12 -from dataclasses import asdict, dataclass

13 -from pathlib import Path

14 -from typing import Any

15 -from urllib.parse import quote

16 -

17 -if not sys.version_info >= (3, 10):

18 - raise RuntimeError("This script requires Python 3.10 or higher.")

19 -

20 -INDEX_HTML_TEMPLATE = """<!DOCTYPE html>

21 -<html>

22 - <meta name="pypi:repository-version" content="1.0">

23 - <body>

24 -{items}

25 - </body>

26 -</html>

27 -"""

28 -

29 -

30 -@dataclass

31 -class WheelFileInfo:

32 - package_name: str

33 - version: str

34 - build_tag: str | None

35 - python_tag: str

36 - abi_tag: str

37 - platform_tag: str

38 - variant: str | None

39 - filename: str

40 -

41 -

42 -def parse_from_filename(file: str) -> WheelFileInfo:

43 - """

44 - Parse wheel file name to extract metadata.

45 -

46 - The format of wheel names:

47 - {package_name}-{version}(-{build_tag})?-{python_tag}-{abi_tag}-{platform_tag}.whl

48 - All versions could contain a variant like '+cu129' or '.cpu' or `.rocm` (or not).

49 - Example:

50 - vllm-0.11.0-cp38-abi3-manylinux1_x86_64.whl

51 - vllm-0.10.2rc2+cu129-cp38-abi3-manylinux2014_aarch64.whl

52 - vllm-0.11.1rc8.dev14+gaa384b3c0-cp38-abi3-manylinux2014_aarch64.whl

53 - vllm-0.11.1rc8.dev14+gaa384b3c0.cu130-cp38-abi3-manylinux1_x86_64.whl

54 - """

55 - wheel_file_re = re.compile(

56 - r"^(?P<package_name>.+)-(?P<version>[^-]+?)(-(?P<build_tag>[^-]+))?-(?P<python_tag>[^-]+)-(?P<abi_tag>[^-]+)-(?P<platform_tag>[^-]+)\.whl$"

57 - )

58 - match = wheel_file_re.match(file)

59 - if not match:

60 - raise ValueError(f"Invalid wheel file name: {file}")

61 -

62 - package_name = match.group("package_name")

63 - version = match.group("version")

64 - build_tag = match.group("build_tag")

65 - python_tag = match.group("python_tag")

66 - abi_tag = match.group("abi_tag")

67 - platform_tag = match.group("platform_tag")

68 -

69 - # extract variant from version

70 - variant = None

71 - if "dev" in version:

72 - ver_after_dev = version.split("dev")[-1]

73 - if "." in ver_after_dev:

74 - variant = ver_after_dev.split(".")[-1]

75 - version = version.removesuffix("." + variant)

76 - else:

77 - if "+" in version:

78 - version, variant = version.split("+")

79 -

80 - return WheelFileInfo(

81 - package_name=package_name,

82 - version=version,

83 - build_tag=build_tag,

84 - python_tag=python_tag,

85 - abi_tag=abi_tag,

86 - platform_tag=platform_tag,

87 - variant=variant,

88 - filename=file,

89 - )

90 -

91 -

92 -def generate_project_list(subdir_names: list[str]) -> str:

93 - """

94 - Generate project list HTML content linking to each project & variant sub-directory.

95 - """

96 - href_tags = []

97 - for name in sorted(subdir_names):

98 - name = name.strip("/").strip(".")

99 - href_tags.append(f' <a href="{name}/">{name}/</a><br/>')

100 - return INDEX_HTML_TEMPLATE.format(items="\n".join(href_tags))

101 -

102 -

103 -def generate_package_index_and_metadata(

104 - wheel_files: list[WheelFileInfo], wheel_base_dir: Path, index_base_dir: Path

105 -) -> tuple[str, str]:

106 - """

107 - Generate package index HTML content for a specific package, linking to actual wheel files.

108 - """

109 - href_tags = []

110 - metadata = []

111 - for file in sorted(wheel_files, key=lambda x: x.filename):

112 - relative_path = (

113 - wheel_base_dir.relative_to(index_base_dir, walk_up=True) / file.filename

114 - )

115 - # handle with '+' in URL, and avoid double-encoding '/' and already-encoded '%2B'

116 - # NOTE: this is AWS S3 specific behavior!

117 - file_path_quoted = quote(relative_path.as_posix(), safe=":%/")

118 - href_tags.append(f' <a href="{file_path_quoted}">{file.filename}</a><br/>')

119 - file_meta = asdict(file)

120 - file_meta["path"] = file_path_quoted

121 - metadata.append(file_meta)

122 - index_str = INDEX_HTML_TEMPLATE.format(items="\n".join(href_tags))

123 - metadata_str = json.dumps(metadata, indent=2)

124 - return index_str, metadata_str

125 -

126 -

127 -def generate_index_and_metadata(

128 - whl_files: list[str],

129 - wheel_base_dir: Path,

130 - index_base_dir: Path,

131 - default_variant: str | None = None,

132 - alias_to_default: str | None = None,

133 -):

134 - """

135 - Generate index for all wheel files.

136 -

137 - Args:

138 - whl_files (list[str]): List of wheel files (must be directly under `wheel_base_dir`).

139 - wheel_base_dir (Path): Base directory for wheel files.

140 - index_base_dir (Path): Base directory to store index files.

141 - default_variant (str | None): The default variant name, if any.

142 - alias_to_default (str | None): Alias variant name for the default variant, if any.

143 -

144 - First, parse all wheel files to extract metadata.

145 - We need to collect all wheel files for each variant, and generate an index for it (in a sub-directory).

146 - The index for the default variant (if any) is generated in the root index directory.

147 -

148 - If `default_variant` is provided, all wheels must have variant suffixes, and the default variant index

149 - is purely a copy of the corresponding variant index, with only the links adjusted.

150 - Otherwise, all wheels without variant suffixes are treated as the default variant.

151 -

152 - If `alias_to_default` is provided, an additional alias sub-directory is created, it has the same content

153 - as the default variant index, but the links are adjusted accordingly.

154 -

155 - Index directory structure:

156 - index_base_dir/ (hosted at wheels.vllm.ai/{nightly,$commit,$version}/)

157 - index.html # project list, linking to "vllm/" and other packages, and all variant sub-directories

158 - vllm/

159 - index.html # package index, pointing to actual files in wheel_base_dir (relative path)

160 - metadata.json # machine-readable metadata for all wheels in this package

161 - cpu/ # cpu variant sub-directory

162 - index.html

163 - vllm/

164 - index.html

165 - metadata.json

166 - cu129/ # cu129 is actually the alias to default variant

167 - index.html

168 - vllm/

169 - index.html

170 - metadata.json

171 - cu130/ # cu130 variant sub-directory

172 - index.html

173 - vllm/

174 - index.html

175 - metadata.json

176 - ...

177 -

178 - metadata.json stores a dump of all wheel files' metadata in a machine-readable format:

179 - [

180 - {

181 - "package_name": "vllm",

182 - "version": "0.10.2rc2",

183 - "build_tag": null,

184 - "python_tag": "cp38",

185 - "abi_tag": "abi3",

186 - "platform_tag": "manylinux2014_aarch64",

187 - "variant": "cu129",

188 - "filename": "vllm-0.10.2rc2+cu129-cp38-abi3-manylinux2014_aarch64.whl",

189 - "path": "../vllm-0.10.2rc2%2Bcu129-cp38-abi3-manylinux2014_aarch64.whl" # to be concatenated with the directory URL and URL-encoded

190 - },

191 - ...

192 - ]

193 - """

194 -

195 - parsed_files = [parse_from_filename(f) for f in whl_files]

196 -

197 - if not parsed_files:

198 - print("No wheel files found, skipping index generation.")

199 - return

200 -

201 - # Group by variant

202 - variant_to_files: dict[str, list[WheelFileInfo]] = {}

203 - for file in parsed_files:

204 - variant = file.variant or "default"

205 - if variant not in variant_to_files:

206 - variant_to_files[variant] = []

207 - variant_to_files[variant].append(file)

208 -

209 - print(f"Found variants: {list(variant_to_files.keys())}")

210 -

211 - # sanity check for default variant

212 - if default_variant:

213 - if "default" in variant_to_files:

214 - raise ValueError(

215 - "All wheel files must have variant suffixes when `default_variant` is specified."

216 - )

217 - if default_variant not in variant_to_files:

218 - raise ValueError(

219 - f"Default variant '{default_variant}' not found among wheel files."

220 - )

221 -

222 - if alias_to_default:

223 - if "default" not in variant_to_files:

224 - # e.g. only some wheels are uploaded to S3 currently

225 - print(

226 - "[WARN] Alias to default variant specified, but no default variant found."

227 - )

228 - elif alias_to_default in variant_to_files:

229 - raise ValueError(

230 - f"Alias variant name '{alias_to_default}' already exists among wheel files."

231 - )

232 - else:

233 - variant_to_files[alias_to_default] = variant_to_files["default"].copy()

234 - print(f"Alias variant '{alias_to_default}' created for default variant.")

235 -

236 - # Generate index for each variant

237 - subdir_names = set()

238 - for variant, files in variant_to_files.items():

239 - if variant == "default":

240 - variant_dir = index_base_dir

241 - else:

242 - variant_dir = index_base_dir / variant

243 - subdir_names.add(variant)

244 -

245 - variant_dir.mkdir(parents=True, exist_ok=True)

246 -

247 - # gather all package names in this variant

248 - packages = set(f.package_name for f in files)

249 - if variant == "default":

250 - # these packages should also appear in the "project list"

251 - # generate after all variants are processed

252 - subdir_names = subdir_names.union(packages)

253 - else:

254 - # generate project list for this variant directly

255 - project_list_str = generate_project_list(sorted(packages))

256 - with open(variant_dir / "index.html", "w") as f:

257 - f.write(project_list_str)

258 -

259 - for package in packages:

260 - # filter files belonging to this package only

261 - package_files = [f for f in files if f.package_name == package]

262 - package_dir = variant_dir / package

263 - package_dir.mkdir(parents=True, exist_ok=True)

264 - index_str, metadata_str = generate_package_index_and_metadata(

265 - package_files, wheel_base_dir, package_dir

266 - )

267 - with open(package_dir / "index.html", "w") as f:

268 - f.write(index_str)

269 - with open(package_dir / "metadata.json", "w") as f:

270 - f.write(metadata_str)

271 -

272 - # Generate top-level project list index

273 - project_list_str = generate_project_list(sorted(subdir_names))

274 - with open(index_base_dir / "index.html", "w") as f:

275 - f.write(project_list_str)

276 -

277 -

278 -if __name__ == "__main__":

279 - """

280 - Arguments:

281 - --version <version> : version string for the current build (e.g., commit hash)

282 - --current-objects <path_to_json> : path to JSON file containing current S3 objects listing in this version directory

283 - --output-dir <output_directory> : directory to store generated index files

284 - --alias-to-default <alias_variant_name> : (optional) alias variant name for the default variant

285 - """

286 -

287 - parser = argparse.ArgumentParser(

288 - description="Process nightly build wheel files to generate indices."

289 - )

290 - parser.add_argument(

291 - "--version",

292 - type=str,

293 - required=True,

294 - help="Version string for the current build (e.g., commit hash)",

295 - )

296 - parser.add_argument(

297 - "--current-objects",

298 - type=str,

299 - required=True,

300 - help="Path to JSON file containing current S3 objects listing in this version directory",

301 - )

302 - parser.add_argument(

303 - "--output-dir",

304 - type=str,

305 - required=True,

306 - help="Directory to store generated index files",

307 - )

308 - parser.add_argument(

309 - "--alias-to-default",

310 - type=str,

311 - default=None,

312 - help="Alias variant name for the default variant",

313 - )

314 -

315 - args = parser.parse_args()

316 -

317 - version = args.version

318 - if "/" in version or "\\" in version:

319 - raise ValueError("Version string must not contain slashes.")

320 - current_objects_path = Path(args.current_objects)

321 - output_dir = Path(args.output_dir)

322 - if not output_dir.exists():

323 - output_dir.mkdir(parents=True, exist_ok=True)

324 -

325 - # Read current objects JSON

326 - with open(current_objects_path) as f:

327 - current_objects: dict[str, list[dict[str, Any]]] = json.load(f)

328 -

329 - # current_objects looks like from list_objects_v2 S3 API:

330 - """

331 - "Contents": [

332 - {

333 - "Key": "e2f56c309d2a28899c68975a7e104502d56deb8f/vllm-0.11.2.dev363+ge2f56c309-cp38-abi3-manylinux1_x86_64.whl",

334 - "LastModified": "2025-11-28T14:00:32+00:00",

335 - "ETag": "\"37a38339c7cdb61ca737021b968075df-52\"",

336 - "ChecksumAlgorithm": [

337 - "CRC64NVME"

338 - ],

339 - "ChecksumType": "FULL_OBJECT",

340 - "Size": 435649349,

341 - "StorageClass": "STANDARD"

342 - },

343 - ...

344 - ]

345 - """

346 -

347 - # Extract wheel file keys

348 - wheel_files = []

349 - for item in current_objects.get("Contents", []):

350 - key: str = item["Key"]

351 - if key.endswith(".whl"):

352 - wheel_files.append(key.split("/")[-1]) # only the filename is used

353 -

354 - print(f"Found {len(wheel_files)} wheel files for version {version}: {wheel_files}")

355 -

356 - # Generate index and metadata, assuming wheels and indices are stored as:

357 - # s3://vllm-wheels/{version}/<wheel files>

358 - # s3://vllm-wheels/<anything>/<index files>

359 - wheel_base_dir = Path(output_dir).parent / version

360 - index_base_dir = Path(output_dir)

361 -

362 - generate_index_and_metadata(

363 - whl_files=wheel_files,

364 - wheel_base_dir=wheel_base_dir,

365 - index_base_dir=index_base_dir,

366 - default_variant=None,

367 - alias_to_default=args.alias_to_default,

368 - )

369 - print(f"Successfully generated index and metadata in {output_dir}")

.buildkite/scripts/upload-wheels.sh

+52 -69

@@ -2,28 +2,6 @@

3 set -ex

5 -# ======== part 0: setup ========

6 -

7 -BUCKET="vllm-wheels"

8 -INDICES_OUTPUT_DIR="indices"

9 -DEFAULT_VARIANT_ALIAS="cu129" # align with vLLM_MAIN_CUDA_VERSION in vllm/envs.py

10 -PYTHON=${PYTHON_PROG:=python3} # try to read from env var, otherwise use python3

11 -SUBPATH=$BUILDKITE_COMMIT

12 -S3_COMMIT_PREFIX="s3://$BUCKET/$SUBPATH/"

13 -

14 -# detect if python3.10+ is available

15 -has_new_python=$($PYTHON -c "print(1 if __import__('sys').version_info >= (3,10) else 0)")

16 -if [[ "$has_new_python" -eq 0 ]]; then

17 - # use new python from docker

18 - docker pull python:3-slim

19 - PYTHON="docker run --rm -v $(pwd):/app -w /app python:3-slim python3"

20 -fi

21 -

22 -echo "Using python interpreter: $PYTHON"

23 -echo "Python version: $($PYTHON --version)"

24 -

25 -# ========= part 1: collect, rename & upload the wheel ==========

26 -

5 # Assume wheels are in artifacts/dist/*.whl

6 wheel_files=(artifacts/dist/*.whl)

@@ -32,69 +10,74 @@ if [[ ${#wheel_files[@]} -ne 1 ]]; then

10 echo "Error: Expected exactly one wheel file in artifacts/dist/, but found ${#wheel_files[@]}"

11 exit 1

12 fi

13 +

14 +# Get the single wheel file

15 wheel="${wheel_files[0]}"

37 -# current build image uses ubuntu 20.04, which corresponds to manylinux_2_31

38 -# refer to https://github.com/mayeut/pep600_compliance?tab=readme-ov-file#acceptable-distros-to-build-wheels

39 -manylinux_version="manylinux_2_31"

17 +# Detect architecture and rename 'linux' to appropriate manylinux version

18 +arch=$(uname -m)

19 +if [[ $arch == "x86_64" ]]; then

20 + manylinux_version="manylinux1"

21 +elif [[ $arch == "aarch64" ]]; then

22 + manylinux_version="manylinux2014"

23 +else

24 + echo "Warning: Unknown architecture $arch, using manylinux1 as default"

25 + manylinux_version="manylinux1"

26 +fi

28 # Rename 'linux' to the appropriate manylinux version in the wheel filename

42 -if [[ "$wheel" != *"linux"* ]]; then

43 - echo "Error: Wheel filename does not contain 'linux': $wheel"

44 - exit 1

45 -fi

29 new_wheel="${wheel/linux/$manylinux_version}"

30 mv -- "$wheel" "$new_wheel"

31 wheel="$new_wheel"

49 -echo "Renamed wheel to: $wheel"

33 # Extract the version from the wheel

34 version=$(unzip -p "$wheel" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)

53 -echo "Version in wheel: $version"

54 -pure_version="${version%%+*}"

55 -echo "Pure version (without variant): $pure_version"

56 -

57 -# copy wheel to its own bucket

58 -aws s3 cp "$wheel" "$S3_COMMIT_PREFIX"

35 +echo "Version: $version"

36 +

37 +normal_wheel="$wheel" # Save the original wheel filename

38 +

39 +# If the version contains "dev", rename it to v1.0.0.dev for consistency

40 +if [[ $version == *dev* ]]; then

41 + suffix="${version##*.}"

42 + if [[ $suffix == cu* ]]; then

43 + new_version="1.0.0.dev+${suffix}"

44 + else

45 + new_version="1.0.0.dev"

46 + fi

47 + new_wheel="${wheel/$version/$new_version}"

48 + # use cp to keep both files in the artifacts directory

49 + cp -- "$wheel" "$new_wheel"

50 + wheel="$new_wheel"

51 + version="$new_version"

52 +fi

60 -# ========= part 2: generate and upload indices ==========

61 -# generate indices for all existing wheels in the commit directory

62 -# this script might be run multiple times if there are multiple variants being built

63 -# so we need to guarantee there is little chance for "TOCTOU" issues

64 -# i.e., one process is generating indices while another is uploading a new wheel

65 -# so we need to ensure no time-consuming operations happen below

54 +# Upload the wheel to S3

55 +python3 .buildkite/generate_index.py --wheel "$normal_wheel"

67 -# list all wheels in the commit directory

68 -echo "Existing wheels on S3:"

69 -aws s3 ls "$S3_COMMIT_PREFIX"

70 -obj_json="objects.json"

71 -aws s3api list-objects-v2 --bucket "$BUCKET" --prefix "$SUBPATH/" --delimiter / --output json > "$obj_json"

72 -mkdir -p "$INDICES_OUTPUT_DIR"

57 +# generate index for this commit

58 +aws s3 cp "$wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"

59 +aws s3 cp "$normal_wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"

74 -# call script to generate indicies for all existing wheels

75 -# this indices have relative paths that could work as long as it is next to the wheel directory in s3

76 -# i.e., the wheels are always in s3://vllm-wheels/<commit>/

77 -# and indices can be placed in /<commit>/, or /nightly/, or /<version>/

78 -if [[ ! -z "$DEFAULT_VARIANT_ALIAS" ]]; then

79 - alias_arg="--alias-to-default $DEFAULT_VARIANT_ALIAS"

61 +if [[ $normal_wheel == *"cu129"* ]]; then

62 + # only upload index.html for cu129 wheels (default wheels) as it

63 + # is available on both x86 and arm64

64 + aws s3 cp index.html "s3://vllm-wheels/$BUILDKITE_COMMIT/vllm/index.html"

65 + aws s3 cp "s3://vllm-wheels/nightly/index.html" "s3://vllm-wheels/$BUILDKITE_COMMIT/index.html"

66 else

81 - alias_arg=""

67 + echo "Skipping index files for non-cu129 wheels"

68 fi

84 -$PYTHON .buildkite/scripts/generate-nightly-index.py --version "$SUBPATH" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" $alias_arg

85 -

86 -# copy indices to /<commit>/ unconditionally

87 -echo "Uploading indices to $S3_COMMIT_PREFIX"

88 -aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "$S3_COMMIT_PREFIX"

70 +# generate index for nightly

71 +aws s3 cp "$wheel" "s3://vllm-wheels/nightly/"

72 +aws s3 cp "$normal_wheel" "s3://vllm-wheels/nightly/"

90 -# copy to /nightly/ only if it is on the main branch and not a PR

91 -if [[ "$BUILDKITE_BRANCH" == "main" && "$BUILDKITE_PULL_REQUEST" == "false" ]]; then

92 - echo "Uploading indices to overwrite /nightly/"

93 - aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/nightly/"

74 +if [[ $normal_wheel == *"cu129"* ]]; then

75 + # only upload index.html for cu129 wheels (default wheels) as it

76 + # is available on both x86 and arm64

77 + aws s3 cp index.html "s3://vllm-wheels/nightly/vllm/index.html"

78 +else

79 + echo "Skipping index files for non-cu129 wheels"

80 fi

96 -# copy to /<pure_version>/ only if it does not have "dev" in the version

97 -if [[ "$version" != *"dev"* ]]; then

98 - echo "Uploading indices to overwrite /$pure_version/"

99 - aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/$pure_version/"

100 -fi

82 +aws s3 cp "$wheel" "s3://vllm-wheels/$version/"

83 +aws s3 cp index.html "s3://vllm-wheels/$version/vllm/index.html"

docs/getting_started/installation/cpu.md

+1 -14

@@ -46,23 +46,10 @@ vLLM is a Python library that supports the following CPU variants. Select your C

47 ### Pre-built wheels

49 -Please refer to the instructions for [pre-built wheels on GPU](./gpu.md#pre-built-wheels).

50 -

51 -When specifying the index URL, please make sure to use the `cpu` variant subdirectory.

52 -For example, the nightly build index is: `https://wheels.vllm.ai/nightly/cpu/`.

49 +Currently, there are no pre-built CPU wheels.

51 ### Build wheel from source

56 -#### Set up using Python-only build (without compilation) {#python-only-build}

57 -

58 -Please refer to the instructions for [Python-only build on GPU](./gpu.md#python-only-build), and replace the build commands with:

59 -

60 -```bash

61 -VLLM_USE_PRECOMPILED=1 VLLM_PRECOMPILED_WHEEL_VARIANT=cpu VLLM_TARGET_DEVICE=cpu uv pip install --editable .

62 -```

63 -

64 -#### Full build (with compilation) {#full-build}

65 -

53 === "Intel/AMD x86"

55 --8<-- "docs/getting_started/installation/cpu.x86.inc.md:build-wheel-from-source"

docs/getting_started/installation/gpu.cuda.inc.md

+40 -33

@@ -26,50 +26,43 @@ uv pip install vllm --torch-backend=auto

27 ??? console "pip"

28 ```bash

29 - # Install vLLM with CUDA 12.9.

30 - pip install vllm --extra-index-url https://download.pytorch.org/whl/cu129

29 + # Install vLLM with CUDA 12.8.

30 + pip install vllm --extra-index-url https://download.pytorch.org/whl/cu128

31 ```

33 -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.

33 +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.

35 !!! note

36 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.

38 -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:

38 +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:

40 ```bash

41 -# Install vLLM with a specific CUDA version (e.g., 13.0).

41 +# Install vLLM with a specific CUDA version (e.g., 11.8 or 12.6).

42 export VLLM_VERSION=$(curl -s https://api.github.com/repos/vllm-project/vllm/releases/latest | jq -r .tag_name | sed 's/^v//')

43 -export CUDA_VERSION=130 # or other

44 -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}

43 +export CUDA_VERSION=118 # or 126

44 +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}

45 ```

47 #### Install the latest code

49 -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:

50 -

51 -* `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.

52 -* `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.

53 -

54 -To install from nightly index, run:

49 +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`.

51 ```bash

52 uv pip install -U vllm \

53 --torch-backend=auto \

59 - --extra-index-url https://wheels.vllm.ai/nightly # add variant subdirectory here if needed

54 + --extra-index-url https://wheels.vllm.ai/nightly

55 ```

62 -!!! warning "`pip` caveat"

63 -

64 - 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).

65 -

66 - 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).

67 -

57 +??? console "pip"

58 ```bash

69 - 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!)

70 - 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

59 + pip install -U vllm \

60 + --pre \

61 + --extra-index-url https://wheels.vllm.ai/nightly

62 ```

64 + `--pre` is required for `pip` to consider pre-released versions.

65 +

66 ##### Install specific revisions

68 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

71 export VLLM_COMMIT=72d9c316d3f6ede485146fe5aabd4e61dbc59069 # use full commit hash from the main branch

72 uv pip install vllm \

73 --torch-backend=auto \

81 - --extra-index-url https://wheels.vllm.ai/${VLLM_COMMIT} # add variant subdirectory here if needed

74 + --extra-index-url https://wheels.vllm.ai/${VLLM_COMMIT}

75 ```

77 +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.

78 +

79 +??? note "pip"

80 + If you want to access the wheels for previous commits (e.g. to bisect the behavior change,

81 + performance regression), due to the limitation of `pip`, you have to specify the full URL of the

82 + wheel file by embedding the commit hash in the URL:

83 +

84 + ```bash

85 + export VLLM_COMMIT=33f460b17a54acb3b6cc0b03f4a17876cff5eafd # use full commit hash from the main branch

86 + pip install https://wheels.vllm.ai/${VLLM_COMMIT}/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl

87 + ```

88 +

89 + Note that the wheels are built with Python 3.8 ABI (see [PEP

90 + 425](https://peps.python.org/pep-0425/) for more details about ABI), so **they are compatible

91 + with Python 3.8 and later**. The version string in the wheel file name (`1.0.0.dev`) is just a

92 + placeholder to have a unified URL for the wheels, the actual versions of wheels are contained in

93 + the wheel metadata (the wheels listed in the extra index url have correct versions). Although we

94 + don't support Python 3.8 any more (because PyTorch 2.5 dropped support for Python 3.8), the

95 + wheels are still built with Python 3.8 ABI to keep the same wheel name as before.

96 +

97 # --8<-- [end:pre-built-wheels]

98 # --8<-- [start:build-wheel-from-source]

87 -#### Set up using Python-only build (without compilation) {#python-only-build}

100 +#### Set up using Python-only build (without compilation)

101

102 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:

103

@@ -108,24 +121,18 @@ This command will do the following:

121 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.

122

123 ```bash

111 -export VLLM_PRECOMPILED_WHEEL_COMIMT=$(git rev-parse HEAD~1) # or earlier commit on main

112 -export VLLM_USE_PRECOMPILED=1

124 +export VLLM_COMMIT=72d9c316d3f6ede485146fe5aabd4e61dbc59069 # use full commit hash from the main branch

125 +export VLLM_PRECOMPILED_WHEEL_LOCATION=https://wheels.vllm.ai/${VLLM_COMMIT}/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl

126 uv pip install --editable .

127 ```

128

116 -There are more environment variables to control the behavior of Python-only build:

117 -

118 -* `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.

119 -* `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.

120 -* `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.

121 -

129 You can find more information about vLLM's wheels in [Install the latest code](#install-the-latest-code).

130

131 !!! note

132 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.

133 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.

134

128 -#### Full build (with compilation) {#full-build}

135 +#### Full build (with compilation)

136

137 If you want to modify C++ or CUDA code, you'll need to build vLLM from source. This can take several minutes:

138

docs/getting_started/installation/gpu.md

+1 -1

@@ -52,7 +52,7 @@ vLLM is a Python library that supports the following GPU variants. Select your G

53 --8<-- "docs/getting_started/installation/gpu.xpu.inc.md:set-up-using-python"

55 -### Pre-built wheels {#pre-built-wheels}

55 +### Pre-built wheels

57 === "NVIDIA CUDA"

setup.py

+26 -90

@@ -310,6 +310,9 @@ class cmake_build_ext(build_ext):

310 class precompiled_build_ext(build_ext):

311 """Disables extension building when using precompiled binaries."""

312

313 + def run(self) -> None:

314 + assert _is_cuda(), "VLLM_USE_PRECOMPILED is only supported for CUDA builds"

315 +

316 def build_extensions(self) -> None:

317 print("Skipping build_ext: using precompiled extensions.")

318 return

@@ -319,17 +322,14 @@ class precompiled_wheel_utils:

322 """Extracts libraries and other files from an existing wheel."""

323

324 @staticmethod

322 - def extract_precompiled_and_patch_package(

323 - wheel_url_or_path: str, download_filename: str | None

324 - ) -> dict:

325 + def extract_precompiled_and_patch_package(wheel_url_or_path: str) -> dict:

326 import tempfile

327 import zipfile

328

329 temp_dir = None

330 try:

331 if not os.path.isfile(wheel_url_or_path):

331 - # use provided filename first, then derive from URL

332 - wheel_filename = download_filename or wheel_url_or_path.split("/")[-1]

332 + wheel_filename = wheel_url_or_path.split("/")[-1]

333 temp_dir = tempfile.mkdtemp(prefix="vllm-wheels")

334 wheel_path = os.path.join(temp_dir, wheel_filename)

335 print(f"Downloading wheel from {wheel_url_or_path} to {wheel_path}")

@@ -648,102 +648,38 @@ package_data = {

648 ]

649 }

650

651 -

652 -def _fetch_metadata_for_variant(

653 - commit: str, variant: str | None

654 -) -> tuple[list[dict], str]:

655 - variant_dir = f"{variant}/" if variant is not None else ""

656 - repo_url = f"https://wheels.vllm.ai/{commit}/{variant_dir}vllm/"

657 - meta_url = repo_url + "metadata.json"

658 - logger.info("Trying to fetch metadata from {}", meta_url)

659 - from urllib.request import urlopen

660 -

661 - with urlopen(meta_url) as resp:

662 - # urlopen raises HTTPError on unexpected status code

663 - wheels = json.loads(resp.read().decode("utf-8"))

664 - return wheels, repo_url

665 -

666 -

651 # If using precompiled, extract and patch package_data (in advance of setup)

652 if envs.VLLM_USE_PRECOMPILED:

669 - # Attempts:

670 - # 1. user-specified wheel location (can be either local or remote, via

671 - # VLLM_PRECOMPILED_WHEEL_LOCATION)

672 - # 2. user-specified variant from nightly repo (current main commit via

673 - # VLLM_PRECOMPILED_WHEEL_VARIANT)

674 - # 3. the variant corresponding to VLLM_MAIN_CUDA_VERSION from nightly repo

675 - # 4. the default variant from nightly repo (current main commit)

653 + assert _is_cuda(), "VLLM_USE_PRECOMPILED is only supported for CUDA builds"

654 wheel_location = os.getenv("VLLM_PRECOMPILED_WHEEL_LOCATION", None)

655 if wheel_location is not None:

656 wheel_url = wheel_location

679 - download_filename = None

680 - logger.info("Using user-specified precompiled wheel location: %s", wheel_url)

657 else:

658 import platform

659

660 arch = platform.machine()

685 - # try to fetch the wheel metadata from the nightly wheel repo

686 - main_variant = envs.VLLM_MAIN_CUDA_VERSION.replace(".", "")

687 - variant = os.getenv("VLLM_PRECOMPILED_WHEEL_VARIANT", main_variant)

688 - commit = os.getenv(

689 - "VLLM_PRECOMPILED_WHEEL_COMMIT",

690 - precompiled_wheel_utils.get_base_commit_in_main_branch(),

691 - )

692 - logger.info(

693 - "Using precompiled wheel commit %s with variant %s", commit, variant

661 + if arch == "x86_64":

662 + wheel_tag = "manylinux1_x86_64"

663 + elif arch == "aarch64":

664 + wheel_tag = "manylinux2014_aarch64"

665 + else:

666 + raise ValueError(f"Unsupported architecture: {arch}")

667 + base_commit = precompiled_wheel_utils.get_base_commit_in_main_branch()

668 + wheel_url = f"https://wheels.vllm.ai/{base_commit}/vllm-1.0.0.dev-cp38-abi3-{wheel_tag}.whl"

669 + nightly_wheel_url = (

670 + f"https://wheels.vllm.ai/nightly/vllm-1.0.0.dev-cp38-abi3-{wheel_tag}.whl"

671 )

695 - try_default = False

696 - wheels, repo_url, download_filename = None, None, None

672 + from urllib.request import urlopen

673 +

674 try:

698 - wheels, repo_url = _fetch_metadata_for_variant(commit, variant)

699 - except Exception:

700 - logger.warning(

701 - "Failed to fetch precompiled wheel metadata for variant %s",

702 - variant,

703 - exc_info=True,

704 - )

705 - try_default = True # try outside handler to keep the stacktrace simple

706 - if try_default:

707 - logger.info("Trying the default variant")

708 - wheels, repo_url = _fetch_metadata_for_variant(commit, None)

709 - # if this also fails, then we have nothing more to try / cache

710 - assert wheels is not None and repo_url is not None, (

711 - "Failed to fetch precompiled wheel metadata"

712 - )

713 - # The metadata.json has the following format:

714 - # see .buildkite/scripts/generate-nightly-index.py for details

715 - """[{

716 -"package_name": "vllm",

717 -"version": "0.11.2.dev278+gdbc3d9991",

718 -"build_tag": null,

719 -"python_tag": "cp38",

720 -"abi_tag": "abi3",

721 -"platform_tag": "manylinux1_x86_64",

722 -"variant": null,

723 -"filename": "vllm-0.11.2.dev278+gdbc3d9991-cp38-abi3-manylinux1_x86_64.whl",

724 -"path": "../vllm-0.11.2.dev278%2Bgdbc3d9991-cp38-abi3-manylinux1_x86_64.whl"

725 -},

726 -...]"""

727 - for wheel in wheels:

728 - # TODO: maybe check more compatibility later? (python_tag, abi_tag, etc)

729 - if wheel.get("package_name") == "vllm" and arch in wheel.get(

730 - "platform_tag", ""

731 - ):

732 - logger.info("Found precompiled wheel metadata: %s", wheel)

733 - if "path" not in wheel:

734 - raise ValueError(f"Wheel metadata missing path: {wheel}")

735 - wheel_url = repo_url + wheel["path"]

736 - download_filename = wheel.get("filename")

737 - logger.info("Using precompiled wheel URL: %s", wheel_url)

738 - break

739 - else:

740 - raise ValueError(

741 - f"No precompiled vllm wheel found for architecture {arch} "

742 - f"from repo {repo_url}. All available wheels: {wheels}"

743 - )

744 - patch = precompiled_wheel_utils.extract_precompiled_and_patch_package(

745 - wheel_url, download_filename

746 - )

675 + with urlopen(wheel_url) as resp:

676 + if resp.status != 200:

677 + wheel_url = nightly_wheel_url

678 + except Exception as e:

679 + print(f"[warn] Falling back to nightly wheel: {e}")

680 + wheel_url = nightly_wheel_url

681 +

682 + patch = precompiled_wheel_utils.extract_precompiled_and_patch_package(wheel_url)

683 for pkg, files in patch.items():

684 package_data.setdefault(pkg, []).extend(files)

685

vllm/envs.py

+4 -3

@@ -74,7 +74,7 @@ if TYPE_CHECKING:

74 VLLM_MEDIA_CONNECTOR: str = "http"

75 VLLM_MM_INPUT_CACHE_GIB: int = 4

76 VLLM_TARGET_DEVICE: str = "cuda"

77 - VLLM_MAIN_CUDA_VERSION: str = "12.9"

77 + VLLM_MAIN_CUDA_VERSION: str = "12.8"

78 MAX_JOBS: str | None = None

79 NVCC_THREADS: str | None = None

80 VLLM_USE_PRECOMPILED: bool = False

@@ -445,9 +445,10 @@ environment_variables: dict[str, Callable[[], Any]] = {

445 # Target device of vLLM, supporting [cuda (by default),

446 # rocm, cpu]

447 "VLLM_TARGET_DEVICE": lambda: os.getenv("VLLM_TARGET_DEVICE", "cuda").lower(),

448 - # Main CUDA version of vLLM. This follows PyTorch but can be overridden.

448 + # Main CUDA version of vLLM, supporting [12.6, 12.8, 12.9],

449 + # 12.8 is the default. This follows PyTorch but can be overridden.

450 "VLLM_MAIN_CUDA_VERSION": lambda: os.getenv("VLLM_MAIN_CUDA_VERSION", "").lower()

450 - or "12.9",

451 + or "12.8",

452 # Maximum number of compilation jobs to run in parallel.

453 # By default this is the number of CPUs

454 "MAX_JOBS": lambda: os.getenv("MAX_JOBS", None),