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

[AMD] Update dsv3.2 AMD GPU docs and unify ROCm TileLang build (#17783)

Browse Source 
Co-authored-by: wufann <715544327@qq.com>
This commit is contained in:
Hubert Lu
2026-01-26 21:10:32 -08:00
committed by GitHub
co-authored by wufann
parent a723d1c5ef
commit df42f4d386
2 changed files with 81 additions and 88 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docker/rocm.Dockerfile
+71 -87
View File
@@ -1,8 +1,7 @@
# Usage (to build SGLang ROCm docker image): # Usage (to build SGLang ROCm docker image):
# docker build --build-arg SGL_BRANCH=v0.5.6.post2 --build-arg GPU_ARCH=gfx942 -t v0.5.6.post2-rocm630-mi30x -f rocm.Dockerfile . # docker build --build-arg SGL_BRANCH=v0.5.8 --build-arg GPU_ARCH=gfx942 -t v0.5.8-rocm630-mi30x -f rocm.Dockerfile .
# docker build --build-arg SGL_BRANCH=v0.5.6.post2 --build-arg GPU_ARCH=gfx942-rocm700 -t v0.5.6.post2-rocm700-mi30x -f rocm.Dockerfile . # docker build --build-arg SGL_BRANCH=v0.5.8 --build-arg GPU_ARCH=gfx942-rocm700 -t v0.5.8-rocm700-mi30x -f rocm.Dockerfile .
# docker build --build-arg SGL_BRANCH=v0.5.6.post2 --build-arg GPU_ARCH=gfx950 -t v0.5.6.post2-rocm700-mi35x -f rocm.Dockerfile . # docker build --build-arg SGL_BRANCH=v0.5.8 --build-arg GPU_ARCH=gfx950 -t v0.5.8-rocm700-mi35x -f rocm.Dockerfile .
# Default base images # Default base images
ARG BASE_IMAGE_942="rocm/sgl-dev:vllm20250114" ARG BASE_IMAGE_942="rocm/sgl-dev:vllm20250114"
@@ -65,13 +64,8 @@ ARG LLVM_COMMIT="6520ace8227ffe2728148d5f3b9872a870b0a560"
ARG MOONCAKE_REPO="https://github.com/kvcache-ai/Mooncake.git" ARG MOONCAKE_REPO="https://github.com/kvcache-ai/Mooncake.git"
ARG MOONCAKE_COMMIT="b6a841dc78c707ec655a563453277d969fb8f38d" ARG MOONCAKE_COMMIT="b6a841dc78c707ec655a563453277d969fb8f38d"
ARG TILELANG_REPO="https://github.com/HaiShaw/tilelang.git" ARG TILELANG_REPO="https://github.com/tile-ai/tilelang.git"
ARG TILELANG_BRANCH="dsv32-mi35x" ARG TILELANG_COMMIT="ebf4a7cb8881432165ae8760e99d209d905c704a"
ARG TILELANG_COMMIT="ae938cf885743f165a19656d1122ad42bb0e30b8"
ARG TILELANG_GFX942_REPO="https://github.com/tile-ai/tilelang.git"
ARG TILELANG_GFX942_BRANCH="main"
ARG TILELANG_GFX942_COMMIT="2d8d3676eda18bd3d8e6fa783399ff96d3cd4ded"
ARG FHT_REPO="https://github.com/jeffdaily/fast-hadamard-transform.git" ARG FHT_REPO="https://github.com/jeffdaily/fast-hadamard-transform.git"
ARG FHT_BRANCH="rocm" ARG FHT_BRANCH="rocm"
@@ -204,11 +198,12 @@ RUN find /sgl-workspace/sglang/python/sglang/srt/layers/quantization/configs/ \
ENV PATH="/root/.cargo/bin:${PATH}" ENV PATH="/root/.cargo/bin:${PATH}"
RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y \ RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y \
&& rustc --version && cargo --version && rustc --version && cargo --version
ENV CARGO_BUILD_JOBS=4
# Build and install sgl-model-gateway # Build and install sgl-model-gateway
RUN python3 -m pip install --no-cache-dir setuptools-rust \ RUN python3 -m pip install --no-cache-dir setuptools-rust \
&& cd /sgl-workspace/sglang/sgl-model-gateway/bindings/python \ && cd /sgl-workspace/sglang/sgl-model-gateway/bindings/python \
&& cargo build --release \ && /bin/bash -lc 'ulimit -n 8192 && cargo build --release' \
&& python3 -m pip install --no-cache-dir . \ && python3 -m pip install --no-cache-dir . \
&& rm -rf /root/.cache && rm -rf /root/.cache
@@ -219,82 +214,71 @@ ENV LIBGL_ALWAYS_INDIRECT=1
RUN echo "LC_ALL=en_US.UTF-8" >> /etc/environment RUN echo "LC_ALL=en_US.UTF-8" >> /etc/environment
RUN /bin/bash -lc 'set -euo pipefail; \ RUN /bin/bash -lc 'set -euo pipefail; \
# Build TileLang for gfx950 and gfx942-rocm700
if [ "${GPU_ARCH:-}" != "gfx950" ] && [ "${GPU_ARCH:-}" != "gfx942-rocm700" ]; then \
echo "[TileLang] Skipping (GPU_ARCH=${GPU_ARCH:-unset})"; \
exit 0; \
fi; \
echo "[TileLang] Building TileLang for ${GPU_ARCH}"; \ echo "[TileLang] Building TileLang for ${GPU_ARCH}"; \
if [ "$GPU_ARCH" = "gfx950" ]; then \ # System dependencies (NO llvm-dev to avoid llvm-config-16 shadowing)
\ apt-get update && apt-get install -y --no-install-recommends \
# System dependencies (NO llvm-dev to avoid llvm-config-16 shadowing) build-essential git wget curl ca-certificates gnupg \
apt-get update && apt-get install -y --no-install-recommends \ libgtest-dev libgmock-dev \
build-essential git wget curl ca-certificates gnupg \ libprotobuf-dev protobuf-compiler libgflags-dev libsqlite3-dev \
libgtest-dev libgmock-dev \ python3 python3-dev python3-setuptools python3-pip python3-apt \
libprotobuf-dev protobuf-compiler libgflags-dev libsqlite3-dev \ gcc libtinfo-dev zlib1g-dev libedit-dev libxml2-dev \
python3 python3-dev python3-setuptools python3-pip \ cmake ninja-build pkg-config libstdc++6 software-properties-common \
gcc libtinfo-dev zlib1g-dev libedit-dev libxml2-dev \ && rm -rf /var/lib/apt/lists/*; \
cmake ninja-build pkg-config libstdc++6 \ \
&& rm -rf /var/lib/apt/lists/*; \ # Prefer the container venv
\ VENV_PY="/opt/venv/bin/python"; \
# Build GoogleTest static libs (Ubuntu package ships sources only) VENV_PIP="/opt/venv/bin/pip"; \
cmake -S /usr/src/googletest -B /tmp/build-gtest -DBUILD_GTEST=ON -DBUILD_GMOCK=ON -DCMAKE_BUILD_TYPE=Release && \ if [ ! -x "$VENV_PY" ]; then VENV_PY="python3"; fi; \
cmake --build /tmp/build-gtest -j"$(nproc)" && \ if [ ! -x "$VENV_PIP" ]; then VENV_PIP="pip3"; fi; \
cp -v /tmp/build-gtest/lib/*.a /usr/lib/x86_64-linux-gnu/ && \ \
rm -rf /tmp/build-gtest; \ # Build GoogleTest static libs (Ubuntu package ships sources only)
\ cmake -S /usr/src/googletest -B /tmp/build-gtest -DBUILD_GTEST=ON -DBUILD_GMOCK=ON -DCMAKE_BUILD_TYPE=Release && \
# Keep setuptools < 80 (compat with base image) cmake --build /tmp/build-gtest -j"$(nproc)" && \
python3 -m pip install --upgrade "setuptools>=77.0.3,<80" wheel cmake ninja && \ cp -v /tmp/build-gtest/lib/*.a /usr/lib/x86_64-linux-gnu/ && \
python3 -m pip cache purge || true; \ rm -rf /tmp/build-gtest; \
\ \
# Locate ROCm llvm-config; fallback to installing LLVM 18 if missing # Keep setuptools < 80 (compat with base image)
LLVM_CONFIG_PATH=""; \ "$VENV_PIP" install --upgrade "setuptools>=77.0.3,<80" wheel cmake ninja scikit-build-core && \
for p in /opt/rocm/llvm/bin/llvm-config /opt/rocm/llvm-*/bin/llvm-config /opt/rocm-*/llvm*/bin/llvm-config; do \ "$VENV_PIP" cache purge || true; \
if [ -x "$p" ]; then LLVM_CONFIG_PATH="$p"; break; fi; \ \
done; \ # Locate ROCm llvm-config; fallback to installing LLVM 18 if missing
if [ -z "$LLVM_CONFIG_PATH" ]; then \ LLVM_CONFIG_PATH=""; \
echo "[TileLang] ROCm llvm-config not found; installing LLVM 18..."; \ for p in /opt/rocm/llvm/bin/llvm-config /opt/rocm/llvm-*/bin/llvm-config /opt/rocm-*/llvm*/bin/llvm-config; do \
curl -fsSL https://apt.llvm.org/llvm.sh -o /tmp/llvm.sh; \ if [ -x "$p" ]; then LLVM_CONFIG_PATH="$p"; break; fi; \
chmod +x /tmp/llvm.sh; \ done; \
/tmp/llvm.sh 18; \ if [ -z "$LLVM_CONFIG_PATH" ]; then \
LLVM_CONFIG_PATH="$(command -v llvm-config-18)"; \ echo "[TileLang] ROCm llvm-config not found; installing LLVM 18..."; \
if [ -z "$LLVM_CONFIG_PATH" ]; then echo "ERROR: llvm-config-18 not found after install"; exit 1; fi; \ curl -fsSL https://apt.llvm.org/llvm-snapshot.gpg.key | gpg --dearmor -o /etc/apt/keyrings/llvm.gpg; \
fi; \ echo "deb [signed-by=/etc/apt/keyrings/llvm.gpg] http://apt.llvm.org/jammy/ llvm-toolchain-jammy-18 main" > /etc/apt/sources.list.d/llvm.list; \
echo "[TileLang] Using LLVM_CONFIG at: $LLVM_CONFIG_PATH"; \ apt-get update; \
export PATH="$(dirname "$LLVM_CONFIG_PATH"):/usr/local/bin:${PATH}"; \ apt-get install -y --no-install-recommends llvm-18; \
export LLVM_CONFIG="$LLVM_CONFIG_PATH"; \ rm -rf /var/lib/apt/lists/*; \
\ LLVM_CONFIG_PATH="$(command -v llvm-config-18)"; \
# Optional shim for tools that expect llvm-config-16 if [ -z "$LLVM_CONFIG_PATH" ]; then echo "ERROR: llvm-config-18 not found after install"; exit 1; fi; \
mkdir -p /usr/local/bin && \ fi; \
printf "#!/usr/bin/env bash\nexec \"%s\" \"\$@\"\n" "$LLVM_CONFIG_PATH" > /usr/local/bin/llvm-config-16 && \ echo "[TileLang] Using LLVM_CONFIG at: $LLVM_CONFIG_PATH"; \
chmod +x /usr/local/bin/llvm-config-16; \ export PATH="$(dirname "$LLVM_CONFIG_PATH"):/usr/local/bin:${PATH}"; \
\ export LLVM_CONFIG="$LLVM_CONFIG_PATH"; \
# TVM Python bits need Cython \
python3 -m pip install --no-cache-dir "cython>=0.29.36,<3.0"; \ # Optional shim for tools that expect llvm-config-16
\ mkdir -p /usr/local/bin && \
# Clone + pin TileLang (bundled TVM), then build printf "#!/usr/bin/env bash\nexec \"%s\" \"\$@\"\n" "$LLVM_CONFIG_PATH" > /usr/local/bin/llvm-config-16 && \
git clone --recursive --branch "${TILELANG_BRANCH}" "${TILELANG_REPO}" /opt/tilelang && \ chmod +x /usr/local/bin/llvm-config-16; \
cd /opt/tilelang && \ \
git fetch --depth=1 origin "${TILELANG_COMMIT}" || true && \ # TVM Python bits need Cython + z3 before configure
git checkout -f "${TILELANG_COMMIT}" && \ "$VENV_PIP" install --no-cache-dir "cython>=0.29.36,<3.0" "apache-tvm-ffi>=0.1.6" "z3-solver>=4.13.0"; \
git submodule update --init --recursive && \ \
export CMAKE_ARGS="-DLLVM_CONFIG=${LLVM_CONFIG} ${CMAKE_ARGS:-}" && \ # Clone + pin TileLang (bundled TVM), then build
bash ./install_rocm.sh; \ git clone --recursive "${TILELANG_REPO}" /opt/tilelang && \
else \ cd /opt/tilelang && \
# Build GoogleTest static libs (Ubuntu package ships sources only) git fetch --depth=1 origin "${TILELANG_COMMIT}" || true && \
apt-get install -y libgtest-dev libgmock-dev && \ git checkout -f "${TILELANG_COMMIT}" && \
cmake -S /usr/src/googletest -B /tmp/build-gtest -DBUILD_GTEST=ON -DBUILD_GMOCK=ON -DCMAKE_BUILD_TYPE=Release && \ git submodule update --init --recursive && \
cmake --build /tmp/build-gtest -j && \ export CMAKE_ARGS="-DUSE_CUDA=OFF -DUSE_ROCM=ON -DROCM_PATH=/opt/rocm -DLLVM_CONFIG=${LLVM_CONFIG} -DSKBUILD_SABI_VERSION= ${CMAKE_ARGS:-}" && \
cp -v /tmp/build-gtest/lib/*.a /usr/lib/x86_64-linux-gnu/ && \ "$VENV_PIP" install -e . -v --no-build-isolation --no-deps; \
rm -rf /tmp/build-gtest; \ if [ -f pyproject.toml ]; then sed -i "/^[[:space:]]*\"torch/d" pyproject.toml || true; fi; \
# Build TileLang for gfx942-rocm700 "$VENV_PIP" cache purge || true; \
git clone --branch "${TILELANG_GFX942_BRANCH}" "${TILELANG_GFX942_REPO}" /opt/tilelang && \ "$VENV_PY" -c "import tilelang; print(tilelang.__version__)"'
cd /opt/tilelang && \
git checkout -f "${TILELANG_GFX942_COMMIT}" && \
git submodule update --init --recursive && \
sed -i "/^[[:space:]]*\"torch/d" pyproject.toml && \
USE_ROCM=1 USE_CUDA=0 pip install -e . -v ; \
fi'
# ----------------------- # -----------------------
# Hadamard-transform (HIP build) # Hadamard-transform (HIP build)
docs/basic_usage/deepseek_v32.md
+10 -1
View File
@@ -16,7 +16,13 @@ Note: This document is originally written for the usage of [DeepSeek-V3.2-Exp](h
docker pull lmsysorg/sglang:latest docker pull lmsysorg/sglang:latest
# MI350/MI355 # MI350/MI355
docker pull lmsysorg/sglang:dsv32-rocm docker pull lmsysorg/sglang:v0.5.8-rocm700-mi35x
# MI300
# v0.5.8-rocm700-mi30x does not include PR #17504. Prefer the newest MI30x ROCm
# image tag from Docker Hub when available, or build from source (below).
docker pull lmsysorg/sglang:v0.5.8-rocm700-mi30x
# NPUs # NPUs
docker pull lmsysorg/sglang:dsv32-a2 docker pull lmsysorg/sglang:dsv32-a2
@@ -45,6 +51,9 @@ python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --ep
# Launch with Pure TP # Launch with Pure TP
python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8
# Launch with TP on MI30x/MI35x
python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --nsa-prefill-backend tilelang --nsa-decode-backend tilelang
``` ```
### Configuration Tips ### Configuration Tips