// Build configuration, CMake options, cross-compilation and packaging. Use when building MindSpore Lite, configuring CMake, cross-compiling for ARM/iOS/MCU, packaging release archives, or troubleshooting build errors.
| name | lite-build |
| description | Build configuration, CMake options, cross-compilation and packaging. Use when building MindSpore Lite, configuring CMake, cross-compiling for ARM/iOS/MCU, packaging release archives, or troubleshooting build errors. |
| paths | ["build.sh","build.bat","scripts/build/**","mindspore-lite/CMakeLists.txt","cmake/**"] |
build.sh (Linux/macOS) / build.bat (Windows)
-> scripts/build/build_lite.sh
-> CMake (mindspore-lite/CMakeLists.txt)
-> cmake/package_lite.cmake (packaging)
# End-side ARM64 (most common)
bash build.sh -I arm64 -j8
# End-side ARM32
bash build.sh -I arm32 -j8
# x86 debug build
bash build.sh -I x86_64 -j8
# Cloud-side inference (ExtendRT)
bash build.sh -I x86_64 -e cpu -a x64 -j8
# iOS
bash build.sh -I arm64 -T ios -j8
# Micro Cortex-M (IoT bare metal)
bash build.sh -I cortex-m -j8
Options defined in mindspore-lite/CMakeLists.txt, set via build.sh flags or passed directly.
| CMake Variable | Default | Description |
|---|---|---|
MSLITE_ENABLE_SSE | ON (x86) | Enable SSE instruction set |
MSLITE_ENABLE_AVX | ON (x86) | Enable AVX instruction set |
MSLITE_ENABLE_AVX512 | OFF | Enable AVX-512 |
MSLITE_ENABLE_FP16 | ON (ARM) | Enable FP16 operators |
MSLITE_ENABLE_INT8 | ON | Enable INT8 quantization operators |
MSLITE_ENABLE_CONTROL_FLOW | ON | Support control flow operators (If/While) |
MSLITE_ENABLE_DYNAMIC_THREAD | OFF | Dynamic thread scheduling |
MSLITE_ENABLE_WEIGHT_DECODE | ON | Quantized weight decoding |
| CMake Variable | Default | Description |
|---|---|---|
MSLITE_ENABLE_GPU_OPENCL | OFF | OpenCL GPU backend |
MSLITE_ENABLE_GPU_VULKAN | OFF | Vulkan GPU backend |
MSLITE_ENABLE_GPU_TENSORRT | OFF | TensorRT GPU backend |
MSLITE_ENABLE_NPU | OFF | HiSilicon NPU delegate |
MSLITE_ENABLE_COREML | OFF | Apple CoreML delegate |
MSLITE_ENABLE_TRAIN | OFF | Device-side training support |
MSLITE_ENABLE_MICRO | OFF | Micro code generation |
MSLITE_ENABLE_RISCV | OFF | RISC-V support |
| CMake Variable | Default | Description |
|---|---|---|
MSLITE_ENABLE_CLOUD_INFERENCE | OFF | Cloud-side inference (ExtendRT) |
MSLITE_ENABLE_ACL | OFF | Ascend ACL backend |
MSLITE_ENABLE_ACL_TDT | OFF | Ascend TDT data transfer |
MSLITE_ENABLE_SHARED_MUTEX | ON | Shared mutex (cloud-side concurrency) |
MSLITE_ENABLE_WEIGHT_SHARE | OFF | Multi-model weight sharing |
MSLITE_ENABLE_AUTO_PARALLEL | OFF | Automatic parallel inference |
MSLITE_ENABLE_CLOUD_FUSION_INFERENCE | OFF | Cloud-side fusion inference |
| CMake Variable | Default | Description |
|---|---|---|
MSLITE_ENABLE_TESTCASES | OFF | Compile UT test cases |
MSLITE_ENABLE_DEBUG | OFF | Debug mode (-g, no optimization) |
MSLITE_ENABLE_PROFILE | OFF | Performance profiling |
export ANDROID_NDK=/path/to/android-ndk-r21e
bash build.sh -I arm64 -j8 # ARM64
bash build.sh -I arm32 -j8 # ARM32
NDK r21e+ required, r25c recommended. Version mismatch is the most common build failure cause.
bash build.sh -I arm64 -T ios -j8 # Device
bash build.sh -I x86_64 -T ios -j8 # Simulator
Uses cmake/lite_ios.cmake toolchain file.
bash build.sh -I arm64 -e cpu -a aarch64 -j8
export CCACHE_DIR=/path/to/ccache
export CMAKE_C_COMPILER_LAUNCHER=ccache
export CMAKE_CXX_COMPILER_LAUNCHER=ccache
cmake -G Ninja .. && ninja -j8
ninja's incremental build detection is more precise than make.
rm -rf build/CMakeCache.txt # Reconfigure only
rm -rf build/ # Full cleanup
Packaging logic in cmake/package_lite.cmake (~961 lines).
mindspore-lite-{version}-{os}-{arch}.tar.gz
runtime/
include/api/ # C++ API headers
include/c_api/ # C API headers
lib/libmindspore-lite.so
third_party/
tools/
converter/ # Model conversion tool
benchmark/ # Performance benchmark tool
cropper/ # Library cropping tool
micro/ # Micro code generation (if enabled)
| Scenario | Package Name Template |
|---|---|
| Android ARM64 | mindspore-lite-{version}-android-aarch64.tar.gz |
| Android ARM32 | mindspore-lite-{version}-android-armv7.tar.gz |
| Linux x86_64 | mindspore-lite-{version}-linux-x64.tar.gz |
| Windows x86_64 | mindspore-lite-{version}-win-x64.zip |
| Cloud Linux x86_64 | mindspore-lite-{version}-linux-x64.tar.gz |
| Cloud Linux ARM64 | mindspore-lite-{version}-linux-aarch64.tar.gz |
| Error | Solution |
|---|---|
Could not find Protobuf | Install protobuf 3.x or use auto-download in cmake/external_libs/ |
protobuf version mismatch 3.x vs 4.x | MindSpore Lite requires protobuf 3.x; specify path: cmake -DPROTOBUF_INCLUDE_DIR=... |
android-ndk-rXX does not support XXX | Use NDK r21e+, check ANDROID_NDK env var |
| FlatBuffers version mismatch | Use third_party/flatbuffers/ bundled version |
| File | Purpose |
|---|---|
build.sh | Main build entry point |
build.bat | Windows build entry point |
scripts/build/build_lite.sh | Lite build orchestration |
scripts/build/default_options.sh | Default build configuration |
scripts/build/option_proc_lite.sh | Lite option processing |
scripts/build/process_options.sh | Command line argument parsing |
scripts/build/usage.sh | Help information |
Cloud-side inference with ExtendRT and Ascend backends. Use for server-side inference, Ascend 310/910 deployment, ModelParallelRunner for concurrent serving, ModelGroup for weight sharing, distributed inference, or .mindir format loading.
Model conversion pipeline, parser development, optimization passes and quantization. Use when converting models to .ms, writing parser code, implementing optimizer passes, or configuring quantization.
实现开源模型从PyTorch→ONNX→MindIR→MindSpore Lite的端到端导出/验证/部署/性能评测。用户要求模型拆分导出、精度对齐、MindIR转换或部署工具链时调用。
Code formatting, naming conventions, security checks and CI verification. Use when running clang-format, checking code style, writing secure code for model parsing, reviewing code quality, or configuring CI/Jenkins pipelines.
Debugging, unit testing, benchmarking and performance analysis. Use when running gtest, benchmark tools, profiling latency or accuracy, diagnosing operator precision issues, delegate fallback, or memory leaks.
Device-side inference with LiteRT, NNACL and hardware delegates. Use for mobile/IoT inference, Android/iOS integration, NPU/GPU/CoreML delegates, Micro codegen for MCU, on-device training, or C/C++/Java/Python API usage with .ms models.