MNN端侧部署 on huluhuluu

MNN Backend 介绍

Thu, 26 Feb 2026 23:00:00 +0800

MNN 介绍

本文将介绍MNN框架的后端支持，并介绍后端有关的核心代码与功能。

MNN 介绍
- 1. 后端

1. 后端

MNN是阿里巴巴开源的高效轻量级深度学习推理框架，提供了较为全面的后端支持。主要包括:

架构/精度	类型	Normal	FP16	BF16	Int8
CPU	Native	B	C	B	B
CPU	x86/x64-SSE4.1	A	C	C	A
CPU	x86/x64-AVX2	S	C	C	A
CPU	x86/x64-AVX512	S	C	C	S
CPU	ARMv7a	S	S（ARMv8.2）	C	S
CPU	ARMv8	S	S（ARMv8.2）	S（ARMv8.6）	S
GPU	OpenCL	A	S	C	S
GPU	Vulkan	A	A	C	A
GPU	Metal	A	S	C	S
GPU	CUDA	A	S	C	A
NPU	CoreML	A	C	C	C
NPU	HIAI	A	C	C	C
NPU	NNAPI	B	B	C	B
NPU	QNN	C	B	C	C

S（Support and work well）：深度优化，推荐使用
A（Support and work well）：支持良好，可以使用
B（Support but has bug or not optimized）：支持但有bug或未优化，不推荐使用
C（Not Support）：不支持

1.1 CPU后端系列

1.1.1 x86/x64-SSE4.1后端

技术背景： SSE4.1（Streaming SIMD Extensions 4.1）是Intel在2007年推出的SIMD指令集扩展，提供了47条新指令，主要用于加速多媒体和浮点运算。

适用设备：一般是较老CPU

Intel Core系列处理器（2008年后）
AMD处理器（部分支持）
服务器和桌面PC
部分笔记本电脑

技术特点：

128位SIMD寄存器
支持打包整数和浮点运算
提供点积、混合、提取和插入操作

1.1.2 x86/x64-AVX2后端

技术背景： AVX2（Advanced Vector Extensions 2）是Intel在2013年推出的256位SIMD指令集，是AVX的增强版本，提供了更宽的向量寄存器和更多的指令。

适用设备：主流CPU

Intel Haswell架构及以后的处理器
AMD Excavator架构及以后的处理器
现代服务器和高性能工作站
游戏PC和高端笔记本

1.1.3 x86/x64-AVX512后端

技术背景： AVX512是Intel最新的512位SIMD指令集，首次出现在Xeon Phi处理器中，后来扩展到Xeon和Core系列。

适用设备：主流高性能CPU

Intel Xeon Scalable处理器
Intel Core i7/i9高端处理器（部分型号）
高性能计算服务器
AI训练和推理专用硬件

1.1.4 ARMv7a后端

技术背景： ARMv7-A是ARM公司的32位架构，广泛应用于早期智能手机和嵌入式设备。支持NEON SIMD指令集。

适用设备：一般是比较老旧的移动端CPU

早期Android手机（2010-2015年）
Raspberry Pi 2
部分嵌入式开发板
工业控制设备

1.1.5 ARMv8后端

技术背景： ARMv8-A是ARM的64位架构，是现代移动设备的主流架构，提供了更强的性能和更大的内存寻址空间。

适用设备：当前主流的移动端CPU

现代智能手机（iPhone 5s以后，Android旗舰机）
平板电脑
ARM服务器（如AWS Graviton）
Apple M系列芯片设备
Raspberry Pi 3/4

1.2 GPU后端系列

1.2.1 OpenCL后端

技术背景： OpenCL（Open Computing Language）是Khronos Group制定的开放标准，用于异构计算平台的并行编程。

适用设备：除专用的GPU后端外，主流的GPU后端

支持OpenCL的GPU（NVIDIA、AMD、Intel）
部分移动GPU（Adreno、Mali、PowerVR）
FPGA设备
DSP处理器

1.2.2 Vulkan后端

技术背景： Vulkan是Khronos Group开发的低开销、跨平台的3D图形和计算API，提供了更直接的GPU控制。

适用设备：大部分GPU支持的后端

现代GPU（NVIDIA GTX 900系列以后）
AMD GCN架构GPU
Intel集成显卡（部分支持）
Android设备（API 24+）

1.2.3 Metal后端

技术背景： Metal是Apple开发的低级图形和计算API，专为Apple设备优化，提供了接近硬件的性能。

适用设备：ios设备的GPU后端

iPhone iPad Mac等

1.2.4 CUDA后端

技术背景： CUDA（Compute Unified Device Architecture）是NVIDIA开发的并行计算平台和编程模型，专为NVIDIA GPU设计。

适用设备：NVIDIA GPU的专用后端

NVIDIA GeForce系列GPU
NVIDIA Quadro专业卡
NVIDIA Tesla计算卡
NVIDIA Jetson嵌入式平台

1.2.5 OpenGL后端

技术背景：OpenGL（Open Graphics Library）是一个图形渲染 API 标准，可以利用 GPU 并行计算能力处理非图形任务

适用设备：大部分GPU支持的后端

支持OpenCL的GPU（NVIDIA等）
部分移动GPU（Adreno等）

1.3 专用加速器

1.3.1 CoreML后端

技术背景： Core ML是Apple的机器学习框架，能够利用Apple设备的Neural Engine进行AI推理加速。

适用设备：当前主流的apple ai加速器

iPhone（A11芯片以后，支持Neural Engine）
iPad Pro（A12芯片等）
Mac（M1芯片等）

1.3.2 HIAI后端

技术背景： HIAI（Huawei Intelligent Acceleration Infrastructure）是华为开发的AI加速平台，利用华为麒麟芯片的NPU。

适用设备：当前主流的华为 ai加速器

华为/荣耀手机（麒麟970以后）
华为平板电脑
华为笔记本电脑（部分型号）

1.3.3 NNAPI后端

技术背景： NNAPI（Neural Networks API）是Google在Android 8.1中引入的API，为Android设备提供统一的AI加速接口。

适用设备：在 Android 15 中已弃用。在专用供应商驱动程序的Android设备运行，缺乏时运行在 CPU 上

Android 8.1+设备

1.3.4 QNN后端

技术背景： QNN（Qualcomm Neural Network SDK）是高通开发的AI推理SDK，专为高通Snapdragon平台优化。

适用设备：主流的高通芯片NPU后端

搭载Snapdragon处理器的设备
支持Hexagon DSP的设备
高通AI开发套件

1.3.5 NeuroPilot 后端

技术背景： NeuroPilot SDK是联发科自研的专用 AI 加速硬件单元SDK，集成于天玑（Dimensity）系列 SoC 中。主要通过 NNAPI 间接调用 APU 硬件加速

适用设备：主流的天玑芯片APU后端

搭载天玑系列处理器的设备
支持 NeuroPilot SDK 且 Android 10+（API 29+）的联发科平台设备

1.4 后端相关目录结构

MNN的后端实现集中在source/backend/目录中：

source/backend/
├── cpu/          # CPU通用后端
|    └───x86_x64
|    		└─── avx/          # x86 AVX2优化
|    		└─── avx512/       # x86 AVX512优化
|    		└─── avxfma/       # x86 AVXfma 支持Fused Multiply-Add
|	 		└─── sse/          # x86 SSE优化
|    └─── arm
|    		└───arm32		   # 32为arm指令支持 如ARMv7-A后端
|    		└───arm64		   # 64为arm指令支持
├── arm82/        # ARMv8.2 支持
|    └─── asm
|    		└───arm32		   # 32为arm指令兼容
|    		└───arm64		   # 64为arm指令支持
├── nnapi/        # NNAPI后端
├── metal/        # Metal后端（iOS/macOS）
├── opencl/       # OpenCL后端
├── opengl/       # OpenGL后端
├── vulkan/       # Vulkan后端
├── cuda/         # CUDA后端
├── tensorrt
├── coreml/       # CoreML后端
├── hiai/         # HIAI后端
└── qnn/          # QNN后端
└── neuropilot/   # neuropilot后端

相应的后端需要设置指定的编译宏，例如-DMNN_OPENCL=true表示启用OPENCL后端支持，但是具体调用需要在程序中设置，例如MNN-LLM需要设置… 具体见后续的代码梳理

1.4.1 后端文件说明

MNN 的后端系统由以下几个核心类组成，大致执行顺序如下：

// 1. 创建 下面的关系只表示时间上的顺序依赖关系 如Backend需要调用Runtime创建
Runtime (运行时, 各类后端继承Runtime, 如class CPURuntime : public Runtime)
  ↓ 创建
Backend (后端, 各类后端继承Backend, 如class CPUBackend : public Backend)
  ↓ 创建
Execution (执行器, 大部分算子实现是通过继承Execution实现,如:class CPUAttention : public Execution;
			小部分为汇编优化过的代码, 如source/backend/cpu/x86_x64/avx512/_AVX512_MNNGemmFloatUnit16x8.S)
			
// 2. 重新计算输出形状 在/workspace/code/MNN/source/shape/SizeComputer.hpp内完成 

// 3. 根据输入长度变化 执行的resize操作 下面的关系只表示时间上的顺序依赖关系
Backend::onResizeBegin (resize的准备阶段)
  ↓
Execution::onResize (算子的resize)
  ↓
Backend::onResizeEnd (resize的收尾阶段)

// 4. 执行算子  下面的关系只表示时间上的顺序依赖关系
Backend::onExecuteBegin (execute的准备阶段)
  ↓
Execution::onExecute (算子的execute)
  ↓
Backend::onExecuteEnd (execute的收尾阶段)

1.4.1.1 Runtime - 运行时抽象层

位置：source/core/Backend.hpp

Runtime 是硬件运行时的抽象层，负责管理整个后端的生命周期和资源配置。

核心职责：

创建 Backend 实例
管理线程池和内存分配器
提供垃圾回收机制
支持异步执行和并发控制

关键接口：

class Runtime : public NonCopyable {
public:
    // 创建 Backend
    virtual Backend* onCreate(const BackendConfig* config, Backend* origin) const = 0;
    // 重置运行时配置
    virtual void onReset(int numberThread, const BackendConfig* config, bool full);
    
    // 垃圾回收
    virtual void onGabageCollect(int level) = 0;
    // 获取内存使用量
    virtual float onGetMemoryInMB();
	
    // 主要是cpuruntime使用 cpu后端实现了线程池(source/backend/cpu/ThreadPool.hpp)
    virtual void onConcurrencyBegin() const = 0;
    virtual void onConcurrencyEnd() const = 0;
	
    // 执行优化, mnn以算子图形式运行 下面是不同的运行方式
    enum CompilerType {
        Compiler_Geometry = 0, // 部分执行几何计算，分解形变算子，但不分解 BatchMatMul / Gather 等算子
        Compiler_Origin = 1, // 直接使用原始算子，不进行分解
        Compiler_Loop = 2, // 完全执行几何计算，仅此模式下，可以在算子不支持时自动回退到CPU计算
    };
private:    
    // 记录运行时信息 如kvcacheSizeLimit cpuIds等
    RuntimeHint mHint;
};

1.4.1.2 Backend - 后端抽象基类

位置：source/core/Backend.hpp

Backend 是所有硬件后端的基类，定义了后端必须实现的接口。

核心职责：

管理内存分配和释放
创建 Execution
处理张量缓冲区操作
支持多种存储策略

存储类型枚举：

enum StorageType {
    STATIC,           // 不可重用内存，分配后立即释放
    DYNAMIC,          // 可重用内存，优先重用已有内存
    DYNAMIC_SEPERATE, // 不可重用内存，但延迟释放
    DYNAMIC_IN_EXECUTION // 执行时动态分配
};

关键接口：

class Backend : public NonCopyable {
public:
    // 创建 Execution 
    virtual Execution* onCreate(const std::vector<Tensor*>& inputs,
                                const std::vector<Tensor*>& outputs,
                                const MNN::Op* op) = 0;

    // 执行算子Resize 的开始/结束的一些准备/收尾工作
    virtual void onResizeBegin();
    virtual ErrorCode onResizeEnd() = 0;
    // 执行算子execute 的开始/结束的一些准备/收尾工作
    // 例如OpenCLBackend中 
    virtual void onExecuteBegin() const = 0;
    virtual void onExecuteEnd() const = 0;

    // 内存管理
    virtual MemObj* onAcquire(const Tensor* tensor, StorageType storageType) = 0;
    virtual bool onClearBuffer() = 0;
    virtual void onCopyBuffer(const Tensor* srcTensor, const Tensor* dstTensor) const = 0;
    
    // 把数据映射到指定后端 / 从指定后端映射回数据
    virtual void* onMapTensor(Tensor::MapType mtype, Tensor::DimensionType dtype, const Tensor* srcTensor);
    virtual bool onUnmapTensor(Tensor::MapType mtype, Tensor::DimensionType dtype, const Tensor* dstTensor, void* mapPtr);
    
    // 利用backend后端传递指针, llm中主要传递kvmeta(kv cache有关信息)
    void setMetaPtr(void* ptr);

private:
    // 枚举类 记录执行后端 如cpu opencl为: MNN_FORWARD_CPU MNN_FORWARD_OPENCL
    const MNNForwardType mType;
};

1.4.1.3 Execution - 执行器抽象基类

位置：source/core/Execution.hpp

Execution 是算子执行的抽象基类，每个算子都有对应的 Execution 实现。

核心职责：

响应输入输出张量的形状变化
执行算子计算
支持克隆和共享权重

关键接口：

class Execution : public NonCopyable {
public:
    // 根据输入seq_len长度变化 调整执行时用到的值
    virtual ErrorCode onResize(const std::vector<Tensor*>& inputs,
                               const std::vector<Tensor*>& outputs);

    // 执行计算
    virtual ErrorCode onExecute(const std::vector<Tensor*>& inputs,
                                const std::vector<Tensor*>& outputs) = 0;

    // 克隆执行器 这在llm推理中很有用 例如prefill和decode之间的seq_len输入长度不一样 就会触发克隆 mnn默认保存不同输入长度的计算图
    // 这里通常不克隆权重, 权重在Op* op中 这里会通过指针引用
    virtual bool onClone(Backend* bn, const Op* op, Execution** dst);

    // 获取后端 例如在CPUAttention中利用这里获取kvmeta信息(mMeta = (KVMeta*)(backend->getMetaPtr());)
    Backend* backend() const;
};

1.4.2 后端调用设置

除了后端架构中的 Runtime、Backend、Execution 等核心类，MNN中还有部分特定基础算子的优化代码，例如source/backend/cpu/x86_x64/avx512/_AVX512_MNNGemmFloatUnit16x8.S 等，前者是汇编代码，后者是opencl的代码。

这里指的算子一般都是底层支持的基础算子，会在更上层的算子中被调用，例如CPUAttention中会调用core->Int8GemmKernel;以使用int8矩阵乘算子

1.4.2.1 Core Function

下面以CPU的后端在实例化过程中调用的不同后端的指令集为例说明，在source/backend/cpu/compute/CommonOptFunction.h中定义了CoreFunctions

struct CoreFunctions {
    // CPU 特性标志
    bool supportFp16arith = false;  // 支持 FP16 算术运算
    bool supportSDot = false;       // 支持 ARM S-Dot 指令
    bool supportI8mm = false;       // 支持 ARM I8MM 指令
    bool supportSME2 = false;       // 支持 ARM SME2 指令
    int  smeCoreNumber = 0;         // SME2 核心数量

    // 矩阵乘法相关函数
    void(*MNNGetMatMulPackMode)(int* eP, int *lP, int* hP);
    void(*MNNPackC4ForMatMul_A)(float* destOrigin, float const** sourceGroup,
                                 const int32_t* info, const int32_t* el);
    void(*MNNPackForMatMul_B)(float* dest, const float* source,
                              size_t h, size_t kernelsize, size_t ic, bool transpose);
    void(*MNNPackedMatMul)(float* C, const float* A, const float* B,
                           const size_t* parameter, const float* postParameters,
                           const float* bias, const float* k, const float* b);
    void(*MNNPackedMatMulRemain)(float* C, const float* A, const float* B,
                                 size_t eSize, const size_t* parameter,
                                 const float* postParameters, const float* bias,
                                 const float* k, const float* b);
    // 激活函数
    void(*MNNReluInt8)(int8_t* dst, const int8_t* src, size_t size, ssize_t zeroPoint);
    void(*MNNHardSwish)(float* dst, const float* src, size_t size);
    void(*MNNGelu)(float* dst, const float* src, size_t size, float* parameters);
    const float *beta, float epsilon, size_t size, bool RMSNorm);

    // 其他函数...
};

MNN 使用全局单例 gCoreFunction 来存储当前平台的最优函数实现，在程序启动时会选择各个函数的最优实现代码

// 在source/backend/cpu/CPUBackend.cpp 中创建runtime时 会进入到不同指令集的初始化注册过程
{
#ifdef MNN_SUPPORT_BF16
extern void registerBF16Backend();
#endif
#ifdef ENABLE_ARMV82
extern void registerArm82RuntimeCreator();
#endif
void registerCPURuntimeCreator() {
    MNNCoreFunctionInit();
    CPUBackend::initCreatorMap();
    registerCPUOps();
#ifdef MNN_SUPPORT_BF16
    registerBF16Backend();
#endif
#ifdef MNN_USE_ARMV82
    registerArm82RuntimeCreator();
#endif
    // TODO: Merge _initCoreFunction MNNFunctionInit and cpuinfo_arm_init
    MNNInsertExtraRuntimeCreator(MNN_FORWARD_CPU, new CPURuntimeCreator);
}

// 下面以armv82的初始化为例 在Arm82Functions::init()中
Arm82Functions::init(){
    
    // 部分特殊优化的底层代码
    FUNC_PTR_ASSIGN(gInstance->MNNPackC4ForMatMul_A, Arm82MNNPackForMatMul_A);
    /* 在声明中 这个代码是外部导入的, 路径位置source/backend/arm82/asm/arm64/Arm82MNNPackForMatMul_A.S
    extern "C" {
    // (UP_DIV(l,8), e, 8) -> (UP_DIV(e,eP), l, eP)
        void Arm82MNNPackForMatMul_A(float* destOrigin, float const** sourceGroup, const int32_t* info, const int32_t* el);
    }
    */
    
    // 其它代码...
}

在cpu后端使用时通过调用后端的core function执行底层算子，如

// source/backend/cpu/CPUAttention.cpp
ErrorCode CPUAttention::onExecute(const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs) {
    auto gcore  = static_cast<CPUBackend *>(backend())->functions();
    auto core   = static_cast<CPUBackend*>(backend())->int8Functions();
    // 其它代码...
 	
    // 调用后端的core functions终端矩阵乘法
    gcore->MNNPackedMatMul(...);
    
    // 其它代码...
}

其它后端的底层代码调用可能不同，如opencl后端的.cl代码source/backend/opencl/execution/cl/attention_buf.cl等通过运行时加载 Kernel 源码方式使用。

MNN 核心类介绍

Thu, 26 Feb 2026 23:00:00 +0800

MNN 介绍

MNN框架作为一个高性能的深度学习推理引擎，核心设计围绕着几个关键的抽象类展开。其中，Varp，Expr，Op等类是整个框架的基石，它们不仅定义了数据的表示方式，还构建了计算图的基本结构。本文将深入分析这些核心类的设计理念、实现细节以及它们之间的相互关系。

MNN 介绍
- 1. MNN核心类

1. MNN核心类

MNN 中的核心类主要包括：

类名	职责
`VARP`	智能指针包装的 Variable，表示表达式中的变量节点
`Variable`	表达式图中的变量节点，持有张量数据或计算信息
`Expr`	表达式边，表示一个计算操作以及输入节点和输出节点
`Tensor`	张量数据容器，存储实际的多维数据
`Op`	算子描述符，定义计算的类型和参数、模型权重等
`Pipeline`	计算流水线，TODO:
`Session`	执行会话，推理数据的持有者, TODO
`Interpreter`	解释器, 模型数据的持有者
`Executor`	执行器， TODO
`ExecutorScope`	执行器作用域， TODO

1.1 类之间关系

**逻辑依赖关系： **VARP和Expr是MNN计算图格式的核心，分别表示计算图的节点和边，Op类是整个计算图的核心，表示计算图的边Expr的计算操作。关系图如下：

VARP (Variable Ptr)
    ↓ 指向
Variable
    ↓ 包含
Expr ──→ Op (操作描述)
    ↓
Tensor (存储数据)

每个边Expr都有属性std::vector<VARP> mInputs;指示输入来自的节点; 每个VARP节点有属性EXPRP mFrom; // typedef std::shared_ptr<Expr> EXPRP;指示输入的边，同时VARP的节点的Tensor数据也存放在入边mFrom中，数据读取和写入在Variable::readMap和Variable::writeMap中，需要转换成对应数据格式的指针进行读取/写入。一个简单的计算图可以如下：

Expr (mFrom) -> Tensor x(data)        Expr (mFrom) -> Tensor y(data) 
         	│									    │
         	│									    │
         VARP x                    				  VARP y
            ↓                        	  		    ↓
            └────────────────────┬──────────────────┘
                                 │
                				 │ Expr -------> Op(Add)
	                             │   └─--------> Tensor z
	                             ↓
	             			  VARP z

执行依赖关系： MNN的计算图有两种模式，Defer(延迟计算)模式或Eager(立即计算)模式：Defer模式下，调用表达式相关API不直接计算，而是搭建模型，在需要获取输出值时才执行；Eager模式下，直接进行计算，对应地无法搭建模型。下面以Eager模式为例，梳理MNN中表达式计算顺序，更具体的代码分析见readmap详解:

用户定义VARP计算代码 // 如：VARP x = _Input({2, 3});VARP y = _Input({2, 3});VARP z = _Add(x, y);
    ↓
varp->readmap()  // 触发计算
    ↓
ExecutorScope::Current()->computeInfo() //计算节点信息  ─→ SizeComputer::computeOutputSize() // 动态计算中间节点/输出的形状
    ↓
ExecutorScope::Current()->makeCache() // 计算缓存, 可复用, 按数据依赖顺序准备中间节点的Tensor 以及Session会话的计算信息
    ↓
Executor::ComputeCache::compute() // 按数据依赖顺序执行计算缓存中: 算子的resize和执行。 算子执行会从计算缓存的mSession进入下一步
    ↓
Session::run() // 这里继续进入Session::mPipelines的执行
    ↓
Pipeline::execute() // 这里执行execution->onExecute 进入算子的后端执行
    ↓
后端执行计算	// 计算结果存在tensor中 最后从readmap读取为指针
    ↓
获得指针

这当中主要配置的信息是在ExecutorScope::Current()->makeCache()中配置Session，后续的执行大多依赖Session的数据

1.2 VARP类

VARP本质是 Variable 的智能指针包装类，利用Variable指针的地址重载了比较运算符，核心代码如下：

// express/Expr.cpp
class MNN_PUBLIC VARP {
public:
    // 构造函数
    VARP();
    VARP(std::shared_ptr<Variable> c);
    VARP(Variable* c);
    
    // 获取内部 Variable 变量
    Variable* get() const;
    
    // 重载运算符
    bool operator==(const VARP& var) const = 0;
    bool operator<(const VARP& var) const = 0;
    bool operator<=(const VARP& var) const = 0;
	
    // ... 其它代码  

    // 类型标记
    enum InputType {
        INPUT = 0,      // 输入变量
        CONSTANT = 1,   // 常量变量
        TRAINABLE = 2   // 可训练变量
    };
    // 固定变量类型
    bool fix(InputType type) const;
    // 设置数据格式
    void setOrder(Dimensionformat format); // enum Dimensionformat { NHWC, NC4HW4, NCHW };
    
private:
    std::shared_ptr<Variable> mContent;
};

1.2.1 Variable 类

描述算子图的一个节点，主要刻画了对节点和数据的部分操作，核心代码如下：

// express/Expr.cpp
class MNN_PUBLIC Variable {
public:
    // 获取节点名称
    const std::string& name() const;
    void setName(const std::string& name);
    
    // 获取数据信息
    const Info* getInfo();
	struct Info {
        Dimensionformat order;  // 数据格式（NCHW/NHWC/NC4HW4）
        INTS dim;              // 维度
        halide_type_t type;    // 数据类型
        size_t size;           // 元素总数
        void syncSize();       // 同步大小
    };
    
    // 设置和获取(输入边 以及 数据是“输入边”的第几个输出Tensor)
	void setExpr(EXPRP expr, int index);
    std::pair<EXPRP, int> expr() const;
    
    // 获取 Tensor
    const Tensor* getTensor() const;
    
    // 调整大小
    bool resize(INTS dims);
    
    // 读取/保存节点到文件
	static std::vector<VARP> load(const char* fileName);
    static void save(const std::vector<VARP>& vars, NetT* dest);
    
    // 获得数据的指针映射
    template <typename T>
    const T* readMap();
    template <typename T>
    T* writeMap();
};

1.2.2 readMap详解

接下来详细解释readMap是怎么执行的，TODO

1.3 Expr类

Expr表示计算图的边，核心属性包括边的输入(VARP) 输出(Tensor) 以及计算算子(Op)，这里的输出是Tensor格式，被输出节点VARP通过边引用，核心代码如下:

// express/Expr.cpp
class MNN_PUBLIC Expr {
public:
    struct Inside;
    enum MemoryType {
        COPY,
        MOVE,
        REF
    };
    // 多种构造函数 使用各种信息构造Expr类
    static EXPRP create(Tensor* tensor, bool own = false);
    static EXPRP create(Variable::Info&& info, const void* ptr, VARP::InputType type, MemoryType copy = COPY);
    static EXPRP create(const OpT* op, std::vector<VARP> inputs, int outputSize = 1);
    static EXPRP create(std::shared_ptr<BufferStorage> extra, std::vector<VARP>&& inputs, int outputSize = 1);
    static EXPRP create(std::unique_ptr<OpT>&& op, std::vector<VARP> inputs, int outputSize = 1);
    
    // 获取/设置节点信息
    const Op* get() const;
    const std::vector<VARP>& inputs() const = 0;
    int outputSize() const = 0;
    void setName(const std::string& name);
    const std::string& name() const = 0;
    const std::string& outputName(int index);
    static void replace(EXPRP oldExpr, EXPRP newExpr);
    VARP::InputType inputType() const;
    
    // 判断当前操作依赖的输入节点 以及输入节点往前依赖的所有节点的信息是否正确, 过程中会推导节点的shape
    bool requireInfo();
    
    // 遍历，过程中会调用用户传入的 回调函数
    void visitOutputs(const std::function<bool(EXPRP, int)>& visit);
    static void visit(EXPRP expr, const std::function<bool(EXPRP)>& before, const std::function<bool(EXPRP)>& after);
   

private:
    const Op* mOp;				// 表示的计算算子
    std::vector<VARP> mInputs;	// 输入节点
	
    // 节点信息
    VARP::InputType mType;
    std::string mName;
    std::vector<std::string> mOutputNames;
    // 其中inside记录了很多更多有关边的信息
    std::shared_ptr<Inside> mInside = nullptr;
    struct Expr::Inside {
		// 构造 析构函数
        Inside(int outputSize);
        Inside(Tensor* tensor, bool own = false);
        ~ Inside();
        
        std::vector<Variable::Info> mOutputInfos;	// 输出tensor的信息 与tensor一一对应
        std::vector<Tensor*> mOutputTensors;	// 边对应的输出tensor	
        
        std::shared_ptr<Executor::ComputeCache> mCache; // 计算cache缓存 在1.1 类之间关系提到 通过cache 
        int mCacheOffset = 0;	// 用来索引tensor
        
        // 标记位
        Executor::Requirement mReq;	
        bool mInfoDirty = true;
        bool mContentDirty = true;
        bool mOwnTensor = true;
        
        Tensor* mHostTensor = nullptr;			// 设备内存（GPU/NPU）中的tensor
        std::shared_ptr<Backend> mHoldBackend;	// 持有数据的后端
    };
};

1.3.1 requireInfo 详解

TODO:

1.4 Tensor类

张量数据类，包括数据指针，数据格式，数据维度等信息，所有属性存储在下面两个结构体对象中，

// include/MNN/Tensor.hpp
class MNN_PUBLIC Tensor{
// 其它代码
private:
    halide_buffer_t mBuffer;
    struct InsideDescribe* mDescribe;
};

1.4.1 数据格式

MNN支持常见的数据格式，其中N C H W分别表示批次大小通道数高度宽度。这个格式是对图片格式的兼容，在大模型推理中输入embedding的shape是(batch_size, seq_len, hidden_size)，依次是输入、序列长度和隐藏层大小，不需要太关注这个NCHW，（大模型推理过程中的shape变化可以参考这篇介绍）。

// include/MNN/Tensor.hpp
// 维度类型
    enum DimensionType {
        TENSORFLOW,  // TensorFlow 格式：NHWC
        CAFFE,       // Caffe 格式：NCHW
        CAFFE_C4     // Caffe 格式：NC4HW4（4通道对齐）
    };

值得一提的是 MNN默认只支持1条数据输入，输入embedding维度是(seq_len, hidden_size)，但是除了Attention算子的大部分算子都做了batch维度的适配，对MNN中大模型推理感兴趣可以看这篇介绍。

我尝试在MNN框架上做了Chunk prefill，把不同输入请求合并在seq_len维度上，并在Attention算子中展开:传送门

1.4.2 底层数据格式

从更底层出发，这个数据格式信息由下面结构体存储，对应Tensor类的halide_buffer_t mBuffer属性，其中存储了数据的指针、数据类型、数据维度等信息，核心代码如下：

// include/MNN/HalideRuntime.h
/**
 * The raw representation of an image passed around by generated
 * Halide code. It includes some stuff to track whether the image is
 * not actually in main memory, but instead on a device (like a
 * GPU). For a more convenient C++ wrapper, use Halide::Buffer<T>. */
typedef struct halide_buffer_t {
    /** A device-handle for e.g. GPU memory used to back this buffer. */
    uint64_t device; // 设备句柄

    /** The interface used to interpret the above handle. */
    const struct halide_device_interface_t *device_interface; // 接口指针

    /** A pointer to the start of the data in main memory. In terms of
     * the Halide coordinate system, this is the address of the min
     * coordinates (defined below). */
    uint8_t* host; // 指针 指向数据

    /** flags with various meanings. */
    uint64_t flags;	

    /** The type of each buffer element. */
    struct halide_type_t type;	// 数据类型

    /** The dimensionality of the buffer. */
    int32_t dimensions;	// 数据维度，如[batch_size, seq_len, hidden_size]大小的Tensor数据就是2维

    /** The shape of the buffer. Halide does not own this array - you
     * must manage the memory for it yourself. */
    halide_dimension_t *dim; // 数据各维度的数值, 如如[batch_size, seq_len, hidden_size]大小的Tensor数据 dim[0]就表示batch_size维度的信息

    /** Pads the buffer up to a multiple of 8 bytes */
    void *padding; // 用来对齐内存
} halide_buffer_t;

其中的数据类型halide_type_t type通过数据占比特数和数据的性质判断数据的类型，例如code = halide_type_float并且bits = 16表示半精度浮点数。核心代码如下：

// include/MNN/HalideRuntime.h
/** A runtime tag for a type in the halide type system. Can be ints,
 * unsigned ints, or floats of various bit-widths (the 'bits'
 * field). Can also be vectors of the same (by setting the 'lanes'
 * field to something larger than one). This struct should be
 * exactly 32-bits in size. */
struct halide_type_t {
    /** The basic type code: signed integer, unsigned integer, or floating point. */
    
    // 这个code表示数据的性质 见本代码块最下方结构体，例如 halide_type_int = 0, 表示 signed integers 有符号整型
#ifndef _MSC_VER
    HALIDE_ATTRIBUTE_ALIGN(1) halide_type_code_t code; // halide_type_code_t
#else
    HALIDE_ATTRIBUTE_ALIGN(1) uint8_t code; // halide_type_code_t
#endif

    /** The number of bits of precision of a single scalar value of this type. */
    HALIDE_ATTRIBUTE_ALIGN(1) uint8_t bits;	// 数据占用比特数

    /** How many elements in a vector. This is 1 for scalar types. */
    HALIDE_ATTRIBUTE_ALIGN(2) uint16_t lanes; // 一次处理的数据宽度 用于SIMD

    // 构造函数
#ifdef __cplusplus
    /** Construct a runtime representation of a Halide type from:
     * code: The fundamental type from an enum.
     * bits: The bit size of one element.
     * lanes: The number of vector elements in the type. */
    HALIDE_ALWAYS_INLINE halide_type_t(halide_type_code_t code, uint8_t bits, uint16_t lanes = 1)
        : code(code), bits(bits), lanes(lanes) {
    }
    /** Default constructor is required e.g. to declare halide_trace_event
     * instances. */
    HALIDE_ALWAYS_INLINE halide_type_t() : code((halide_type_code_t)0), bits(0), lanes(0) {}

    // 重载了对比函数
    /** Compare two types for equality. */
    HALIDE_ALWAYS_INLINE bool operator==(const halide_type_t &other) const {
        return (code == other.code &&
                bits == other.bits &&
                lanes == other.lanes);
    }
    HALIDE_ALWAYS_INLINE bool operator!=(const halide_type_t &other) const {
        return !(*this == other);
    }
    
	// 单个数据占据的内存字节数, 按8bit向上对齐
    /** Size in bytes for a single element, even if width is not 1, of this type. */
    HALIDE_ALWAYS_INLINE int bytes() const { return (bits + 7) / 8; }
#endif
};


typedef enum halide_type_code_t
{
    halide_type_int = 0,   //!< signed integers
    halide_type_uint = 1,  //!< unsigned integers
    halide_type_float = 2, //!< IEEE floating point numbers
    halide_type_handle = 3, //!< opaque pointer type (void *)
    halide_type_bfloat = 4  //!< floating point numbers in the bfloat format
} halide_type_code_t;

这里数据维度信息halide_dimension_t *dim 包含该维度的元素个数(extend)和在该维度移动一步时内存地址的偏移量信息(stride，通常在TensorUtils::setLinearLayout中利用extend计算stride)。核心代码如下：

// include/MNN/HalideRuntime.h
typedef struct halide_dimension_t { 
    // extend: 该维度的元素个数
    // stride: 在该维度移动一步时内存地址的偏移量信息
    int32_t min, extent, stride;

    // Per-dimension flags. None are defined yet (This is reserved for future use).
    uint32_t flags;

#ifdef __cplusplus
    HALIDE_ALWAYS_INLINE halide_dimension_t() : min(0), extent(0), stride(0), flags(0) {}
    HALIDE_ALWAYS_INLINE halide_dimension_t(int32_t m, int32_t e, int32_t s, uint32_t f = 0) :
        min(m), extent(e), stride(s), flags(f) {}

    HALIDE_ALWAYS_INLINE bool operator==(const halide_dimension_t &other) const {
        return (min == other.min) &&
            (extent == other.extent) &&
            (stride == other.stride) &&
            (flags == other.flags);
    }

    HALIDE_ALWAYS_INLINE bool operator!=(const halide_dimension_t &other) const {
        return !(*this == other);
    }
#endif
} halide_dimension_t;

1.4.2 核心接口

Tensor类的核心接口包括设置/获取数据信息、把数据映射到执行设备、调整大小等，MNN文档有详细介绍, 需要用到时可以自己查看，比较常用的调试接口是打印数据和打印形状，这里打印数据中有根据数据底层的bits和code信息自动转换成对应数据类型的指针进行打印的转化。

// include/MNN/Tensor.hpp
class MNN_PUBLIC Tensor {
public:
    /**
        * @brief print tensor data. for DEBUG use only.
        */
    void print() const;

    /**
        *@brief print tensor shape
        */
    void printShape() const;
}

1.5 Op类

Op 是算子的描述类，定义了神经网络中各种操作的类型和参数。MNN 使用内存高效的FlatBuffers库来序列化/反序列化来表示Op等信息，核心结构体是OpT，在需要修改算子信息时通常使用该结构体，通常代码中使用的就是该结构体，核心代码如下：

// schema/current/MNN_generated.h
   struct OpT : public flatbuffers::NativeTable {
     std::vector<int32_t> inputIndexes;      // 输入张量索引列表
     OpParameterUnion main;                   // 算子参数（联合体）
     std::string name;                        // 算子名称
     std::vector<int32_t> outputIndexes;     // 输出张量索引列表
     OpType type;                             // 算子类型枚举
     MNN_DATA_FORMAT defaultDimentionFormat; // 默认数据格式（NHWC/NCHW等）
     std::string externalPath;               // 外部权重路径
   };

这里OpType 是一个枚举类型，用于表示不同的算子类型，例如 Conv2D, Add, Relu 等。MNN 中定义了多个算子类型，每个类型对应一个具体的算子实现，部分代码如下：

// schema/current/MNN_generated.h
enum OpType {
    OpType_AbsVal = 0,
    OpType_QuantizedAdd = 1,
    OpType_ArgMax = 2,
    OpType_AsString = 3,
    OpType_InstanceNorm = 4,
    // ...
};

继续往底层是flatbuffers::Table的数据接口，提供了获取输入索引、算子类型等信息的接口，可以通过unpack操作转换为上层的OpT结构体，主要在读取、解析、构建算子的Execution时用到Op结构体，核心代码如下：

struct Op FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
  typedef OpT NativeTableType;
  static const flatbuffers::TypeTable *MiniReflectTypeTable() {
    return OpTypeTable();
  }
  // 获取输入索引列表
  const flatbuffers::Vector<int32_t> *inputIndexes() const {
    return GetPointer<const flatbuffers::Vector<int32_t> *>(4);
  }
  // 获取参数类型
  OpParameter main_type() const {
    return static_cast<OpParameter>(GetField<uint8_t>(6, 0));
  }
  // 获取指针
  const void *main() const {
    return GetPointer<const void *>(8);
  }
  
  // 通过一堆main_as_XXX函数根据算子类型转换成对应的参数结构体指针，例如main_as_ArgMax算子就转换成ArgMax结构体指针
  template<typename T> const T *main_as() const;
  const QuantizedAdd *main_as_QuantizedAdd() const {
    return main_type() == OpParameter_QuantizedAdd ? static_cast<const QuantizedAdd *>(main()) : nullptr;
  }
  const ArgMax *main_as_ArgMax() const {
    return main_type() == OpParameter_ArgMax ? static_cast<const ArgMax *>(main()) : nullptr;
  }
  // ... 其它代码

  // 通过FlatBuffers进行序列化和反序列化
  OpT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const;
  void UnPackTo(OpT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const;
  static flatbuffers::Offset<Op> Pack(flatbuffers::FlatBufferBuilder &_fbb, const OpT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr);
};

接下来各个算子都有一个自己的参数解析类，通过判断main_type()的参数类型来转换成对应的参数结构体指针，例如main_as_ArgMax()函数会判断参数类型是否是ArgMax，如果是就转换成ArgMax结构体指针，这里OpParameter是一个枚举类型，用于表示不同的算子参数类型，例如 QuantizedAdd, ArgMax, InstanceNorm 等，部分代码如下：

// schema/current/MNN_generated.h
enum OpParameter {
  OpParameter_NONE = 0,
  OpParameter_QuantizedAdd = 1,
  OpParameter_ArgMax = 2,
  // ... 其它代码 
};

以ArgMax为例，算子可以通过main_as_XXX转换成对应算子读取时使用的底层Flatbuffer结构体，可以继续通过UnPack函数转换成修改/写入时用到的ArgMaxT结构体，核心代码如下：

// schema/current/CaffeOp_generated.h
struct ArgMax FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
  typedef ArgMaxT NativeTableType;
  static const flatbuffers::TypeTable *MiniReflectTypeTable() {
    return ArgMaxTypeTable();
  }
  // 获取各种算子的参数
  int32_t outMaxVal() const {
    return GetField<int32_t>(4, 0);
  }
  int32_t topK() const {
    return GetField<int32_t>(6, 0);
  }
  int32_t axis() const {
    return GetField<int32_t>(8, 0);
  }
  int32_t softmaxThreshold() const {
    return GetField<int32_t>(10, 0);
  }
  bool Verify(flatbuffers::Verifier &verifier) const {
    return VerifyTableStart(verifier) &&
           VerifyField<int32_t>(verifier, 4) &&
           VerifyField<int32_t>(verifier, 6) &&
           VerifyField<int32_t>(verifier, 8) &&
           VerifyField<int32_t>(verifier, 10) &&
           verifier.EndTable();
  }
  // 序列化和反序列化
  ArgMaxT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const;
  void UnPackTo(ArgMaxT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const;
  static flatbuffers::Offset<ArgMax> Pack(flatbuffers::FlatBufferBuilder &_fbb, const ArgMaxT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr);
};

反序列化后的ArgMaxT结构体就是我们平时上层使用的算子参数结构体，属性包含了算子参数的具体数值，例如outMaxVal, topK, axis, softmaxThreshold等，核心代码如下：

struct ArgMaxT : public flatbuffers::NativeTable {
  typedef ArgMax TableType;
  int32_t outMaxVal;
  int32_t topK;
  int32_t axis;
  int32_t softmaxThreshold;
  ArgMaxT()
      : outMaxVal(0),
        topK(0),
        axis(0),
        softmaxThreshold(0) {
  }
};

从Op类再往上层就是后端算子的具体实现，通过继承Execution类实现，该类解释可见后端介绍，例如：

// source/backend/cpu/CPUArgMax.hpp
class CPUArgMax : public Execution {
public:
    enum ArgMinOrMax {
        ARGMIN,
        ARGMAX
    };
    CPUArgMax(Backend *backend, ArgMinOrMax mode, int topk, int outMaxVal, int softmaxThreshold, int axis);
    virtual ~CPUArgMax() = default;
    virtual ErrorCode onResize(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) override;
    virtual ErrorCode onExecute(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) override;

private:
    Tensor mInputBuffer;
    Tensor mOutputBuffer;
    int mTopk;
    int mOutMaxVal;
    int mSoftmaxThreshold;
    int mAxis;
    int mNum;
    int mDim;
    int mKeyExtent;
    bool mFromNHWC;
    ArgMinOrMax mMode;
};

在构造不同后端的算子实现时，会把Op结构体中的参数传入到成员变量中，例如：topK, outMaxVal, softmaxThreshold, axis等，后续在执行时就可以直接使用这些参数进行计算了，核心代码如下：

// source/backend/cpu/CPUArgMax.cpp
class CPUArgMaxCreator : public CPUBackend::Creator {
public:
    virtual Execution *onCreate(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs,
                                const MNN::Op *op, Backend *backend) const {
        auto argMax = op->main_as_ArgMax();
        if (op->type() == OpType_ArgMin) {
            return new CPUArgMax(backend, CPUArgMax::ArgMinOrMax::ARGMIN,
                    argMax->topK(), argMax->outMaxVal(), argMax->softmaxThreshold(), argMax->axis());
        } else {
            return new CPUArgMax(backend, CPUArgMax::ArgMinOrMax::ARGMAX,
                    argMax->topK(), argMax->outMaxVal(), argMax->softmaxThreshold(), argMax->axis());
        }
    }
};

总结一下：读取算子时通常用Op结构体，修改/写入时用OpT结构体。阅读代码时可以查看带T的结构体名称，例如ArgMaxT、ConvolutionT，便于阅读。

MNN中会把常见的线性层转换为卷积Convolution算子，常用算子UnaryOp和BinaryOp分别表示一元和二元算子。

常用接口：可以使用EnumNameXXX方式获取OpType、OpParameter等枚举类型的字符串名称，便于调试，例如：

// schema/current/MNN_generated.h
inline const char *EnumNameOpType(OpType e);
inline const char *EnumNameOpParameter(OpParameter e);

// schema/current/TensorflowOp_generated.h
inline const char *EnumNameBinaryOpOperation(BinaryOpOperation e);

这里的获取都是从一个静态数组中取值实现的，例如二元操作的各个名称存储在静态数组中，并且通过EnumNameBinaryOpOperation函数根据枚举值获取对应的名称字符串，核心代码如下：

// schema/current/TensorflowOp_generated.h
inline const char * const *EnumNamesBinaryOpOperation() {
  static const char * const names[] = {
    "ADD",
    "SUB",
    "MUL",
    "DIV",
    "MAX_TEMP",
    "MIN_TEMP",
    "POW",
    "REALDIV",
    "MINIMUM",
    "MAXIMUM",
    "GREATER",
    "GREATER_EQUAL",
    "LESS",
    "FLOORDIV",
    "SquaredDifference",
    "EQUAL",
    "LESS_EQUAL",
    "FLOORMOD",
    "",
    "MOD",
    "ATAN2",
    "LOGICALOR",
    "NOTEQUAL",
    "BITWISE_AND",
    "BITWISE_OR",
    "BITWISE_XOR",
    "LOGICALXOR",
    "LEFTSHIFT",
    "RIGHTSHIFT",
    nullptr
  };
  return names;
}
inline const char *EnumNameBinaryOpOperation(BinaryOpOperation e) {
  if (e < BinaryOpOperation_ADD || e > BinaryOpOperation_RIGHTSHIFT) return "";
  const size_t index = static_cast<int>(e);
  return EnumNamesBinaryOpOperation()[index];
}

1.5.1 GEMM转卷积算子的理解

MNN中会把常见的线性层转换为$1\times1$的卷积Convolution算子，

TODO:

1.6 Pipeline类

TODO:

1.7 Session类

TODO:

1.8 Executor & ExecutorScope类

MNN端侧部署教程

Thu, 26 Feb 2026 23:00:00 +0800

MNN端侧部署教程

本教程介绍 MNN 端侧推理部署的完整流程，从环境配置到核心概念讲解，再到 LLM 部署实践。

教程概览

阶段	内容	文章数
环境配置	远程调试、交叉编译	2 篇
核心概念	Backend、工厂模式、核心类	3 篇
MNN-LLM	配置、加载、推理	3 篇

1. 环境配置

搭建端侧开发调试环境的完整指南。

文章	说明	状态
远程ADB环境配置	通过端口转发链路，让内网服务器直连本地手机	✅ 完成
交叉编译环境配置	Android NDK 配置、Clangd 配置、远程调试	✅ 完成

2. 核心概念

深入理解 MNN 框架的设计理念与核心组件。

文章	说明	状态
Backend 介绍	CPU/OpenCL/Vulkan 等后端的作用和选择	✅ 完成
工厂模式介绍	MNN 中工厂模式的设计与应用	📝 TODO
核心类介绍	VARP、Expr、Op 等关键类的设计	🚧 WIP

3. MNN-LLM

端侧大语言模型部署实践。

文章	说明	状态
LLM 配置	模型配置、量化配置	📝 TODO
LLM 加载流程	模型文件到推理就绪的完整过程	📝 TODO
LLM 推理流程	Token 处理、KV Cache 管理	📝 TODO

交叉编译开发环境配置与远程调试

Thu, 26 Feb 2026 23:00:00 +0800

交叉编译开发环境配置与远程调试

本文以MNN框架为例介绍交叉编译环境中的常见设置。

交叉编译开发环境配置与远程调试

1. ANDROID_NDK配置

可执行文件在手机侧CPU上运行，需要源代码在编译过程中与ANDROID_NDK 进行交叉编译

步骤1：下载最新版软件包, 获取地址(选择对应平台最新稳定版,点击后出现 “Android 软件开发套件许可协议”，勾选后鼠标右键下载按钮可以复制下载链接)

# 服务器端运行
# 到指定目录 下载 解压 android_ndk 
cd ~ && mkdir android_ndk && cd ~/android_ndk
wget https://googledownloads.cn/android/repository/android-ndk-r29-linux.zip
unzip android-ndk-r29-linux.zip # 这个文件名和下载的版本有关
# 删除压缩包
rm android-ndk-r29-linux.zip # 这个文件名和下载的版本有关

# 进入并记住解压后的目录
# 在目录用"pwd"命令可以获取绝对路径
cd android-ndk-r29 # 这个目录和下载的版本有关 可以通过ls -lah 显示目录

步骤2：配置环境变量

# 服务器端运行
# 通常会配置ANDROID_NDK_ROOT的环境变量
# 但MNN中在project/android/build_64.sh 中使用的是$ANDROID_NDK环境变量 所以都设置一下
➜ export ANDROID_NDK="$HOME/android_ndk/android-ndk-r29" # 环境变量
➜ $ANDROID_NDK/ndk-build --version	# 验证环境变量
GNU Make 4.3
Built for x86_64-pc-linux-gnu
Copyright (C) 1988-2020 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.

# 在命令行单次export只会在当前命令行有效 新开的终端不生效
# 需要把环境变量写到shell的配置文件中 bash对应~/.bashrc zsh对应~/.zshrc
# !!注意这里重定向是重定向文件末尾 '>>' 符号
echo 'export ANDROID_NDK="$HOME/android_ndk/android-ndk-r29"' >> ~/.zshrc # 环境变量
echo 'export ANDROID_NDK_ROOT="$HOME/android_ndk/android-ndk-r29"'  >> ~/.zshrc # 环境变量
echo 'export PATH="$ANDROID_NDK:$PATH"' >> ~/.zshrc
# 验证
➜  ndk-build
Android NDK: Could not find application project directory !
Android NDK: Please define the NDK_PROJECT_PATH variable to point to it.
/root/android_ndk/android-ndk-r29/build/core/build-local.mk:151: *** Android NDK: Aborting    .  Stop.

编译测试，MNN官方文档

# 服务器端运行
# 克隆仓库
git clone https://github.com/alibaba/MNN.git
# 进入仓库
cd MNN && cd project/android
# 主库编译
# 可以根据开发服务器cpu适当调高 project/android/build_64.sh里编译的线程数
# make -j256 # 原脚本4线程编译慢
mkdir build64 && cd build64 && ../build_64.sh
# 无报错，编译通过

1.1 logcat

MNN中的MNN_PRINT MNN_ERROR等信息在android调试上需要通过logcat捕获

#include <android/log.h>
#define MNN_ERROR(format, ...) __android_log_print(ANDROID_LOG_ERROR, "MNNJNI", format, ##__VA_ARGS__)
#define MNN_PRINT(format, ...) __android_log_print(ANDROID_LOG_INFO, "MNNJNI", format, ##__VA_ARGS__)
#endif

在连接了adb的终端可以使用， logcat参考

1
2
3

adb logcat # 所有日志 可以用grep过滤
adb logcat -c # 清除缓冲区
adb logcat -s <Tag>:<Level> # 仅显示指定标签的日志 <Tag>日志的标签 <Level>日志级别 如D(Debug) W(Warning) E(Error)等

2 VSCode Clangd配置

平时使用IDEA进行cpp开发时，可以智能补全代码、快速跳转到定义/声明，交叉编译环境下可以选择使用Clangd实现。

步骤1：VSCode安装clangd插件
- 在插件商店搜索下载clangd
- 下载 C/C++插件，并且禁用该插件的补全，（这个插件的补全会与clangd插件冲突，但是cmake tools依赖该插件
  
  在设置中搜索Intelli Sense Engine，并且设置为disable

步骤2：添加编译选项

# 同上编译位置 MNN目录位置可能不同
cd MNN && cd project/android/build64
# 添加编译选项
../build_64.sh -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
# 该命令会在编译目录生成compile_commands.json 文件
# 记录了项目中每个源文件的完整编译命令， 给clangd提供编译信息

步骤3：配置clangd

配置settings.json，在项目根目录中创建文件夹.vscode，在文件夹.vscode中创建文件settings.json

// path: MNN/.vscode/settings.json 用于为当前项目配置编辑器行为和开发环境设置的文件
{
    "clangd.arguments": [
        "-j=16",	// 索引时使用16线程
        "--header-insertion=never", // 禁用自动插入头文件
        "--compile-commands-dir=${workspaceFolder}/project/android/build64", // compile_commands.json 文件路径，可用绝对路径 
        "--query-driver=/root/android_ndk/android-ndk-r29/toolchains/llvm/prebuilt/linux-x86_64/bin/clang*", //解析 NDK 的 Clang 编译器, 这个路径在compile_commands.json中的个文件编译命令"command"可以找到，带上*号表示通配符, 这个路径也和NDK安装路径有关 因人而异
        "--clang-tidy" // 代码静态分析
    ],
    "clangd.path": "/root/android_ndk/android-ndk-r29/toolchains/llvm/prebuilt/linux-x86_64/bin/clangd", //cland路径
}

这里使用ANDROID_NDK中的clangd.path

# 不同版本路径可能略有不同, 可以通过find命令查找，
# 使用找到的clangd的绝对路径
➜  cd $ANDROID_NDK      
➜  find . -type f -iname "clangd"
./toolchains/llvm/prebuilt/linux-x86_64/bin/clangd

在vscode中使用ctrl+shift+p调出命面板，搜索选择clangd: Restart language server或者Developer: Reload Window，可以在vscode中进行正常代码跳转

配置.clangd,在MNN根目录下创建文件.clangd

# 消除一些无意义的报错
Diagnostics:
  UnusedIncludes: None # 禁用“未使用头文件”警告
  Suppress:	# 屏蔽指定的警告/提示 按需使用
    - "macro_is_not_used"
    - "pp_including_mainfile_in_preamble"
    - "misleading-indentation"

Suppress屏蔽的报错信息可以把鼠标靠近报错的地方查找，例如：括号中的pp_file_not_found

部分无法索引的可能因为编译宏未打开，例如llm.hpp需要打开-DMNN_BUILD_LLM=true才会编译，查找文件是否编译可以直接在compile_commands.json中搜索对应文件。

步骤4：最后的目录结构

MNN/
├── .vscode/
│   └── settings.json
├── .clangd
└── ...（其他项目文件和目录，如 include/, source/, transformers/ 等）

3 调试

3.1 服务器调试

代码出错一直靠printf打印需要麻烦且看不到调用栈，vscode中有直接在服务器上进行逻辑调试的插件。

注意：这里服务器调试是在服务器上运行，所以编译指令与前面在移动手机运行不同，最终运行是在服务器运行

步骤1：在vscode中安装插件cmake tools
步骤2：安装完成后侧边栏多出cmake功能区
步骤3：点击选择工具包，选择在服务器测试时的编译器，如果没有编译器，需自行安装gcc clang等
步骤4：选择好工具包后，会自动扫描根目录的cmakelist,进行配置，也可以执行点击按钮配置，下图中的按钮是配置、生成所有文件

步骤5：配置cmake 生成构建系统, 如果需要添加编译宏需要在.vscode/settings.json中添加如下参数

{
    // 下面内容和前面的clagnd设置放在一个{}中
    "cmake.configureArgs": [ // 编译宏
        "-DMNN_LOW_MEMORY=true", 
        "-DMNN_CPU_WEIGHT_DEQUANT_GEMM=true",
        "-DMNN_BUILD_LLM=true", // 编译MNN-LLM模块
        "-DMNN_SUPPORT_TRANSFORMER_FUSE=true", // attention模块融合成算子 而不是进一步拆成算子
    ],
    "cmake.buildArgs" : ["-j128"], // 编译线程
    "cmake.generator": "Unix Makefiles", // 使用make构建
}

步骤6：在项目大纲的配置中选择”使用cmake调试器配置所有项目“，然后打开控制面板选择Developer: Reload Window; 重新配置的对象就会刷新在左侧项目大纲
步骤7：在项目大纲这里选择想调试/生成的文件，右键鼠标可以选择调试

如果调试的可执行文件执行时需要args，需要在.vscode/settings.json中添加参数，

这里以llm_demo为例，运行llm_demo的命令是./llm_demo /data/HUGGINGFACE/Qwen3-4B-Instruct-2507-Eagle3-MNN/llm.mnn ，如果follow这个步骤需要自行下载Qwen3-4B-Instruct-2507-Eagle3-MN模型，并且复制到llm.mnn的路径
1 2 3 4 5

{ // 添加到.vscode/settings.json "cmake.debugConfig": { "args": ["/data/HUGGINGFACE/Qwen3-4B-Instruct-2507-Eagle3-MNN/llm.mnn"], } }
步骤8：在程序中打断点按照上一步开始调试。如果没有调试器，需自行安装gdb lldb等；这一步需要c/c++插件，未安装见[VSCode Clangd配置](#2 VSCode Clangd配置), 需要禁用部分与clangd冲突功能。

在终端输入”hello“后成功卡在断点，说明配置成功。

说明：cmake tools的配置编译产物在根目录的build文件夹下

3.2 远程调试（launch模式）

部分在指定手机的后端运行的代码无法在服务器调试，需要通过远程调试。远程调试需要建立服务器到手机的通信链路，如果链路带宽有限启动调试会很慢，所以部分与运行后端无关的逻辑推荐服务器调试，并且推荐的服务器到手机的通信链路使用TailScale/EasyTier隧道的方式

原理：服务器上通过ANDROID_NDK目录下的lldb 借助服务器->手机的通信链路与手机上运行的ANDROID_NDK目录下的lldb-server通信，进行远程调试。总结需要完成的事情是: (1)找ANDROID_NDK目录下的lldb lldb-server文件， (2)做两个端口的通信转发链路，在手机运行lldb-server (3)编译把所需运行文件推送到手机 (4)下载配置CodeLLDB插件 (5) 配置调试参数开始调试
步骤1：VSCode安装CodeLLDB插件官方文档

步骤2：查找lldb lldb-server。lldb和lldb-server是一对，调试时服务器上的lldb会连接设备端中启动的lldb-server。lldb文档

# 服务器端运行
➜  cd $ANDROID_NDK                     
➜  find . -type f -iname "lldb"  
./toolchains/llvm/prebuilt/linux-x86_64/bin/lldb
➜  find . -type f -iname "lldb-server"
./toolchains/llvm/prebuilt/linux-x86_64/lib/clang/21/lib/linux/arm/lldb-server
./toolchains/llvm/prebuilt/linux-x86_64/lib/clang/21/lib/linux/i386/lldb-server
./toolchains/llvm/prebuilt/linux-x86_64/lib/clang/21/lib/linux/riscv64/lldb-server
./toolchains/llvm/prebuilt/linux-x86_64/lib/clang/21/lib/linux/x86_64/lldb-server
./toolchains/llvm/prebuilt/linux-x86_64/lib/clang/21/lib/linux/aarch64/lldb-server

步骤3：构建通信链路，lldb与lldb-server之间的通信需要启动额外的通信端口，所以构建通信链路时需要开启两个端口

# 本地运行
# 为了简化说明 adb调试端口统一设置为30303 lldb通信端口设置为31313

# adb 远程调试端口应该设置为30303
adb tcpip 30303 # 插入usb连接手机 启动无线adb调试，部分机型可能需要在开发者模式设置允许无线调试
adb connect ip_P:30303  # 验证无线adb连接，手机的ip可能会随着网络变化，可以在手机设置中搜索ip查看，或者路由器页面查看，或者adb shell ip a查看

# 下面的host_name是本地通过SSH连接服务器时config中的Host别名
# 可以修改为本地L通过ssh连接服务器S的user ip port, 不是需要转发的ip和端口

# 构建adb调试端口转发
ssh -N -R 0.0.0.0:30303:localhost:30303 host_name # 将服务器30303端口转发至本地30303端口
adb forward tcp:30303 tcp:30303 # 将本地30303端口转发给adb设备30303端口

# 构建lldb通信端口转发
ssh -N -R 0.0.0.0:31313:localhost:31313 host_name # 将服务器31313端口转发至本地31313端口
adb forward tcp:31313 tcp:31313 # 将本地31313端口转发给adb设备31313端口

步骤4：给手机端推送并启动lldb-server，由于adb连接根目录下有许多系统文件不root没有权限所以通常文件推送到/data/local/tmp目录下运行

# 服务器运行

# 1. 连接adb
# sudo apt install adb -y 若无adb
adb connect 127.0.0.1:30303 # 连接adb
# 2. 查看指令集，大部分手机是arm64-v8a架构，需要aarch64 版本的lldb-server
➜  adb shell getprop ro.product.cpu.abi
arm64-v8a

# 3. 把lldb-server推送给手机端 根据上面的查找路径 推送
➜  adb shell mkdir /data/local/tmp/mnn-test/   # 创建路径
➜  adb push $ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/lib/clang/21/lib/linux/aarch64/lldb-server /data/local/tmp/mnn-test/ # !!注意末尾的 / 表示目录

# 4. 进入adb 运行lldb-server
➜  adb shell  
RE602CL1:/ $ cd /data/local/tmp/mnn-test/                                           
RE602CL1:/data/local/tmp/mnn-test $ chmod a+x lldb-server               
RE602CL1:./lldb-server platform --server --listen "*:31313" --log-channels "lldb process" # 这里端口要对应上lldb链路的端口

# 说明可以通过下面命令清理/查看adb转发的端口列表
# adb forward --remove-all
# adb forward --list

# 5. 测试
# 另起一个shell 启动lldb
➜ cd $ANDROID_NDK
➜ export PATH=$(pwd)/toolchains/llvm/prebuilt/linux-x86_64/python3/bin:$PATH # 临时设置python路径，后面调试无需设置
➜ export LD_LIBRARY_PATH=$(pwd)/toolchains/llvm/prebuilt/linux-x86_64/python3/lib:$LD_LIBRARY_PATH # 临时设置环境变量，后面调试无需设置
# 具体路径可以根据缺失的包查找 如下
# ➜ find . -type f -iname "libpython3.11.so.1.0"
#./toolchains/llvm/prebuilt/linux-x86_64/python3/lib/libpython3.11.so.1.0
➜ adb connect 127.0.0.1:30303 # 连接adb
already connected to 127.0.0.1:30303
➜ ./toolchains/llvm/prebuilt/linux-x86_64/bin/lldb # 启动lldb
No entry for terminal type "xterm-256color";
using dumb terminal settings.
(lldb) version
lldb version 21.0.0 (/mnt/disks/build-disk/src/googleplex-android/llvm-r563880-release/out/llvm-project/lldb revision 5e96669f06077099aa41290cdb4c5e6fa0f59349)
  clang revision 5e96669f06077099aa41290cdb4c5e6fa0f59349
  llvm revision 5e96669f06077099aa41290cdb4c5e6fa0f59349
(lldb) platform select remote-android	# 选择平台
  Platform: remote-android
 Connected: no
(lldb) platform connect connect://localhost:31313 # 连接设备
  Platform: remote-android
    Triple: aarch64-unknown-linux-android
OS Version: 35 (6.6.30-android15-8-g013ec21bba94-abogki383916444-4k)
  Hostname: localhost
 Connected: yes
WorkingDir: /data/local/tmp/mnn-test
    Kernel: #1 SMP PREEMPT Tue Dec 17 23:36:49 UTC 2024
(lldb) quit # 以上流程验证成功说明链路测试成功，!!!!!quit退出当前lldb!!!!!

# -------------------------------
# 连接成功后lldb-server侧出现连接建立提示
# RE602CL1:/data/local/tmp/mnn-test $ ./lldb-server platform --server --listen "*:31313"
# Connection established.

步骤5：配置.vscode/launch.json文件，设置调试参数,依旧以llm_demo为例，部分参数需要和前面的设置一致。这里使用launch模式无法使用标准输入，样例使用llm_demo的另一个用法–文件输入。需要标准输入要使用attach模式在adb窗口运行进程并attach给lldb。

{
    "version": "0.2.0",
    "configurations": [
        {
            "name": "Android Native Debug (CodeLLDB)",
            "type": "lldb",   
            "request": "launch",
            "program": "${workspaceFolder}/project/android/build64/llm_demo", // 要调试的可执行文件
            "initCommands": [
                "platform select remote-android", // 选择远程android平台，参考lldb文档https://lldb.llvm.org/use/remote.html
                "platform connect connect://127.0.0.1:31313", // 连接lldb-server
                "settings set target.inherit-env false", // 禁用继承本地环境变量
                "platform settings -w /data/local/tmp/mnn-test", // 设置调试目录
                "platform status", // 验证连接状态
            ],
            "preRunCommands": [
                "settings set target.env-vars LD_LIBRARY_PATH=/data/local/tmp/mnn-test", //设置环境变量，MNN端侧运行需要设置环境变量
            ],
            "args": [
                "/data/local/tmp/mnn-test/Qwen3-4B-Instruct-2507-Eagle3-MNN/llm.mnn", // 调试参数
                "hello.txt
            ],
        }
    ]
}

步骤6：编译，并且推送代码和所需的动态链接库。!!注意这里一定要把编译修改为Debug才能运行

# 服务器运行
# 移动到编译目录
cd MNN && cd project/android/build64
# 编译类型为从Release改为Debug
vim ../build_64.sh # -DCMAKE_BUILD_TYPE=Debug \ # 从Release改为Debug
# 编译
../build_64.sh -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DMNN_LOW_MEMORY=true -DMNN_CPU_WEIGHT_DEQUANT_GEMM=true -DMNN_BUILD_LLM=true -DMNN_SUPPORT_TRANSFORMER_FUSE=true

# 推送可执行文件和动态库，需要的动态链接库可以运行测试 缺啥补啥
adb connect 127.0.0.1:30303
adb push llm_demo /data/local/tmp/mnn-test/
adb push /data/HUGGINGFACE/Qwen3-4B-Instruct-2507-Eagle3-MNN/ /data/local/tmp/mnn-test/Qwen3-4B-Instruct-2507-Eagle3-MNN/
adb push libllm.so /data/local/tmp/mnn-test/
adb push libMNN_Express.so /data/local/tmp/mnn-test/
adb push libMNN.so /data/local/tmp/mnn-test/
# 使用launch模式 无法使用标准输入 使用文件方式输入
echo hello > hello.txt && adb push hello.txt /data/local/tmp/mnn-test/ 

# 测试能否运行
adb shell # 进入adb
cd /data/local/tmp/mnn-test/ # 进入目录
export LD_LIBRARY_PATH=/data/local/tmp/mnn-test # 临时设置环境变量
./llm_demo /data/local/tmp/mnn-test/Qwen3-4B-Instruct-2507-Eagle3-MNN/llm.mnn # 运行测试

# 测试如下 输入hello 有输出生成 表示能够编译运行
CPU Group: [ 0  1  2  3  4  5 ], 384000 - 3532800
CPU Group: [ 6  7 ], 1017600 - 4320000
(last_midr & (CPUINFO_ARM_MIDR_IMPLEMENTER_MASK | CPUINFO_ARM_MIDR_PART_MASK))=0x 51000010 in _getInfoArm, 1234 
The device supports: i8sdot:1, fp16:1, i8mm: 1, sve2: 0, sme2: 0
config path is /data/local/tmp/mnn-test/Qwen3-4B-Instruct-2507-Eagle3-MNN/llm.mnn
main, 267, cost time: 5396.304199 ms
Prepare for tuning opt Begin
Prepare for tuning opt End
main, 275, cost time: 475.540009 ms

User: hello

A: Hello! How can I help you today? 😊

# 调试前应该有下面文件
130|RE602CL1:/data/local/tmp/mnn-test $ ls -lah
total 33K
drwxrwxrwx 3 shell shell 3.3K 2026-01-30 15:53 .
drwxrwx--x 7 shell shell 3.3K 2026-01-29 17:18 ..
drwxrwxr-x 3 shell shell 3.3K 2026-01-30 12:34 Qwen3-4B-Instruct-2507-Eagle3-MNN
-rw-rw-rw- 1 shell shell    6 2026-01-30 15:53 hello.txt
-rwxrwxrwx 1 shell shell 2.6M 2026-01-30 12:39 libMNN.so
-rwxrwxrwx 1 shell shell 714K 2026-01-30 12:40 libMNN_Express.so
-rwxrwxrwx 1 shell shell 1.4M 2026-01-30 12:40 libllm.so
-rwxrwxrwx 1 shell shell  28M 2025-10-03 04:45 lldb-server
-rwx------ 1 shell shell  30K 2026-01-30 12:40 llm_demo

步骤7：断点调试（记得把前面测试的lldb进程结束释放端口，这里设置的参数不同断点位置也修改了下）

开始调试后，会通过通信加载大量符号表，然后停在断点处

交叉编译环境，codelldb插件使用的lldb可能与ANDROID_NDK目录下的lldb-server版本不匹配,但是目前测试暂时没遇到，修改插件使用的lldb可能比较麻烦，包括lldb.library等应该都要修改（前面测试lldb用到的环境变量）

3.3 远程调试（Attach模式尝试失败）

理论上“launch模式不能同步标准输入，改用attach模式可以在adb启动调试进程并通过lldb直接attach到该进程，随后调试程序在adb中应该可以正常进行stdio行为。”但lldb命令行模式可以成功，使用vscode插件测试无法成功。

步骤1：同样的构建lldb与lldb-server链路并且在adb执行lldb-server

步骤2：配置launch.json

{
    "version": "0.2.0",
    "inputs": [
        {
            "id": "PidAttach",
            "type": "promptString",
            "description": "Enter the PID to attach to",
            "default": ""
        }
    ],
    "configurations": [
        {
            "name": "Android LLDB Attach",
            "type": "lldb",
            "request": "attach",
            "pid": "${input:PidAttach}",
            "program": "${workspaceFolder}/project/android/build64/llm_demo", 
            "initCommands": [
                "platform select remote-android",
                "platform connect connect://127.0.0.1:31313",
                "settings set target.inherit-env false",
                "platform settings -w /data/local/tmp/mnn-test",
                // "platform status",
                // "platform process list",  // 查看进程列表
                "settings set target.disable-stdio true", // 禁用lldb的stdio拦截 如果希望使用stdio参数应该也要禁用
            ],
        }
    ]
}

步骤3：启动调试，先在adb内启动llm_demo

adb shell # 进入adb
cd /data/local/tmp/mnn-test/ # 进入目录
export LD_LIBRARY_PATH=/data/local/tmp/mnn-test # 临时设置环境变量
./llm_demo /data/local/tmp/mnn-test/Qwen3-4B-Instruct-2507-Eagle3-MNN/llm.mnn # 运行测试

然后设置断点开始调试，失败log:

# 另起一个adb发现llm_demo应该成功被attach
RE602CL1:/ $ ps -a                                                                    
USER           PID  PPID        VSZ    RSS WCHAN            ADDR S NAME                
shell         9239  9226    7086452 2820732 ptrace_st+         0 t llm_demo
shell         9243  9221    8732304   6912 do_sys_po+          0 S lldb-server
shell         9246  9243    8697360   7552 do_select           0 S lldb-server
shell         9250  9243    8700968  10536 do_sys_po+          0 S lldb-server
shell         9285  9232    2289024   5864 0                   0 R ps

# 在vscode输出中找到lldb(需要在settings.json中设置"lldb.verboseLogging": true)
# 正常执行命令 并且加载了部分符号表
# 但是调试没有正常开始

vscode log

  Verbose logging: on  (Use "lldb.verboseLogging" setting to change)
  Platform: linux x64
  Initial debug configuration: {
    name: 'Android LLDB Attach',
    type: 'lldb',
    request: 'attach',
    pid: '${input:PidAttach}',
    program: '${workspaceFolder}/project/android/build64/llm_demo',
    initCommands: [
      'platform select remote-android',
      'platform connect connect://127.0.0.1:31313',
      'settings set target.inherit-env false',
      'platform settings -w /data/local/tmp/mnn-test',
      'platform status',
      'platform process list',
      'settings show target.disable-stdio',
      'settings set target.disable-stdio true',
      'settings show target.disable-stdio',
      'settings set target.preload-symbols false'
    ],
    __configurationTarget: 6
  }
  Resolved debug configuration: {
    name: 'Android LLDB Attach',
    type: 'lldb',
    request: 'attach',
    pid: '9239',
    program: '/workspace/code/MNN/project/android/build64/llm_demo',
    initCommands: [
      'platform select remote-android',
      'platform connect connect://127.0.0.1:31313',
      'settings set target.inherit-env false',
      'platform settings -w /data/local/tmp/mnn-test',
      'platform status',
      'platform process list',
      'settings show target.disable-stdio',
      'settings set target.disable-stdio true',
      'settings show target.disable-stdio',
      'settings set target.preload-symbols false'
    ],
    __configurationTarget: 6,
    relativePathBase: '/workspace/code/MNN',
    _adapterSettings: {
      displayFormat: 'auto',
      showDisassembly: 'auto',
      dereferencePointers: true,
      suppressMissingSourceFiles: true,
      evaluationTimeout: 5,
      consoleMode: 'commands',
      sourceLanguages: null,
      scriptConfig: {},
      evaluateForHovers: true,
      commandCompletions: true
    }
  }
  Launching adapter
  liblldb: 
  lldbServer: 
  environment: {}
  settings: { evaluateForHovers: true, commandCompletions: true }
  [0.10 INFO codelldb] Loaded "/root/.vscode-server/extensions/vadimcn.vscode-lldb-1.12.1/lldb/lib/liblldb.so", version="lldb version 21.1.7-codelldb"
  [0.376 DEBUG codelldb] Connecting to 127.0.0.1:35049
  [0.376 DEBUG codelldb] New debug session
  [0.402 DEBUG codelldb::python] Compiling code: codelldb.interface.update_adapter_settings("""{"displayFormat":null,"showDisassembly":null,"dereferencePointers":null,"containerSummary":null,"evaluationTimeout":null,"summaryTimeout":null,"suppressMissingSourceFiles":null,"consoleMode":null,"sourceLanguages":null,"scriptConfig":null,"evaluateForHovers":true,"commandCompletions":true}""", globals())
  [0.402 DEBUG codelldb::python] Created code object at 0x7f6fc0597110
  [0.402 DEBUG codelldb::python] Evaluating code object at 0x7f6fc0597110
  [0.402 DEBUG codelldb::python] Evaluation result: (void) result =

  [0.416 DEBUG codelldb::python] Dropping object at 0x7f6fc0597110
  [0.434 DEBUG codelldb::dap_codec] --> {"command":"initialize","arguments":{"clientID":"vscode","clientName":"Visual Studio Code","adapterID":"lldb","pathFormat":"path","linesStartAt1":true,"columnsStartAt1":true,"supportsVariableType":true,"supportsVariablePaging":true,"supportsRunInTerminalRequest":true,"locale":"zh-cn","supportsProgressReporting":true,"supportsInvalidatedEvent":true,"supportsMemoryReferences":true,"supportsArgsCanBeInterpretedByShell":true,"supportsMemoryEvent":true,"supportsStartDebuggingRequest":true,"supportsANSIStyling":true},"type":"request","seq":1}
  [0.435 DEBUG codelldb::dap_codec] <-- {"seq":1,"type":"response","request_seq":1,"success":true,"command":"initialize","body":{"exceptionBreakpointFilters":[{"default":true,"filter":"cpp_throw","label":"C++: on throw","supportsCondition":true},{"default":false,"filter":"cpp_catch","label":"C++: on catch","supportsCondition":true}],"supportTerminateDebuggee":true,"supportsCancelRequest":true,"supportsClipboardContext":true,"supportsCompletionsRequest":true,"supportsConditionalBreakpoints":true,"supportsConfigurationDoneRequest":true,"supportsDataBreakpointBytes":true,"supportsDataBreakpoints":true,"supportsDelayedStackTraceLoading":true,"supportsDisassembleRequest":true,"supportsEvaluateForHovers":true,"supportsExceptionFilterOptions":true,"supportsExceptionInfoRequest":true,"supportsFunctionBreakpoints":true,"supportsGotoTargetsRequest":true,"supportsHitConditionalBreakpoints":true,"supportsInstructionBreakpoints":true,"supportsLogPoints":true,"supportsModulesRequest":true,"supportsReadMemoryRequest":true,"supportsRestartRequest":true,"supportsSetVariable":true,"supportsStepInTargetsRequest":true,"supportsSteppingGranularity":true,"supportsWriteMemoryRequest":true}}
  [0.479 DEBUG codelldb::dap_codec] --> {"command":"attach","arguments":{"name":"Android LLDB Attach","type":"lldb","request":"attach","pid":"9239","program":"/workspace/code/MNN/project/android/build64/llm_demo","initCommands":["platform select remote-android","platform connect connect://127.0.0.1:31313","settings set target.inherit-env false","platform settings -w /data/local/tmp/mnn-test","platform status","platform process list","settings show target.disable-stdio","settings set target.disable-stdio true","settings show target.disable-stdio","settings set target.preload-symbols false"],"__configurationTarget":6,"relativePathBase":"/workspace/code/MNN","_adapterSettings":{"displayFormat":"auto","showDisassembly":"auto","dereferencePointers":true,"suppressMissingSourceFiles":true,"evaluationTimeout":5,"consoleMode":"commands","sourceLanguages":null,"scriptConfig":{},"evaluateForHovers":true,"commandCompletions":true},"__sessionId":"d48f9fb7-6d82-457d-8328-c1b37b556795"},"type":"request","seq":2}
  [0.479 DEBUG codelldb::python] Compiling code: codelldb.interface.update_adapter_settings("""{"displayFormat":"auto","showDisassembly":"auto","dereferencePointers":true,"containerSummary":null,"evaluationTimeout":5.0,"summaryTimeout":null,"suppressMissingSourceFiles":true,"consoleMode":"commands","sourceLanguages":null,"scriptConfig":{},"evaluateForHovers":true,"commandCompletions":true}""", globals())
  [0.479 DEBUG codelldb::python] Created code object at 0x7f6fc0597110
  [0.479 DEBUG codelldb::python] Evaluating code object at 0x7f6fc0597110
  [0.479 DEBUG codelldb::python] Evaluation result: (void) result =

  [0.480 DEBUG codelldb::python] Dropping object at 0x7f6fc0597110
  [0.480 DEBUG codelldb::dap_codec] <-- {"seq":2,"type":"event","event":"output","body":{"category":"console","output":"Console is in 'commands' mode, prefix expressions with '?'.\n"}}
  [0.480 DEBUG codelldb::dap_codec] <-- {"seq":3,"type":"event","event":"output","body":{"category":"console","output":"Executing script: initCommands\n"}}
  [0.480 DEBUG codelldb::debug_session] platform select remote-android -> SuccessFinishResult, Error:  Success
  Output Message:
    Platform: remote-android
   Connected: no

  [0.480 DEBUG codelldb::dap_codec] <-- {"seq":4,"type":"event","event":"output","body":{"category":"console","output":"  Platform: remote-android\n Connected: no\n\n"}}
  [1.37 DEBUG codelldb::debug_session] platform connect connect://127.0.0.1:31313 -> SuccessFinishResult, Error:  Success
  Output Message:
    Platform: remote-android
      Triple: aarch64-unknown-linux-android
  OS Version: 35 (6.6.30-android15-8-g013ec21bba94-abogki383916444-4k)
    Hostname: localhost
   Connected: yes
  WorkingDir: /data/local/tmp/mnn-test
      Kernel: #1 SMP PREEMPT Tue Dec 17 23:36:49 UTC 2024

  [1.37 DEBUG codelldb::dap_codec] <-- {"seq":5,"type":"event","event":"output","body":{"category":"console","output":"  Platform: remote-android\n    Triple: aarch64-unknown-linux-android\nOS Version: 35 (6.6.30-android15-8-g013ec21bba94-abogki383916444-4k)\n  Hostname: localhost\n Connected: yes\nWorkingDir: /data/local/tmp/mnn-test\n    Kernel: #1 SMP PREEMPT Tue Dec 17 23:36:49 UTC 2024\n\n"}}
  [1.37 DEBUG codelldb::debug_session] settings set target.inherit-env false -> SuccessFinishResult, Error:  Success
  [1.137 DEBUG codelldb::debug_session] platform settings -w /data/local/tmp/mnn-test -> SuccessFinishNoResult, Error:  Success
  [1.233 DEBUG codelldb::debug_session] platform status -> SuccessFinishResult, Error:  Success
  Output Message:
    Platform: remote-android
      Triple: aarch64-unknown-linux-android
  OS Version: 35 (6.6.30-android15-8-g013ec21bba94-abogki383916444-4k)
    Hostname: localhost
   Connected: yes
  WorkingDir: /data/local/tmp/mnn-test
      Kernel: #1 SMP PREEMPT Tue Dec 17 23:36:49 UTC 2024

  [1.234 DEBUG codelldb::dap_codec] <-- {"seq":6,"type":"event","event":"output","body":{"category":"console","output":"  Platform: remote-android\n    Triple: aarch64-unknown-linux-android\nOS Version: 35 (6.6.30-android15-8-g013ec21bba94-abogki383916444-4k)\n  Hostname: localhost\n Connected: yes\nWorkingDir: /data/local/tmp/mnn-test\n    Kernel: #1 SMP PREEMPT Tue Dec 17 23:36:49 UTC 2024\n\n"}}
  [1.947 DEBUG codelldb::debug_session] platform process list -> SuccessFinishNoResult, Error:  Success
  Output Message:
  6 matching processes were found on "remote-android"

  PID    PARENT USER       TRIPLE                         NAME
  ====== ====== ========== ============================== ============================
  3264   1      shell                                     
  9221   3264   shell      aarch64-unknown-linux-android  sh
  9226   3264   shell      aarch64-unknown-linux-android  sh
  9232   3264   shell      aarch64-unknown-linux-android  sh
  9239   9226   shell      aarch64-unknown-linux-android  llm_demo
  10599  9221   shell      aarch64-unknown-linux-android  lldb-server

  [1.947 DEBUG codelldb::debug_session] settings show target.disable-stdio -> SuccessFinishResult, Error:  Success
  Output Message:
  target.disable-stdio (boolean) = false

  [1.947 DEBUG codelldb::dap_codec] <-- {"seq":7,"type":"event","event":"output","body":{"category":"console","output":"6 matching processes were found on \"remote-android\"\n\nPID    PARENT USER       TRIPLE                         NAME\n====== ====== ========== ============================== ============================\n3264   1      shell                                     \n9221   3264   shell      aarch64-unknown-linux-android  sh\n9226   3264   shell      aarch64-unknown-linux-android  sh\n9232   3264   shell      aarch64-unknown-linux-android  sh\n9239   9226   shell      aarch64-unknown-linux-android  llm_demo\n10599  9221   shell      aarch64-unknown-linux-android  lldb-server\n\n"}}
  [1.947 DEBUG codelldb::debug_session] settings set target.disable-stdio true -> SuccessFinishResult, Error:  Success
  [1.947 DEBUG codelldb::debug_session] settings show target.disable-stdio -> SuccessFinishResult, Error:  Success
  Output Message:
  target.disable-stdio (boolean) = true

  [1.947 DEBUG codelldb::debug_session] settings set target.preload-symbols false -> SuccessFinishResult, Error:  Success
  [1.947 DEBUG codelldb::dap_codec] <-- {"seq":8,"type":"event","event":"output","body":{"category":"console","output":"target.disable-stdio (boolean) = false\n\n"}}
  [1.947 DEBUG codelldb::dap_codec] <-- {"seq":9,"type":"event","event":"output","body":{"category":"console","output":"target.disable-stdio (boolean) = true\n\n"}}
  INFO(Python) 13:09:06 lang_support: languages: set()
  [2.51 DEBUG codelldb::dap_codec] <-- {"seq":10,"type":"event","event":"initialized"}
  [2.94 DEBUG codelldb::dap_codec] --> {"command":"setBreakpoints","arguments":{"source":{"name":"llm_demo.cpp","path":"/workspace/code/MNN/transformers/llm/engine/demo/llm_demo.cpp"},"lines":[197,246,265,301],"breakpoints":[{"line":197},{"line":246},{"line":265},{"line":301}],"sourceModified":false},"type":"request","seq":3}
  [2.104 DEBUG codelldb::debug_session] Debug event: 0x55938b1e6058 Event: broadcaster = 0x55938b1cafb0 (lldb.target), type = 0x00000001 (breakpoint-changed), data = {bkpt: 1 type: breakpoint added}
  [2.104 DEBUG codelldb::debug_session] Debug event: 0x55938bbb0cd8 Event: broadcaster = 0x55938b1cafb0 (lldb.target), type = 0x00000001 (breakpoint-changed), data = {bkpt: 2 type: breakpoint added}
  [2.104 DEBUG codelldb::dap_codec] <-- {"seq":11,"type":"response","request_seq":3,"success":true,"command":"setBreakpoints","body":{"breakpoints":[{"id":1,"line":198,"message":"Resolved locations: 0","verified":true},{"id":2,"line":246,"message":"Resolved locations: 0","verified":true},{"id":3,"line":265,"message":"Resolved locations: 0","verified":true},{"id":4,"line":301,"message":"Resolved locations: 0","verified":true}]}}
  [2.104 DEBUG codelldb::debug_session] Debug event: 0x55938be74b38 Event: broadcaster = 0x55938b1cafb0 (lldb.target), type = 0x00000001 (breakpoint-changed), data = {bkpt: 3 type: breakpoint added}
  [2.104 DEBUG codelldb::debug_session] Debug event: 0x55938be759d8 Event: broadcaster = 0x55938b1cafb0 (lldb.target), type = 0x00000001 (breakpoint-changed), data = {bkpt: 4 type: breakpoint added}
  [2.104 DEBUG codelldb::dap_codec] --> {"command":"setFunctionBreakpoints","arguments":{"breakpoints":[]},"type":"request","seq":4}
  [2.104 DEBUG codelldb::dap_codec] --> {"command":"setInstructionBreakpoints","arguments":{"breakpoints":[]},"type":"request","seq":5}
  [2.104 DEBUG codelldb::dap_codec] --> {"command":"setExceptionBreakpoints","arguments":{"filters":[],"filterOptions":[{"filterId":"cpp_throw"},{"filterId":"cpp_catch"}]},"type":"request","seq":6}
  [2.104 DEBUG codelldb::dap_codec] --> {"command":"setDataBreakpoints","arguments":{"breakpoints":[]},"type":"request","seq":7}
  [2.104 DEBUG codelldb::dap_codec] <-- {"seq":12,"type":"response","request_seq":4,"success":true,"command":"setFunctionBreakpoints","body":{"breakpoints":[]}}
  [2.104 DEBUG codelldb::dap_codec] <-- {"seq":13,"type":"response","request_seq":5,"success":true,"command":"setInstructionBreakpoints","body":{"breakpoints":[]}}
  [2.104 DEBUG codelldb::debug_session] Debug event: 0x55938be759d8 Event: broadcaster = 0x55938b1cafb0 (lldb.target), type = 0x00000001 (breakpoint-changed), data = {bkpt: 5 type: breakpoint added}
  [2.104 DEBUG codelldb::debug_session] Debug event: 0x55938be74b38 Event: broadcaster = 0x55938b1cafb0 (lldb.target), type = 0x00000001 (breakpoint-changed), data = {bkpt: 6 type: breakpoint added}
  [2.104 DEBUG codelldb::dap_codec] <-- {"seq":14,"type":"response","request_seq":6,"success":true,"command":"setExceptionBreakpoints","body":{"breakpoints":[{"id":5,"message":"Resolved locations: 0","verified":false},{"id":6,"message":"Resolved locations: 0","verified":false}]}}
  [2.104 DEBUG codelldb::dap_codec] <-- {"seq":15,"type":"response","request_seq":7,"success":true,"command":"setDataBreakpoints","body":{"breakpoints":[]}}
  [2.208 DEBUG codelldb::dap_codec] --> {"command":"configurationDone","type":"request","seq":8}
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/8C68F140-AAA2-F707-F730-F2EE4E510120/libandroidfw.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/47A4DF2E-C980-CA7B-EE2C-5162D1311A67/libpowermanager.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/9DF5DFA7-C0A9-BEE1-18E8-59821ACCE119/libhidlbase.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/B91287F0-90CF-9F69-4DDD-2F1A515321DD/libandroid_runtime.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/59856B9D-1F82-D3A6-C71D-66834CA846B9/libGLESv2.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F4D570AA-256C-18CF-EA9D-AE414A89DA7D/libm.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/E5A3DA1A-4375-273B-68D0-BD446CDE5B09/libPlatformProperties.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/DE46E75F-F2ED-1E2C-471D-4636BABCB461/libgralloctypes.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/59EDC081-6CC9-8412-738C-D7D9FD088A51/android.hardware.graphics.bufferqueue@2.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F7541D18-8FBA-9039-6D97-EC2354A99C0C/libandroid.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/B18C8A03-80E8-9331-80E2-E5D0A8F17968/libhardware.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/236EA1BC-94E0-AFD5-BC7F-F1431CE6E873/libui.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/9C5275F2-9E5F-BCAB-CA91-F471539F6C33/libharfbuzz_ng.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/49F04F9F-4BAB-7644-EBE0-CD12D5CBA422/android.media.audiopolicy-aconfig-cc.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F643E105-5895-E50E-6646-2991C8BD651A/libsensor.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D4EFBE1C-EE05-5ADB-7BC0-6563275B325A/libactivitymanager_aidl.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F4788000-0CAD-34AE-85A3-A6D6EB47AF56/libnetd_client.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/E9F6E11C-F330-D7B5-DA0A-E574B08DE2BF/libappfuse.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/1ABE5FF0-44ED-C7D5-EDEC-7513BEFDFB08/libhwui.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/6616E752-0D49-82C8-B16F-490C2FCE7B07/libxml2.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D7BFEF10-7332-8147-237F-DCD213DDB0F4/libEGL.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/5A4C3962-7C1C-0C27-B85E-2B07FF8406E8/android.hardware.configstore@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/5794558E-3A70-DE27-7ACE-8D96CC135A6E/libprocessgroup.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/E53513AD-07B9-A2D3-8E10-BB9640873A51/android.os.flags-aconfig-cc.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/17FB6C42-A328-8C80-F6B1-2E8DCE44BA87/android.hardware.common.fmq-V1-ndk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/BCF22BAF-8100-0E78-7DBF-DA0634CDCAE3/libdng_sdk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/53E0091D-25A7-8880-2D2D-3A5324F79B52-7DF4913F/libc++.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F813C105-4BF3-8946-F3DD-ACFA5A0F5D92/libaconfig_storage_read_api_cc.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D2388CD6-B4A2-3C03-AAC6-A43F69DBE3E7/libprotobuf-cpp-lite.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F51A883E-B8A3-E395-D776-028AAFD911B2/libincfs.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/1D37DF90-6550-A769-0656-A118D5D3E48C/libtinyxml2.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/71BEA50E-B26B-2551-6A93-71624A6F6C48/libapexsupport.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/75F07E30-8942-51A2-09D9-EFD05765443E/libcutils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/ACCECA4F-7FDF-AAB5-8CA4-D22F02EFCFA2/android.hardware.graphics.allocator@2.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/268C25BF-EE33-DEFE-E41F-04FA65438E3F/android.hardware.graphics.mapper@3.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/7249C1EE-880D-5DC5-6D20-218B90D0EE70/liblog.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D1D3BF8F-E6FE-CD69-21AC-714B5EF42E28/libutils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/FD865493-6014-8D9B-9146-148D898ACE8B/android.hardware.graphics.bufferqueue@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/4FC00C98-6358-5F87-0F76-2CDD96F8D6FE/android.hardware.graphics.common@1.1.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D6FD5669-9855-0295-6CD6-3C15A6BA2463/libnativewindow.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F100ABB4-5F8D-9443-385E-41FAA8C60894/libbinder_ndk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/10880913-0278-27E8-B7B9-E2B0073332E6/libpermission.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/E7298183-6700-D01A-9E07-A7EFE0463AB4/libgui.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D3BDEF9F-76E0-0746-698A-FDB95AFBC4E3/android.hardware.configstore@1.1.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/65F76722-B6BA-9085-2F32-C65B2917B63B/android.media.audio.common.types-V4-cpp.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/622B1765-4662-9AB1-9F5E-C1DFBBC18E6A/av-types-aidl-cpp.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/AE1416DE-9C2C-8E2A-03B1-9B4A77B111D1/libandroid_net.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/92739A8D-5622-5793-8128-28A5C6293AF2/libminikin.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/6A5F8F94-CFE5-65DC-C001-95D7A3CA9602/libdebuggerd_client.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/FB11B3F2-0030-F996-41AD-EBC1949142B6/libsqlite.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D33A3544-5B95-B61F-40F2-1B097B982B2E/android.hardware.configstore-utils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/93CCDE3F-366F-7E38-3EC0-899B87F4C45F/libselinux.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/89BD714E-5DEF-C9AE-88C7-2982F3EF3AC8/libnativeloader_lazy.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D4B2FFC5-F138-6CF1-32E1-2AB8C6DCFB37/liboplusplugin.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/3E69A5D9-5456-0A62-90D7-0565B0B436B0/libultrahdr.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/90259EF0-41B0-1885-8D40-258E8170B217/aconfig_text_flags_c_lib.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/070EE97F-074D-C13F-D642-DC3627F7CE57/libz.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D89F718A-0F70-5B6B-21A7-444F4CB496F3/libheif.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/EEA2E831-DF57-26D5-5A9E-100BA0E291AF/libft2.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/6E0FBED8-8045-2E8A-276A-ED99974D2903/lib-platform-compat-native-api.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/7926FF37-5B46-6646-59D5-04CBC606B997/libpackagelistparser.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F05BB986-1975-74CF-310A-A7F2244D981F/android.hardware.power@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/217D4A46-760A-940D-FED2-1538F547262D/libbinder.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/B8FA905C-387B-6EC6-2B8C-29E3B0D5E1C1/android.hardware.power@1.1.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/E1E52A3B-FD01-7616-6655-977ED25478CF/android.hardware.power@1.2.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/48EDFCD7-4765-28A8-2472-D6BA77AA16C1/android.hardware.power@1.3.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/7F0888CC-ACBD-22DC-6197-E024D61DB019/com.android.media.audioclient-aconfig-cc.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/27AA0829-9412-F262-58C6-9E6226585DEA/android.hardware.graphics.common@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D1234763-4E7A-21ED-F65D-AE44CB959CDB/libaudiomanager.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/5A118426-37D9-7B8C-323D-5ACA6BB7AA92/android.hidl.token@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/4EB1AD8B-2324-C684-BF89-FDBAA325836A/framework-permission-aidl-cpp.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/22F4901A-3743-4C39-8C4E-5F0BEA118079/effect-aidl-cpp.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/5B965CF4-265D-81F2-1F99-DA9B6B0EBF34/shared-file-region-aidl-cpp.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/68D733EC-E14A-6DDF-08A5-0C37D3EE00E8/android.hardware.camera.common@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F891BE37-E405-66BF-2B9B-8C8C5474A680/android.hardware.memtrack@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/4D0532B3-F1D1-B4FC-ED69-FC5BD7A829BC/android.hardware.memtrack-V1-ndk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/CE04558B-0A23-351D-CADD-5E511B6B8953/libETC1.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/BC2C7F87-8C6B-0064-210F-D4977F2C66F0/libandroid_runtime_lazy.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/4B758857-1C68-D480-E403-0D92344242F7/libcgrouprc.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/A8BBDFEA-E174-30B5-7E6A-5BABCBA908B1/libbpf_bcc.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/24CF4CEC-1D73-44C8-ED75-5884E4953E22/libdatasource.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/4D599915-9EC2-D153-B954-35161286B527/libmediadrm.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/8A8683E2-3717-CC18-680C-E0EFE97EAA96/libmedia_omx.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/A96DAF57-6898-C714-C5EC-DC88B80C294B/libmedia_jni_utils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/4E0635FF-666C-8B48-8CCF-5566C7FDBD40/libstagefright.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/EEBAA949-2A40-455A-8475-F7E9A76F87D2/libstagefright_foundation.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/3C0DB02C-271B-A2E2-CC7B-63E904E5638E/libmediandk_utils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F1D853BA-297B-CA2A-3014-65C02F8CF647/android.hardware.drm-V1-ndk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D7E60A1B-02A0-A89D-5AB7-E1883A8B5067/libSurfaceFlingerProp.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/0C7C0800-AF68-65E6-4893-CE47B8AB6DA3/camera_platform_flags_c_lib.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/4E788058-CF19-E541-C41C-22663B333A4D/libunwindstack.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/3B0FA744-EEA5-F951-8170-DE832FD9F1CE/libziparchive.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/FCDD8634-77D9-C35A-04F7-E46E76ADF01C/libpcre2.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/CFB68FE0-E192-4651-97FC-C0D3F944F7A0/libmedia_codeclist.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/1EC03020-A00F-FCAD-59B4-6CDF97537D05/liboplusremotedisplay.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/CEAF7077-8476-ED6E-BBBB-10D160D6BD60/liboplusremotedisplayclient.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/1F7EED6E-4DD5-6CBB-45B6-E89615257E9D/android.hardware.drm@1.4.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/2158C75E-4988-C285-1896-357F99993475/android.hardware.media.omx@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/AC68D018-F02B-14CC-1B51-F5E83226129C/capture_state_listener-aidl-cpp.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/01DF55F6-1428-95DC-DBAA-4FE0203DE907/libaudio_aidl_conversion_common_cpp.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/63A840F8-69D6-5DFC-BD42-B430C2F559F9/libcodec2.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/75FC895D-0EDC-4A29-C95B-D983A003758F/libaudioutils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F4EB4AAA-F8B1-F58E-6040-561179D5E38E/libmedia_omx_client.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F8CCC3F7-ABD2-6816-EE77-C2BBBC1EF11F/libnblog.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F4D585C7-43F3-C29F-70F0-60B139280236/libshmemcompat.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/BC87EF18-DF8F-7AA9-77EE-59BE3A33D4BD/libcamera_metadata.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/8D5546FC-C984-4C53-AE4D-A68376649C07/packagemanager_aidl-cpp.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/3BC123B3-2E0D-1DA2-4608-CBE1EFDFEDF3/liboplusmmdebug.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/8C6E0D1F-F75D-74B0-94BC-781C562A7092/libnativedisplay.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/52C82393-C710-2821-9339-0A5C5CE8BF91/libbpf_minimal.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F862C6CB-6B56-928F-7B24-D2A3B49ADA5E/liboplusloadframe.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/364C8677-368A-1BA8-E6F8-CBCBF5B8FF74/libjpegencoder.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/61F0425A-F9C9-47CE-4E65-7AC31E7BAAC8/android.hardware.media.bufferpool@2.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/9B34299E-EB2E-93CF-3C80-9A6B855ABDC5/liblzma.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/BC304671-6DBD-7AFD-44CB-44531D3F3C53/libmediadrmmetrics_lite.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/C1BED4E7-F50E-454B-3115-B12789484E9C/android.hardware.drm@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/16EE5061-D6A1-1B5D-3C3A-58B595FE99E1/android.hardware.drm@1.1.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/59C7A5B8-100F-C786-A237-534275FFF031/vendor.dolby.dvs-V1-ndk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D8133337-22F8-4E5C-FC7E-81580D92885D/android.system.suspend-V1-ndk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D0870D5B-92DD-9BD8-990C-2A6C5D189A33/android.hardware.drm@1.2.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/6576446C-95B5-8FC8-12D0-19864EAF1A6E/libcodec2_client.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/56EE6975-54BD-8AF9-256B-E8F25D7B5093/android.hardware.drm@1.3.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/773E4036-29F6-0510-7FE9-D09F3ACBFC17/mediametricsservice-aidl-cpp.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/B13CB8E9-12AE-139B-DFD0-F590ADC67617/libhidlallocatorutils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/42F83B19-A621-BCAE-5536-880C786C936D/audiopolicy-aidl-cpp.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/8571E439-E7CC-207A-B9E8-3981166AA953/android.hidl.memory@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/C62C6178-578C-7D26-455F-B1F590F8624C/android.hardware.cas@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/94E17DA2-D20D-C000-3672-D0C4933C9BF3/libcodec2_vndk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/470D87DF-0FBB-48DE-058A-BF31ED48D4BA/libcodec2_hal_common.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/090FF4C1-571A-EF9F-4679-C5600FF8886B/libsfplugin_ccodec.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/A11669AA-BD01-1D70-2046-55568F4BC3B6/libsfplugin_ccodec_utils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/39D4A1F6-8339-7BE8-174D-C8C26569B3D3/libstagefright_codecbase.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D488A480-3BB5-3EAF-655D-4B4D6793C0A9/libstagefright_omx_utils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/5141AA36-4CA6-DEB3-0E83-899F3A211574/libhidlmemory.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/C85E596D-528C-8413-61FD-1543C6A468A4/android.hidl.allocator@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/4FDA2AC3-923D-7AFB-EA12-7D310D09CC33/android.hardware.cas.native@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/45386C91-5C0E-7873-D765-C358266F5F5B/aconfig_mediacodec_flags_c_lib.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/6CF112A3-E61A-25C9-DF8C-4B0C67589CF7/libspeexresampler.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/C7F41251-6A79-5DC6-5A07-3BF6241A7C9C/com.android.media.audio-aconfig-cc.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F9CB190B-FDC3-6AC0-2DFE-AF1A0887DDA2/libjpegdecoder.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/1FA53D37-49E7-8B23-E3B4-20BD59C68B49/libshmemutil.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/30BB277C-5D0B-993A-92A2-FB5C746FE461/libandroidicu.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F6613968-B4EC-FA82-4056-2754492D3640/android.os.flags-aconfig-cc-host.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/05E97885-D5B8-746E-8213-2727CC62D273/android.hardware.media.bufferpool2-V2-ndk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/91696B8B-0757-00FC-B0F2-BFA26B9DCA0C/android.hardware.media.c2-V1-ndk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D69A31D4-3CA9-6A0F-9FF7-695A6AAF01B8/libdmabufheap.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/A608E5AE-1FC9-711B-200A-F2FB29BE6197/libbsproxy.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F7332E1E-E32D-AAD0-791F-F4D8CFC6877D/libstagefright_bufferpool@2.0.1.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/7CF0AFB8-6366-A80A-9D9A-A2350B697FBC/libstagefright_aidl_bufferpool2.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/44DB7622-A843-97FB-D032-22840478D1E4/android.hardware.media.c2@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D471166E-5DD7-A800-8B7E-465E56DD000E/liboplusvideoboostclient.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/26B88FA8-5E98-04BD-532B-C2BA5C3E86B6/libstagefright_bufferqueue_helper.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/62F6DFE5-993D-3537-C5B0-6D67D64BFB3C/libstagefright_graphicbuffersource_aidl.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/E392C887-912B-D578-93AE-F162D3B45C0A/libstagefright_omx.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/FD199BBD-8790-6261-A65B-5208D1BD269F/libstagefright_surface_utils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D15566A0-561C-7926-F6FC-29385CEE8E55/libstagefright_xmlparser.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/C7A683E7-6644-0421-38F2-BA9883A3AC3A/libstagefright_framecapture_utils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/1412FD82-5729-D5FE-9E90-E3DAB35F0785/android.hidl.memory.token@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/29E601BA-E033-65DD-3170-6DD393D601FC/libopluscfgpolicy.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/A051CC35-7764-8479-301A-98F0F8E6766D/android.hardware.media@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F595F888-ADB0-0AEE-048E-C2D74AE8DC79/android.hidl.safe_union@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/36E68CBB-F317-919C-458E-33CC44855754/libmediautils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/3733209E-B9F5-F655-C2A7-5B82D7C640A8/libcodec2_aidl_client.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/B41BE3A5-3B4A-FC6D-DC12-E9C67A361AE2/android.hardware.media.c2@1.1.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/BCED8136-6039-1212-C20A-492C58E80466/android.hardware.media.c2@1.2.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/37E43EAB-0240-B0A8-A9F0-EF72DC2D5B17/libcodec2_hidl_client@1.0.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/B7503E25-16EE-CD1A-8742-9B5B1A01F5BB/libcodec2_hidl_client@1.1.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/DCBAF5C2-CC21-0D9D-BFCD-FDD334B0431D/libcodec2_hidl_client@1.2.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/5C9EE94B-207A-672D-08E7-48164014FB06/libstagefright_httplive.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/977C3587-F5F1-6B11-C61B-35D997D61738/libicu.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/4F4D5F18-A6BE-4344-EA17-DC9DC7411427/libicuuc.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/63DA7563-739E-6B10-0281-6046A3D1FF9D/libatlasservice.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/5BA0567C-E2DB-159D-2FBC-A7A1553C861A/libostatslog.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/44104D0C-22BC-EB23-188C-DC8D9879E00E/libinputextensionsimpl.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/2D06A995-3D42-EDC9-126A-602DEA5F176E/libdrmframeworkcommon.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/C3618A49-D5FE-BCC0-33B1-8F81144623B9/libion.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/5CE798D7-1505-31D7-04C9-931E7E821859/libskjpegencoderextimpl.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/55CC7435-9BEE-63CE-66EF-B7E49A8E25C3/vendor.oplus.hardware.urcc-V1-ndk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/0369ED75-A4C8-871F-CAC0-CA5058B89AF4/graphicbuffersource-aidl-ndk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/7725EFB8-4388-9F2F-AC5E-381796F1A207/libstatssocket.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/4438F18D-386D-940E-CB37-6C59BB81938B/libstatspull.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/3DACF6CB-307A-C2BD-6BC6-41B4705D805E/libaudioclientextimpl.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/8724936D-12B0-0A45-E855-DE82BD2BF9F5/vendor.oplus.hardware.performance-V1-ndk.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/2A91ADBC-2C8D-8343-B339-45A3530E3D97/libiatlasservice.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/FB393EC0-4529-BBE4-7DE1-CDC2B623642A/libimmlistservice.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/80072B39-258E-25AD-482F-BCA4E522757A/liboplusaudioDump.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/145692BC-7B48-FA7D-4DEE-9F3647C9F400/libvibrator.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/86AFE71B-BEDF-86F9-96DD-27031077E951/libmmlistparser.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/2472CDC5-FA7D-3D3C-FBF6-4F6F6CB161FA/liboplusavenhancements.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/7401D122-F709-E469-F346-2FA68B905BB8/libavenhancements.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/0B7AB220-8D7B-1F08-5CC8-36C34FBDA133/liboplusstagefright.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/7EF0AF17-81DE-160C-6E4D-62A930D5CDAD/libmediaplayerservice.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/1883C9A5-0D74-FFBB-8BEB-975550D98460/liboplusmediaplayerservice.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/F33FF539-76AE-26D3-0A5F-7307FE509EEE/liboplusutils.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/4BDDFB25-F00A-C2AA-33E7-F04B4FEAB5DE/liboplussfplugin_ccodec.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/0D38A646-14CF-FFED-7FDC-57868CBD5B7C/audio-permission-aidl-cpp.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/6150C4DB-DC98-5D10-3D67-8BFF53218D0B/libnbaio.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/BC19C49B-365B-EC00-D092-32719922C4E7/libicui18n.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/6C75C77B-BF27-8F15-9C20-8538923D24D5/liboplusmultimediaconfig.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/1DDB624D-F713-CBA4-5D72-9CE6CD7B0E43/libdrmframework.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/8BC19754-BC9E-4207-00A5-8416B1FF940F/liboplusvideoboostservice.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/7D4D1F92-679E-170A-1058-DD60CB619122/liboplus_multimedia_kernel_event.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/BE5A17FD-B8C5-DB79-ECCF-C69F19FB4B1B/libstatslog.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/602B6754-2D58-D7AE-C164-2BDE3EB18001/liboplus-uah-client.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/D5C4EBD4-46F8-8579-13D9-E6959D84C26B/liboplus_imageprocessing.so No LZMA support found for reading .gnu_debugdata section
  warning: (aarch64) /root/.lldb/module_cache/remote-android/.cache/B325B13A-EDD2-831A-56E3-75A6A4C9332A/libolc.so No LZMA support found for reading .gnu_debugdata section

3.4 内存错误检测工具

AddressSanitizer（ASan） 是 CMake 构建系统中用于启用内存检测工具，能够高效识别多种难以察觉的内存问题，包括：堆/栈/全局缓冲区溢出、释放后使用（use-after-free）、内存分配/释放不匹配、越界访问等。

开启下面编译选项，在运行时发生内存越界过程程序会立刻停止并且打印报告，帮助更好定位内存问题。

`1`	`-DCMAKE_CXX_FLAGS=-fsanitize=address, -DCMAKE_C_FLAGS=-fsanitize=address`

例如在重新开发CPUAttention算子时，触发的下面内存越界log：

=================================================================
==271723==ERROR: AddressSanitizer: heap-buffer-overflow on address 0x61500004ad4f at pc 0x7f65d4c2fabe bp 0x7f65d08f72b0 sp 0x7f65d08f72a8
WRITE of size 16 at 0x61500004ad4f thread T3
    #0 0x7f65d4c2fabd in void _AVX_MNNPackednMatMulRemainCommon<float>(float*, float const*, float const*, unsigned long, unsigned long const*) /workspace/code/MNN/source/backend/cpu/x86_x64/avxfma/../avx/GemmFunction.hpp:815:9
    #1 0x7f65d4c2cc84 in _AVX_MNNPackedMatMulRemainFMA /workspace/code/MNN/source/backend/cpu/x86_x64/avxfma/GemmAVX2FMA.cpp:58:5
    #2 0x7f65d4772393 in MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_1::operator()(int) const /workspace/code/MNN/source/backend/cpu/CPUAttention.cpp:615:25
    #3 0x7f65d476b03d in void std::__invoke_impl<void, MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_1&, int>(std::__invoke_other, MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_1&, int&&) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/invoke.h:61:14
    #4 0x7f65d476af81 in std::enable_if<__and_<std::is_void<void>, std::__is_invocable<MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_1&, int> >::value, void>::type std::__invoke_r<void, MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_1&, int>(MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_1&, int&&) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/invoke.h:154:7
    #5 0x7f65d476aab1 in std::_Function_handler<void (int), MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_1>::_M_invoke(std::_Any_data const&, int&&) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/std_function.h:290:9
    #6 0x7f65d608b921 in std::function<void (int)>::operator()(int) const /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/std_function.h:590:9
    #7 0x7f65d4778e9f in MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_2::operator()(int) const /workspace/code/MNN/source/backend/cpu/CPUAttention.cpp:757:9
    #8 0x7f65d4778e3d in void std::__invoke_impl<void, MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_2&, int>(std::__invoke_other, MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_2&, int&&) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/invoke.h:61:14
    #9 0x7f65d4778d91 in std::enable_if<__and_<std::is_void<void>, std::__is_invocable<MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_2&, int> >::value, void>::type std::__invoke_r<void, MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_2&, int>(MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_2&, int&&) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/invoke.h:154:7
    #10 0x7f65d4778c81 in std::_Function_handler<void (int), MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_2>::_M_invoke(std::_Any_data const&, int&&) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/std_function.h:290:9
    #11 0x7f65d608b921 in std::function<void (int)>::operator()(int) const /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/std_function.h:590:9
    #12 0x7f65d49c0274 in MNN::ThreadPool::ThreadPool(int)::$_1::operator()() const /workspace/code/MNN/source/backend/cpu/ThreadPool.cpp:63:29
    #13 0x7f65d49bff9c in void std::__invoke_impl<void, MNN::ThreadPool::ThreadPool(int)::$_1>(std::__invoke_other, MNN::ThreadPool::ThreadPool(int)::$_1&&) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/invoke.h:61:14
    #14 0x7f65d49bff2c in std::__invoke_result<MNN::ThreadPool::ThreadPool(int)::$_1>::type std::__invoke<MNN::ThreadPool::ThreadPool(int)::$_1>(MNN::ThreadPool::ThreadPool(int)::$_1&&) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/invoke.h:96:14
    #15 0x7f65d49bff04 in void std::thread::_Invoker<std::tuple<MNN::ThreadPool::ThreadPool(int)::$_1> >::_M_invoke<0ul>(std::_Index_tuple<0ul>) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/std_thread.h:259:13
    #16 0x7f65d49bfed4 in std::thread::_Invoker<std::tuple<MNN::ThreadPool::ThreadPool(int)::$_1> >::operator()() /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/std_thread.h:266:11
    #17 0x7f65d49bfdf8 in std::thread::_State_impl<std::thread::_Invoker<std::tuple<MNN::ThreadPool::ThreadPool(int)::$_1> > >::_M_run() /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/std_thread.h:211:13
    #18 0x7f65d3ec8252  (/usr/lib/x86_64-linux-gnu/libstdc++.so.6+0xdc252) (BuildId: e37fe1a879783838de78cbc8c80621fa685d58a2)
    #19 0x7f65d3b50ac2 in start_thread nptl/./nptl/pthread_create.c:442:8
    #20 0x7f65d3be1a03 in __clone misc/../sysdeps/unix/sysv/linux/x86_64/clone.S:100

0x61500004ad4f is located 7 bytes to the right of 456-byte region [0x61500004ab80,0x61500004ad48)
allocated by thread T0 here:
    #0 0x55656862c61e in malloc (/workspace/code/MNN/build/llm_demo+0xac61e) (BuildId: 5f2185528b18e8c779d7ac797dd7f7b8df7019f2)
    #1 0x7f65d43b126c in MNNMemoryAllocAlign /workspace/code/MNN/source/core/MNNMemoryUtils.cpp:24:30
    #2 0x7f65d4355009 in MNN::DefaultAllocator::onAlloc(unsigned long, unsigned long) /workspace/code/MNN/source/core/BufferAllocator.cpp:59:25
    #3 0x7f65d4347bfb in MNN::EagerBufferAllocator::alloc(unsigned long, bool, unsigned long) /workspace/code/MNN/source/core/BufferAllocator.cpp:223:30
    #4 0x7f65d4787cd4 in MNN::CPUBackend::allocBuffer(unsigned long, MNN::Tensor*, MNN::Backend::StorageType) /workspace/code/MNN/source/backend/cpu/CPUBackend.cpp:591:49
    #5 0x7f65d4bc55e9 in MNN::AVX2Backend::onAcquire(MNN::Tensor const*, MNN::Backend::StorageType) /workspace/code/MNN/source/backend/cpu/x86_x64/AVX2Backend.cpp:344:16
    #6 0x7f65d4340670 in MNN::Backend::onAcquireBuffer(MNN::Tensor const*, MNN::Backend::StorageType) /workspace/code/MNN/source/core/Backend.cpp:133:22
    #7 0x7f65d47600a6 in MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&) /workspace/code/MNN/source/backend/cpu/CPUAttention.cpp:301:16
    #8 0x7f65d4435b17 in MNN::Pipeline::execute() /workspace/code/MNN/source/core/Pipeline.cpp:1165:40
    #9 0x7f65d449dfb1 in MNN::Session::run() const /workspace/code/MNN/source/core/Session.cpp:249:28
    #10 0x7f65d5413db8 in MNN::Express::StaticModule::_execute() /workspace/code/MNN/express/module/StaticModule.cpp:545:26
    #11 0x7f65d5414580 in MNN::Express::StaticModule::onForward(std::vector<MNN::Express::VARP, std::allocator<MNN::Express::VARP> > const&) /workspace/code/MNN/express/module/StaticModule.cpp:582:16
    #12 0x7f65d53d95db in MNN::Express::PipelineModule::onForward(std::vector<MNN::Express::VARP, std::allocator<MNN::Express::VARP> > const&) /workspace/code/MNN/express/module/PipelineModule.cpp:240:57
    #13 0x7f65d53add01 in MNN::Express::NetModule::onForward(std::vector<MNN::Express::VARP, std::allocator<MNN::Express::VARP> > const&) /workspace/code/MNN/express/module/Module.cpp:224:32
    #14 0x7f65d5dcce5e in MNN::Transformer::Llm::generate(std::vector<std::vector<int, std::allocator<int> >, std::allocator<std::vector<int, std::allocator<int> > > > const&, int) /workspace/code/MNN/transformers/llm/engine/src/llm.cpp:866:56
    #15 0x7f65d5dd0f52 in MNN::Transformer::Llm::response(std::vector<std::vector<std::pair<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > >, std::allocator<std::pair<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > > > >, std::allocator<std::vector<std::pair<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > >, std::allocator<std::pair<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > > > > > > const&, std::ostream*, char const*, int) /workspace/code/MNN/transformers/llm/engine/src/llm.cpp:1034:5
    #16 0x55656866b0de in chat(MNN::Transformer::Llm*) /workspace/code/MNN/transformers/llm/engine/demo/llm_demo.cpp:284:18
    #17 0x55656866cb2b in main /workspace/code/MNN/transformers/llm/engine/demo/llm_demo.cpp:349:9
    #18 0x7f65d3ae5d8f in __libc_start_call_main csu/../sysdeps/nptl/libc_start_call_main.h:58:16

Thread T3 created by T0 here:
    #0 0x556568615a9c in __interceptor_pthread_create (/workspace/code/MNN/build/llm_demo+0x95a9c) (BuildId: 5f2185528b18e8c779d7ac797dd7f7b8df7019f2)
    #1 0x7f65d3ec8328 in std::thread::_M_start_thread(std::unique_ptr<std::thread::_State, std::default_delete<std::thread::_State> >, void (*)()) (/usr/lib/x86_64-linux-gnu/libstdc++.so.6+0xdc328) (BuildId: e37fe1a879783838de78cbc8c80621fa685d58a2)
    #2 0x7f65d49bfb50 in void __gnu_cxx::new_allocator<std::thread>::construct<std::thread, MNN::ThreadPool::ThreadPool(int)::$_1>(std::thread*, MNN::ThreadPool::ThreadPool(int)::$_1&&) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/ext/new_allocator.h:162:23
    #3 0x7f65d49bf55c in void std::allocator_traits<std::allocator<std::thread> >::construct<std::thread, MNN::ThreadPool::ThreadPool(int)::$_1>(std::allocator<std::thread>&, std::thread*, MNN::ThreadPool::ThreadPool(int)::$_1&&) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/alloc_traits.h:516:8
    #4 0x7f65d49bf85d in void std::vector<std::thread, std::allocator<std::thread> >::_M_realloc_insert<MNN::ThreadPool::ThreadPool(int)::$_1>(__gnu_cxx::__normal_iterator<std::thread*, std::vector<std::thread, std::allocator<std::thread> > >, MNN::ThreadPool::ThreadPool(int)::$_1&&) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/vector.tcc:449:4
    #5 0x7f65d49b6fbe in void std::vector<std::thread, std::allocator<std::thread> >::emplace_back<MNN::ThreadPool::ThreadPool(int)::$_1>(MNN::ThreadPool::ThreadPool(int)::$_1&&) /usr/bin/../lib/gcc/x86_64-linux-gnu/11/../../../../include/c++/11/bits/vector.tcc:121:4
    #6 0x7f65d49b6703 in MNN::ThreadPool::ThreadPool(int) /workspace/code/MNN/source/backend/cpu/ThreadPool.cpp:58:18
    #7 0x7f65d49b52bf in MNN::ThreadPool::init(int, unsigned long, MNN::ThreadPool*&) /workspace/code/MNN/source/backend/cpu/ThreadPool.cpp:26:35
    #8 0x7f65d477d3a7 in MNN::CPURuntime::_resetThreadPool() const /workspace/code/MNN/source/backend/cpu/CPUBackend.cpp:120:25
    #9 0x7f65d477f872 in MNN::CPURuntime::onReset(int, MNN::BackendConfig const*, bool) /workspace/code/MNN/source/backend/cpu/CPUBackend.cpp:219:5
    #10 0x7f65d51719f6 in MNN::Express::Executor::_getOrCreateRuntime(MNNForwardType, MNN::BackendConfig const*, int, bool) /workspace/code/MNN/express/Executor.cpp:57:23
    #11 0x7f65d51772d5 in MNN::Express::Executor::RuntimeManager::createRuntimeManager(MNN::ScheduleConfig const&) /workspace/code/MNN/express/Executor.cpp:324:20
    #12 0x7f65d5db9405 in MNN::Transformer::Llm::initRuntime() /workspace/code/MNN/transformers/llm/engine/src/llm.cpp:205:27
    #13 0x7f65d5dbae4e in MNN::Transformer::Llm::load() /workspace/code/MNN/transformers/llm/engine/src/llm.cpp:254:5
    #14 0x55656866ca94 in main /workspace/code/MNN/transformers/llm/engine/demo/llm_demo.cpp:337:25
    #15 0x7f65d3ae5d8f in __libc_start_call_main csu/../sysdeps/nptl/libc_start_call_main.h:58:16

SUMMARY: AddressSanitizer: heap-buffer-overflow /workspace/code/MNN/source/backend/cpu/x86_x64/avxfma/../avx/GemmFunction.hpp:815:9 in void _AVX_MNNPackednMatMulRemainCommon<float>(float*, float const*, float const*, unsigned long, unsigned long const*)
Shadow bytes around the buggy address:
  0x0c2a80001550: fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa
  0x0c2a80001560: fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa
  0x0c2a80001570: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x0c2a80001580: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x0c2a80001590: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
=>0x0c2a800015a0: 00 00 00 00 00 00 00 00 00[fa]fa fa fa fa fa fa
  0x0c2a800015b0: fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa fa
  0x0c2a800015c0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x0c2a800015d0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x0c2a800015e0: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
  0x0c2a800015f0: 00 00 00 00 00 00 00 00 00 fa fa fa fa fa fa fa
Shadow byte legend (one shadow byte represents 8 application bytes):
  Addressable:           00
  Partially addressable: 01 02 03 04 05 06 07 
  Heap left redzone:       fa
  Freed heap region:       fd
  Stack left redzone:      f1
  Stack mid redzone:       f2
  Stack right redzone:     f3
  Stack after return:      f5
  Stack use after scope:   f8
  Global redzone:          f9
  Global init order:       f6
  Poisoned by user:        f7
  Container overflow:      fc
  Array cookie:            ac
  Intra object redzone:    bb
  ASan internal:           fe
  Left alloca redzone:     ca
  Right alloca redzone:    cb
==271723==ABORTING

日志大致分为四段：

第一段发生越界写入的位置，也就是log中WRITE of size 16 at 0x61500004ad4f thread T3，根据开发需求在该部分log的调用栈中找到出现内存越界写入位置的代码：

    #2 0x7f65d4772393 in MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&)::$_1::operator()(int) const /workspace/code/MNN/source/backend/cpu/CPUAttention.cpp:615:25

这里发生越界写入位置的是我代码中的这一行

gcore->MNNPackedMatMulRemain((float*)(qkPacked + (loop_e * eP * mPack) * mBytes), (float*)(qReordered + loop_e * qStride0), (float*)keyPtr, remain, shapeParameters, nullptr, nullptr, nullptr, nullptr);

往下对应，发现是中间的计算缓冲区内存开小了

第二段被越界写入的内存分配的位置，对应log中的allocated by thread T0 here: 对应日志中可以看到，内存分配位置在算子CPUAttention::onExecute中的第301行，可以根据需求继续顺着调用栈向上或者向下查代码

#7 0x7f65d47600a6 in MNN::CPUAttention::onExecute(std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&, std::vector<MNN::Tensor*, std::allocator<MNN::Tensor*> > const&) /workspace/code/MNN/source/backend/cpu/CPUAttention.cpp:301:16

第三段是发生越界写入的线程被启动的位置，MNN中为CPU设计了线程池，根据需求可以查看对应线程的创建日志。
第四段是堆缓冲区溢出的错误报告

远程ADB环境配置

Thu, 26 Feb 2026 23:00:00 +0800

1. 远程ADB环境配置

当开发服务器位于内网、手机在开发者本地时，“编译-传输-测试”、调试比较麻烦。

实际上adb提供了远程连接方式，可以通过端口转发链路，让内网开发服务器直接adb connect本地手机。

本文以将介绍远程ADB环境的配置

1. 远程ADB环境配置

1.1 网络拓扑

现有网络拓扑：本地机器可以通过跳板机连接开发服务器，同时可以连接测试手机

[开发服务器 S, Server] 
       ▲
       │ (内网，仅J可达)
[跳板机 J, Jump] ←──┐
       ▲       (Frp穿透)
       |           │
[本地机器 L, Local] ┘
       │ (adb无线/USB)
       ▼
[测试手机 P, Phone]

预期网络拓扑：开发服务器可以直接连接测试手机

[开发服务器 S, Server] 
       │ (内网，仅J可达)
       ▼
[跳板机 J, Jump] ←──┐
       │       (Frp穿透)
       ▼           │
[本地机器 L, Local] ┘
       │ (adb无线/USB)
       ▼
[测试手机 P, Phone]

S (Server)：代码开发服务器，IP为 $ip_S$，端口 $port_S$
J (Jump)：跳板机
L (Local)：开发者本地机器（Windows/Linux/WSL）
P (Phone)：测试手机，已开启ADB over TCP，IP为 $ip_P$，SSH端口 $port_P$

1.2 原理

构建端口转发链路：S($ip_S$: $port_S$) → L($ip_L$: $port_L$) → P($ip_P$: $port_P$)

1.3 本地ADB环境配置

手机端开启ADB网络调试，需要自行安装配置adb， adb介绍

1
2
3

# 通过USB临时连接手机后执行
adb tcpip port_P	# 插入usb连接手机 启动无线adb调试，部分机型可能需要在开发者模式设置允许无线调试
adb connect ip_P:port_P  # 验证无线adb连接，手机的ip可能会随着网络变化，可以在手机设置中搜索ip查看，或者路由器页面查看

这里注意WSL中USB设备需由Windows层接管，需要把USB访问权给WSL，参考文档

1.4 构建服务器->本地转发链路(S->L)

方案1: SSH端口转发

步骤1：服务器S → 本地L（在本地L建立远程端口转发）

# 本地L 执行
# 后面user@ip -p port 是本地L通过ssh连接服务器S的user ip port, 不是需要转发的ip和端口
ssh -N -R 0.0.0.0:port_S:localhost:port_L user@ip -p port 

# 如果本地配置了config, (参考配置链接: https://zhuanlan.zhihu.com/p/1984208939030684698)
# 假设别名是host_name, 可以修改成下面形式
ssh -N -R 0.0.0.0:port_S:localhost:port_L host_name

# -N: 只用来转发数据
# -R: 反向转发
# 0.0.0.0: 监听跳板机的所有网卡

步骤2：测试链路连通，参考

# 每一步可以通过netcat测试链路连通
# netcat安装: https://www.trae.cn/article/706600706

# 例如: 测试 服务器S → 本地L 连通性
# 服务器 服务器S端口port_S是否已经转发给本地L端口port_L
echo "hello" | nc 127.0.0.1 port_S
# 本地
nc -l -p port_L
# 测试成功本地输出hello

方案2: TailScale/EasyTier隧道 (推荐)

直接用TailScale/EasyTier打通服务器到本地隧道，TailScale会尝试直连打通隧道，打通成功网速不受限于Frp穿透机器，比较推荐。

步骤1：本地L和服务器S分别部署TailScale/EasyTier, TailScale参考教程， EasyTier参考教程下面以EasyTier为例，注意EasyTier需要root权限并且docker默认不具备TUN 支持，建议联系集群管理员提供隧道通信方式

本地L安装：从github-relase界面，选择对应版本，如winodws选择easytier-gui_2.5.0_x64-setup.exe 随后手动安装

打开EasyTier, 左下角可以切换中文模式

设置网络名称和密码，然后点击创建新网络，只有网络名和密码匹配才会加入同一个网络，所以网络名设置不要太简单。

滑倒底下运行网络，本地节点就启动了

服务器S节点安装

# 安装
wget -O- https://raw.githubusercontent.com/EasyTier/EasyTier/main/script/install.sh | sudo bash -s install
# 配置，需要和前面一致的网络名和密码, 把下面passwd替换密码
sudo easytier-core -d --network-name test --network-secret "passwd" -p tcp://public.easytier.cn:11010

# 查看节点，这里docker不具备
easytier-cli peer

结果如下，这里的公共服务器可以替换为自己的公网服务器。

ipv4	hostname	cost	lat(ms)	loss	rx	tx	tunnel	NAT	version
$ip_{S1}$	—–	Local	-	-	-	-	-	Symmetric	2.4.5-4c4d172e
	PublicServer_hubei-public.easyti	p2p	9.42	0.0%	82.91 kB	8.78 kB	tcp	Restricted	2.5.0-db059c17~
$ip_{L1}$	—–	relay(2)	18.00	0.0%	0 B	0 B		Symmetric	2.5.0-88a45d11~

在本地的gui界面可以看见两个节点和对应IP, 这里测试的docker不支持TUN所以无法打通，自己搭建需要有docker权限且支持TUN

步骤2：登录TailScale，获取本地L和服务器S分配的ip, 分别记录为$ip_{L1}$ 和$ip_{S1}$。EasyTier直接在界面中获取。
步骤3：建立本地L到服务器S的SSH连接，其余配置一样，只需要修改原来通过SSH连接到服务器S的ip为$ip_{S1}$
步骤4：同[方案1](#方案1: SSH端口转发)，在本地建立到服务器的反向端口转发
注：同理，可以建立服务器到本地的SSH连接，然后在服务器开本地端口转发，ssh转发参考

1.5 构建本地->手机的转发链路(L->P)

1
2

# 本地L执行, 将本地L端口转发至 手机P的adb连接端口
adb forward tcp:port_L tcp:port_P

1.6 测试

# 服务器S执行, 已经将服务器端口port_S转发给手机adb端口
# sudo apt install adb -y 若服务器无adb
adb connect 127.0.0.1:port_S
adb devices # 正常显示连接设备

MNN端侧部署 on huluhuluu

MNN Backend 介绍

MNN 介绍

1. 后端

1.1 CPU后端系列

1.1.1 x86/x64-SSE4.1后端

1.1.2 x86/x64-AVX2后端

1.1.3 x86/x64-AVX512后端

1.1.4 ARMv7a后端

1.1.5 ARMv8后端

1.2 GPU后端系列

1.2.1 OpenCL后端

1.2.2 Vulkan后端

1.2.3 Metal后端

1.2.4 CUDA后端

1.2.5 OpenGL后端

1.3 专用加速器

1.3.1 CoreML后端

1.3.2 HIAI后端

1.3.3 NNAPI后端

1.3.4 QNN后端

1.3.5 NeuroPilot 后端

1.4 后端相关目录结构

1.4.1 后端文件说明

1.4.1.1 Runtime - 运行时抽象层

1.4.1.2 Backend - 后端抽象基类

1.4.1.3 Execution - 执行器抽象基类

1.4.2 后端调用设置

1.4.2.1 Core Function

MNN 核心类介绍

MNN 介绍

1. MNN核心类

1.1 类之间关系

1.2 VARP类

1.2.1 Variable 类

1.2.2 readMap详解

1.3 Expr类

1.3.1 requireInfo 详解

1.4 Tensor类

1.4.1 数据格式

1.4.2 底层数据格式

1.4.2 核心接口

1.5 Op类

1.5.1 GEMM转卷积算子的理解

1.6 Pipeline类

1.7 Session类

1.8 Executor & ExecutorScope类

MNN端侧部署教程

MNN端侧部署教程

教程概览

1. 环境配置

2. 核心概念

3. MNN-LLM

相关链接

交叉编译开发环境配置与远程调试

交叉编译开发环境配置与远程调试

1. ANDROID_NDK配置

1.1 logcat

2 VSCode Clangd配置

3 调试

3.1 服务器调试

3.2 远程调试（launch模式）

3.3 远程调试（Attach模式 尝试失败）

3.4 内存错误检测工具

远程ADB环境配置

1. 远程ADB环境配置

1.1 网络拓扑

1.2 原理

1.3 本地ADB环境配置

1.4 构建服务器->本地转发链路(S->L)

方案1: SSH端口转发

方案2: TailScale/EasyTier隧道 (推荐)

1.5 构建本地->手机的转发链路(L->P)

1.6 测试

3.3 远程调试（Attach模式尝试失败）