[WIP] How to Install OpenVINO GenAI on Ubuntu 25.04 with oneMKL and jemalloc for Agentic Coding

How to Install OpenVINO GenAI on Ubuntu 25.04 with oneMKL and System jemalloc for Agentic Coding

This guide provides a step-by-step process to install OpenVINO GenAI from source on Ubuntu 25.04 (Plucky Puffin)—an unsupported distribution—using a Python virtual environment. It includes performance optimizations with Intel oneMKL (via OneAPI) and jemalloc from Ubuntu’s package repository, linked directly at build time using -ljemalloc. This setup is ideal for agentic coding applications requiring high inference throughput and memory efficiency.

⚠️ Note: Ubuntu 25.04 is not officially supported by OpenVINO. This guide uses the GitHub source build method for full control over compiler and library integration.

---

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:

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

✅ Tip: Install a supported Python version:

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

---

Step 1: Set Up Python Virtual Environment

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

---

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

Install Intel’s optimized math libraries.

Add Intel Repository

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

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

sudo apt update
sudo apt install -y intel-basekit

Source Environment

source /opt/intel/oneapi/setvars.sh

🔗 Intel oneAPI Base Toolkit

---

Step 3: Install System jemalloc (from Ubuntu)

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

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.

Clone OpenVINO

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

Patch CMake to Link -ljemalloc

Edit the main CMakeLists.txt to add jemalloc as a link library:

# Backup original
cp CMakeLists.txt CMakeLists.txt.bak

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

# --- Custom: Link jemalloc directly ---
find_library(JEMALLOC_LIB jemalloc PATHS /usr/lib /usr/lib/x86_64-linux-gnu NO_DEFAULT_PATH)
if(JEMALLOC_LIB)
    message(STATUS "jemalloc found: ${JEMALLOC_LIB}")
    # Link jemalloc to core runtime
    set_property(TARGET ngraph_frontend_manager APPEND PROPERTY INTERFACE_LINK_LIBRARIES ${JEMALLOC_LIB})
    set_property(TARGET ov_runtime APPEND PROPERTY INTERFACE_LINK_LIBRARIES ${JEMALLOC_LIB})
    set_property(TARGET openvino_c APPEND PROPERTY INTERFACE_LINK_LIBRARIES ${JEMALLOC_LIB})
else()
    message(WARNING "jemalloc not found. Install libjemalloc-dev.")
endif()
# ---
EOF

✅ This links jemalloc to key OpenVINO targets: ov_runtime, ngraph_frontend_manager, and openvino_c.

---

Step 5: Install Build Dependencies

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

---

Step 6: Build OpenVINO with oneMKL and jemalloc

mkdir -p build && cd build

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

cmake .. \
    -DCMAKE_BUILD_TYPE=Release \
    -DENABLE_PYTHON=ON \
    -DENABLE_WHEEL=ON \
    -DPYTHON_EXECUTABLE=$(which python) \
    -DENABLE_MKL_DNN=ON \
    -DENABLE_INTEL_CPU_MATH=ON \
    -DENABLE_OV_ONNX_FRONTEND=ON \
    -DENABLE_OV_IR_FRONTEND=ON \
    -DENABLE_OV_GENAI=ON

Compile:

make --jobs=$(nproc)

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

---

Step 7: Install OpenVINO into Virtual Environment

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

Verify:

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

---

Step 8: Install OpenVINO GenAI

pip install openvino-genai

Verify:

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:

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

Should show:

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

Or use objdump:

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

Should output:

NEEDED               libjemalloc.so.2

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

---

Step 10: Install Agentic Coding Dependencies

pip install \
    optimum-intel \
    huggingface-hub \
    transformers \
    llama-index \
    llama-index-llms-openvino-genai \
    jupyter

---

Troubleshooting

1. jemalloc Not Found During Link

Ensure libjemalloc-dev is installed:

dpkg -L libjemalloc-dev | grep libjemalloc.so

If not found, reinstall:

sudo apt install --reinstall libjemalloc-dev

2. Linking Fails with “undefined reference”

The patch assumes the library is named libjemalloc.so. Verify:

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

If the name differs (e.g., libjemalloc.so.2), update the find_library call:

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

3. oneMKL Not Used

Check CMake output for:

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

Ensure source /opt/intel/oneapi/setvars.sh was run before CMake.

---

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!