MIGraphX Field Guide

Practical AMD GPU inference documentation for builders using MIGraphX, ROCm, ONNX, C++, Python, Docker, and production services.

MIGraphX is AMD's graph compiler and inference runtime for optimized model execution on AMD GPUs through ROCm. Use it when you want the AMD equivalent of a compiled inference backend instead of running models through an unoptimized CPU or generic runtime path.

This guide is written as an implementation manual. It focuses on the operations you need in real projects:

install ROCm and MIGraphX
validate an AMD GPU runtime
compile ONNX models for the gpu target
save and load compiled programs
call MIGraphX from the command line, C++, and Python
run in Docker containers
benchmark latency and throughput
design an AMD backend beside TensorRT

Core Mental Model

model.onnx
  -> MIGraphX parser
  -> graph optimization and lowering
  -> compile for target: gpu
  -> saved compiled program
  -> runtime execution with input tensors

For a C++ application, the practical flow usually looks like this:

build mode:
  parse ONNX -> compile for gpu -> save program

run mode:
  load program -> allocate input/output arguments -> eval/run

That maps cleanly to services that already have a TensorRT-style build and run split.

Where To Start

Quick start

Install, validate, compile, and run a model.

C++ API

Embed MIGraphX in a native inference service.

Argus backend notes

Port a TensorRT video pipeline to an AMD backend.

Docker

Run MIGraphX inside ROCm containers.

What This Guide Is Not

This is not a replacement for AMD's official reference. It is a field guide that reorganizes the official material into workflows. When exact API signatures matter, prefer the upstream reference linked from the source links page.