Implementation:Alibaba MNN PyMNN Build Setup

Field	Value
implementation_name	PyMNN_Build_Setup
schema_version	0.1.0
workflow	Python_Model_Inference
implementation_type	API_Doc
domain	Deep_Learning_Inference
scope	Building and installing the MNN Python package from source or PyPI
source_file	pymnn/CMakeLists.txt:L1-235
related_patterns	CMake_Build_Systems, Python_Extension_Modules, Cross_Platform_Compilation
last_updated	2026-02-10 14:00 GMT

Summary

This implementation covers two approaches for installing MNN's Python bindings: (1) installing a pre-built wheel from PyPI using pip, and (2) building from source using CMake for custom configurations. The build system is defined in pymnn/CMakeLists.txt and produces a native Python extension module (mnnpybridge) that links against the MNN inference engine.

API Signatures

Method 1: pip install (Pre-built Wheel)

pip install MNN

Method 2: Build from Source (CMake)

cd /path/to/MNN
mkdir pymnn_build && cd pymnn_build
cmake -DPYMNN_EXPR_API=ON -DPYMNN_OPENCV_API=ON .. && make

Method 3: setup.py (Development Install)

cd /path/to/MNN/pymnn/pip_package
python setup.py install

Parameters

Parameter	Type	Default	Description
PYMNN_EXPR_API	CMake option	ON	Enables the expression API binding (MNN.expr module) for tensor manipulation and computation graph operations
PYMNN_OPENCV_API	CMake option	ON	Enables the OpenCV-compatible image processing module (MNN.cv)
PYMNN_TRAIN_API	CMake option	OFF	Enables the training API for model fine-tuning and custom module training
PYMNN_NUMPY_USABLE	CMake option	OFF	Enables interoperability with numpy arrays, allowing Var.read() to return numpy.ndarray
PYMNN_LLM_API	CMake option	OFF	Enables the large language model API (MNN.llm)
PYMNN_AUDIO_API	CMake option	OFF	Enables the audio processing API (MNN.audio)
PYMNN_IMGCODECS	CMake option	OFF	Enables image codec functions (imread, imwrite) for mobile builds
MNN_BUILD_SHARED_LIBS	CMake option	ON	Controls whether MNN is built as a shared (.so/.dylib) or static (.a) library
MNN_WIN_RUNTIME_MT	CMake option	OFF	On Windows, controls /MT vs /MD runtime linkage

Inputs

• MNN source tree -- The complete MNN repository containing C++ source, headers, and the pymnn bridge code
• Python development headers -- Required for compiling the native extension module
• CMake 3.4.1+ -- Build system generator
• C++11-compatible compiler -- GCC, Clang, or MSVC

Outputs

• Installed MNN Python package with the following submodules:
  • MNN.expr -- Tensor operations, data format conversion, model loading
  • MNN.nn -- Module loading, runtime management, neural network layers
  • MNN.cv -- OpenCV-compatible image processing (imread, resize, cvtColor, etc.)
  • MNN.numpy -- Numpy-compatible array creation and manipulation

Code Example

# After installation, verify the package is working
import MNN
import MNN.expr as expr
import MNN.nn as nn
import MNN.cv as cv
import MNN.numpy as np

# Check basic tensor creation
x = expr.const([1.0, 2.0, 3.0, 4.0], [2, 2], expr.NCHW, expr.float)
print(x.shape)   # [2, 2]
print(x.dtype)   # dtype.float

Build System Details

The CMake build (pymnn/CMakeLists.txt) performs the following:

• Lines 7-21: Define all optional feature flags as CMake options with their defaults
• Lines 37-41: Create the mnnpybridge library target (shared or static based on MNN_BUILD_SHARED_LIBS)
• Lines 52-103: Set compile definitions based on enabled features (PYMNN_EXPR_API, PYMNN_OPENCV_API, etc.)
• Lines 114-139: Configure platform-specific compiler flags (MSVC flags on Windows, GCC/Clang flags on Unix)
• Lines 149-226: Link against MNN, Python, and optional libraries (numpy, AliNNPython, MNNOpenCV, MNN_LLM, MNNAudio)

# Key CMake options from pymnn/CMakeLists.txt:L7-21
option(PYMNN_EXPR_API "MNN expr API be exposed" ON)
option(PYMNN_NUMPY_USABLE "Build based on numpy" OFF)
option(PYMNN_TRAIN_API "MNN train API be exposed" OFF)
option(PYMNN_OPENCV_API "MNN OpenCV API be exposed" ON)
option(PYMNN_IMGCODECS "MNN IMGCODECS API be exposed" OFF)
option(PYMNN_AUDIO_API "MNN Audio API be exposed" OFF)
option(PYMNN_LLM_API "MNN LLM API be exposed" OFF)

Edge Cases and Limitations

• Mobile platforms: On Android, some features like imread/imwrite are not available by default; enable PYMNN_IMGCODECS to include them
• numpy dependency: PYMNN_NUMPY_USABLE is OFF by default on mobile; Var.read() returns numpy.ndarray only when enabled
• Windows 32-bit: Requires special handling via the BUILD_ARCH=x86 setup.py flag
• GPU backend libraries: Building with CUDA or OpenCL support requires the appropriate SDK to be installed and library paths set

Related Pages

• Principle:Alibaba_MNN_PyMNN_Installation
• Environment:Alibaba_MNN_CPU_Build_Environment
• Environment:Alibaba_MNN_Python_Export_Environment