Python Sample Usage Guide

Directory Structure

├── README.md  Sample usage guide
├── config
│   ├── model_generator.py  Script to generate models needed for test cases
│   ├── add_func.json  UDF configuration file for test cases
│   ├── add_graph.json  Computation graph compilation configuration file for test cases
│   ├── data_flow_deploy_info.json  Deployment location configuration file for sample_udf_python
│   ├── multi_model_deploy.json  Deployment location configuration file for sample_multiple_model
│   ├── sample_npu_model_deploy_info.json  Deployment location configuration file for sample_npu_model
│   ├── sample_pytorch_deploy_info.json  Deployment location configuration file for sample_pytorch
│   └── invoke_func.json  UDF call NN compilation configuration file for test cases
├── sample_base.py  This sample demonstrates basic DataFlow API graph construction, including construction and execution of three types of nodes: UDF, GraphPp, and UDF executing NN inference
├── sample_udf_python.py  This sample demonstrates python dataflow calling udf python process
├── sample_exception.py  This sample demonstrates enabling exception reporting
├── sample_pytorch.py  This sample demonstrates DataFlow combined with pytorch for online model inference
├── sample_npu_model.py  This sample demonstrates DataFlow online inference, with udf using npu_model to implement model sinking and data sinking scenarios
├── sample_multiple_model.py  This sample constructs pytorch model as funcPp through decorator, and concatenates it with GraphPp executing onnx/pb models
├── sample_perf.py  This sample is a performance profiling sample
├── udf_py
│   ├── udf_add.py  Implement udf multi-func functionality using python
│   └── udf_control.py  Implement udf functionality using python, used to control which func in udf_add is actually executed
└── udf_py_ws_sample  Complete sample for explaining python udf implementation
    ├── CMakeLists.txt  UDF python complete project cmake file sample
    ├── func_add.json   UDF python complete project configuration file sample
    ├── src_cpp
    │   └── func_add.cpp  UDF python complete project C++ source file sample
    └── src_python
        └── func_add.py  UDF python complete project python source file sample

Environment Preparation

• Refer to Environment Preparation to download and install driver/firmware/CANN software packages;

• Python version requirement: python3.12. The specific version should match the python version used when compiling the dataflow wheel package. If you need to use a different python version, you can refer to Compilation to recompile the ge_compiler package and install it;

• The model generation script model_generator.py in the config directory depends on tensorflow-cpu and onnx, which need to be installed via pip3 install tensorflow-cpu and pip3 install onnx;

• sample_pytorch.py, sample_npu_model.py samples depend on torch_npu and torchvision packages. torch_npu needs to install corresponding torch and torch_npu packages according to the actual environment (it is recommended to use version 2.1.0 or above, Acquisition Method). torchvision has a matching relationship with torch, and should be installed using pip3 install torchvision after torch is installed.

Running Samples

The numa_config.json file configures device resource information required by dflow. You can use the script in the tools directory to auto-generate a single-node configuration. For custom configuration, refer to numa_config Field Description and Samples:

# Generate single-node numa_config.json (auto-detects chip type and device count, uses all devices by default)
# You can also specify a device list: python3 ../tools/create_numa_config.py --device-list 0,1,2
# If npu-smi is unavailable, the script prints a warning and uses defaults (Ascend910B1, 8 devices); specify --soc-version and --total-dev if they do not match the actual environment
python3 ../tools/create_numa_config.py
# After generation, numa_config.json will be created in the current directory

# Execute the tensorflow raw model generation script in the config directory:
python3 config/model_generator.py
# After execution, add.pb model and simple_model.onnx model will be generated in the config directory

# Optional
export ASCEND_GLOBAL_LOG_LEVEL=3       #0 debug 1 info 2 warn 3 error. Default is error level if not set
# Required
source {HOME}/Ascend/cann/set_env.sh #{HOME} is the CANN software package installation directory, please replace according to the actual installation path
export RESOURCE_CONFIG_PATH=./numa_config.json # Please specify according to the actual generated file path

python3.12 sample_base.py
python3.12 sample_udf_python.py
python3.12 sample_exception.py
python3.12 sample_pytorch.py
python3.12 sample_npu_model.py
python3.12 sample_multiple_model.py
python3.12 sample_perf.py

# Unset this environment variable to prevent affecting non-dflow test cases
unset RESOURCE_CONFIG_PATH