Implementation:Alibaba MNN CMake Build Quantools

Field	Value
Implementation Name	CMake_Build_Quantools
Type	API Doc
Topic	Model_Compression
Workflow	Model_Compression
Description	CMake build commands and options for compiling MNN model compression tools
Source File(s)	CMakeLists.txt:L49,L51
Last Updated	2026-02-10 14:00 GMT

API Signature

cmake -DMNN_BUILD_CONVERTER=ON -DMNN_BUILD_QUANTOOLS=ON .. && make -j8

Source Definition

From CMakeLists.txt (lines 49 and 51):

option(MNN_BUILD_QUANTOOLS "Build Quantized Tools or not" OFF)
option(MNN_BUILD_CONVERTER "Build Converter" OFF)

Both options default to OFF and must be explicitly enabled to compile the compression tool binaries.

Parameters

Parameter	Type	Default	Description
-DMNN_BUILD_CONVERTER	BOOL	OFF	Enables compilation of the MNNConvert binary. This tool handles model format conversion (ONNX, TF, Caffe, TFLite to MNN), weight quantization (2-8 bit), FP16 half-precision storage, and HQQ quantization.
-DMNN_BUILD_QUANTOOLS	BOOL	OFF	Enables compilation of the quantized.out binary. This tool performs offline INT8 quantization using calibration data with KL-divergence, ADMM, or EMA feature quantization methods.
-DMNN_LOW_MEMORY	BOOL	OFF	Optional. Enables dynamic weight dequantization support in the MNN runtime library for accelerated inference on weight-quantized models. Defined at CMakeLists.txt:L79.

Inputs

• MNN source tree -- The complete MNN repository checked out from source control.
• CMake 3.6+ -- Required build system generator.
• C/C++ toolchain -- A compiler supporting C99 and C++11 standards.

Outputs

Output	Description
MNNConvert	Command-line binary for model conversion and weight quantization. Supports --weightQuantBits, --fp16, --hqq, --weightQuantBlock, and --compressionParamsFile flags.
quantized.out	Command-line binary for offline INT8 quantization. Takes a float MNN model, a destination path, and a JSON calibration configuration file as positional arguments.
mnnconvert (Python)	Python wrapper for MNNConvert, installed via pip install MNN.
mnnquant (Python)	Python wrapper for quantized.out, installed via pip install MNN.
mnn (Python)	Unified Python CLI entry point for all MNN tools.

Usage Example

C++ Build (Production)

cd MNN
mkdir build && cd build
cmake .. -DMNN_BUILD_CONVERTER=ON -DMNN_BUILD_QUANTOOLS=ON
make -j8

Python Install (Experimentation)

pip install MNN

After installation, both mnnconvert and mnnquant are available on the command line.

Verify Build Outputs

# Weight quantization tool
./MNNConvert --help

# Offline INT8 quantization tool
./quantized.out
# Usage: ./quantized.out src.mnn dst.mnn preTreatConfig.json

Related Pages

• Principle:Alibaba_MNN_Compression_Tool_Setup
• Environment:Alibaba_MNN_CPU_Build_Environment