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 12, 18:30 UTC
Update Dockerfile to use gcc-toolset-14 and fix test case failures on power (ppc64le) (#28957)
Signed-off-by: Bhagyashri <Bhagyashri.Gaikwad2@ibm.com>
Bhagyashri committed
Nov 21, 2025 at 17:54 UTC
2b1b3dfa4b02456b11b2bdbcd0857ddb96214a71
3 files changed
+27
-19
.buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh
+6
-4
@@ -25,20 +25,22 @@ function cpu_tests() {
25
26
# offline inference
27
podman exec -it "$container_id" bash -c "
28
+ export TORCH_COMPILE_DISABLE=1
29
set -xve
30
python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m" >> $HOME/test_basic.log
31
32
# Run basic model test
33
podman exec -it "$container_id" bash -c "
34
+ export TORCH_COMPILE_DISABLE=1
35
set -evx
36
pip install pytest pytest-asyncio einops peft Pillow soundfile transformers_stream_generator matplotlib
35
- pip install sentence-transformers datamodel_code_generator
37
+ pip install sentence-transformers datamodel_code_generator tblib
38
39
# Note: disable Bart until supports V1
40
# pytest -v -s tests/models/language/generation/test_bart.py -m cpu_model
39
- pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-openai-community/gpt2]
40
- pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-facebook/opt-125m]
41
- pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-google/gemma-1.1-2b-it]
41
+ pytest -v -s tests/models/language/generation/test_common.py::test_models[False-False-5-32-openai-community/gpt2]
42
+ pytest -v -s tests/models/language/generation/test_common.py::test_models[False-False-5-32-facebook/opt-125m]
43
+ pytest -v -s tests/models/language/generation/test_common.py::test_models[False-False-5-32-google/gemma-1.1-2b-it]
44
pytest -v -s tests/models/language/pooling/test_classification.py::test_models[float-jason9693/Qwen2.5-1.5B-apeach]
45
# TODO: Below test case tests/models/language/pooling/test_embedding.py::test_models[True-ssmits/Qwen2-7B-Instruct-embed-base] fails on ppc64le. Disabling it for time being.
46
# pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model" >> $HOME/test_rest.log
docker/Dockerfile.ppc64le
+19
-13
@@ -8,8 +8,8 @@ FROM registry.access.redhat.com/ubi9/ubi-minimal:${BASE_UBI_IMAGE_TAG} AS openbl
8
9
ARG MAX_JOBS
10
ARG OPENBLAS_VERSION=0.3.30
11
-RUN microdnf install -y dnf && dnf install -y gcc-toolset-13 make wget unzip \
12
- && source /opt/rh/gcc-toolset-13/enable \
11
+RUN microdnf install -y dnf && dnf install -y gcc-toolset-14 make wget unzip \
12
+ && source /opt/rh/gcc-toolset-14/enable \
13
&& wget https://github.com/OpenMathLib/OpenBLAS/releases/download/v$OPENBLAS_VERSION/OpenBLAS-$OPENBLAS_VERSION.zip \
14
&& unzip OpenBLAS-$OPENBLAS_VERSION.zip \
15
&& cd OpenBLAS-$OPENBLAS_VERSION \
@@ -57,7 +57,7 @@ COPY --from=openblas-builder /tmp/control /dev/null
57
RUN --mount=type=bind,from=openblas-builder,source=/OpenBLAS-$OPENBLAS_VERSION/,target=/openblas/,rw \
58
dnf install -y openssl-devel \
59
&& dnf install -y \
60
- git tar gcc-toolset-13 automake libtool \
60
+ git tar gcc-toolset-14 automake libtool \
61
pkgconfig xsimd zeromq-devel kmod findutils protobuf* \
62
libtiff-devel libjpeg-devel zlib-devel freetype-devel libwebp-devel \
63
harfbuzz-devel libraqm-devel libimagequant-devel libxcb-devel \
@@ -84,7 +84,7 @@ ARG _GLIBCXX_USE_CXX11_ABI=1
84
ARG OPENBLAS_VERSION=0.3.30
85
86
RUN --mount=type=cache,target=/root/.cache/uv \
87
- source /opt/rh/gcc-toolset-13/enable && \
87
+ source /opt/rh/gcc-toolset-14/enable && \
88
git clone --recursive https://github.com/pytorch/pytorch.git -b v${TORCH_VERSION} && \
89
cd pytorch && \
90
uv pip install -r requirements.txt && \
@@ -97,7 +97,7 @@ ARG TORCHVISION_VERSION=0.22.0
97
ARG TORCHVISION_USE_NVJPEG=0
98
ARG TORCHVISION_USE_FFMPEG=0
99
RUN --mount=type=cache,target=/root/.cache/uv \
100
- source /opt/rh/gcc-toolset-13/enable && \
100
+ source /opt/rh/gcc-toolset-14/enable && \
101
git clone --recursive https://github.com/pytorch/vision.git -b v${TORCHVISION_VERSION} && \
102
cd vision && \
103
MAX_JOBS=${MAX_JOBS:-$(nproc)} \
@@ -113,7 +113,7 @@ ARG USE_ROCM=0
113
ARG USE_CUDA=0
114
ARG TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_FFMPEG=1
115
RUN --mount=type=cache,target=/root/.cache/uv \
116
- source /opt/rh/gcc-toolset-13/enable && \
116
+ source /opt/rh/gcc-toolset-14/enable && \
117
git clone --recursive https://github.com/pytorch/audio.git -b v${TORCHAUDIO_VERSION} && \
118
cd audio && \
119
MAX_JOBS=${MAX_JOBS:-$(nproc)} \
@@ -130,7 +130,7 @@ ARG MAX_JOBS
130
ARG PYARROW_PARALLEL
131
ARG PYARROW_VERSION=21.0.0
132
RUN --mount=type=cache,target=/root/.cache/uv \
133
- source /opt/rh/gcc-toolset-13/enable && \
133
+ source /opt/rh/gcc-toolset-14/enable && \
134
git clone --recursive https://github.com/apache/arrow.git -b apache-arrow-${PYARROW_VERSION} && \
135
cd arrow/cpp && \
136
mkdir build && cd build && \
@@ -162,7 +162,7 @@ ARG OPENCV_VERSION=86
162
ARG OPENCV_PATCH=97f3f39
163
ARG ENABLE_HEADLESS=1
164
RUN --mount=type=cache,target=/root/.cache/uv \
165
- source /opt/rh/gcc-toolset-13/enable && \
165
+ source /opt/rh/gcc-toolset-14/enable && \
166
git clone --recursive https://github.com/opencv/opencv-python.git -b ${OPENCV_VERSION} && \
167
cd opencv-python && \
168
sed -i -E -e 's/"setuptools.+",/"setuptools",/g' pyproject.toml && \
@@ -196,7 +196,7 @@ ARG MAX_JOBS
196
ARG NUMBA_VERSION=0.61.2
197
198
# Clone all required dependencies
199
-RUN dnf install ninja-build llvm15 llvm15-devel -y && source /opt/rh/gcc-toolset-13/enable && export PATH=$PATH:/usr/lib64/llvm15/bin && \
199
+RUN dnf install ninja-build llvm15 llvm15-devel -y && source /opt/rh/gcc-toolset-14/enable && export PATH=$PATH:/usr/lib64/llvm15/bin && \
200
git clone --recursive https://github.com/numba/numba.git -b ${NUMBA_VERSION} && \
201
cd ./numba && \
202
if ! grep '#include "dynamic_annotations.h"' numba/_dispatcher.cpp; then \
@@ -211,6 +211,9 @@ RUN dnf install ninja-build llvm15 llvm15-devel -y && source /opt/rh/gcc-toolset
211
212
FROM base-builder AS vllmcache-builder
213
214
+ENV LLVM_CONFIG=/usr/lib64/llvm15/bin/llvm-config
215
+ENV PATH=/usr/lib64/llvm15/bin:$PATH
216
+
217
COPY --from=torch-builder /tmp/control /dev/null
218
COPY --from=arrow-builder /tmp/control /dev/null
219
COPY --from=cv-builder /tmp/control /dev/null
@@ -225,10 +228,13 @@ ARG GRPC_PYTHON_BUILD_SYSTEM_OPENSSL=1
228
RUN --mount=type=cache,target=/root/.cache/uv \
229
dnf install llvm15 llvm15-devel -y && \
230
rpm -ivh --nodeps https://mirror.stream.centos.org/9-stream/CRB/ppc64le/os/Packages/protobuf-lite-devel-3.14.0-16.el9.ppc64le.rpm && \
228
- source /opt/rh/gcc-toolset-13/enable && \
231
+ source /opt/rh/gcc-toolset-14/enable && \
232
git clone https://github.com/huggingface/xet-core.git && cd xet-core/hf_xet/ && \
233
uv pip install maturin && \
234
uv build --wheel --out-dir /hf_wheels/
235
+
236
+ENV CXXFLAGS="-fno-lto -Wno-error=free-nonheap-object" \
237
+ CFLAGS="-fno-lto"
238
RUN --mount=type=cache,target=/root/.cache/uv \
239
--mount=type=bind,from=torch-builder,source=/torchwheels/,target=/torchwheels/,ro \
240
--mount=type=bind,from=arrow-builder,source=/arrowwheels/,target=/arrowwheels/,ro \
@@ -236,7 +242,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
242
--mount=type=bind,from=numa-builder,source=/numactl/,target=/numactl/,rw \
243
--mount=type=bind,from=numba-builder,source=/numbawheels/,target=/numbawheels/,ro \
244
--mount=type=bind,src=.,dst=/src/,rw \
239
- source /opt/rh/gcc-toolset-13/enable && \
245
+ source /opt/rh/gcc-toolset-14/enable && \
246
export PATH=$PATH:/usr/lib64/llvm15/bin && \
247
uv pip install /opencvwheels/*.whl /arrowwheels/*.whl /torchwheels/*.whl /numbawheels/*.whl && \
248
sed -i -e 's/.*torch.*//g' /src/pyproject.toml /src/requirements/*.txt && \
@@ -260,7 +266,7 @@ FROM base-builder AS lapack-builder
266
ARG MAX_JOBS
267
ARG LAPACK_VERSION=3.12.1
268
RUN git clone --recursive https://github.com/Reference-LAPACK/lapack.git -b v${LAPACK_VERSION} \
263
- && cd lapack && source /opt/rh/gcc-toolset-13/enable \
269
+ && cd lapack && source /opt/rh/gcc-toolset-14/enable \
270
&& cmake -B build -S . \
271
&& cmake --build build -j ${MAX_JOBS:-$(nproc)}
272
@@ -299,7 +305,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
305
--mount=type=bind,from=openblas-builder,source=/OpenBLAS-$OPENBLAS_VERSION/,target=/openblas/,rw \
306
rpm -ivh https://dl.fedoraproject.org/pub/epel/epel-release-latest-9.noarch.rpm && \
307
microdnf install --nodocs -y \
302
- libomp tar findutils openssl llvm15 llvm15-devel \
308
+ libomp libicu tar findutils openssl llvm15 llvm15-devel \
309
pkgconfig xsimd g++ gcc-fortran libsndfile \
310
libtiff libjpeg openjpeg2 zlib zeromq \
311
freetype lcms2 libwebp tcl tk utf8proc \
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 >= 1.3.0, < 1.4.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64" or platform_machine == "s390x"
22
+llguidance >= 1.3.0, < 1.4.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64" or platform_machine == "s390x" or platform_machine == "ppc64le"
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 == 0.1.27; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64" or platform_machine == "s390x"
27
+xgrammar == 0.1.27; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64" or platform_machine == "s390x" or platform_machine == "ppc64le"
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