FuseMatMulAndAddPass Python Sample Usage Guide

【免费下载链接】ge GE（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

This directory provides a pure Python version of pattern_base_pass/1_fuse_matmul_add_pass, with logic consistent with C++ FuseMatMulAndAddPass:

• Pattern 0: MatMul(a, b) → Add(..., c), graph inputs 0/1/2 correspond to a/b/c
• Pattern 1: BatchMatMulV2(a, b) → Add(..., c), same three-input topology
• Replacement: GEMM(r_a, r_b, r_c, alpha=1, beta=1) (scalar 1.0 aligns with C++ CreateScalar(1))
• Inherits PatternFusionPass, implements patterns() / meet_requirements() / replacement(); phase is BeforeInferShape

No additional matcher configuration like enable_const_value_match() enabled, consistent with C++ default matching behavior.

Directory Structure

python/
├── README.md                     // Python sample documentation
├── CMakeLists.txt                // Build script for generating es_all Python ES API
├── src
│   ├── python_fuse_matmul_add_pass.py // Python pass implementation file

Prerequisites

• CANN environment variables configured via source ${ASCEND_PATH}/set_env.sh, see C++ sample README environment configuration steps
• GE Python package importable (includes ge.passes and pass loading chain)

run package typically includes ge_py wheel, no need to install separate ge_py-*.whl. Python pass runtime loads pybind11-based precompiled binary components. CANN package provides artifacts matching current Python version; if no match, fallback compilation occurs automatically. Fallback compilation requires pybind11 installed in current Python environment. If execution reports missing BatchMatMulV2, GEMM etc. ES API, follow "ES API Missing Handling (Optional)" below to generate and load es_all.

Usage

Commands below default to execution in 1_fuse_matmul_add_pass/python directory.

1. Set environment variable for GE to load this Python pass during compilation:

export ASCEND_GE_PY_PASS_PATH=$PWD/src/python_fuse_matmul_add_pass.py

1. Follow C++ sample README program execution section for verification.

2. Notes:

   • This sample is not a standalone execution script, directly running python src/python_fuse_matmul_add_pass.py won't trigger pass execution
   • If execution reports missing BatchMatMulV2, GEMM etc. ES API, generate and load es_all first per below, then rerun

ES API Missing Handling (Optional)

If execution reports missing BatchMatMulV2, GEMM etc. ES API, generate es_all via this Python directory's CMakeLists.txt:

cmake -S . -B build
cmake --build build --target build_es_all -j$(nproc)

Install generated Python package and ensure current Python process can find the package and corresponding dynamic library:

pip install --force-reinstall --upgrade --target ./build/whl_package ./build/es_output/whl/es_all-1.0.0-py3-none-any.whl
export PYTHONPATH="$PWD/build/whl_package:${PYTHONPATH:-}"
export LD_LIBRARY_PATH="$PWD/build/es_output/lib64:${LD_LIBRARY_PATH:-}"

Expected Behavior

Similar to C++ sample, when pattern/replacement invoked:

Define pattern for FuseMatMulAndAddPass
Define replacement for FuseMatMulAndAddPass

【免费下载链接】ge GE（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