Build and release sgl-deep-ep wheels (#33498)

This commit is contained in:
Baizhou Zhang
2026-08-06 02:07:36 -07:00
committed by GitHub
parent 8b29c90218
commit ba9074035a
4 changed files with 615 additions and 0 deletions
+261
View File
@@ -0,0 +1,261 @@
name: Release sgl-deep-ep
on:
workflow_dispatch:
inputs:
version:
description: "Public wheel version (for example, 0.1.0 or 0.1.1rc0)"
type: string
required: true
target:
description: "CUDA target"
type: choice
required: false
default: "all"
options:
- "all"
- "cu129"
- "cu130"
packaging-ref:
description: "DeepEP ref containing the shared sgl_deep_ep overlay"
type: string
required: false
default: "sgl-deepep-packaging"
concurrency:
group: release-sgl-deep-ep
cancel-in-progress: false
env:
TORCH_VERSION: "2.11.0"
jobs:
build-cu129-matrix:
if: |
github.repository == 'sgl-project/sglang' &&
(inputs.target == 'all' || inputs.target == 'cu129')
strategy:
fail-fast: false
matrix:
python-version: ["3.10", "3.12"]
arch: [x86_64, aarch64]
include:
- arch: x86_64
runner: x64-kernel-build-node
source-branch: sgl-deepep-x86
- arch: aarch64
runner: arm-kernel-build-node
source-branch: sgl-deepep-cu12-arm
runs-on: ${{ matrix.runner }}
steps:
- name: Clean workspace
run: |
docker run --rm -v "${{ github.workspace }}:/workspace" alpine:3 \
sh -c 'rm -rf /workspace/..?* /workspace/.[!.]* /workspace/*' || true
- name: Checkout SGLang
uses: actions/checkout@v4
- name: Checkout DeepEP implementation
uses: actions/checkout@v4
with:
repository: sgl-project/DeepEP
ref: ${{ matrix.source-branch }}
path: DeepEP-source
submodules: recursive
- name: Checkout DeepEP packaging overlay
uses: actions/checkout@v4
with:
repository: sgl-project/DeepEP
ref: ${{ inputs['packaging-ref'] || 'sgl-deepep-packaging' }}
path: DeepEP-packaging
sparse-checkout: sgl_deep_ep
- name: Set wheel version
env:
RELEASE_VERSION: ${{ inputs.version }}
run: |
printf '%s' "${RELEASE_VERSION}" > DeepEP-packaging/sgl_deep_ep/VERSION
- name: Build wheel
run: |
chmod +x scripts/build_sgl_deepep.sh DeepEP-packaging/sgl_deep_ep/build_sgl_deep_ep.sh
scripts/build_sgl_deepep.sh \
"${{ matrix.python-version }}" \
"12.9" \
"${{ github.workspace }}/DeepEP-source" \
"${{ github.workspace }}/DeepEP-packaging/sgl_deep_ep" \
"${{ matrix.arch }}"
- name: Upload wheel
uses: actions/upload-artifact@v4
with:
name: deepep-wheel-cu129-py${{ matrix.python-version }}-${{ matrix.arch }}
path: DeepEP-source/dist/*.whl
if-no-files-found: error
build-cu130-matrix:
if: |
github.repository == 'sgl-project/sglang' &&
(inputs.target == 'all' || inputs.target == 'cu130')
strategy:
fail-fast: false
matrix:
python-version: ["3.10", "3.11", "3.12", "3.13"]
arch: [x86_64, aarch64]
include:
- arch: x86_64
runner: x64-kernel-build-node
source-branch: sgl-deepep-x86
- arch: aarch64
runner: arm-kernel-build-node
source-branch: sgl-deepep-arm
runs-on: ${{ matrix.runner }}
steps:
- name: Clean workspace
run: |
docker run --rm -v "${{ github.workspace }}:/workspace" alpine:3 \
sh -c 'rm -rf /workspace/..?* /workspace/.[!.]* /workspace/*' || true
- name: Checkout SGLang
uses: actions/checkout@v4
- name: Checkout DeepEP implementation
uses: actions/checkout@v4
with:
repository: sgl-project/DeepEP
ref: ${{ matrix.source-branch }}
path: DeepEP-source
submodules: recursive
- name: Checkout DeepEP packaging overlay
uses: actions/checkout@v4
with:
repository: sgl-project/DeepEP
ref: ${{ inputs['packaging-ref'] || 'sgl-deepep-packaging' }}
path: DeepEP-packaging
sparse-checkout: sgl_deep_ep
- name: Set wheel version
env:
RELEASE_VERSION: ${{ inputs.version }}
run: |
printf '%s' "${RELEASE_VERSION}" > DeepEP-packaging/sgl_deep_ep/VERSION
- name: Build wheels
run: |
chmod +x scripts/build_sgl_deepep.sh DeepEP-packaging/sgl_deep_ep/build_sgl_deep_ep.sh
scripts/build_sgl_deepep.sh \
"${{ matrix.python-version }}" \
"13.0" \
"${{ github.workspace }}/DeepEP-source" \
"${{ github.workspace }}/DeepEP-packaging/sgl_deep_ep" \
"${{ matrix.arch }}"
- name: Upload CUDA-tagged wheel
uses: actions/upload-artifact@v4
with:
name: deepep-wheel-cu130-py${{ matrix.python-version }}-${{ matrix.arch }}
path: DeepEP-source/dist/*.whl
if-no-files-found: error
- name: Upload PyPI wheel
uses: actions/upload-artifact@v4
with:
name: deepep-pypi-cu130-py${{ matrix.python-version }}-${{ matrix.arch }}
path: DeepEP-source/dist-pypi/*.whl
if-no-files-found: error
release-whl:
needs: [build-cu129-matrix, build-cu130-matrix]
if: |
always() &&
github.repository == 'sgl-project/sglang' &&
(
(inputs.target == 'cu129' && needs.build-cu129-matrix.result == 'success') ||
(inputs.target == 'cu130' && needs.build-cu130-matrix.result == 'success') ||
(
inputs.target == 'all' &&
needs.build-cu129-matrix.result == 'success' &&
needs.build-cu130-matrix.result == 'success'
)
)
runs-on: ubuntu-latest
steps:
- name: Checkout SGLang
uses: actions/checkout@v4
- name: Download CUDA-tagged wheels
uses: actions/download-artifact@v4
with:
path: dist
merge-multiple: true
pattern: deepep-wheel-*
- name: Upload release assets
uses: softprops/action-gh-release@v2
with:
tag_name: v${{ inputs.version }}
repository: sgl-project/whl
token: ${{ secrets.GH_PAT_FOR_WHL_RELEASE }}
fail_on_unmatched_files: true
files: dist/*.whl
- name: Clone wheel index
env:
WHL_TOKEN: ${{ secrets.GH_PAT_FOR_WHL_RELEASE }}
run: |
git clone --branch gh-pages \
"https://oauth2:${WHL_TOKEN}@github.com/sgl-project/whl.git" sgl-whl
- name: Update wheel indexes
run: |
if [[ "${{ inputs.target }}" == "all" || "${{ inputs.target }}" == "cu129" ]]; then
python3 scripts/update_deepep_whl_index.py --cuda 129
fi
if [[ "${{ inputs.target }}" == "all" || "${{ inputs.target }}" == "cu130" ]]; then
python3 scripts/update_deepep_whl_index.py --cuda 130
fi
- name: Push wheel indexes
env:
RELEASE_VERSION: ${{ inputs.version }}
run: |
cd sgl-whl
git config --local user.name "sglang-bot"
git config --local user.email "sglangbot@gmail.com"
git add -A
if git diff --cached --quiet; then
echo "Wheel indexes are already up to date"
else
git commit -m "update sgl-deep-ep whl indexes for v${RELEASE_VERSION}"
git push origin gh-pages
fi
release-pypi:
needs: [build-cu129-matrix, build-cu130-matrix]
if: |
always() &&
github.repository == 'sgl-project/sglang' &&
needs.build-cu130-matrix.result == 'success' &&
(
inputs.target == 'cu130' ||
(inputs.target == 'all' && needs.build-cu129-matrix.result == 'success')
)
runs-on: ubuntu-latest
steps:
- name: Download PyPI wheels
uses: actions/download-artifact@v4
with:
path: dist-pypi
merge-multiple: true
pattern: deepep-pypi-cu130-*
- name: Upload to PyPI
env:
TWINE_PASSWORD: ${{ secrets.SGL_DEEP_EP_PYPI_TOKEN }}
TWINE_USERNAME: __token__
run: |
python3 -m pip install --upgrade twine "packaging>=24.2"
python3 -m twine upload --skip-existing dist-pypi/*.whl
+75
View File
@@ -0,0 +1,75 @@
ARG BASE_IMAGE=pytorch/manylinux2_28-builder
ARG CUDA_VERSION=13.0
FROM ${BASE_IMAGE}:cuda${CUDA_VERSION}
ARG ARCHITECTURE=x86_64
ARG CUDA_TAG=cu130
ARG CUDA_VERSION=13.0
ARG GDRCOPY_VERSION=2.5.1
ARG PYTHON_TAG=cp312-cp312
ARG TORCH_VERSION=2.11.0
ENV CUDA_HOME=/usr/local/cuda
ENV GDRCOPY_HOME=/usr/local
ENV LD_LIBRARY_PATH=/usr/local/lib:/usr/local/lib64:/usr/local/cuda/lib64:${LD_LIBRARY_PATH}
ENV PATH=/opt/python/${PYTHON_TAG}/bin:${PATH}
ENV PYTHON_BIN=/opt/python/${PYTHON_TAG}/bin/python
# These mirror the build and RDMA dependencies used by ci_install_deepep.sh.
# The image builds only GDRCopy's user-space library: the host owns gdrdrv and
# passes /dev/gdrdrv through at runtime.
RUN yum install -y --nogpgcheck \
cmake \
curl \
gcc \
gcc-c++ \
git \
infiniband-diags \
libfabric \
libfabric-devel \
libibverbs \
libibverbs-devel \
libibverbs-utils \
librdmacm \
librdmacm-devel \
make \
openssh-server \
patchelf \
perftest \
pkgconfig \
rdma-core \
wget \
&& yum clean all \
&& rm -rf /var/cache/yum
RUN set -eux; \
if [ "${ARCHITECTURE}" = aarch64 ]; then cuda_target=sbsa; else cuda_target="${ARCHITECTURE}"; fi; \
cuda_stub="/usr/local/cuda-${CUDA_VERSION}/targets/${cuda_target}-linux/lib/stubs/libcuda.so"; \
test -f "${cuda_stub}"; \
mkdir -p /usr/lib64 "/usr/lib/${ARCHITECTURE}-linux-gnu"; \
ln -sf "${cuda_stub}" /usr/lib64/libcuda.so; \
ln -sf "${cuda_stub}" "/usr/lib/${ARCHITECTURE}-linux-gnu/libcuda.so"
RUN git clone --depth 1 --branch "v${GDRCOPY_VERSION}" \
https://github.com/NVIDIA/gdrcopy.git /opt/gdrcopy \
&& make -C /opt/gdrcopy CUDA="${CUDA_HOME}" prefix=/usr/local lib_install \
&& ldconfig \
&& test -f /usr/local/include/gdrapi.h \
&& ldconfig -p | grep -q libgdrapi
RUN --mount=type=cache,id=sgl-deep-ep-pip-${CUDA_TAG}-${PYTHON_TAG}-${ARCHITECTURE},target=/root/.cache/pip \
set -eux; \
"${PYTHON_BIN}" -m pip uninstall -y deep-ep sgl-deep-ep || true; \
"${PYTHON_BIN}" -m pip install --upgrade pip; \
"${PYTHON_BIN}" -m pip install --force-reinstall \
"torch==${TORCH_VERSION}" \
--index-url "https://download.pytorch.org/whl/${CUDA_TAG}"; \
"${PYTHON_BIN}" -m pip install \
"auditwheel>=6.0" \
build \
ninja \
packaging \
setuptools \
wheel; \
"${PYTHON_BIN}" -c 'import torch; assert torch.__version__.startswith("2.11.0"); print(torch.__version__, torch.version.cuda)'
+197
View File
@@ -0,0 +1,197 @@
#!/usr/bin/env bash
# Build sgl-deep-ep wheels in a CUDA-versioned manylinux container.
#
# Usage:
# build_sgl_deepep.sh <python-version> <cuda-version> <deepep-source> <packaging-overlay> [architecture]
#
# Writes CUDA-tagged wheels to <deepep-source>/dist. CUDA 13 builds also write
# PyPI-ready wheels without the local CUDA version to <deepep-source>/dist-pypi.
set -euo pipefail
usage() {
cat <<'EOF'
Usage: build_sgl_deepep.sh <python-version> <cuda-version> <deepep-source> <packaging-overlay> [architecture]
python-version: 3.10, 3.11, 3.12, or 3.13
cuda-version: 12.9 or 13.0
deepep-source: checkout of the selected DeepEP implementation branch
packaging-overlay: path to the shared DeepEP sgl_deep_ep directory
architecture: x86_64 or aarch64 (defaults to the current machine)
EOF
}
if [[ $# -lt 4 || $# -gt 5 ]]; then
usage >&2
exit 2
fi
PYTHON_VERSION="$1"
CUDA_VERSION="$2"
DEEPEP_SOURCE="$(cd "$3" && pwd)"
PACKAGING_OVERLAY="$(cd "$4" && pwd)"
ARCHITECTURE="${5:-$(uname -m)}"
if [[ "${ARCHITECTURE}" == arm64 ]]; then
ARCHITECTURE=aarch64
fi
case "${PYTHON_VERSION}" in
3.10|3.11|3.12|3.13) ;;
*)
echo "Unsupported Python version: ${PYTHON_VERSION}" >&2
exit 2
;;
esac
case "${CUDA_VERSION}" in
12.9)
CUDA_TAG=cu129
;;
13.0)
CUDA_TAG=cu130
;;
*)
echo "Unsupported CUDA version: ${CUDA_VERSION}" >&2
exit 2
;;
esac
case "${ARCHITECTURE}" in
x86_64)
BASE_IMAGE=pytorch/manylinux2_28-builder
;;
aarch64)
BASE_IMAGE=pytorch/manylinuxaarch64-builder
;;
*)
echo "Unsupported architecture: ${ARCHITECTURE}" >&2
exit 2
;;
esac
for required_file in setup.py deep_ep/__init__.py; do
if [[ ! -f "${DEEPEP_SOURCE}/${required_file}" ]]; then
echo "DeepEP source is missing ${required_file}: ${DEEPEP_SOURCE}" >&2
exit 1
fi
done
for required_file in build_sgl_deep_ep.sh setup.py VERSION; do
if [[ ! -f "${PACKAGING_OVERLAY}/${required_file}" ]]; then
echo "Packaging overlay is missing ${required_file}: ${PACKAGING_OVERLAY}" >&2
exit 1
fi
done
if ! command -v docker >/dev/null; then
echo "docker is required to build sgl-deep-ep" >&2
exit 1
fi
PYTHON_TAG="cp${PYTHON_VERSION//.}-cp${PYTHON_VERSION//.}"
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
REPOSITORY_ROOT="$(cd "${SCRIPT_DIR}/.." && pwd)"
DOCKERFILE="${REPOSITORY_ROOT}/docker/sgl-deep-ep.Dockerfile"
IMAGE_TAG="sgl-deep-ep-builder:cuda${CUDA_VERSION}-${PYTHON_TAG}-${ARCHITECTURE}"
DIST_DIR="${DEEPEP_SOURCE}/dist"
PYPI_DIST_DIR="${DEEPEP_SOURCE}/dist-pypi"
mkdir -p "${DIST_DIR}" "${PYPI_DIST_DIR}"
echo "----------------------------------------"
echo "Python: ${PYTHON_VERSION} (${PYTHON_TAG})"
echo "CUDA: ${CUDA_VERSION} (${CUDA_TAG})"
echo "Architecture: ${ARCHITECTURE}"
echo "Base image: ${BASE_IMAGE}:cuda${CUDA_VERSION}"
echo "DeepEP source: ${DEEPEP_SOURCE}"
echo "Packaging overlay: ${PACKAGING_OVERLAY}"
echo "Builder image: ${IMAGE_TAG}"
echo "----------------------------------------"
docker build \
--file "${DOCKERFILE}" \
--build-arg BASE_IMAGE="${BASE_IMAGE}" \
--build-arg CUDA_VERSION="${CUDA_VERSION}" \
--build-arg CUDA_TAG="${CUDA_TAG}" \
--build-arg PYTHON_TAG="${PYTHON_TAG}" \
--build-arg ARCHITECTURE="${ARCHITECTURE}" \
--build-arg TORCH_VERSION="${TORCH_VERSION:-2.11.0}" \
--tag "${IMAGE_TAG}" \
--network=host \
"${REPOSITORY_ROOT}/docker"
docker run --rm \
--network=host \
--env ARCHITECTURE="${ARCHITECTURE}" \
--env CUDA_TAG="${CUDA_TAG}" \
--env CUDA_VERSION="${CUDA_VERSION}" \
--env MAX_JOBS="${MAX_JOBS:-8}" \
--volume "${DEEPEP_SOURCE}:/deepep:ro" \
--volume "${PACKAGING_OVERLAY}:/packaging:ro" \
--volume "${DIST_DIR}:/output/dist" \
--volume "${PYPI_DIST_DIR}:/output/dist-pypi" \
"${IMAGE_TAG}" \
bash -euo pipefail -c '
find /output/dist -maxdepth 1 -type f -name "sgl_deep_ep-*.whl" -delete
find /output/dist-pypi -maxdepth 1 -type f -name "sgl_deep_ep-*.whl" -delete
raw_dir="$(mktemp -d -t sgl-deep-ep-raw.XXXXXX)"
trap '\''rm -rf -- "${raw_dir}"'\'' EXIT
bash /packaging/build_sgl_deep_ep.sh \
/deepep /packaging "${raw_dir}" "${CUDA_VERSION}" "${ARCHITECTURE}"
shopt -s nullglob
raw_wheels=("${raw_dir}"/*.whl)
if [[ ${#raw_wheels[@]} -ne 1 ]]; then
echo "Expected exactly one raw wheel, found ${#raw_wheels[@]}" >&2
exit 1
fi
auditwheel repair \
--plat "manylinux_2_28_${ARCHITECTURE}" \
--wheel-dir /output/dist \
--exclude libcuda.so.1 \
--exclude libcudart.so.12 \
--exclude libcudart.so.13 \
--exclude libc10.so \
--exclude libc10_cuda.so \
--exclude libtorch.so \
--exclude libtorch_cpu.so \
--exclude libtorch_cuda.so \
--exclude libtorch_python.so \
--exclude libnvshmem_host.so.1 \
--exclude libnvshmem_host.so.2 \
--exclude libnvshmem_host.so.3 \
--exclude libnccl.so.2 \
--exclude libgdrapi.so.2 \
--exclude libnvToolsExt.so.1 \
"${raw_wheels[0]}"
if [[ "${CUDA_TAG}" == cu130 ]]; then
tagged_wheels=(/output/dist/sgl_deep_ep-*+cu130-*.whl)
if [[ ${#tagged_wheels[@]} -ne 1 ]]; then
echo "Expected exactly one CUDA 13 wheel, found ${#tagged_wheels[@]}" >&2
exit 1
fi
unpack_root="$(mktemp -d -t sgl-deep-ep-pypi.XXXXXX)"
python -m wheel unpack "${tagged_wheels[0]}" --dest "${unpack_root}"
unpacked="$(find "${unpack_root}" -mindepth 1 -maxdepth 1 -type d | head -1)"
dist_info="$(find "${unpacked}" -maxdepth 1 -type d -name "*.dist-info" | head -1)"
metadata="${dist_info}/METADATA"
original_version="$(sed -n "s/^Version:[[:space:]]*//p" "${metadata}" | head -1)"
public_version="${original_version%+cu130}"
if [[ "${original_version}" == "${public_version}" ]]; then
echo "CUDA 13 wheel metadata lacks the +cu130 local version" >&2
exit 1
fi
sed -i "s/^Version:.*/Version: ${public_version}/" "${metadata}"
old_dist_info="$(basename "${dist_info}")"
new_dist_info="${old_dist_info/${original_version}/${public_version}}"
mv "${dist_info}" "$(dirname "${dist_info}")/${new_dist_info}"
python -m wheel pack "${unpacked}" --dest-dir /output/dist-pypi
fi
ls -lh /output/dist/*.whl
if [[ "${CUDA_TAG}" == cu130 ]]; then
ls -lh /output/dist-pypi/*.whl
fi
'
+82
View File
@@ -0,0 +1,82 @@
"""Update the PEP 503 indexes for sgl-deep-ep release wheels."""
import argparse
import hashlib
import pathlib
import re
SUPPORTED_CUDA_VERSIONS = ("129", "130")
WHEEL_PATTERN = re.compile(
r"^sgl_deep_ep-(?P<version>[0-9][^-]*)-[^-]+-[^-]+-[^-]+\.whl$"
)
ANCHOR_PATTERN = re.compile(r'^<a href="[^"]+">(?P<filename>[^<]+)</a><br>$')
def _sha256(path: pathlib.Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as wheel_file:
for chunk in iter(lambda: wheel_file.read(1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
def _ensure_root_link(cuda_root: pathlib.Path) -> None:
root_index = cuda_root / "index.html"
lines = root_index.read_text().splitlines() if root_index.exists() else []
doctype = [line for line in lines if line.startswith("<!DOCTYPE")]
anchors = [line for line in lines if line.startswith("<a ")]
link = '<a href="sgl-deep-ep/">sgl-deep-ep</a>'
if link not in anchors:
anchors.append(link)
content = [*(doctype or ["<!DOCTYPE html>"]), *sorted(set(anchors))]
root_index.write_text("\n".join(content) + "\n")
def update_wheel_index(
cuda_version: str,
wheel_dir: pathlib.Path,
repository: pathlib.Path,
) -> None:
if cuda_version not in SUPPORTED_CUDA_VERSIONS:
raise ValueError(f"Unsupported CUDA version: {cuda_version}")
cuda_root = repository / f"cu{cuda_version}"
index_dir = cuda_root / "sgl-deep-ep"
index_dir.mkdir(exist_ok=True, parents=True)
_ensure_root_link(cuda_root)
index_path = index_dir / "index.html"
entries = {}
if index_path.exists():
for line in index_path.read_text().splitlines():
match = ANCHOR_PATTERN.match(line)
if match:
entries[match.group("filename")] = line
suffix = f"+cu{cuda_version}"
release_base = "https://github.com/sgl-project/whl/releases/download"
for path in sorted(wheel_dir.glob("*.whl")):
match = WHEEL_PATTERN.match(path.name)
if not match or suffix not in match.group("version"):
continue
public_version = match.group("version").split("+", 1)[0]
url = f"{release_base}/v{public_version}/{path.name}#sha256={_sha256(path)}"
entries[path.name] = f'<a href="{url}">{path.name}</a><br>'
content = ["<!DOCTYPE html>", *(entries[name] for name in sorted(entries))]
index_path.write_text("\n".join(content) + "\n")
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--cuda", required=True, choices=SUPPORTED_CUDA_VERSIONS)
parser.add_argument("--wheel-dir", type=pathlib.Path, default=pathlib.Path("dist"))
parser.add_argument(
"--repository", type=pathlib.Path, default=pathlib.Path("sgl-whl")
)
args = parser.parse_args()
update_wheel_index(args.cuda, args.wheel_dir, args.repository)
if __name__ == "__main__":
main()