A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 18:15 UTC
[Hardware][IBM Z] Optimize s390x Dockerfile (#28023)
Signed-off-by: Rehan Khan <Rehan.Khan7@ibm.com>
R3hankhan committed
Nov 6, 2025 at 00:55 UTC
e04492449eeb1ca945ce08b2740ea75bddd0c8a9
3 files changed
+14
-60
docker/Dockerfile.s390x
+11
-57
@@ -14,7 +14,7 @@ ENV LANG=C.UTF-8 \
14
15
# Install development utilities
16
RUN microdnf install -y \
17
- which procps findutils tar vim git gcc gcc-gfortran g++ make patch zlib-devel \
17
+ which procps findutils tar vim git gcc-toolset-14 gcc-toolset-14-libatomic-devel patch zlib-devel \
18
libjpeg-turbo-devel libtiff-devel libpng-devel libwebp-devel freetype-devel harfbuzz-devel \
19
openssl-devel openblas openblas-devel autoconf automake libtool cmake numpy libsndfile \
20
clang llvm-devel llvm-static clang-devel && \
@@ -85,40 +85,15 @@ RUN curl https://sh.rustup.rs -sSf | sh -s -- -y && \
85
rustup default stable && \
86
rustup show
87
88
-FROM python-install AS torch
89
-ARG TORCH_VERSION=2.7.0
90
-ENV export _GLIBCXX_USE_CXX11_ABI=1
91
-ENV CARGO_HOME=/root/.cargo
92
-ENV RUSTUP_HOME=/root/.rustup
93
-ENV PATH="$CARGO_HOME/bin:$RUSTUP_HOME/bin:$PATH"
94
-
95
-WORKDIR /tmp
96
-
97
-RUN --mount=type=cache,target=/root/.cache/uv \
98
- --mount=type=bind,from=rust,source=/root/.cargo,target=/root/.cargo,rw \
99
- --mount=type=bind,from=rust,source=/root/.rustup,target=/root/.rustup,rw \
100
- git clone https://github.com/pytorch/pytorch.git && \
101
- cd pytorch && \
102
- git checkout v2.7.0 && \
103
- git submodule sync && \
104
- git submodule update --init --recursive && \
105
- uv pip install cmake ninja && \
106
- uv pip install -r requirements.txt && \
107
- python setup.py bdist_wheel
108
-
109
-
88
FROM python-install AS torch-vision
89
# Install torchvision
112
-ARG TORCH_VERSION=2.7.0
113
-ARG TORCH_VISION_VERSION=v0.20.1
90
+ARG TORCH_VISION_VERSION=v0.23.0
91
WORKDIR /tmp
92
RUN --mount=type=cache,target=/root/.cache/uv \
116
- --mount=type=bind,from=torch,source=/tmp/pytorch/dist,target=/tmp/torch-wheels/ \
93
git clone https://github.com/pytorch/vision.git && \
94
cd vision && \
95
git checkout $TORCH_VISION_VERSION && \
120
- TORCH_WHL_FILE=$(ls /tmp/torch-wheels/*.whl | head -n 1) && \
121
- uv pip install -v $TORCH_WHL_FILE && \
96
+ uv pip install torch==2.8.0 --index-url https://download.pytorch.org/whl/cpu && \
97
python setup.py bdist_wheel
98
99
FROM python-install AS hf-xet-builder
@@ -199,26 +174,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
174
if ! grep '#include "dynamic_annotations.h"' numba/_dispatcher.cpp; then \
175
sed -i '/#include "internal\/pycore_atomic.h"/i\#include "dynamic_annotations.h"' numba/_dispatcher.cpp; \
176
fi && python setup.py bdist_wheel
202
-
203
-# Edit aws-lc-sys to support s390x
204
-FROM python-install AS aws-lc-sys-editor
205
-WORKDIR /tmp
206
-ENV CARGO_HOME=/root/.cargo
207
-ENV RUSTUP_HOME=/root/.rustup
208
-ENV PATH="$CARGO_HOME/bin:$RUSTUP_HOME/bin:$PATH"
209
-ARG AWS_LC_VERSION=v0.30.0
210
-RUN --mount=type=cache,target=/root/.cache/uv \
211
- --mount=type=bind,from=rust,source=/root/.cargo,target=/root/.cargo,rw \
212
- --mount=type=bind,from=rust,source=/root/.rustup,target=/root/.rustup,rw \
213
- git clone --recursive https://github.com/aws/aws-lc-rs.git && \
214
- cd aws-lc-rs && \
215
- git checkout tags/aws-lc-sys/${AWS_LC_VERSION} && \
216
- git submodule sync && \
217
- git submodule update --init --recursive && \
218
- cd aws-lc-sys && \
219
- sed -i '682 s/strncmp(buf, "-----END ", 9)/memcmp(buf, "-----END ", 9)/' aws-lc/crypto/pem/pem_lib.c && \
220
- sed -i '712 s/strncmp(buf, "-----END ", 9)/memcmp(buf, "-----END ", 9)/' aws-lc/crypto/pem/pem_lib.c && \
221
- sed -i '747 s/strncmp(buf, "-----END ", 9)/memcmp(buf, "-----END ", 9)/' aws-lc/crypto/pem/pem_lib.c
177
178
# Build Outlines Core
179
FROM python-install AS outlines-core-builder
@@ -226,17 +181,17 @@ WORKDIR /tmp
181
ENV CARGO_HOME=/root/.cargo
182
ENV RUSTUP_HOME=/root/.rustup
183
ENV PATH="$CARGO_HOME/bin:$RUSTUP_HOME/bin:$PATH"
229
-ARG OUTLINES_CORE_VERSION=0.2.10
184
+COPY requirements/common.txt /tmp/requirements/common.txt
185
+ARG OUTLINES_CORE_VERSION
186
RUN --mount=type=cache,target=/root/.cache/uv \
187
--mount=type=bind,from=rust,source=/root/.cargo,target=/root/.cargo,rw \
188
--mount=type=bind,from=rust,source=/root/.rustup,target=/root/.rustup,rw \
233
- --mount=type=bind,from=aws-lc-sys-editor,source=/tmp/aws-lc-rs/aws-lc-sys,target=/tmp/aws-lc-sys,rw \
189
+ OUTLINES_CORE_VERSION=${OUTLINES_CORE_VERSION:-$(grep -E '^outlines_core\s*==\s*[0-9.]+' /tmp/requirements/common.txt | grep -Eo '[0-9.]+')} && \
190
+ if [ -z "${OUTLINES_CORE_VERSION}" ]; then echo "ERROR: Could not determine outlines_core version"; exit 1; fi && \
191
git clone https://github.com/dottxt-ai/outlines-core.git && \
192
cd outlines-core && \
193
git checkout tags/${OUTLINES_CORE_VERSION} && \
194
sed -i "s/version = \"0.0.0\"/version = \"${OUTLINES_CORE_VERSION}\"/" Cargo.toml && \
238
- echo '[patch.crates-io]' >> Cargo.toml && \
239
- echo 'aws-lc-sys = { path = "/tmp/aws-lc-sys" }' >> Cargo.toml && \
195
uv pip install maturin && \
196
python -m maturin build --release --out dist
197
@@ -245,13 +200,15 @@ FROM python-install AS vllm-cpu
200
ARG PYTHON_VERSION
201
202
# Set correct library path for torch and numactl
248
-ENV LD_LIBRARY_PATH="/opt/vllm/lib64/python${PYTHON_VERSION}/site-packages/torch/lib:/usr/local/lib:$LD_LIBRARY_PATH"
203
+ENV LD_LIBRARY_PATH="/opt/vllm/lib64/python${PYTHON_VERSION}/site-packages/torch/lib:/usr/local/lib:/opt/rh/gcc-toolset-14/root/usr/lib64:$LD_LIBRARY_PATH"
204
ENV C_INCLUDE_PATH="/usr/local/include:$C_INCLUDE_PATH"
205
ENV UV_LINK_MODE=copy
206
ENV CARGO_HOME=/root/.cargo
207
ENV RUSTUP_HOME=/root/.rustup
253
-ENV PATH="$CARGO_HOME/bin:$RUSTUP_HOME/bin:$PATH"
208
ENV GRPC_PYTHON_BUILD_SYSTEM_OPENSSL=1
209
+ENV PCP_DIR=/opt/rh/gcc-toolset-14/root
210
+ENV PKG_CONFIG_PATH="/opt/rh/gcc-toolset-14/root/usr/lib64/pkgconfig:/usr/local/lib/pkgconfig/"
211
+ENV PATH="${VIRTUAL_ENV:+${VIRTUAL_ENV}/bin}:/opt/rh/gcc-toolset-14/root/usr/bin:/usr/local/bin:$CARGO_HOME/bin:$RUSTUP_HOME/bin:$PATH"
212
213
COPY . /workspace/vllm
214
WORKDIR /workspace/vllm
@@ -266,7 +223,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
223
--mount=type=bind,from=pyarrow,source=/tmp/arrow/python/dist,target=/tmp/arrow-wheels \
224
--mount=type=bind,from=torch-vision,source=/tmp/vision/dist,target=/tmp/vision-wheels/ \
225
--mount=type=bind,from=hf-xet-builder,source=/tmp/hf-xet/dist,target=/tmp/hf-xet-wheels/ \
269
- --mount=type=bind,from=torch,source=/tmp/pytorch/dist,target=/tmp/torch-wheels/ \
226
--mount=type=bind,from=numba-builder,source=/tmp/llvmlite/dist,target=/tmp/llvmlite-wheels/ \
227
--mount=type=bind,from=numba-builder,source=/tmp/numba/dist,target=/tmp/numba-wheels/ \
228
--mount=type=bind,from=outlines-core-builder,source=/tmp/outlines-core/dist,target=/tmp/outlines-core/dist/ \
@@ -274,7 +230,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
230
ARROW_WHL_FILE=$(ls /tmp/arrow-wheels/pyarrow-*.whl) && \
231
VISION_WHL_FILE=$(ls /tmp/vision-wheels/*.whl) && \
232
HF_XET_WHL_FILE=$(ls /tmp/hf-xet-wheels/*.whl) && \
277
- TORCH_WHL_FILE=$(ls /tmp/torch-wheels/*.whl) && \
233
LLVM_WHL_FILE=$(ls /tmp/llvmlite-wheels/*.whl) && \
234
NUMBA_WHL_FILE=$(ls /tmp/numba-wheels/*.whl) && \
235
OUTLINES_CORE_WHL_FILE=$(ls /tmp/outlines-core/dist/*.whl) && \
@@ -282,7 +237,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
237
$ARROW_WHL_FILE \
238
$VISION_WHL_FILE \
239
$HF_XET_WHL_FILE \
285
- $TORCH_WHL_FILE \
240
$LLVM_WHL_FILE \
241
$NUMBA_WHL_FILE \
242
$OUTLINES_CORE_WHL_FILE \
requirements/common.txt
+2
-2
@@ -19,12 +19,12 @@ pillow # Required for image processing
19
prometheus-fastapi-instrumentator >= 7.0.0
20
tiktoken >= 0.6.0 # Required for DBRX tokenizer
21
lm-format-enforcer == 0.11.3
22
-llguidance >= 0.7.11, < 0.8.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64"
22
+llguidance >= 0.7.11, < 0.8.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64" or platform_machine == "s390x"
23
outlines_core == 0.2.11
24
# required for outlines backend disk cache
25
diskcache == 5.6.3
26
lark == 1.2.2
27
-xgrammar == 0.1.25; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64"
27
+xgrammar == 0.1.25; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64" or platform_machine == "s390x"
28
typing_extensions >= 4.10
29
filelock >= 3.16.1 # need to contain https://github.com/tox-dev/filelock/pull/317
30
partial-json-parser # used for parsing partial JSON outputs
requirements/cpu.txt
+1
-1
@@ -7,7 +7,7 @@ numba == 0.61.2; platform_machine != "s390x" # Required for N-gram speculative d
7
packaging>=24.2
8
setuptools>=77.0.3,<80.0.0
9
--extra-index-url https://download.pytorch.org/whl/cpu
10
-torch==2.8.0+cpu; platform_machine == "x86_64"
10
+torch==2.8.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x"
11
torch==2.8.0; platform_system == "Darwin"
12
torch==2.8.0; platform_machine == "ppc64le" or platform_machine == "aarch64"
13