Prerequisites

Ensure your system meets:

• OS: Ubuntu 25.04 (or beta)
• Python: 3.9–3.12 (avoid 3.13 unless confirmed compatible)
• Disk: ≥15 GB free
• RAM: 16 GB recommended
• Internet: Required

Install build tools:

1
sudo apt update
2
sudo apt install -y build-essential cmake git python3-dev python3-venv wget libelf-dev

✅ Tip: Install a supported Python version:

1
sudo apt install python3.11 python3.11-venv python3.11-dev

Step 1: Set Up Python Virtual Environment

1
mkdir openvino-genai-env && cd openvino-genai-env
2
python3.11 -m venv venv
3
source venv/bin/activate
4
pip install --upgrade pip

Step 2: Install Intel oneAPI Base Toolkit (for oneMKL)

Install Intel’s optimized math libraries.

1
wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | sudo tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null
2

3
echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | sudo tee /etc/apt/sources.list.d/oneapi.list
4

5
sudo apt update
6
sudo apt install -y intel-basekit

1
source /opt/intel/oneapi/setvars.sh

Step 3: Install System jemalloc (from Ubuntu)

Use the Ubuntu-provided jemalloc instead of building from source.

1
sudo apt install -y libjemalloc-dev

This installs:

• Header: /usr/include/jemalloc/jemalloc.h
• Library: /usr/lib/x86_64-linux-gnu/libjemalloc.so

The libjemalloc2 is already installed because it is used by many applications such as Firefox.

Step 4: Modify OpenVINO to Link jemalloc Directly

To link jemalloc directly into the OpenVINO runtime (not via LD_PRELOAD), we must patch the CMake build configuration.

1
git clone -b 2025.2.0 https://github.com/openvinotoolkit/openvino.git

1
# Backup original
2
cp CMakeLists.txt CMakeLists.txt.bak
3

4
# Append jemalloc linking logic
5
cat >> CMakeLists.txt << 'EOF'
6

7
# --- Custom: Link jemalloc directly ---
8
find_library(JEMALLOC_LIB jemalloc PATHS /usr/lib /usr/lib/x86_64-linux-gnu NO_DEFAULT_PATH)
9
if(JEMALLOC_LIB)
10
    message(STATUS "jemalloc found: ${JEMALLOC_LIB}")
11
    # Link jemalloc to core runtime
12
    set_property(TARGET ngraph_frontend_manager APPEND PROPERTY INTERFACE_LINK_LIBRARIES ${JEMALLOC_LIB})
13
    set_property(TARGET ov_runtime APPEND PROPERTY INTERFACE_LINK_LIBRARIES ${JEMALLOC_LIB})
14
    set_property(TARGET openvino_c APPEND PROPERTY INTERFACE_LINK_LIBRARIES ${JEMALLOC_LIB})
15
else()
16
    message(WARNING "jemalloc not found. Install libjemalloc-dev.")
17
endif()
18
# ---
19
EOF

Step 5: Install Build Dependencies

1
chmod +x install_build_dependencies.sh
2
sed -i 's/ubuntu24.04/ubuntu25.04/g' install_build_dependencies.sh || echo "Patch not needed or failed."
3
./install_build_dependencies.sh

Step 6: Build OpenVINO with oneMKL and jemalloc

1
mkdir -p build && cd build
2

3
source /opt/intel/oneapi/setvars.sh > /dev/null 2>&1
4

5
cmake .. \
6
    -DCMAKE_BUILD_TYPE=Release \
7
    -DENABLE_PYTHON=ON \
8
    -DENABLE_WHEEL=ON \
9
    -DPYTHON_EXECUTABLE=$(which python) \
10
    -DENABLE_MKL_DNN=ON \
11
    -DENABLE_INTEL_CPU_MATH=ON \
12
    -DENABLE_OV_ONNX_FRONTEND=ON \
13
    -DENABLE_OV_IR_FRONTEND=ON \
14
    -DENABLE_OV_GENAI=ON

Compile:

1
make --jobs=$(nproc)

The build will now statically link libjemalloc.so into the OpenVINO runtime libraries.

Step 7: Install OpenVINO into Virtual Environment

1
WHEEL_PATH=$(find . -name "openvino*.whl" | head -n 1)
2
pip install "$WHEEL_PATH"

Verify:

1
python -c "import openvino; print('OpenVINO version:', openvino.__version__)"

Step 8: Install OpenVINO GenAI

1
pip install openvino-genai

Verify:

1
python -c "from openvino import genai; print('OpenVINO GenAI imported')"

🔗 OpenVINO GenAI PyPI

Step 9: Verify jemalloc is Linked

Check if jemalloc symbols are embedded in the OpenVINO library:

1
# Find the core OpenVINO shared library
2
find venv/lib/python*/site-packages/openvino -name "*.so" -exec ldd {} \; | grep jemalloc

Should show:

1
libjemalloc.so.2 => /usr/lib/x86_64-linux-gnu/libjemalloc.so.2

Or use objdump:

1
objdump -p venv/lib/python*/site-packages/openvino/*.so | grep NEEDED | grep jemalloc

Should output:

1
NEEDED               libjemalloc.so.2

✅ This confirms jemalloc is directly linked, not just preloaded.

Step 10: Install Agentic Coding Dependencies

1
pip install \
2
    optimum-intel \
3
    huggingface-hub \
4
    transformers \
5
    llama-index \
6
    llama-index-llms-openvino-genai \
7
    jupyter

Troubleshooting

1
dpkg -L libjemalloc-dev | grep libjemalloc.so

1
sudo apt install --reinstall libjemalloc-dev

1
ls /usr/lib/x86_64-linux-gnu/libjemalloc*

1
find_library(JEMALLOC_LIB jemalloc.so.2 PATHS /usr/lib/x86_64-linux-gnu)

1
-- MKL-DNN enabled
2
-- Intel Math Kernel Library: YES

Conclusion

You’ve successfully built OpenVINO GenAI with:

• ✅ Intel oneMKL via oneAPI for accelerated math
• ✅ Ubuntu’s system jemalloc linked directly via -ljemalloc
• ✅ No need for LD_PRELOAD—memory allocation is built-in
• ✅ Ready for agentic coding with optimized performance

This configuration ensures that all OpenVINO-managed memory allocations go through jemalloc, reducing fragmentation and improving concurrency in agent workloads.

📚 Official Docs:

• OpenVINO 2025 Installation
• OpenVINO GenAI GitHub
• Optimum-Intel

Happy coding with your optimized AI agents!