@tduncan / Nvidia-Optimization-Stack-VLLM
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
xpu 286 lines 11.6 KB
Raw
1 ######################### RUST BUILD IMAGE #########################
2 # Build the Rust frontend (`vllm-rs`) in a dedicated stage so the main image
3 # doesn't need the rust toolchain. Runs in parallel with vllm-base.
4 FROM ubuntu:22.04 AS rust-build-cache
5 ENV DEBIAN_FRONTEND=noninteractive
6
7 RUN apt-get update -y \
8 && apt-get install -y --no-install-recommends \
9 ca-certificates curl git build-essential python3 python3-pip \
10 && rm -rf /var/lib/apt/lists/*
11
12 WORKDIR /workspace
13
14 COPY requirements/build/rust.txt requirements/build/rust.txt
15 RUN python3 -m pip install --no-cache-dir -r requirements/build/rust.txt
16
17 # Copy only the Rust build inputs; tools/build_rust.sh publishes artifacts needed
18 # by the wheel build stage.
19 COPY rust rust
20 COPY rust-toolchain.toml rust-toolchain.toml
21 COPY tools/build_rust.py tools/build_rust.py
22 COPY tools/build_rust.sh tools/build_rust.sh
23
24 # Cap cargo parallelism to avoid exhausting the CI host's open-file limit
25 # (rustc spawns enough concurrent processes to hit RLIMIT_NOFILE otherwise).
26 ENV CARGO_BUILD_JOBS=4
27
28 # Build with a stable version so this target snapshot is keyed only by Rust
29 # inputs. The child stage below relinks the crates that embed the exact version.
30 RUN --mount=type=cache,target=/root/.cargo/registry,sharing=locked \
31 --mount=type=cache,target=/root/.cargo/git,sharing=locked \
32 SETUPTOOLS_SCM_PRETEND_VERSION="0.0.0+docker.cache" bash tools/build_rust.sh
33
34 # Relink with the exact Git-derived package version.
35 FROM rust-build-cache AS rust-build
36
37 RUN --mount=type=cache,target=/root/.cargo/registry,sharing=locked \
38 --mount=type=cache,target=/root/.cargo/git,sharing=locked \
39 --mount=type=bind,source=.git,target=.git \
40 SETUPTOOLS_SCM_PRETEND_METADATA="{dirty=false}" bash tools/build_rust.sh
41
42 FROM ubuntu:24.04 AS vllm-base
43
44 ENV DEBIAN_FRONTEND=noninteractive
45
46 WORKDIR /workspace/
47
48 ARG PYTHON_VERSION=3.12
49 ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/xpu"
50
51 RUN apt-get update -y && \
52 apt-get install -y --no-install-recommends \
53 build-essential \
54 curl \
55 ffmpeg \
56 git \
57 gpg \
58 libsndfile1 \
59 libsm6 \
60 libxext6 \
61 libgl1 \
62 lsb-release \
63 libaio-dev \
64 numactl \
65 wget \
66 vim \
67 ca-certificates \
68 python3.12 \
69 python3.12-dev \
70 python3-pip && \
71 rm -rf /var/lib/apt/lists/*
72
73 # Install UMD
74 RUN mkdir neo && \
75 cd neo && \
76 wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.38.2/intel-igc-core-2_2.38.2+22051_amd64.deb && \
77 wget https://github.com/intel/intel-graphics-compiler/releases/download/v2.38.2/intel-igc-opencl-2_2.38.2+22051_amd64.deb && \
78 wget https://github.com/intel/compute-runtime/releases/download/26.27.39122.11/intel-ocloc_26.27.39122.11-0_amd64.deb && \
79 wget https://github.com/intel/compute-runtime/releases/download/26.27.39122.11/intel-opencl-icd_26.27.39122.11-0_amd64.deb && \
80 wget https://github.com/intel/compute-runtime/releases/download/26.27.39122.11/libigdgmm12_22.10.0_amd64.deb && \
81 wget https://github.com/intel/compute-runtime/releases/download/26.27.39122.11/libze-intel-gpu1_26.27.39122.11-0_amd64.deb && \
82 wget https://github.com/oneapi-src/level-zero/releases/download/v1.32.0/libze1_1.32.0+u24.04_amd64.deb && \
83 wget https://github.com/oneapi-src/level-zero/releases/download/v1.32.0/libze-dev_1.32.0+u24.04_amd64.deb && \
84 dpkg -i *.deb && \
85 cd .. && \
86 rm -rf neo
87
88 # Install XPU-Manager payloads without registering conflicting libze package
89 # dependencies in dpkg.
90 # Note that ocl-icd-opencl-dev is required by DLE upgrade but xpu-smi, should
91 # can be removed in future.
92 RUN apt-get update -y && \
93 apt-get install -y --no-install-recommends software-properties-common && \
94 add-apt-repository -y ppa:kobuk-team/intel-graphics && \
95 apt-get update -y && \
96 apt-get install -y --no-install-recommends libhwloc15 libpciaccess0 libigsc1 intel-metrics-library && \
97 apt-get install -y ocl-icd-opencl-dev && \
98 mkdir xpu-m && \
99 cd xpu-m && \
100 wget https://github.com/intel/xpumanager/releases/download/v2.1.0/xpu-smi_2.1.0+26.33.6468cec-1.24.04_amd64.deb && \
101 apt install -y ./xpu-smi_2.1.0+26.33.6468cec-1.24.04_amd64.deb --install-suggests && \
102 cd .. && \
103 rm -rf xpu-m && \
104 add-apt-repository --remove -y ppa:kobuk-team/intel-graphics && \
105 apt-get update -y
106
107 ENV PATH="/root/.local/bin:$PATH"
108 ENV VIRTUAL_ENV="/opt/venv"
109 ENV UV_PYTHON_INSTALL_DIR=/opt/uv/python
110 RUN curl -LsSf https://astral.sh/uv/install.sh | sh \
111 && uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
112 ENV PATH="$VIRTUAL_ENV/bin:$PATH"
113
114 SHELL ["bash", "-c"]
115 CMD ["bash", "-c", "source /root/.bashrc && exec bash"]
116
117 WORKDIR /workspace/vllm
118
119 ENV UV_HTTP_TIMEOUT=500
120
121 # Configure package index for XPU
122 ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
123 ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
124 ENV UV_INDEX_STRATEGY="unsafe-best-match"
125 ENV UV_LINK_MODE="copy"
126
127 RUN --mount=type=cache,target=/root/.cache/uv \
128 --mount=type=bind,src=requirements/common.txt,target=/workspace/vllm/requirements/common.txt \
129 --mount=type=bind,src=requirements/xpu.txt,target=/workspace/vllm/requirements/xpu.txt \
130 uv pip install --upgrade pip
131
132
133 ENV LD_LIBRARY_PATH=/opt/venv/lib:/usr/local/lib
134 CMD ["/bin/bash"]
135
136 ######################### UCX + NIXL BUILD STAGE #########################
137 # Build UCX and NIXL in a dedicated stage so compiler/autotools layers are
138 # never included in the final runtime image (mirrors ROCm's build_rixl stage).
139 ######################### WHEEL BUILD STAGE #########################
140 # Produces a standalone dist/*.whl artifact (used by the release pipeline's
141 # "Build wheel - XPU" step), independent of the runtime image below.
142 FROM vllm-base AS vllm-build
143
144 ARG GIT_REPO_CHECK=0
145
146 RUN --mount=type=cache,target=/root/.cache/uv \
147 --mount=type=bind,src=requirements/common.txt,target=/workspace/vllm/requirements/common.txt \
148 --mount=type=bind,src=requirements/xpu.txt,target=/workspace/vllm/requirements/xpu.txt \
149 uv pip install grpcio-tools protobuf nanobind && \
150 uv pip install -r /workspace/vllm/requirements/xpu.txt
151
152 # Keep source-dependent layers near the end so frequent code-only changes
153 # don't invalidate heavy dependency layers.
154 COPY . .
155
156 # Drop the pre-built Rust artifacts into the source tree. setup.py detects
157 # them and ships them as-is, skipping the local Rust build.
158 COPY --from=rust-build /workspace/vllm/vllm-rs vllm/vllm-rs
159 COPY --from=rust-build /workspace/vllm/_rust_*.so vllm/
160
161 RUN --mount=type=bind,source=.git,target=.git \
162 if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi
163
164 RUN --mount=type=cache,target=/root/.cache/uv \
165 --mount=type=bind,source=.git,target=.git \
166 VLLM_TARGET_DEVICE=xpu python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38
167
168 FROM vllm-base AS ucx-nixl-build
169
170 # Keep UCX_VERSION aligned with the UCX that NIXL itself builds against for this
171 # NIXL_VERSION (nixl contrib/Dockerfile.manylinux: ARG UCX_REF), so XPU runs the
172 # same UCX commit as the released nixl-cu* wheels rather than an untested pairing.
173 ARG UCX_VERSION=v1.21.x
174 ARG NIXL_VERSION=v1.3.2
175
176 # Build-time only: compiler, autotools, and verbs dev headers
177 RUN apt-get update -y && apt-get install -y --no-install-recommends \
178 build-essential \
179 autoconf \
180 automake \
181 libtool \
182 pkg-config \
183 libibverbs-dev \
184 librdmacm-dev \
185 && rm -rf /var/lib/apt/lists/*
186
187 # Build UCX and produce a NIXL wheel so the final image needs no compiler.
188 # patchelf (installed via uv) is used by the NIXL wheel build to rewrite
189 # RPATH entries, making the wheel portable across stages.
190 RUN --mount=type=cache,target=/root/.cache/uv \
191 git clone --depth 1 --branch "${UCX_VERSION}" https://github.com/openucx/ucx /tmp/ucx_source && \
192 cd /tmp/ucx_source && \
193 bash autogen.sh && \
194 ./configure --prefix=/opt/ucx --with-ze=yes --enable-examples --enable-mt && \
195 make CFLAGS="-Wno-error=incompatible-pointer-types" -j"$(nproc)" && make install && \
196 git clone --depth 1 --branch "${NIXL_VERSION}" https://github.com/ai-dynamo/nixl /tmp/nixl_source && \
197 cd /tmp/nixl_source && \
198 uv pip install --upgrade meson pybind11 patchelf && \
199 uv pip install -r requirements.txt && \
200 PKG_CONFIG_PATH=/opt/ucx/lib/pkgconfig \
201 LD_LIBRARY_PATH=/opt/ucx/lib \
202 python -m pip wheel --no-deps . -w /tmp/nixl_wheels/ && \
203 find /opt/ucx -type f \( -name '*.a' -o -name '*.la' \) -delete && \
204 rm -rf /opt/ucx/{include,share,etc} /opt/ucx/lib/cmake \
205 /tmp/ucx_source /tmp/nixl_source
206
207 FROM vllm-base AS vllm-openai
208 ENV UV_COMPILE_BYTECODE=1
209
210 # Must match the ucx-nixl-build stage above: ARG does not cross stages, so a stale
211 # value here installs a meta package that disagrees with the wheel built there.
212 ARG NIXL_VERSION=v1.3.2
213
214 # Copy compiled UCX runtime libraries and the pre-built NIXL wheel.
215 # No compiler or autotools are installed in this stage.
216 COPY --from=ucx-nixl-build /opt/ucx /opt/ucx
217 COPY --from=ucx-nixl-build /tmp/nixl_wheels /tmp/nixl_wheels
218
219 ENV LD_LIBRARY_PATH=/opt/ucx/lib:${LD_LIBRARY_PATH}
220
221 # Install RDMA runtime libraries (no build tools) and the pre-built NIXL wheel.
222 # Do not uninstall/reinstall large Python packages here to avoid extra layer
223 # churn; final package resolution remains in the later app install step.
224 RUN --mount=type=cache,target=/root/.cache/uv \
225 apt-get update -y && apt-get install -y --no-install-recommends \
226 rdma-core \
227 libibverbs1 \
228 librdmacm1 \
229 libibumad3 \
230 libibmad5 \
231 libmlx5-1 \
232 libmlx4-1 \
233 ibverbs-providers \
234 librdmacm1t64 \
235 && rm -rf /var/lib/apt/lists/* \
236 && uv pip install --no-deps /tmp/nixl_wheels/nixl*.whl \
237 && uv pip install --no-deps nixl==${NIXL_VERSION} \
238 && rm -rf /tmp/nixl_wheels
239
240 RUN --mount=type=cache,target=/root/.cache/uv \
241 --mount=type=bind,src=requirements/common.txt,target=/workspace/vllm/requirements/common.txt \
242 --mount=type=bind,src=requirements/xpu.txt,target=/workspace/vllm/requirements/xpu.txt \
243 --mount=type=bind,src=requirements/test/xpu.txt,target=/workspace/vllm/requirements/test/xpu.txt \
244 uv pip install grpcio-tools protobuf nanobind && \
245 uv pip install -r /workspace/vllm/requirements/xpu.txt && \
246 uv pip install --no-build-isolation -r /workspace/vllm/requirements/test/xpu.txt
247
248 # Keep source-dependent layers near the end so frequent code-only changes
249 # don't invalidate heavy dependency and UCX/NIXL layers.
250 COPY . .
251
252 # Drop the pre-built Rust artifacts into the source tree. setup.py detects
253 # them and ships them as-is, skipping the local Rust build.
254 COPY --from=rust-build /workspace/vllm/vllm-rs vllm/vllm-rs
255 COPY --from=rust-build /workspace/vllm/_rust_*.so vllm/
256
257 ARG GIT_REPO_CHECK=0
258 RUN --mount=type=bind,source=.git,target=.git \
259 if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi
260
261 ENV VLLM_TARGET_DEVICE=xpu
262 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
263
264 RUN --mount=type=cache,target=/root/.cache/uv \
265 --mount=type=bind,source=.git,target=.git \
266 uv pip install --no-build-isolation --no-deps .
267
268 RUN --mount=type=cache,target=/root/.cache/uv \
269 uv pip install -e tests/vllm_test_utils
270
271 # Replace 2026.1.2 DLE packages to 2026.1.3
272 RUN --mount=type=cache,target=/root/.cache/uv \
273 --mount=type=bind,source=.git,target=.git \
274 uv pip install \
275 "intel-cmplr-lib-rt==2026.1.1" \
276 "intel-cmplr-lib-ur==2026.1.1" \
277 "intel-cmplr-lic-rt==2026.1.1" \
278 "intel-sycl-rt==2026.1.1" \
279 "oneccl-devel==2022.1.2; platform_system == 'Linux' and platform_machine == 'x86_64'" \
280 "oneccl==2022.1.2; platform_system == 'Linux' and platform_machine == 'x86_64'" \
281 "dpcpp-cpp-rt==2026.1.1" \
282 "intel-opencl-rt==2026.1.1" \
283 "intel-openmp==2026.1.1" \
284 "intel-pti==1.1.0"
285
286 ENTRYPOINT ["vllm", "serve"]