### Configure NNAL Environment

Source: https://github.com/cann/ascend-transformer-boost/blob/master/ops_customize/README.md

Sources the environment setup script for NNAL. Assumes installation path is ${HOME}/Ascend.

```shell
source ${HOME}/Ascend/nnal/atb/set_env.sh
```

--------------------------------

### mlapa_demo.cpp Example Configuration

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/op_demo/multi_latent_attention/README.md

Illustrates the parameter settings for the mlapa_demo.cpp example. Note the qkScale value represents the headSize before MLA's rope projection.

```cpp
#define headNum 128
#define qkScale (1.0/sqrt(576))
#define kvHeadNum 1
#define maskType UNDEFINED
#define calcType CALC_TYPE_UNDEFINED
#define cacheMode INT8_NZCACHE
```

--------------------------------

### mlapa_ds_demo.cpp Example Configuration

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/op_demo/multi_latent_attention/README.md

Shows the parameter configuration for the mlapa_ds_demo.cpp example. The qkScale value is a specific floating-point number.

```cpp
#define headNum 128
#define qkScale 0.1352667747812271
#define kvHeadNum 1
#define maskType UNDEFINED
#define calcType CALC_TYPE_UNDEFINED
#define cacheMode KROPE_CTKV
```

--------------------------------

### Compiling Operation Example

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/日志与调试.md

Navigate to the specific operation's demo directory and execute the build script to compile the operation example. This is part of running ATB acceleration library op_demo calls.

```bash
bash build.sh
```

--------------------------------

### Install msprobe and atb_probe Module (8.5.0+)

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/日志与调试.md

For CANN 8.5.0 and later, clone the msprobe repository, install necessary packages, and build the atb_probe module. This prepares the environment for data dumping.

```bash
git clone https://gitcode.com/Ascend/msprobe.git
cd msprobe

pip install setuptools wheel

python3 setup.py bdist_wheel --include-mod=atb_probe --no-check
cd ./dist
pip install ./mindstudio_probe*.whl
```

--------------------------------

### Install NPU Driver with Debug Parameter

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/日志与调试.md

Install the NPU driver using the --debug parameter. Refer to the official documentation for specific installation instructions.

```bash
./Ascend-hdk-<chip_type>-npu-driver_<version>_linux-<arch>.run --debug

```

--------------------------------

### Install CANN Toolkit

Source: https://github.com/cann/ascend-transformer-boost/blob/master/README.md

Execute the CANN toolkit installer script to perform the installation. Ensure the toolkit version matches the ATB version.

```shell
./Ascend-cann-toolkit_${VERSION}_linux-$(arch).run --install
```

--------------------------------

### Execute the Example

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/atb_aclnn/README.md

Run the compiled test_model executable to perform the integrated ACLNN and ATB operation test.

```sh
./build/test_model
```

--------------------------------

### Install NPU Driver with Full Debugging Support

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/日志与调试.md

Install the NPU driver with the --full parameter and enable the debug channel by writing 1 to /proc/debug_switch. This method is recommended for versions after CANN 8.1.RC1 with drivers after 25.0.RC1.

```bash
./Ascend-hdk-<chip_type>-npu-driver_<version>_linux-<arch>.run --full

```

--------------------------------

### Navigate to atb-aclnn Directory

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/atb_aclnn/README.md

Change the current directory to the atb-aclnn directory to begin the setup process.

```sh
cd  ${用例所在目录}/atb-aclnn
```

--------------------------------

### Install torch_atb WHL Package

Source: https://github.com/cann/ascend-transformer-boost/blob/master/README.md

Install the torch_atb Python package from a generated WHL file. This is part of the ATB Python API setup.

```sh
pip3 install torch_atb-{version}-py3-none-any.whl
```

--------------------------------

### Set up ATB Environment Variables

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/atb_aclnn/README.md

Source the set_env.sh script from the ATB installation directory to configure ATB-specific environment variables. Replace ${nnal安装目录} with the actual installation path.

```sh
source ${nnal安装目录}/atb/set_env.sh
```

--------------------------------

### Install CANN Toolkit

Source: https://github.com/cann/ascend-transformer-boost/blob/master/ops_customize/README.md

Installs the CANN toolkit. Ensure you replace $(version) and $(arch) with your specific version and architecture.

```shell
chmod +x Ascend-cann-toolkit_$(version)_linux-$(arch).run
./Ascend-cann-toolkit_$(version)_linux-$(arch).run --install
```

--------------------------------

### Install Ascend-cann-nnal run package

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/常见问题与回答.md

Install the Ascend-cann-nnal run package. This requires the CANN environment. You can check package integrity and install it using the provided commands.

```shell
chmod +x 软件包名.run
```

```shell
./软件包名.run --check
```

```shell
./软件包名.run --install
```

--------------------------------

### Install NNAL

Source: https://github.com/cann/ascend-transformer-boost/blob/master/ops_customize/README.md

Installs the NNAL component. Replace $(version) and $(arch) with your specific version and architecture.

```shell
chmod +x Ascend-cann-nnal_$(version)_linux-$(arch).run
./Ascend-cann-nnal_$(version)_linux-$(arch).run --install
```

--------------------------------

### Install Python Dependencies for CANN

Source: https://github.com/cann/ascend-transformer-boost/blob/master/README.md

Install necessary Python third-party libraries for CANN runtime. Use the --user flag unless installing as root.

```shell
pip3 install attrs cython 'numpy>=1.19.2,<=1.24.0' decorator sympy cffi pyyaml pathlib2 psutil protobuf==3.20.0 scipy requests absl-py --user
```

--------------------------------

### Install CANN Operator Package

Source: https://github.com/cann/ascend-transformer-boost/blob/master/README.md

Execute the CANN operator package installer. Ensure the corresponding Toolkit version is installed and environment variables are configured.

```shell
./Ascend-cann-${chip_type}-ops_${VERSION}_linux-$(arch).run --install
```

--------------------------------

### Environment Variable Setup Script

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/编译与构建.md

After installation, the `scripts/set_env.sh` script can be used to set process-level environment variables for the ATB library. These variables are automatically unset when the user process ends.

```shell
source scripts/set_env.sh
```

--------------------------------

### Prepare Environment for Integrated Compilation

Source: https://github.com/cann/ascend-transformer-boost/blob/master/ops_customize/README.md

Sets up environment variables required for compiling the ATB acceleration library with custom operators. Assumes installation path is ${HOME}/Ascend.

```shell
source ${HOME}/Ascend/ascend-toolkit/set_env.sh
source ${HOME}/Ascend/nnal/atb/set_env.sh
export ATB_BUILD_DEPENDENCY_PATH=${ATB_HOME_PATH}
```

--------------------------------

### Compile and Run rope_qwen_demo_0.cpp

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/op_demo/rope/README.md

To compile and run rope_qwen_demo_0.cpp, replace 'rope_demo.cpp' with the new filename in the build script. This example is specific to certain Atlas hardware series.

```sh
# Replace rope_demo.cpp with rope_qwen_demo_0.cpp in build script
```

--------------------------------

### Compile and Run rope_qwen_demo_1.cpp

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/op_demo/rope/README.md

To compile and run rope_qwen_demo_1.cpp, replace 'rope_demo.cpp' with the new filename in the build script. This example is specific to certain Atlas hardware series.

```sh
# Replace rope_demo.cpp with rope_qwen_demo_1.cpp in build script
```

--------------------------------

### Compile and Run rope_qwen_demo_2.cpp

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/op_demo/rope/README.md

To compile and run rope_qwen_demo_2.cpp, replace 'rope_demo.cpp' with the new filename in the build script. This example is specific to certain Atlas hardware series.

```sh
# Replace rope_demo.cpp with rope_qwen_demo_2.cpp in build script
```

--------------------------------

### Project Directory Setup

Source: https://github.com/cann/ascend-transformer-boost/blob/master/src/kernels/CMakeLists.txt

Sets key project directories, including the project root and third-party directory. These variables are used to locate project-specific files and dependencies.

```cmake
set(OPS_PROJECT_ROOT_DIR "${CMAKE_CURRENT_LIST_DIR}")
set(OPS_THIRD_PARTY_DIR "${PROJECT_SOURCE_DIR}/3rdparty")
set(MKI_PACKAGE_DIR "${OPS_THIRD_PARTY_DIR}/mki")
set(MKI_SCRIPT_DIR "${MKI_PACKAGE_DIR}/scripts")
```

--------------------------------

### Set up CANN Environment Variables

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/atb_aclnn/README.md

Source the set_env.sh script to configure the necessary environment variables for the CANN toolkit. Replace ${toolkit安装目录} with the actual installation path.

```sh
source ${toolkit安装目录}/set_env.sh
```

--------------------------------

### Make CANN Toolkit Executable

Source: https://github.com/cann/ascend-transformer-boost/blob/master/README.md

Before running the CANN toolkit installer, make it executable using chmod. Replace ${VERSION} with the specific CANN version.

```shell
chmod +x Ascend-cann-toolkit_${VERSION}_linux-$(arch).run  # 其中${VERSION}表示对应的CANN版本，如9.1.0
```

--------------------------------

### Registering Kernel Functions with add_kernel

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/开发指南.md

This example demonstrates the usage of the `add_kernel` command in CMakeLists.txt to register kernel functions with the Ascend runtime system. It specifies the operator name, SOC version, operator type, kernel file path, and kernel class name.

```cmake
add_kernel(算子名称 [soc_version] 算子类型
    文件路径
    kernel类名称)

```

--------------------------------

### Set ATB_BUILD_DEPENDENCY_PATH for nnal

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/常见问题与回答.md

Set the ATB_BUILD_DEPENDENCY_PATH environment variable to point to the nnal installation. This is required to resolve the 'libtbe_adapter.so does not exist' error. Ensure nnal package versions match toolkit and kernels.

```shell
export ATB_BUILD_DEPENDENCY_PATH={nnal install path}/nnal/atb/latest/atb/cxx_abi_{cxx_abi_version}
```

--------------------------------

### Python ATB Operator Call Example

Source: https://github.com/cann/ascend-transformer-boost/blob/master/README.md

Example demonstrating how to call an ATB operator using the Python API. Ensure PyTorch and torch_npu are installed and torch_atb is imported.

```python
import torch
import torch_atb#导入ATB Python API模块

#创建参数对象
linear_param = torch_atb.LinearParam()
linear_param.has_bias = False

#创建算子对象
op = torch_atb.Operation(linear_param)

#准备输入数据
x = torch.randn(2, 3, dtype=torch.float16).npu()  
y = torch.randn(2, 3, dtype=torch.float16).npu()

#使用forward方法完成操作，并获取输出
outputs = op.forward([x, y]) 
torch.npu.synchronize()
```

--------------------------------

### Build Demo Script

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/op_demo/activation/README.md

Execute this shell script to build the C++ demo. Ensure environment variables are set correctly before running.

```shell
bash build.sh
```

--------------------------------

### Run Quantization Demo

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/op_demo/linear_parallel/README.md

Execute the compiled demo for the quantization scenario.

```sh
./linear_parallel_demo
```

--------------------------------

### Build the Project

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/atb_aclnn/README.md

Execute the build.sh script to compile the project and generate the necessary executable files.

```sh
bash ./build.sh
```

--------------------------------

### Build Script for Addcustom Demo

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/从开发一个简单算子出发.md

Compiles and executes the addcustom_demo.cpp file. It determines the correct C++ ABI flag and links against ATB and Ascend CL libraries.

```shell
cxx_abi=$(python3 -c '
try:
    import torch
    print("1" if torch.compiled_with_cxx11_abi() else "0")
except ImportError:
    print("1")
')

echo "Using cxx_abi=$cxx_abi"

g++ -D_GLIBCXX_USE_CXX11_ABI=$cxx_abi -I "${ATB_HOME_PATH}/include" -I "${ASCEND_HOME_PATH}/include" -L "${ATB_HOME_PATH}/lib" -L "${ASCEND_HOME_PATH}/lib64" \
addcustom_demo.cpp ../demo_util.h -l atb -l ascendcl -o addcustom_demo
./addcustom_demo

```

--------------------------------

### Determine CXX ABI Version

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/常见问题与回答.md

Determine the correct CXX ABI version (0 or 1) to use for the ATB_BUILD_DEPENDENCY_PATH. This is based on your PyTorch installation. If PyTorch is not installed, cxx_abi_1 is used by default.

```python
import torch
print(1 if torch.compiled_with_cxx11_abi() else 0)
```

--------------------------------

### Full Compilation Process for ATB Test Framework

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/测试框架指南.md

This snippet outlines the complete steps to compile the ATB test framework, including setting up CANN environment variables, building the framework, and setting up the ATB runtime environment.

```shell
cd ascend-transformer-boost

# 1. Set CANN environment variables (based on actual installation path)
source /usr/local/Ascend/ascend-toolkit/set_env.sh

# 2. Compile the test framework (including ATB core library + C++ test framework + Python test tools)
bash scripts/build.sh testframework

# 3. Set up ATB runtime environment
source output/atb/set_env.sh
```

--------------------------------

### CSV Test Case Example: Parameter Validation Error

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/测试框架指南.md

A negative test case example demonstrating a parameter validation error. It specifies an expected error code 'C:ERROR_INVALID_PARAM' for the 'LinearOperation' operator.

```default
146|MatmulCError|LinearOperation|{"hasBias":false,"enAccum":true}|0||||0|||||||||||Ascend310P|C:ERROR_INVALID_PARAM
```

--------------------------------

### CSV Test Case Example: Matrix Multiplication with Transpose

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/测试框架指南.md

A positive test case example for matrix multiplication with transpose parameters enabled. It includes specific operator parameters like 'transposeA' and 'transposeB'.

```default
23|MatmulTransposeA|LinearOperation|{"transposeA":true,"transposeB":false}|3|float16;float16;float16|nd;nd;nd|3,2;3,4;1,4|1|float16|nd|2,4|customize;customize;customize|-2,2;-2,2;-2,2||||||Ascend910B,Ascend310P|NO_ERROR
```

--------------------------------

### Configure and Load ATB Probe for Data Collection (8.5.0+)

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/日志与调试.md

Create a config.json file to specify dump parameters. Then, load the atb_probe module using the provided script, pointing to the output directory and configuration file. This prepares for data collection.

```json
{
    	"task": "tensor",
    	"dump_enable": true,
   	 "exec_range": "all",
    	"ids": "0",
    	"op_name": "",
    	"save_child": false,
    	"device": "",
    	"filter_level": 1
    }
```

```bash
MSPROBE_HOME_PATH=/usr/local/lib/python3.11/site-packages
source $MSPROBE_HOME_PATH/msprobe/scripts/atb/load_atb_probe.sh --output=$PWD --config=$PWD/config.json
# 以ATB加速库中的demo为例：
cd ${home_path}/example/op_demo/mla_preprocess && bash build.sh
```

--------------------------------

### CSV Test Case Example: Basic Matrix Multiplication

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/测试框架指南.md

A positive test case example for basic matrix multiplication using the CSV format. It specifies input tensor details, operator parameters, and expected output format.

```default
1|MatmulElewiseAdd|LinearOperation|{}|3|float16;float16;float16|nd;nd;nd|2,3;4,3;4|1|float16|nd|2,4|customize;customize;customize|-2,2;-2,2;-2,2||||||Ascend910B,Ascend310P|NO_ERROR
```

--------------------------------

### Generate Build System with CXX11 ABI Enabled

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/multiStream/README.md

Create a build directory and generate the build system using CMake, enabling the C++11 ABI.

```sh
mkdir build && cd build
cmake .. -DUSE_CXX11_ABI=ON
```

--------------------------------

### Compile and Run Project

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/multiStream/README.md

Compile the project using CMake and then run the executable.

```sh
cmake --build .
./multiStreamDemo
```

--------------------------------

### Configure Executable for Inter Graph Synchronization

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/multiStream/README.md

Modify the CMakeLists.txt file to specify the executable for the inter-graph synchronization example.

```sh
add_executable(multiStreamDemo multiStream_multiGraph_demo.cpp)
```

--------------------------------

### Build ATB Library

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/编译与构建.md

Navigate to the ATB root directory and execute the build script. This process involves fetching and compiling operator libraries (MKI) and then compiling the ATB library itself.

```shell
cd ascend-transformer-boost
bash scripts/build.sh
```

--------------------------------

### Configure Executable for Single Graph Multi-Stream

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/multiStream/README.md

Modify the CMakeLists.txt file to specify the executable for the intra-graph multi-stream parallel example.

```sh
add_executable(multiStreamDemo multiStream_singleGraph_demo.cpp)
```

--------------------------------

### SelfAttentionParam Configuration for Encoder

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/doxygen/atb_document.txt

Illustrates the configuration of SelfAttentionParam for the PA_ENCODER calculation type. This setup is used for the full pass stage.

```cpp
SelfAttentionParam.calcType = PA_ENCODER;
```

--------------------------------

### Transpose Operation Example

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/doxygen/atb_document.txt

Demonstrates how to use the TransposeParam to permute the dimensions of a tensor. Ensure the input and output tensors are correctly defined.

```cpp
intensor0:
 tensor([[0, 1],
        [2, 3]])
atb::infer::TransposeParam transposeParam;
transposeParam.perm  ={1, 0};
out_tensor:
   tensor([[0, 2],
          [1, 3]])
```

--------------------------------

### Compile ATB Acceleration Library

Source: https://github.com/cann/ascend-transformer-boost/blob/master/README.md

Navigate to the ATB directory and run the build script to compile the acceleration library. Source the environment script after compilation.

```sh
cd ascend-transformer-boost
bash scripts/build.sh
source output/atb/set_env.sh
```

--------------------------------

### Get Tiling Data for FastSoftMax

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/开发指南.md

Defines a function to retrieve `FastSoftMaxTilingData` from a given `tiling` address. This is used in ATB operator development.

```C++
inline __aicore__ FastSoftMaxTilingData GetTilingData(const GM_ADDR tiling)
{
    auto tilingDataPointer = reinterpret_cast<const __gm__ FastSoftMaxTilingData *>(tiling);
    FastSoftMaxTilingData tilingData;
    tilingData.batchSize = tilingDataPointer->batchSize;
    tilingData.headNum = tilingDataPointer->headNum;
    return tilingData;
}
```

--------------------------------

### Addcustom Demo C++ Code

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/从开发一个简单算子出发.md

Demonstrates how to create and call the AddcustomOperation. This code prepares input tensors, sets up the operation, executes it, and cleans up resources.

```c++
#include "../demo_util.h"

const int32_t DEVICE_ID = 0;
const uint32_t DIM_0 = 8;
const uint32_t DIM_1 = 512;

/**
 * @brief 准备 Addcustom 输入 tensor（两输入同 shape）
 */
atb::Status PrepareInTensor(atb::Context *contextPtr, aclrtStream stream, atb::SVector<atb::Tensor> &inTensors)
{
    // 输入 x：[8, 512]，int32，全 1
    atb::Tensor x;
    CHECK_STATUS(CreateTensorFromVector(contextPtr, stream, std::vector<int32_t>(DIM_0 * DIM_1, 1), ACL_INT32,
                                        aclFormat::ACL_FORMAT_ND, {DIM_0, DIM_1}, x));

    // 输入 y：[8, 512]，int32，全 2
    atb::Tensor y;
    CHECK_STATUS(CreateTensorFromVector(contextPtr, stream, std::vector<int32_t>(DIM_0 * DIM_1, 2), ACL_INT32,
                                        aclFormat::ACL_FORMAT_ND, {DIM_0, DIM_1}, y));

    inTensors = {x, y};
    return atb::ErrorType::NO_ERROR;
}

/**
 * @brief 创建 AddcustomOperation
 */
atb::Status CreateAddcustomOperation(atb::Operation **addOp)
{
    atb::infer::AddcustomParam param;
    param.addcustomDim = 0;
    CHECK_STATUS(atb::CreateOperation(param, addOp));
    return atb::ErrorType::NO_ERROR;
}

int main(int argc, char **argv)
{
    CHECK_STATUS(aclInit(nullptr));
    int32_t deviceId = 0;
    CHECK_STATUS(aclrtSetDevice(deviceId));

    atb::Context *context = nullptr;
    CHECK_STATUS(atb::CreateContext(&context));
    void *stream = nullptr;
    CHECK_STATUS(aclrtCreateStream(&stream));
    context->SetExecuteStream(stream);

    /* 创建 op */
    atb::Operation *addOp = nullptr;
    CHECK_STATUS(CreateAddcustomOperation(&addOp));

    /* 准备输入输出 */
    atb::VariantPack variantPack;
    CHECK_STATUS(PrepareInTensor(context, stream, variantPack.inTensors));

    atb::Tensor output;
    CHECK_STATUS(CreateTensorFromVector(context, stream, std::vector<int32_t>(DIM_0 * DIM_1, 0),  // 占位即可
                                        ACL_INT32, aclFormat::ACL_FORMAT_ND, {DIM_0, DIM_1}, output));
    variantPack.outTensors = {output};

    /* Setup -> Execute */
    uint64_t workspaceSize = 0;
    CHECK_STATUS(addOp->Setup(variantPack, workspaceSize, context));
    uint8_t *workspacePtr = nullptr;
    if (workspaceSize > 0) {
        CHECK_STATUS(aclrtMalloc((void **)&workspacePtr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST));
    }
    CHECK_STATUS(addOp->Execute(variantPack, workspacePtr, workspaceSize, context));
    CHECK_STATUS(aclrtSynchronizeStream(stream));

    /* 释放资源 */
    for (auto &t : variantPack.inTensors) aclrtFree(t.deviceData);
    for (auto &t : variantPack.outTensors) aclrtFree(t.deviceData);
    if (workspacePtr) {
        aclrtFree(workspacePtr);
    }
    atb::DestroyOperation(addOp);
    aclrtDestroyStream(stream);
    DestroyContext(context);
    aclFinalize();
    std::cout << "Addcustom demo success!" << std::endl;
    return 0;
}

```

--------------------------------

### Set Environment Script

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/op_demo/linear_parallel/README.md

Source these scripts to set up the necessary environment variables for CANN and nnal packages. Ensure the paths are correct for your installation.

```sh
source [cann安装路径]/set_env.sh
    默认：source /usr/local/Ascend/ascend-toolkit/set_env.sh
```

```sh
source [nnal安装路径]/set_env.sh  
    默认：source /usr/local/Ascend/nnal/atb/set_env.sh  
    ①. 如果使用加速库源码编译，source [加速库源码路径]/output/atb/set_env.sh  
    例如： source ./ascend-transformer-boost/output/atb/set_env.sh
```

--------------------------------

### Create Local Working Directory

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/contributors/gitcode-workflow.md

Use this command to create a directory on your local machine to store the project's code.

```default
mkdir ${your_working_dir}
```

--------------------------------

### ATB算子接入：op_list.yaml 配置

Source: https://github.com/cann/ascend-transformer-boost/blob/master/docs/开发指南.md

在op_list.yaml文件中添加新增算子信息，以确保后续构建能够正确编译新增算子的实现和接口。此配置对于非融合算子需要修改kernels/op_list.yaml。

```yaml
FastSoftMaxOperation:
    FastSoftMaxKernel:
        ascend910b: true

```

--------------------------------

### Configure CANN Environment Variables

Source: https://github.com/cann/ascend-transformer-boost/blob/master/README.md

Source the set_env.sh script to configure the environment variables for the CANN toolkit. The installation path is assumed to be ${HOME}/Ascend.

```shell
source ${HOME}/Ascend/ascend-toolkit/set_env.sh
```

--------------------------------

### Include Directories Setup

Source: https://github.com/cann/ascend-transformer-boost/blob/master/src/kernels/CMakeLists.txt

Configures include directories for the project. This allows the compiler to find header files located in the project root and its tbe_adapter subdirectory.

```cmake
include_directories(
    ${OPS_PROJECT_ROOT_DIR}
    ${OPS_PROJECT_ROOT_DIR}/tbe_adapter/include
)
```

--------------------------------

### Set Environment Script for NNAL

Source: https://github.com/cann/ascend-transformer-boost/blob/master/example/op_demo/paged_cache_load/README.md

Source this script to set up the environment for NNAL. The default path is /usr/local/Ascend/nnal/atb/set_env.sh. If compiling from source, use the provided path.

```sh
source /usr/local/Ascend/nnal/atb/set_env.sh
```

```sh
source ./ascend-transformer-boost/output/atb/set_env.sh
```