mshr-h / tvm-relax-cpp-example

Minimum example for deploying Apache TVM's Relax IR using C++ API

TVM's Relax Deployment Example in C++

This repository provides a minimal example for deploying Apache TVM's Relax IR using the C++ API.

Note: This project supports Apache TVM FFI.

Overview

This project demonstrates how to build and deploy a Relax IR model using TVM's C++ API. It includes instructions for setting up dependencies, building TVM from source, and running the example binary.

Prerequisites

Ensure the following tools and libraries are installed on your system:

• CMake
• Ninja (build system)
• Python (for virtual environment setup)
• LLVM (with llvm-config available in your PATH)

Building TVM from Source

Follow these steps to build TVM from source:

1. Set up a Python virtual environment:

   python3 -m venv .venv
   source .venv/bin/activate

2. Install required Python packages:

   pip install cmake ninja setuptools cython

3. Build TVM: Navigate to the 3rdparty directory and build TVM:

   cd 3rdparty
   ./build-tvm.sh --clean --llvm llvm-config

Building the Example Project

Once TVM is built, follow these steps to build the example project:

1. Configure the build using CMake:

   cmake -B build -G Ninja

2. Build the project using Ninja:

   ninja -C build

Exporting the Relax Library

Before running the example, export the Relax library using the provided Python script:

python export_relax_library.py

Running the Example

After building the project, run the example binary:

./build/main

You'll see the output below:

Found vm_load_executable()
Found vm_initialization()
vm initialized
Found main()
Input array initialized
output: 
  1
  2
  3
  4
  5
  6
  7
  8
  9