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 11, 15:00 UTC
[CI] refine more logic when generating and using nightly wheels & indices, add cuda130 build for aarch64, specify correct manylinux version (#30341)
Signed-off-by: Shengqi Chen <harry-chen@outlook.com>
Shengqi Chen committed
Dec 12, 2025 at 00:42 UTC
305b168a9fc50f322e9c5a07f4fc8c7bbda5f844
4 files changed
+73
-10
.buildkite/release-pipeline.yaml
+18
-3
@@ -15,6 +15,21 @@ steps:
15
env:
16
DOCKER_BUILDKIT: "1"
17
18
+ - label: "Build arm64 wheel - CUDA 13.0"
19
+ depends_on: ~
20
+ id: build-wheel-arm64-cuda-13-0
21
+ agents:
22
+ queue: arm64_cpu_queue_postmerge
23
+ commands:
24
+ # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
25
+ # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
26
+ - "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=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
27
+ - "mkdir artifacts"
28
+ - "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'"
29
+ - "bash .buildkite/scripts/upload-wheels.sh manylinux_2_35"
30
+ env:
31
+ DOCKER_BUILDKIT: "1"
32
+
33
# aarch64 build
34
- label: "Build arm64 CPU wheel"
35
depends_on: ~
@@ -25,7 +40,7 @@ steps:
40
- "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_BUILD_ACL=ON --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
41
- "mkdir artifacts"
42
- "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'"
28
- - "bash .buildkite/scripts/upload-wheels.sh"
43
+ - "bash .buildkite/scripts/upload-wheels.sh manylinux_2_35"
44
env:
45
DOCKER_BUILDKIT: "1"
46
@@ -39,7 +54,7 @@ steps:
54
- "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 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
55
- "mkdir artifacts"
56
- "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"
57
+ - "bash .buildkite/scripts/upload-wheels.sh manylinux_2_31"
58
env:
59
DOCKER_BUILDKIT: "1"
60
@@ -52,7 +67,7 @@ steps:
67
- "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=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
68
- "mkdir artifacts"
69
- "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
55
- - "bash .buildkite/scripts/upload-wheels.sh"
70
+ - "bash .buildkite/scripts/upload-wheels.sh manylinux_2_35"
71
env:
72
DOCKER_BUILDKIT: "1"
73
.buildkite/scripts/generate-nightly-index.py
+11
@@ -372,6 +372,17 @@ if __name__ == "__main__":
372
373
print(f"Found {len(wheel_files)} wheel files for version {version}: {wheel_files}")
374
375
+ # keep only "official" files for a non-nightly version (specifed by cli args)
376
+ PY_VERSION_RE = re.compile(r"^\d+\.\d+\.\d+([a-zA-Z0-9.+-]*)?$")
377
+ if PY_VERSION_RE.match(version):
378
+ # upload-wheels.sh ensures no "dev" is in args.version
379
+ wheel_files = list(
380
+ filter(lambda x: version in x and "dev" not in x, wheel_files)
381
+ )
382
+ print(f"Non-nightly version detected, wheel files used: {wheel_files}")
383
+ else:
384
+ print("Nightly version detected, keeping all wheel files.")
385
+
386
# Generate index and metadata, assuming wheels and indices are stored as:
387
# s3://vllm-wheels/{version}/<wheel files>
388
# s3://vllm-wheels/<anything>/<index files>
.buildkite/scripts/upload-wheels.sh
+8
-4
@@ -34,9 +34,10 @@ if [[ ${#wheel_files[@]} -ne 1 ]]; then
34
fi
35
wheel="${wheel_files[0]}"
36
37
-# current build image uses ubuntu 20.04, which corresponds to manylinux_2_31
37
+# default build image uses ubuntu 20.04, which corresponds to manylinux_2_31
38
+# we also accept params as manylinux tag
39
# refer to https://github.com/mayeut/pep600_compliance?tab=readme-ov-file#acceptable-distros-to-build-wheels
39
-manylinux_version="manylinux_2_31"
40
+manylinux_version="${1:-manylinux_2_31}"
41
42
# Rename 'linux' to the appropriate manylinux version in the wheel filename
43
if [[ "$wheel" != *"linux"* ]]; then
@@ -96,8 +97,11 @@ if [[ "$BUILDKITE_BRANCH" == "main" && "$BUILDKITE_PULL_REQUEST" == "false" ]];
97
aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/nightly/"
98
fi
99
99
-# copy to /<pure_version>/ only if it does not have "dev" in the version
100
+# re-generate and copy to /<pure_version>/ only if it does not have "dev" in the version
101
if [[ "$version" != *"dev"* ]]; then
101
- echo "Uploading indices to overwrite /$pure_version/"
102
+ echo "Re-generating indices for /$pure_version/"
103
+ rm -rf "$INDICES_OUTPUT_DIR/*"
104
+ mkdir -p "$INDICES_OUTPUT_DIR"
105
+ $PYTHON .buildkite/scripts/generate-nightly-index.py --version "$pure_version" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" --comment "version $pure_version" $alias_arg
106
aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/$pure_version/"
107
fi
tests/standalone_tests/python_only_compile.sh
+36
-3
@@ -3,12 +3,45 @@
3
# for users who do not have any compilers installed on their system
4
5
set -e
6
-set -x
6
7
merge_base_commit=$(git merge-base HEAD origin/main)
9
-echo "Current merge base commit with main: $merge_base_commit"
8
+echo "INFO: current merge base commit with main: $merge_base_commit"
9
git show --oneline -s $merge_base_commit
10
11
+# test whether the metadata.json url is valid, retry each 3 minutes up to 5 times
12
+# this avoids cumbersome error messages & manual retries in case the precompiled wheel
13
+# for the given commit is still being built in the release pipeline
14
+meta_json_url="https://wheels.vllm.ai/$merge_base_commit/vllm/metadata.json"
15
+echo "INFO: will use metadata.json from $meta_json_url"
16
+
17
+for i in {1..5}; do
18
+ echo "Checking metadata.json URL (attempt $i)..."
19
+ if curl --fail "$meta_json_url" > metadata.json; then
20
+ echo "INFO: metadata.json URL is valid."
21
+ # check whether it is valid json by python
22
+ if python3 -m json.tool metadata.json; then
23
+ echo "INFO: metadata.json is valid JSON. Proceeding with the test."
24
+ else
25
+ echo "CRITICAL: metadata.json exists but is not valid JSON, please do report in #sig-ci channel!"
26
+ exit 1
27
+ fi
28
+ break
29
+ fi
30
+ # failure handling
31
+ if [ $i -eq 5 ]; then
32
+ echo "ERROR: metadata.json URL is still not valid after 5 attempts."
33
+ echo "ERROR: Please check whether the precompiled wheel for commit $merge_base_commit exists."
34
+ echo " NOTE: If $merge_base_commit is a new commit on main, maybe try again after its release pipeline finishes."
35
+ echo " NOTE: If it fails, please report in #sig-ci channel."
36
+ exit 1
37
+ else
38
+ echo "WARNING: metadata.json URL is not valid. Retrying in 3 minutes..."
39
+ sleep 180
40
+ fi
41
+done
42
+
43
+set -x
44
+
45
cd /vllm-workspace/
46
47
# uninstall vllm
@@ -29,6 +62,6 @@ python3 -c 'import vllm'
62
63
# Check if the clangd log file was created
64
if [ ! -f /tmp/changed.file ]; then
32
- echo "changed.file was not created, python only compilation failed"
65
+ echo "ERROR: changed.file was not created, python only compilation failed"
66
exit 1
67
fi