企业官网建设流程全解析

AddZeroPass Python Sample Usage Guide

【免费下载链接】geGE（Graph Engine）是面向昇腾的图编译器和执行器，提供了计算图优化、多流并行、内存复用和模型下沉等技术手段，加速模型执行效率，减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力，并同时支持 onnx、pb 等主流模型格式的解析与编译。项目地址: https://gitcode.com/cann/ge

Feature Description

This directory provides two pure PythonPatternFusionPasssamples:

• src/python_add_zero_pass_const_value_match.py
• src/python_add_zero_pass.py

The two samples have different purposes:

• python_add_zero_pass_const_value_match.py

  • Demonstrates Python V1PatternMatcherConfigBuilder.enable_const_value_match()
  • Front-loadsAdd(x, 0.0f)to matcher phase via strict const-value-match
  • @patternmethod uses Python expression to describe strictly matchedAdd(x, 0.0f), e.g.,return inputs[0] + 0.0
  • CurrentConstantMatcher::IsMatchvalue matching is strict binary match without floating-point tolerance or cross-dtype normalization
  • Better suited as matcher_config example than fully equivalent version of C++ sample

• python_add_zero_pass.py

  • Main logic aligned with C++ pass sample
  • @patternmethod uses Python expression to describeData + Consttopology, e.g.,return inputs[0] + 0
  • Multi-input pattern can explicitly declare input count viax, y, z = inputs[:3]
  • Multi-pattern pass can declare multiple@patternmethods; legacypatterns(self)returning multiplePattern/Graphstill compatible
  • @patternmethod cannot be used simultaneously withpatterns(self)to avoid ambiguous pattern declaration sources
  • Python pass framework automatically createsGraphBuilder, sets graph outputs, and auto-captures in "visited inputs,returnoutputs" order; legacy explicitGraphBuildersyntax still compatible
  • meet_requirements()explicitly reads matchedConst.valueand judges zero value per same rules as C++ sample
  • Currently supportsDT_FLOAT,DT_DOUBLE,DT_INT32consistent with C++ sample

Recommended expression pattern syntax:

from ge.passes import PatternFusionPass, pattern class PythonAddZeroPass(PatternFusionPass): @pattern def add_zero(self, inputs): return inputs[0] + 0 def replacement(self, inputs): return inputs[0]

Directory Structure

python/ ├── README.md // Python sample documentation ├── CMakeLists.txt // Build script for generating es_all Python ES API ├── src │ ├── python_add_zero_pass.py // Python pass implementation file │ ├── python_add_zero_pass_const_value_match.py // const value match example

Prerequisites

• CANN environment variables configured viasource ${ASCEND_PATH}/set_env.sh, see C++ sample README environment configuration steps
• CANN software package installation per Environment preparation
• Graph compilation flow Python dependencies installed:attrs,decorator,sympy,numpy,psutil,scipy

run package includes GE Python runtimege_pywheel, no need to install separatege_py-*.whl. Python pass runtime loadspybind11-based precompiled binary components. CANN package provides artifacts matching current Python version; if no match, fallback compilation occurs automatically. Fallback compilation requirespybind11installed in current Python environment.

Usage

1. Set Python pass plugin path:

   export ASCEND_GE_PY_PASS_PATH=$(pwd)/src/python_add_zero_pass_const_value_match.py

   For version aligned with C++ main logic:

   export ASCEND_GE_PY_PASS_PATH=$(pwd)/src/python_add_zero_pass.py

2. Follow C++ sample README program execution section for verification. For offline scenarios, replaceatccommand in C++ README withpyatc; both have identical command-line parameters,pyatcruns in current Python interpreter process.

3. Notes:

   • These two samples are not standalone execution scripts, directly runningpython src/python_add_zero_pass.pyorpython src/python_add_zero_pass_const_value_match.pywon't trigger pass execution
   • Expected output prints after GE compilation flow actually loads this Python pass

Expected Logs

After successful execution, logs show similar output (actual input_0 name may differ):

[PythonAddZeroConstValueMatchPass] matched=PythonAddZeroConstValueMatchPass_add_zero_pattern captured=input_0:0 [PythonAddZeroPass] matched=PythonAddZeroPass_add_zero_pattern captured=input_0:0 const_dtype=DT_FLOAT zero=True

【免费下载链接】geGE（Graph Engine）是面向昇腾的图编译器和执行器，提供了计算图优化、多流并行、内存复用和模型下沉等技术手段，加速模型执行效率，减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力，并同时支持 onnx、pb 等主流模型格式的解析与编译。项目地址: https://gitcode.com/cann/ge