How to install numpy, PyTorch and pandas under GraalPy+JVM and x64 architecture? (SOLVED)

[vak]

Are there clear instruction how to install numpy, PyTorch and pandas under GraalPy+JVM+x64? (RESOLVED, see updete below)
(JVM is important, because a Java class should be called from externally supplied JAR)

I am having hard times to install all three packages in GraalPy+JVM.
I tried different OS distributives (ubuntu:22.04, ghcr.io/graalvm/jdk-community:24) without success.

graalpy -m pip install numpy torch pandas triggers the installation from sources and the compilation from sources triggers CMake and it causes the error:

#7 1314.0       -- Building using own protobuf under third_party per request.
#7 1314.0       -- Use custom protobuf build.
#7 1314.0       CMake Error at third_party/protobuf/cmake/CMakeLists.txt:2 (cmake_minimum_required):
#7 1314.0         Compatibility with CMake < 3.5 has been removed from CMake.
#7 1314.0       
#7 1314.0         Update the VERSION argument <min> value.  Or, use the <min>...<max> syntax
#7 1314.0         to tell CMake that the project requires at least <min> but has been updated
#7 1314.0         to work with policies introduced by <max> or earlier.
#7 1314.0       
#7 1314.0         Or, add -DCMAKE_POLICY_VERSION_MINIMUM=3.5 to try configuring anyway.
#7 1314.0       
#7 1314.0       
#7 1314.0       -- Configuring incomplete, errors occurred!
#7 1314.0       [end of output]
#7 1314.0   

I tried to set
ENV CMAKE_ARGS="-DCMAKE_POLICY_VERSION_MINIMUM=3.5"
ENV CMAKE_POLICY_VERSION_MINIMUM=3.5
but they seem to be "shadowed" by isolated pip build, so i tried graalpy -m pip install --no-build-isolation "cmake<3.20", but met many issues on this path as well.

Also i tried these versions:
graalpy -m pip install numpy==1.23.5 torch==1.13.1
because I know that year ago the image was build successfully for them -- still the same CMake error

Instructions are wellcomed, and a Dockerfile for reproducibility even more wellcomed!
🙏

UPDATE: after many-many hours spent, I finally got the working Docker image!

FROM container-registry.oracle.com/graalvm/jdk:24

# Install dependencies for pyenv, GraalPy, and numpy compilation
RUN microdnf update -y && \
    microdnf install -y \
        curl \
        git \
        gcc \
        gcc-c++ \
        make \
        patch \
        openssl-devel \
        bzip2-devel \
        libffi-devel \
        readline-devel \
        sqlite-devel \
        xz-devel \
        && microdnf clean all

# Download and install GraalPy JVM version manually (pyenv only provides Native version)
ARG GRAALPY_VERSION=24.2.1
ARG PLATFORM=linux-amd64

RUN curl -L "https://github.com/oracle/graalpython/releases/download/graal-$GRAALPY_VERSION/graalpy-jvm-$GRAALPY_VERSION-$PLATFORM.tar.gz" \
    -o /tmp/graalpy-jvm.tar.gz && \
    tar -xzf /tmp/graalpy-jvm.tar.gz -C /opt && \
    rm /tmp/graalpy-jvm.tar.gz && \
    mv /opt/graalpy-$GRAALPY_VERSION-$PLATFORM /opt/graalpy

ENV PATH="/opt/graalpy/bin:$PATH"
ENV GRAALPY_HOME="/opt/graalpy"

RUN graalpy --version && \
    python --version

RUN graalpy -m venv /opt/graalpy-venv

# or pin it to version that worked for me: 1.26.4
RUN /opt/graalpy-venv/bin/pip install --upgrade pip && \
    /opt/graalpy-venv/bin/pip install "numpy<2.0"

RUN microdnf update -y && \
    microdnf install -y cmake && \
    microdnf clean all

RUN microdnf --enablerepo=ol10_codeready_builder install ninja-build

RUN /opt/graalpy-venv/bin/pip install pyyaml
RUN /opt/graalpy-venv/bin/pip install typing_extensions

ENV PIP_NO_BUILD_ISOLATION=1
RUN /opt/graalpy-venv/bin/pip install --upgrade setuptools wheel 
ENV CFLAGS="-Wno-error=maybe-uninitialized -Wno-error=uninitialized" \
    CXXFLAGS="-Wno-error=maybe-uninitialized -Wno-error=uninitialized" \
    CMAKE_ARGS="-DCMAKE_CXX_FLAGS='-Wno-error=maybe-uninitialized -Wno-error=uninitialized' -DCMAKE_C_FLAGS='-Wno-error=maybe-uninitialized -Wno-error=uninitialized'"

RUN /opt/graalpy-venv/bin/pip install --no-build-isolation torch==2.2.1

# install like this or pin it to version 2.2.3 that worked for me:
RUN /opt/graalpy-venv/bin/pip install pandas

ENV PATH="/opt/graalpy-venv/bin:$PATH"

WORKDIR /app
CMD ["python"]

not ideal, but finally works!

bash-5.2# python
Python 3.11.7 (Wed Apr 02 19:53:28 GMT 2025)
[Graal, Oracle GraalVM, Java 24.0.1 (amd64)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy, torch, pandas
>>> numpy.__version__
'1.26.4'
>>> torch.__version__
'2.2.1'
>>> pandas.__version__
'2.2.3'
>>>

[msimacek]

GraalPy 24.2 has a patch for torch 2.4.1. That one doesn't download cmake from pip, but uses the one from your system. Could you use that? If not, you can try to create a pip constraints file that pins cmake to older version and use PIP_CONSTRAINTS env var, which should be respected even by the isolated build.

[vak]

Guys, thank you, the issue is resolved and could be closed.
Take the solution (see above) if you need it.