Tune model performance

Tune model performance#

2026-10-12

2 min read time

Applies to Linux

MIGraphX exposes environment variables that change compilation behavior and kernel selection at runtime. Set these variables before you compile or run a model to experiment with performance options without changing application code.

For the complete list of environment variables, including developer-only tracing and debugging options, see MIGraphX environment variables.

Set environment variables#

Export a variable in your shell before running migraphx-driver or your application. Forcing eligible operations through rocMLIR is a common starting point for performance experiments:

export MIGRAPHX_MLIR_USE_SPECIFIC_OPS="attention,convolution,dot,fused"
migraphx-driver perf model.onnx --onnx --gpu

Replace model.onnx with your model path.

Variables that take comma-separated lists must not include spaces between entries. Quote the value when your shell requires it.

Common tuning variables#

The following variables are the most commonly used for model performance tuning. Each entry describes accepted values and default behavior documented in the MIGraphX source tree.

Environment variable

Values

MIGRAPHX_GPU_OPTIONS

JSON object of backend options (for example, {convolution_layout:channels_last}). Quotes around keys and values are optional. Unrecognized options are ignored.

MIGRAPHX_SET_GEMM_PROVIDER

hipblaslt or rocblas. Default: rocblas on gfx90a; hipblaslt on all other architectures.

MIGRAPHX_ENABLE_GEMM_TUNING

1 enables exhaustive GEMM tuning even when --exhaustive-tune is not set. 0 returns to default behavior.

MIGRAPHX_ENABLE_HIP_GEMM_TUNING

1 enables exhaustive hipBLASLt tuning. 0 returns to default behavior.

MIGRAPHX_MLIR_TUNE_EXHAUSTIVE

1 enables exhaustive MLIR tuning. 0 returns to default behavior.

MIGRAPHX_MLIR_USE_SPECIFIC_OPS

Comma-separated list of operations (for example, attention, convolution, dot, fused_dot, fused_convolution, fused). Prefix an entry with ~ to negate it.

MIGRAPHX_DISABLE_MLIR

1 disables rocMLIR. 0 returns to default behavior.

MIGRAPHX_ENABLE_CK

1 enables Composable Kernel. Use with MIGRAPHX_DISABLE_MLIR=1.

MIGRAPHX_USE_FAST_SOFTMAX

1 enables fast softmax optimization. 0 returns to default behavior.

MIGRAPHX_ENABLE_LAYERNORM_FUSION

1 enables layernorm fusion. 0 returns to default behavior.

MIGRAPHX_FLASH_DECODING_ENABLED

1 enables flash decoding for attention fusion. Default: 0.

MIGRAPHX_SKIP_BENCHMARKING

1 skips kernel benchmarking and compiles with the first available solution. 0 returns to default behavior.

Tuning examples#

Force operations through rocMLIR#

Set a comma-separated list of operations that should lower through MLIR:

export MIGRAPHX_MLIR_USE_SPECIFIC_OPS="attention,convolution,dot,fused"
migraphx-driver perf model.onnx --onnx --gpu

Replace model.onnx with your model path.

Select a GEMM provider#

Set the general matrix multiply (GEMM) provider before compilation:

export MIGRAPHX_SET_GEMM_PROVIDER=hipblaslt
migraphx-driver perf model.onnx --onnx --gpu

Replace model.onnx with your model path.

Enable exhaustive GEMM tuning#

Search for the fastest GEMM kernel configuration:

export MIGRAPHX_ENABLE_GEMM_TUNING=1
migraphx-driver perf model.onnx --onnx --gpu --exhaustive-tune

You can also pass --exhaustive-tune on the command line. The environment variable enables exhaustive GEMM tuning even when that flag is omitted.

Measure the effect#

Compare performance before and after changing a variable:

migraphx-driver perf model.onnx --onnx --gpu -n 50

Replace model.onnx with your model path and adjust -n for the number of timing iterations.

See also#

Contents