MIGraphX Field Guide

Installation

Install MIGraphX on bare metal, inside ROCm containers, and in source-build environments.

MIGraphX is part of the ROCm ecosystem. Treat ROCm as the base requirement and MIGraphX as the inference compiler/runtime layered on top.

Linux Packages

After configuring the ROCm package repository for your distribution:

sudo apt update
sudo apt install migraphx

For native development, install headers and CMake package files when available:

sudo apt install migraphx-dev

Validate:

migraphx-driver --help

ROCm Health Check

Before debugging a model, check the GPU runtime:

rocminfo
rocm-smi
hipcc --version

Common failures at this layer:

  • ROCm package version does not support the GPU
  • user is not in the correct render/video group
  • kernel driver is not loaded
  • container is missing device access

Docker

Use a ROCm development image when you want reproducible builds:

docker run --rm -it \
  --device=/dev/kfd \
  --device=/dev/dri \
  --group-add video \
  --ipc=host \
  --security-opt seccomp=unconfined \
  rocm/dev-ubuntu-22.04:latest

Inside the container:

apt update
apt install -y migraphx migraphx-dev
migraphx-driver --help

For more detail, see Docker deployment.

Build From Source

Build from source when you need:

  • a patch not present in your ROCm package
  • development against MIGraphX internals
  • a specific commit for debugging
git clone https://github.com/ROCm/AMDMIGraphX.git
cd AMDMIGraphX

Follow the upstream repository instructions for dependencies and build commands. Source builds are slower to reproduce than package installs, so prefer packages for application integration unless you need a specific change.

Version Pinning

Pin ROCm and MIGraphX together. Avoid mixing a MIGraphX package from one ROCm release with HIP, rocBLAS, MIOpen, or rocBLASLt packages from another release.

Recommended lock file for production:

rocm_version: 6.x
migraphx_package_version: exact package version
gpu: exact card or accelerator
driver: kernel and amdgpu version
model: ONNX SHA256
compile_options: target, precision, input shapes

On this page