Skip to content

refactor: extract shared logic from nightly build scripts #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
maobaolong wants to merge 1 commit into
base: main
Choose a base branch
from
Open

refactor: extract shared logic from nightly build scripts #4

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
173 changes: 23 additions & 150 deletions scripts/build_and_publish_vllm_cpu_nightly.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,12 @@
# 5. Twine-upload to PyPI using the API token from
# /data/home/baoloongmao/pypi_apikey.
#
# The steps that are identical to the macOS build (build_and_publish_
# vllm_cpu_nightly_macos.sh) live under scripts/common/ and are mounted
# read-only into the build container so both scripts execute the exact same
# code -- see the `-v ".../common:/opt/vllm-cpu-nightly-common:ro"` docker
# flag below.
#
# Usage (on the devcloud host):
# bash build_and_publish_vllm_cpu_nightly.sh
#
Expand All @@ -37,16 +43,18 @@
# ----------------------------------------------------------------------------
set -euo pipefail

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
COMMON_DIR="${SCRIPT_DIR}/common"
# shellcheck source=common/helpers.sh
source "${COMMON_DIR}/helpers.sh"

PYPI_TOKEN_FILE="${PYPI_TOKEN_FILE:-/data/home/baoloongmao/pypi_apikey}"
PKG_NAME="${PKG_NAME:-vllm-cpu-nightly}"
VLLM_GIT_REF="${VLLM_GIT_REF:-main}"
SKIP_UPLOAD="${SKIP_UPLOAD:-0}"
WORK_DIR="${WORK_DIR:-/data/vllm-cpu-nightly-build}"
IMAGE="ubuntu:22.04"

log() { printf '\033[1;36m[%s]\033[0m %s\n' "$(date +%H:%M:%S)" "$*"; }
die() { printf '\033[1;31m[ERR]\033[0m %s\n' "$*" >&2; exit 1; }

[[ -f "${PYPI_TOKEN_FILE}" ]] || die "PyPI token file not found: ${PYPI_TOKEN_FILE}"
PYPI_TOKEN="$(tr -d '[:space:]' < "${PYPI_TOKEN_FILE}")"
[[ -n "${PYPI_TOKEN}" ]] || die "PyPI token file is empty"
Expand All @@ -73,6 +81,7 @@ docker run -d --name "${CONTAINER_NAME}" \
-e CC=gcc-12 \
-e CXX=g++-12 \
-v "${HOST_OUT_DIR}:/out" \
-v "${COMMON_DIR}:/opt/vllm-cpu-nightly-common:ro" \
"${IMAGE}" \
sleep infinity >/dev/null

Expand Down Expand Up @@ -124,70 +133,20 @@ run '
set -euo pipefail
cd /src/vllm
. /opt/venv/bin/activate
source /opt/vllm-cpu-nightly-common/helpers.sh

# Determine the upstream version that setuptools-scm would have produced
# for this commit. We deliberately do NOT trust setuptools-scm here:
# a shallow clone has no reachable tag and scm falls back to "0.1.dev1",
# which is too low (PyPI would consider it older than the real 0.x.y
# release line). Instead, fetch all tags and use the latest stable
# release tag (vX.Y.Z, no rc/a/b suffix), then bump the patch by 1
# because vLLM `main` is always post-release.
git fetch --tags --depth=1 origin >/dev/null 2>&1 || true
LAST_STABLE="$(git tag --sort=-v:refname | grep -E "^v[0-9]+\.[0-9]+\.[0-9]+$" | head -1)"
LAST_STABLE="${LAST_STABLE:-v0.0.0}"
BASE_RAW="${LAST_STABLE#v}"
BASE_VER="$(echo "${BASE_RAW}" | awk -F. "{ printf \"%d.%d.%d\", \$1, \$2, \$3+1 }")"
echo "Last stable tag : ${LAST_STABLE}"
echo "Bumped base ver : ${BASE_VER}"
echo "${BASE_VER}" > /src/base_version.txt

NIGHTLY_VER="${BASE_VER}.dev${STAMP_FULL}"
NIGHTLY_VER="$(compute_nightly_version "${STAMP_FULL}")"
echo "${NIGHTLY_VER}" > /src/nightly_version.txt

# Rename the project distribution to PKG_NAME. Only [project].name in
# pyproject.toml needs to change; the import package (`vllm/`) is
# untouched so `import vllm` still works.
python - <<PY
import pathlib, re
pkg = "${PKG_NAME}"
p = pathlib.Path("pyproject.toml")
txt = p.read_text()
new = re.sub(r"^(\s*name\s*=\s*[\"\x27])vllm([\"\x27])",
lambda m: m.group(1) + pkg + m.group(2),
txt, count=1, flags=re.MULTILINE)
assert new != txt, "failed to patch pyproject.toml name"
p.write_text(new)
print(f"Patched name -> {pkg} in pyproject.toml")
PY

# Patch requirements/cpu.txt: PyPI rejects wheels whose dependencies
# carry a PEP 440 local label like `torch==2.11.0+cpu`. Strip the +cpu
# suffix; downstream installers can still pull the cpu wheel with
# `--extra-index-url https://download.pytorch.org/whl/cpu`.
sed -i -E "s/torch==([0-9.]+)\+cpu/torch==\1/g" requirements/cpu.txt
echo "--- patched requirements/cpu.txt ---"
grep -E "^torch" requirements/cpu.txt || true

# Patch csrc/cpu/sgl-kernels/fla.cpp: it initialises a constexpr float from
# std::sqrt, which is not constexpr in libc++ before C++26. GCC accepts it
# as a builtin so Linux is unaffected today, but keep the two build scripts
# in step so a compiler change does not surprise us. D is a template
# parameter, so const is semantically identical and still folds.
FLA_SRC="csrc/cpu/sgl-kernels/fla.cpp"
if [ -f "$FLA_SRC" ] \
&& grep -q "constexpr float scale = 1\.f / std::sqrt" "$FLA_SRC"; then
sed -i -E "s/constexpr( float scale = 1\.f \/ std::sqrt)/const\1/" "$FLA_SRC"
echo "--- patched $FLA_SRC (constexpr -> const) ---"
grep -n "float scale = 1\.f / std::sqrt" "$FLA_SRC" || true
fi
bash /opt/vllm-cpu-nightly-common/patch_vllm_source.sh
'

log "Installing build deps + torch(cpu) inside container..."
run '
set -euo pipefail
. /opt/venv/bin/activate
# Without build isolation we must install everything that
# pyproject.toml [build-system].requires asks for ourselves.
# pyproject.toml [build-system].requires asks forourselves.
pip install "cmake>=3.26.1" ninja "packaging>=24.2" \
"setuptools>=77.0.3,<81.0.0" "setuptools-scm>=8.0" \
"setuptools-rust>=1.9.0" wheel jinja2 >/dev/null
Expand All @@ -202,31 +161,9 @@ fi
# Plain torch (cpu) so setup.py can `import torch`.
pip install "numpy<2" >/dev/null
# Build against exactly the torch this wheel will declare as a dependency.
# This used to be hardcoded to 2.11.0 while the declared version tracked
# upstream requirements/cpu.txt; when upstream moved to 2.13.0 the macOS
# wheel still linked against 2.11 and its _C.abi3.so ended up referencing a
# c10 symbol the installed torch does not export.
#
# requirements/cpu.txt carries one torch line per platform group, each
# guarded by an environment marker, so taking the first match can pick a pin
# meant for another architecture -- pip then evaluates the marker, skips the
# install and still exits 0, leaving the build to fail with "No module named
# torch". Let packaging evaluate the markers and pick the applicable pin.
# Absolute path because this step does not cd into the source tree.
TORCH_REQ="$(python - <<"PYEOF"
from packaging.requirements import Requirement

with open("/src/vllm/requirements/cpu.txt") as f:
for line in f:
line = line.split("#", 1)[0].strip()
if not line.startswith("torch=="):
continue
req = Requirement(line)
if req.marker is None or req.marker.evaluate():
print(req.name + str(req.specifier))
break
PYEOF
)"
# See scripts/common/resolve_torch_requirement.py for why we cannot just
# take the first `torch==` line in requirements/cpu.txt.
TORCH_REQ="$(python3 /opt/vllm-cpu-nightly-common/resolve_torch_requirement.py /src/vllm/requirements/cpu.txt)"
: "${TORCH_REQ:?no applicable torch== pin in /src/vllm/requirements/cpu.txt}"
echo "building against ${TORCH_REQ}"
pip install "${TORCH_REQ}" \
Expand All @@ -248,7 +185,7 @@ run "
export VLLM_VERSION_OVERRIDE='${NIGHTLY_VER}'
# vLLM's setup.py defaults to RelWithDebInfo (-O2 -g). The debug symbols
# pushed this wheel from 96 MiB to 105 MiB, past PyPI's 100 MiB per-file
# limit, and uploads have been failing with an opaque 400 ever since.
# limit, and uploads have been failing with an opaque 400ever since.
# Release keeps the optimisation without the symbols.
export CMAKE_BUILD_TYPE='${CMAKE_BUILD_TYPE:-Release}'
export MAX_JOBS=\$(nproc)
Expand All @@ -259,81 +196,17 @@ run "
twine check /out/*.whl
"

# `twine check` only validates metadata, so it would happily pass a wheel
# built against a different torch than it declares -- the exact defect that
# shipped on 2026-07-30. Compare the two versions directly.
#
# Do NOT try to dlopen the extensions to verify this. vLLM builds several
# ISA-specialised variants (_C, _C_AVX2, _C_AVX512) and the default one is
# compiled with -mavx512* -mamx-*; loading it on a host that lacks those
# instructions, or that has not requested AMX state via arch_prctl, dies with
# SIGILL and tells us nothing about the ABI. vLLM picks a variant at runtime
# after probing the CPU, so "every .so dlopens here" was never the right
# invariant.
log "Verifying the wheel declares the torch it was built against..."
run '
set -euo pipefail
. /opt/venv/bin/activate
python - /out <<"PYEOF"
import glob
import sys
import zipfile

import torch


def normalize(version):
"""Drop the PEP 440 local label, e.g. 2.13.0+cpu -> 2.13.0."""
return version.split("+")[0]


built = normalize(torch.__version__)
wheels = glob.glob(sys.argv[1] + "/*.whl")
if not wheels:
sys.exit("no wheel found in " + sys.argv[1])

for whl in wheels:
with zipfile.ZipFile(whl) as z:
names = [n for n in z.namelist() if n.endswith(".dist-info/METADATA")]
if not names:
sys.exit("no METADATA in " + whl)
text = z.read(names[0]).decode("utf-8", "replace")

declared = set()
for line in text.splitlines():
if line.startswith("Requires-Dist: torch=="):
spec = line.split("==", 1)[1].split(";")[0].strip()
declared.add(normalize(spec))

if not declared:
sys.exit(whl + " declares no pinned torch requirement")
if declared != {built}:
sys.exit(
"ABI mismatch: built against torch %s but the wheel declares %s"
% (built, ", ".join(sorted(declared)))
)
print("torch matches: built against and declares %s" % built)
PYEOF
python3 /opt/vllm-cpu-nightly-common/verify_wheel_torch.py /out
'

log "Wheels staged on host:"
ls -lh "${HOST_OUT_DIR}"

# PyPI rejects files larger than 100 MiB with an opaque "400 Bad Request".
# This wheel sat ~3 MiB below that ceiling for weeks, crossed it on
# 2026-07-20, and every upload failed silently for the next twelve nights.
# Fail here with an actionable message instead.
PYPI_MAX_BYTES=$((100 * 1024 * 1024))
for whl in "${HOST_OUT_DIR}"/*.whl; do
size="$(wc -c <"${whl}")"
printf ' %s: %d MiB\n' "$(basename "${whl}")" "$((size / 1024 / 1024))"
if [[ "${size}" -gt "${PYPI_MAX_BYTES}" ]]; then
echo "[ERROR] $(basename "${whl}") is $((size / 1024 / 1024)) MiB," \
"over PyPI's 100 MiB per-file limit. Shrink the build (check" \
"CMAKE_BUILD_TYPE) or publish somewhere without a size cap." >&2
exit 1
fi
done
check_wheel_size "${HOST_OUT_DIR}"

if [[ "${SKIP_UPLOAD}" == "1" ]]; then
log "SKIP_UPLOAD=1 -> stopping before twine upload."
Expand All @@ -351,4 +224,4 @@ docker exec \
'

log "Done. Index page: https://pypi.org/project/${PKG_NAME}/"
log "Install with: pip install ${PKG_NAME}"
log "Install with: pip install ${PKG_NAME}"
Loading