MIGraphX driver

2025-11-11

4 min read time

Applies to Linux

The MIGraphX driver is a command-line tool that allows you to utilize many of the MIGraphX core functions without having to write a program. It can read, compile, run, and test the performance of a model with randomized data.

It is installed by default when you install MIGraphX. You can find it in /opt/rocm/bin/migraphx-driver or in AMDMIGraphX/build/bin/migraphx-driver after building the source code.

Commands

The table below summarizes the MIGraphX driver commands.

Table 1 commands

Command	Description
op	Prints all operators of MIGraphX when followed by the option --list or -l
params	Prints the input and output parameter shapes
run	Compiles, allocates parameters, evaluates, and prints input graph
read	Loads and prints input graph
compile	Compiles and prints input graph
verify	Runs reference and GPU implementations and checks outputs for consistency
perf	Compiles and runs input graph followed by printing the performance report

Options

The table below summarizes the various options to be used with the MIGraphX driver commands. To learn which options can be used with which commands, see the MIGraphX driver options.

Table 2 commands

Option	Description
–help | -h	Prints help section.
–test	Test MIGraphX with single layer GEMM model.
–onnx	Loads the file as an ONNX graph.
–tf	Loads the file as a tensorflow graph.
–migraphx	Loads the file as a migraphx graph.
–migraphx-json	Loads the file as a migraphx JSON graph.
–batch	Sets batch size for a static model. Sets the batch size at runtime for a dynamic batch model.
–nhwc	Treats tensorflow format as nhwc.
–nchw	Treats tensorflow format as nchw.
–skip-unknown-operators	Skips unknown operators when parsing and continues to parse.
–trim | -t	Trims instructions from the end.
–optimize | -O	Optimizes read
–graphviz | -g	Prints a graphviz representation
–brief	Makes the output brief
–cpp	Prints the program in .cpp format
–json	Prints the program in .json format
–text	Prints the program in .txt format
–binary	Prints the program in binary format
–netron	Prints the program in Netron viewable JSON format
–output | -o	Writes output in a file
–fill0	Fills parameter with 0s
–fill1	Fills parameter with 1s
–input-dim	Sets static dimensions of a parameter
–dyn-input-dim	Sets dynamic dimensions of a parameter
–default-dyn-dim	Sets default dynamic dimension
–gpu	Compiles on the GPU
–cpu	Compiles on the CPU
–ref	Compiles on the reference implementation
–enable-offload-copy	Enables implicit offload copying
–disable-fast-math	Disables fast math optimization
–exhaustive-tune	Enables exhaustive search to find the fastest kernel
–fp16	Quantizes for fp16
–bf16	Quantizes for bf16
–int8	Quantizes for int8
–fp8	Quantize for Float8E4M3FNUZ type
–rms-tol	Sets tolerance for the RMS error (Default: 0.001)
–atol	Sets tolerance for elementwise absolute difference (Default: 0.001)
–rtol	Sets tolerance for elementwise relative difference (Default: 0.001)
–per-instruction | -i	Verifies each instruction
–reduce | -r	Reduces program and verifies
–iterations | -n	Sets the number of iterations to run for perf report
–list | -l	Lists all the MIGraphX operators

Usage

This section demonstrates the usage of MIGraphX driver tool with some commonly used options. Note that these examples use a simple MNIST ConvNet as the input graph for demonstration purposes as models of higher complexity generate considerably larger outputs in most cases.

Option: op

$ /opt/rocm/bin/migraphx-driver op –list

Option: params

$ /opt/rocm/bin/migraphx-driver params simple_graph.pb

Option: run (ONNX file input)

$ /opt/rocm/bin/migraphx-driver run –onnx simple_graph.onnx

Option: read

$ /opt/rocm/bin/migraphx-driver read simple_graph.pb

Option: compile (on GPU, quantized for fp16)

$ /opt/rocm/bin/migraphx-driver compile –gpu –fp16 simple_graph.pb

Option: verify

$ /opt/rocm/bin/migraphx-driver verify simple_graph.pb

Option: perf

$ /opt/rocm/bin/migraphx-driver perf simple_graph.pb