minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Install DeepEP from release wheels (#33932)

Browse Source
This commit is contained in:
Baizhou Zhang
2026-08-07 15:38:44 -07:00
committed by GitHub
parent 115cd7bde1
commit eb3cc879e0
31 changed files with 141 additions and 353 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/ci/cuda/ci_install_deepep.sh
-177
View File
@@ -1,177 +0,0 @@
#!/bin/bash
# Install the dependency in CI.
set -euxo pipefail
# Source (not bash) so that venv activation, $PIP_CMD, $CU_VERSION, $NVCC_VER, and
# $PIP_INSTALL_SUFFIX all propagate into this shell. Without sourcing, the subshell
# exits and this script would fall back to system Python.
#
# Note: any `exit N` or `set -e` trip inside the sourced script terminates *this*
# script too (bash runs sourced commands in the current shell, so `exit` is not
# caught by `if`/`||`). The real error message appears upstream in the log.
# shellcheck disable=SC1091
source scripts/ci/cuda/ci_install_dependency.sh
# In venv mode, PIP_CMD must be set by the sourced script. If it isn't, the
# source chain is broken and we'd silently fall back to system `pip` below —
# exactly the split-install bug the migration is meant to prevent.
if [ -z "${PIP_CMD:-}" ]; then
echo "FATAL:PIP_CMD is unset after sourcing ci_install_dependency.sh"
exit 1
fi
export GDRCOPY_HOME=/usr/src/gdrdrv-2.5.1/
export CUDA_HOME=/usr/local/cuda
GRACE_BLACKWELL=${GRACE_BLACKWELL:-0}
# Detect architecture
ARCH=$(uname -m)
if [ "$ARCH" != "x86_64" ] && [ "$ARCH" != "aarch64" ]; then
echo "Unsupported architecture: $ARCH"
exit 1
fi
if [ "${FORCE_REBUILD_DEEPEP:-0}" = "1" ]; then
echo "FORCE_REBUILD_DEEPEP=1; uninstalling any cached deep_ep before rebuild."
${PIP_UNINSTALL_CMD:-pip uninstall -y} deep_ep ${PIP_UNINSTALL_SUFFIX:-} || true
elif python3 -c "import deep_ep" >/dev/null 2>&1; then
echo "deep_ep is already installed or importable. Skipping installation."
exit 0
fi
# Install system dependencies
# Use fallback logic in case apt fails due to unrelated broken packages on the runner
DEEPEP_SYSTEM_DEPS="curl wget git sudo rdma-core infiniband-diags openssh-server perftest libibumad3 libibverbs-dev libibverbs1 ibverbs-providers ibverbs-utils libnl-3-200 libnl-route-3-200 librdmacm1 build-essential cmake"
apt-get install -y --no-install-recommends $DEEPEP_SYSTEM_DEPS || {
echo "Warning: apt-get install failed, checking if required packages are available..."
for pkg in $DEEPEP_SYSTEM_DEPS; do
if ! dpkg -l "$pkg" 2>/dev/null | grep -q "^ii"; then
echo "ERROR: Required package $pkg is not installed and apt-get failed"
exit 1
fi
done
echo "All required packages are already installed, continuing..."
}
# Install GDRCopy
rm -rf /opt/gdrcopy && mkdir -p /opt/gdrcopy
cd /opt/gdrcopy
git clone https://github.com/NVIDIA/gdrcopy.git .
git checkout v2.5.1
apt-get update || true # May fail due to unrelated broken packages
GDRCOPY_DEPS_1="nvidia-dkms-580"
GDRCOPY_DEPS_2="build-essential devscripts debhelper fakeroot pkg-config dkms"
GDRCOPY_DEPS_3="check libsubunit0 libsubunit-dev python3-venv"
for deps_group in "$GDRCOPY_DEPS_1" "$GDRCOPY_DEPS_2" "$GDRCOPY_DEPS_3"; do
apt-get install -y --no-install-recommends $deps_group || {
echo "Warning: apt-get install failed for '$deps_group', checking if packages are available..."
for pkg in $deps_group; do
if ! dpkg -l "$pkg" 2>/dev/null | grep -q "^ii"; then
echo "ERROR: Required package $pkg is not installed and apt-get failed"
exit 1
fi
done
echo "All required packages from '$deps_group' are already installed, continuing..."
}
done
cd packages
CUDA=/usr/local/cuda ./build-deb-packages.sh
dpkg -i gdrdrv-dkms_*.deb
dpkg -i libgdrapi_*.deb
dpkg -i gdrcopy-tests_*.deb
dpkg -i gdrcopy_*.deb
# Set up library paths based on architecture
LIB_PATH="/usr/lib/$ARCH-linux-gnu"
if [ ! -e "$LIB_PATH/libmlx5.so" ]; then
ln -s $LIB_PATH/libmlx5.so.1 $LIB_PATH/libmlx5.so
fi
apt-get update || true
apt-get install -y --no-install-recommends libfabric-dev || {
if ! dpkg -l libfabric-dev 2>/dev/null | grep -q "^ii"; then
echo "ERROR: Required package libfabric-dev is not installed and apt-get failed"
exit 1
fi
echo "libfabric-dev is already installed, continuing..."
}
# Install DeepEP
DEEPEP_DIR=/root/.cache/deepep
rm -rf ${DEEPEP_DIR}
if [ "$GRACE_BLACKWELL" = "1" ]; then
GRACE_BLACKWELL_DEEPEP_BRANCH=hybrid-ep
git clone https://github.com/deepseek-ai/DeepEP.git -b ${GRACE_BLACKWELL_DEEPEP_BRANCH} ${DEEPEP_DIR} && \
pushd ${DEEPEP_DIR} && \
git checkout d28bd676c2120573c9f1425f0c16c39faa4117e6 && \
sed -i 's/#define NUM_CPU_TIMEOUT_SECS 100/#define NUM_CPU_TIMEOUT_SECS 1000/' csrc/kernels/configs.cuh && \
popd
else
git clone https://github.com/deepseek-ai/DeepEP.git ${DEEPEP_DIR} && \
pushd ${DEEPEP_DIR} && \
git checkout 9af0e0d0e74f3577af1979c9b9e1ac2cad0104ee && \
popd
fi
cd ${DEEPEP_DIR}
if [ "$GRACE_BLACKWELL" = "1" ]; then
# Resolve the toolkit CUDA version. Preference order:
# 1. $NVCC_VER inherited from the sourced ci_install_dependency.sh
# (both scripts agree on the detected value, no re-detection cost).
# 2. Local `nvcc --version` (authoritative — container toolkit).
# 3. `nvidia-smi` (host driver; last resort).
if [ -n "${NVCC_VER:-}" ]; then
CUDA_VERSION="$NVCC_VER"
elif command -v nvcc >/dev/null 2>&1; then
CUDA_VERSION=$(nvcc --version | grep -oP 'release \K[0-9]+\.[0-9]+')
else
CUDA_VERSION=$(nvidia-smi | grep "CUDA Version" | head -n1 | awk '{print $9}' || true)
fi
if [ -z "${CUDA_VERSION:-}" ]; then
echo "FATAL: could not determine CUDA toolkit version (NVCC_VER unset, nvcc missing, nvidia-smi empty)"
exit 1
fi
if [ "$CUDA_VERSION" = "12.8" ]; then
CHOSEN_TORCH_CUDA_ARCH_LIST='10.0'
elif awk -v ver="$CUDA_VERSION" 'BEGIN {exit !(ver > 12.8)}'; then
# CUDA > 12.8 supports sm_103 (Blackwell)
CHOSEN_TORCH_CUDA_ARCH_LIST='10.0;10.3'
else
echo "Unsupported CUDA version for Grace Blackwell: $CUDA_VERSION" && exit 1
fi && \
if [ "${CUDA_VERSION%%.*}" = "13" ]; then \
sed -i "/^ include_dirs = \['csrc\/'\]/a\ include_dirs.append('${CUDA_HOME}/include/cccl')" setup.py; \
fi
TORCH_CUDA_ARCH_LIST="${CHOSEN_TORCH_CUDA_ARCH_LIST}" ${PIP_CMD:-pip} install --no-build-isolation . ${PIP_INSTALL_SUFFIX:-}
else
# CUDA 13.0 puts CCCL headers in /usr/local/cuda/include/cccl/ but nvshmem
# includes them as <cuda/__cccl_config> expecting /usr/local/cuda/include/cuda/.
# Add the cccl path to setup.py include_dirs so the compiler finds them.
NVCC_MAJOR=$(nvcc --version 2>/dev/null | grep -oP 'release \K[0-9]+' || echo "0")
if [ "$NVCC_MAJOR" = "13" ]; then
sed -i "/^ include_dirs = \['csrc\/'\]/a\ include_dirs.append('${CUDA_HOME:-/usr/local/cuda}/include/cccl')" setup.py
fi
# Build for both Hopper (sm_90) and Blackwell (sm_100) so the same wheel
# runs on H200 and B200 runners. Mirrors the CUDA-version-keyed list in
# docker/Dockerfile's DeepEP build stage.
if [ -n "${NVCC_VER:-}" ]; then
CUDA_VERSION="$NVCC_VER"
elif command -v nvcc >/dev/null 2>&1; then
CUDA_VERSION=$(nvcc --version | grep -oP 'release \K[0-9]+\.[0-9]+')
else
CUDA_VERSION=$(nvidia-smi | grep "CUDA Version" | head -n1 | awk '{print $9}' || true)
fi
if [ -z "${CUDA_VERSION:-}" ]; then
echo "FATAL: could not determine CUDA toolkit version (NVCC_VER unset, nvcc missing, nvidia-smi empty)"
exit 1
fi
if [ "$CUDA_VERSION" = "12.8" ]; then
CHOSEN_TORCH_CUDA_ARCH_LIST='9.0;10.0'
elif awk -v ver="$CUDA_VERSION" 'BEGIN {exit !(ver > 12.8)}'; then
# CUDA > 12.8 supports sm_103 (Blackwell)
CHOSEN_TORCH_CUDA_ARCH_LIST='9.0;10.0;10.3'
else
CHOSEN_TORCH_CUDA_ARCH_LIST='9.0'
fi
TORCH_CUDA_ARCH_LIST="${CHOSEN_TORCH_CUDA_ARCH_LIST}" python3 setup.py install
fi
scripts/ci/cuda/ci_install_dependency.sh
+97
View File
@@ -136,7 +136,9 @@ cleanup_stale_shm() {
install_apt_packages() {
CI_APT_PACKAGES=(
python3 python3-pip python3-venv python3-dev git libnuma-dev libssl-dev pkg-config
build-essential cmake rdma-core infiniband-diags perftest libibumad3
libibverbs-dev libibverbs1 ibverbs-providers ibverbs-utils
libfabric-dev libnl-3-200 libnl-route-3-200 librdmacm1
ffmpeg libavcodec-dev libavformat-dev libavutil-dev libswscale-dev
)
@@ -164,6 +166,68 @@ install_apt_packages() {
mark_step_done "${FUNCNAME[0]}"
}
install_gdrcopy() {
# DeepEP tests only run on 4+ GPU hosts. Keep GDRCopy in the shared CUDA
# bootstrap while avoiding a DKMS/package build on the 1- and 2-GPU jobs.
local gpu_count=0
if command -v nvidia-smi >/dev/null 2>&1; then
gpu_count=$(
(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null || true) |
awk 'NF {count++} END {print count + 0}'
)
fi
if [ "${gpu_count}" -lt 4 ]; then
echo "Skipping GDRCopy install on ${gpu_count}-GPU runner"
mark_step_done "${FUNCNAME[0]}"
return
fi
if ldconfig -p 2>/dev/null | grep 'libgdrapi\.so' >/dev/null; then
echo "GDRCopy userspace library is already installed"
mark_step_done "${FUNCNAME[0]}"
return
fi
local gdrcopy_root=/opt/gdrcopy
local gdrcopy_version=2.5.1
local -a gdrcopy_packages=(
nvidia-dkms-580 devscripts debhelper fakeroot dkms
check libsubunit0 libsubunit-dev python3-venv
)
apt-get update || true
apt-get install -y --no-install-recommends "${gdrcopy_packages[@]}" || {
echo "Warning: apt-get failed while installing GDRCopy build dependencies; checking installed packages"
local package
for package in "${gdrcopy_packages[@]}"; do
if ! dpkg -l "${package}" 2>/dev/null | grep -q '^ii'; then
echo "ERROR: Required GDRCopy package ${package} is unavailable"
exit 1
fi
done
}
rm -rf "${gdrcopy_root}"
git clone --branch "v${gdrcopy_version}" --depth 1 \
https://github.com/NVIDIA/gdrcopy.git "${gdrcopy_root}"
(
cd "${gdrcopy_root}/packages"
CUDA=/usr/local/cuda ./build-deb-packages.sh
dpkg -i gdrdrv-dkms_*.deb
dpkg -i libgdrapi_*.deb
dpkg -i gdrcopy-tests_*.deb
dpkg -i gdrcopy_*.deb
)
local lib_path="/usr/lib/${ARCH}-linux-gnu"
if [ ! -e "${lib_path}/libmlx5.so" ] && [ -e "${lib_path}/libmlx5.so.1" ]; then
ln -s "${lib_path}/libmlx5.so.1" "${lib_path}/libmlx5.so"
fi
ldconfig
mark_step_done "${FUNCNAME[0]}"
}
clean_site_packages() {
# Clear torch compilation cache from every location it can be in; sglang
# is not installed yet, so it cannot be asked which one is in use.
@@ -260,6 +324,11 @@ setup_pip_toolchain() {
PIP_UNINSTALL_CMD="uv pip uninstall"
PIP_UNINSTALL_SUFFIX=""
# Remove both the legacy source distribution and the SGLang wheel before
# resolving the pyproject pin. They own the same deep_ep module files, so
# leaving either installed can make pip preserve a mixed installation.
$PIP_UNINSTALL_CMD deep-ep sgl-deep-ep $PIP_UNINSTALL_SUFFIX || true
# sglang-kernel stays: install_sglang_kernel version-gates and reinstalls it.
$PIP_UNINSTALL_CMD sgl-kernel sglang sgl-fa4 flash-attn-4 $PIP_UNINSTALL_SUFFIX || true
@@ -368,6 +437,30 @@ install_pytorch_stack() {
mark_step_done "${FUNCNAME[0]}"
}
install_cuda12_deepep_wheel() {
if [ "$CU_MAJOR" = "13" ]; then
echo "CUDA 13 uses the public sgl-deep-ep wheel declared in python/pyproject.toml"
mark_step_done "${FUNCNAME[0]}"
return
fi
local version
version=$(grep -Po -m1 '"sgl-deep-ep==\K[^"]+' python/pyproject.toml || true)
if [ -z "$version" ]; then
echo "ERROR: python/pyproject.toml must pin sgl-deep-ep"
exit 1
fi
# CUDA 12 wheels intentionally live only on the SGLang wheel index. Their
# local version satisfies the public-version pyproject pin, so the later
# editable SGLang install keeps this CUDA-matched wheel.
$PIP_CMD install "sgl-deep-ep==${version}+${CU_VERSION}" \
--index-url "https://docs.sglang.ai/whl/${CU_VERSION}/" \
--force-reinstall --no-deps $PIP_INSTALL_SUFFIX
mark_step_done "${FUNCNAME[0]}"
}
require_prebuilt_rust_exts() {
# Stages whose download succeeded set this to none. Runs before
# setup_pip_toolchain uninstalls sglang, so clearing it here still reaches
@@ -708,6 +801,8 @@ verify_imports() {
SGLANG_EXPECTED_INIT="${REPO_ROOT}/python/sglang/__init__.py" python3 -c '
import torch
print(torch.version.cuda)
import deep_ep
print(f"deep_ep loads from {deep_ep.__file__}")
import cutlass
import cutlass.cute
@@ -752,6 +847,7 @@ main() {
kill_existing_processes
cleanup_stale_shm
install_apt_packages
install_gdrcopy
clean_site_packages
setup_cargo_cache
require_prebuilt_rust_exts
@@ -759,6 +855,7 @@ main() {
remove_stale_cuda12_nvidia_wheels
uninstall_stale_flashinfer
install_pytorch_stack
install_cuda12_deepep_wheel
install_sglang
# Diffusion B200 CI imports torch inside install_sglang_kernel after removing
# stale CUDA 12 NVIDIA wheels, so opt into one early LD_LIBRARY_PATH refresh.