MIGraphX environment variables

MIGraphX environment variables#

2025-09-19

13 min read time

Applies to Linux

The MIGraphX environment variables can be used by contributors to the MIGraphX code base to customize tuning, verification, and tracing.

Model performance tunable variables#

Model performance tunable variables change the compilation behavior of a model. These are the most commonly used variables.

Environment variable

Values
MIGRAPHX_ENABLE_NHWC
Forces the model to use the NHWC layout.
1: Forces the use of the NHWC layout.
0: Returns to default behavior.
Default: The use of the NHWC layout isn’t forced.
MIGRAPHX_DISABLE_MLIR
When set, the rocMLIR library won’t be used.
1: The rocMLIR library won’t be used.
0: Returns to default behavior.
Default: The rocMLIR library is used.
MIGRAPHX_ENABLE_CK
When set, the Composable Kernel library is used.
Use with MIGRAPHX_DISABLE_MLIR = 1.
1: The Composable Kernel library is used.
0: Returns to default behavior.
Default: Composable Kernel library isn’t used.
MIGRAPHX_SET_GEMM_PROVIDER
Sets the GEMM provider to be either rocBLAS or hipBLASlt.
hipblaslt: hipBLASLt is used as the GEMM provider.
rocblas: rocBLAS is used as the GEMM provider.
Default: rocblas on gfx90a; hipblaslt on all other architectures.
MIGRAPHX_DISABLE_LAYERNORM_FUSION
When set, layernorm fusion isn’t used.
1: Layernorm fusion won’t be used.
0: Returns to default behavior.
Default: Layernorm fusion is used.
MIGRAPHX_DISABLE_MIOPEN_POOLING
When set, MIGraphX pooling is used instead of MIOpen pooling.
1: Use MIGraphX pooling.
0: Returns to default behavior.
Default: MIOpen pooling is used.
MIGRAPHX_USE_FAST_SOFTMAX
Turns on fast softmax optimization to speed up softmax computations.
1: Turns on Softmax optimization.
0: Returns to default behavior.
Default: Softmax optimization is turned off.
MIGRAPHX_DISABLE_FP32_SOFTMAX
Disables upcasting to fp32 when computing softmax for lower precision graphs.
1: Disables forcing full precision computation of softmax
0: Returns to default behavior.
Default: Upcasting to FP32 is turned on.
MIGRAPHX_MLIR_USE_SPECIFIC_OPS
Specifies the MLIR operations to use regardless of GPU architecture.
Takes a comma-separated list of operations. Operations can be any of the following:
attention: Use attention fusion. This is used by default on MI300, but must be specified on other architectures.
convolution: Use MLIR generated kernels for all convolutions. MIOpen is used by default otherwise.
convolution_backwards: Use MLIR generated kernels for backward-convolution. MIOpen is used by default otherwise.
dot: Use MLIR generated kernels for all GEMMs. hipBLASlt is used otherwise.
fused_convolution: Use MLIR generated kernels for convolutions when doing so enables extra fusions.
fused_dot: Use MLIR generated kernels for GEMMs when doing so enables extra fusions.
fused: Equivalent to setting both fused_dot and fused_convolution.
For example: MIGRAPHX_MLIR_USE_SPECIFIC_OP=fuse, convolution,dot.
A tilde (~) can be used to negate an operation.
For example, setting MIGRAPHX_MLIR_USE_SPECIFIC_OP=~convolution specifies that MLIR generated kernels should never be used.
MIGRAPHX_MLIR_TUNE_EXHAUSTIVE
When set, exhaustive tuning for MLIR is used to find the optimal configuration.
1: Exhaustive tuning is used.
0: Returns to default behavior.
Default: No MLIR tuning is used.
MIGRAPHX_ENABLE_HIP_GEMM_TUNING
When set, exhaustive tuning for hipBLASLt is used to find the optimal configuration.
1: Exhaustive tuning is used.
0: Returns to default behavior.
Default: Exhaustive hipBLASLt tuning isn’t used.
MIGRAPHX_ENABLE_MLIR_INPUT_FUSION
Turns on input fusions in MLIR.
1: Turns on input fusions.
0: Returns to default behavior.
Default: Input fusions are turned off.
MIGRAPHX_ENABLE_MLIR_REDUCE_FUSION
Turns on reduction fusions in MLIR.
1: Turns on reduction fusions.
0: Returns to default behavior.
Default: Reduction fusions are turned off.
MIGRAPHX_MLIR_ENABLE_SPLITK
Turns on Split-k performance configurations during MLIR tuning.
1: Turns on Split-k performance configurations.
0: Returns to default behavior.
Default: Split-k performance configurations are turned off.
MIGRAPHX_DISABLE_FP16_INSTANCENORM_CONVERT
When set, FP16 is not converted to FP32 in the InstanceNormalization ONNX operator.
1: FP16 is not converted to F32.
0: Returns to default behavior.
Default: FP16 is converted to F32.
MIGRAPHX_ENABLE_REWRITE_DOT
When set, the rewrite_dot pass is run.
1: Runs the rewrite_dot pass
0: Returns to default behavior.
Default: The rewrite_dot pass isn’t run.
MIGRAPHX_SPLIT_REDUCE_SIZE
Minimum size of a reduction to perform a split reduce.
The minimum size must be an integer.
Set to -1 to disable split reduce.
MIGRAPHX_COPY_LITERALS
When set, literals won’t be stored on the GPU but will only be copied over when needed.
1: Literals are copied over to the GPU as needed.
0: Returns to default behavior.
Default: Literals are stored on the GPU.
MIGRAPHX_VERIFY_ENABLE_ALLCLOSE
When set, the range tolerance is verified using allclose.
1: The range tolerance is verified using allclose.
0: Returns to the default behavior.
Default: Range tolerance isn’t verified.
MIGRAPHX_LOG_CK_GEMM
Turns on printing of Composable Kernel GEMM traces.
1: Composable Kernel GEMM traces will be printed.
0: Returns to default behavior.
Default: Composable Kernel GEMM traces aren’t printed.
MIGRAPHX_CK_DEBUG
When set, -DMIGRAPHX_CK_CHECK=1 is added to the Composable Kernel operator compilation options.
1: -DMIGRAPHX_CK_CHECK=1 is added to the compilation options.
Default: Compilation is run without -DMIGRAPHX_CK_CHECK=1.
MIGRAPHX_TUNE_CK
Turns on tuning for composable kernels.
1: Composable kernel tuning is done.
0: Returns to default behavior.
Default: No tuning is done for composable kernels.

Matching#

Debug settings for matchers. Matchers are responsible for finding optimizations in the graph compilation stage.

Environment variable

Values
MIGRAPHX_TRACE_MATCHES
When set, prints the name of matchers that have found a valid pattern match.
1: Prints the name of the matchers that have found a valid match.
2: When used with MIGRAPHX_TRACE_MATCHES_FOR, prints the names of matchers that have been tried but which have not necessarily found a match.
0: Returns to default behavior.
Default: Nothing is printed.
MIGRAPHX_TRACE_MATCHES_FOR
Turns on the printing of traces for the specified matcher if a string is found in the matcher’s file-name, function-name, or matcher-name.

Takes a string to match.
MIGRAPHX_VALIDATE_MATCHES
When set, module.validate() is used to validate the module after finding matches.
1: Runs module.validate().
0: Returns to default behavior.
Default: module.validate() isn’t run.
MIGRAPHX_TIME_MATCHERS
When set, prints the time spent on a matcher. This helps identify time-consuming patterns.
``1`: Prints the time spent on the matcher.
0: Returns to default behavior.
Default: The time is not printed.

Pass controls#

Debug settings for passes.

Environment variable

Values
MIGRAPHX_TRACE_ELIMINATE_CONTIGUOUS
Turns on the printing of debug statements for eliminate contiguous instruction passes.
1: Debug statements are printed for eliminate contiguous instructions passes.
0: Returns to default behavior.
Default: Debug statements aren’t printed for eliminate contiguous instructions passes.
MIGRAPHX_DISABLE_POINTWISE_FUSION
When set, the fuse_pointwise compile pass isn’t run.
1: The fuse_pointwise compile pass isn’t run.
0: Returns to default behavior.
Default: The fuse_pointwise compile pass is run.
MIGRAPHX_DEBUG_MEMORY_COLORING
Turns on the printing of debug statements for the memory-coloring pass.
1: Debug statements for the memory-coloring pass are printed.
0: Returns to default behavior.
Default: Debug statements for the memory-coloring pass aren’t printed.
MIGRAPHX_TRACE_SCHEDULE
Turns on the printing of debug statements for the schedule pass.
1: Debug statements for the schedule pass are printed.
0: Returns to default behavior.
Default: Debug statements for the memory-coloring pass aren’t printed.
MIGRAPHX_TRACE_PROPAGATE_CONSTANT
Turns on tracing of instructions that have been replaced with a constant.
1: Instructions that have been replaced with a constant are traced.
0: Returns to default behavior.
Default: Instructions that have been replaced with a constant aren’t traced.
MIGRAPHX_DISABLE_DNNL_POST_OPS_WORKAROUND
When set, the DNNL post-ops workaround isn’t used.
1: The DNNL post-ops workaround ins’t used.
0: Returns to default behavior.
Default: The DNNL post-ops workaround is used.
MIGRAPHX_DISABLE_MIOPEN_FUSION
When set, MIOpen fusions aren’t used.
1: MIOpen fusions aren’t used.
0: Returns to default behavior.
Default: MIOpen fusions are used.
MIGRAPHX_DISABLE_SCHEDULE_PASS
When set, the schedule pass isn’t run.
1: The schedule pass isn’t run.
0: Returns to default behavior.
Default: The schedule pass is run.
MIGRAPHX_DISABLE_REDUCE_FUSION
When set, the fuse_reduce pass isn’t run.
1: The fuse_reduce pass isn’t run.
0: Returns to default behavior.
Default: The fuse_reduce pass is run.
MIGRAPHX_TRACE_PASSES
Turns on printing of the compile passes and the program after the passes.
1: Prints the compile passes.
0: Returns to the default behavior.
Default: The compile pass traces aren’t printed.
MIGRAPHX_TIME_PASSES
When set, the compile passes are timed.
1: Compile passes are timed.
0: Returns to the default behavor.
Default: Compile passes aren’t timed.
MIGRAPHX_DISABLE_PASSES
Specifies passes that are to be skipped.
Takes a comma-separated list of passes.
For example:
MIGRAPHX_DISABLE_PASSES=rewrite_pooling,rewrite_gelu.

Compilation tracing#

Environment variable

Values
MIGRAPHX_TRACE_FINALIZE
Turns on printing of graph instructions during the module.finalize() step.
1: Graph instructions will be printed.
0: Returns to default behavior.
Default: Graph instructions won’t be printed.
MIGRAPHX_TRACE_COMPILE
Turns on graph compilation tracing.
1: Turns on graph compilation tracing.
0: Returns to default behavior.
Default: Graph compilation isn’t traced.
MIGRAPHX_TRACE_ONNX_PARSER
Turns on node-by-node tracing for the ONNX parser.
1: Node-by-node tracing is turned on.
0: Returns to the default behavior.
Default: There is no node-by-node tracing of the ONNX parser.
MIGRAPHX_TRACE_EVAL
Turns on model evaluation tracing and sets its tracing level.
1: Print the run instructions and the time taken to complete the evaluation.
2: Print the run instructions, time taken, a snippet of the output, and some statistics.
3: Print the run instructions, time taken, a snippet of the output, and statistics for all output buffers.
MIGRAPHX_TRACE_QUANTIZATION
Turns on the printing of the traces for passes run during quantization.
1: Traces for passes run during quantization will be printed.
0: Returns to default behavior.
Default: The traces for passes run during quantization won’t be printed out.
MIGRAPHX_8BITS_QUANTIZATION_PARAMS
Turns on the printing of the quantization parameters in the main module only.
1: Only the quantization parameters in the main module are printed.
0: Returns to default behavior.
Default:

MLIR#

Environment variable

Values
MIGRAPHX_TRACE_MLIR
Sets the MLIR trace level.
1: MLIR trace failures are printed.
2: MLIR trace failures are printed and all MLIR operations are printed as well.
MIGRAPHX_MLIR_TUNING_DB
The path of the tuning database.

Takes the path to the tuning database.
MIGRAPHX_MLIR_TUNING_CFG
Sets the path to the tuning configuration file to use with rocMLIR tuning scripts.
Takes the path to the configuration file.
For example:
MIGRAPHX_MLIR_TUNING_CFG="path/to/config_file.cfg"
MIGRAPHX_MLIR_TUNE_LIMIT
Sets the maximum number of solutions available for MLIR tuning.
Takes an integer greater than 1.
MIGRAPHX_MLIR_DUMP_TO_MXR
Sets the location to where the MXR files that the MLIR modules are written to are saved.
Takes the path to the directory where the files should be saved.
For example:
MIGRAPHX_MLIR_DUMP_TO_MXR="/path/to/save_mxr_file/
MIGRAPHX_MLIR_DUMP
Sets the the location where the MLIR files that the MLIR modules are written to are saved.
Takes the path to the directory where the files should be saved.
For example:
MIGRAPHX_MLIR_DUMP="/path/to/save_mlir_file/

Testing#

Environment variable

Values
MIGRAPHX_TRACE_TEST_COMPILE
Sets the target to be traced, and turns on printing of the compile trace for verify tests on the given target.
This flag cannot be used if MIGRAPHX_TRACE_COMPILE is used.
cpu: Turns on traces for the CPU target.
GPU: Turns on traces for the GPU target.
Default:
MIGRAPHX_TRACE_TEST
When set, the reference and target programs are printed even if the verify tests pass.
1: The reference and target programs are printed when the verify tests pass.
0: Returns to default behavior.
Default: Reference and target programs aren’t printed if the verify tests pass.
MIGRAPHX_DUMP_TEST
When set, the model that is being verified using test-verify is output to an MXR file.
1: The model that is being verified is output to an MXR file.
0: Returns to default behavior.
Default: The model isn’t output to file.
MIGRAPHX_VERIFY_DUMP_DIFF
When set, writes out the output of the test results, as well as the reference, when they differ.
1: Test results are written out when they differ.
0: Returns to default behavior.
Default: The results and the reference aren’t written out when they differ.

Advanced settings#

Environment variable

Values
MIGRAPHX_TRACE_CMD_EXECUTE
When set, commands run by the MIGraphX process will be printed.
1: Printing of commands is turned on.
0: Returns to default behavior.
Default: Commands aren’t printed.
MIGRAPHX_TRACE_HIPRTC
When set, the HIPRTC options and C++ file used will be printed.
1: HIPRTC options and C++ file will be printed.
0: Returns to default behavior.
Default: HIPRTC options and C++ file aren’t printed.
MIGRAPHX_DEBUG_SAVE_TEMP_DIR
When set, temporary directories won’t be deleted.
1: Temporary directories aren’t deleted.
0: Returns to default behavior.
Default: Temporary directories are deleted.
MIGRAPHX_GPU_DEBUG
When set, the -DMIGRAPHX_DEBUG option is used when compiling GPU kernels. -DMIGRAPHX_DEBUG enables assertions and source location capture.
1: The -DMIGRAPHX_DEBUG option is used when compiling GPU kernels.
Default: Compilation is run without -DMIGRAPHX_DEBUG.
MIGRAPHX_GPU_DEBUG_SYM
When set, the -g option is used when compiling HIPRTC for debugging purposes.
1: The -g option is used when compiling HIPRTC.
Default: Compilation is run without the -g option.
MIGRAPHX_GPU_DUMP_SRC
The compiled HIPRTC source files is written out for further analysis.
1: HIPRTC source files are written out.
0: Returns to default behavior.
Default: HIPRTC source files aren’t written out.
MIGRAPHX_GPU_DUMP_ASM
When set, the hip-clang assembly output is written out for further analysis.
1: The hip-clang assembly output is written out.
0: Returns to default behavior.
Default: The hip-clang assembly output isn’t written out.
MIGRAPHX_GPU_HIP_FLAGS
When set, the hip-clang compiler appends these extra flags for compilation.
Takes a valid string, a valid hip compile option, e.g. “-Wno-error”.
Default: The compiler will not append any extra flags for compilation.
MIGRAPHX_GPU_OPTIMIZE
Sets the GPU compiler optimization mode.
Takes a valid optimization mode such as O3.
Default: No compiler optimization is used.
MIGRAPHX_GPU_COMPILE_PARALLEL
Sets the number of threads to use for parallel GPU code compilation.
Takes a positive integer value.
Default: Compilation is not run in parallel.
MIGRAPHX_TRACE_NARY
When set, the nary device functions used during execution are printed out.
1: The nary device functions are printed out.
0: Returns to default behavior.
Default: nary device functions aren’t printed out.
MIGRAPHX_ENABLE_HIPRTC_WORKAROUNDS
When set, the workarounds for known bugs in HIPRTC are used.
1: HIPRTC workarounds are used.
0: Returns to default behavior.
Default: HIPRTC workarounds aren’t used.
MIGRAPHX_ENABLE_NULL_STREAM
Whem set, a null stream can be used for MIOpen and HIP stream handling.
1: A null stream can be used for stream handling.
0: Returns to default behavior.
Default: A null stream can’t be used for stream handling.
MIGRAPHX_NSTREAMS
Sets the number of HIP streams to use in the GPU.
Takes a positive integer.
Default: one stream will be used.
MIGRAPHX_TRACE_BENCHMARKING
Sets the verbosity of benchmarking traces.
1: Basic trace
2: Detailed trace
3: Compiled traces
MIGRAPHX_PROBLEM_CACHE
Sets the JSON file that the problem cache will be saved to and loaded from.
Takes a fully qualified path to a valid JSON file.
For example:
MIGRAPHX_PROBLEM_CACHE="path/to/cache_file.json"
MIGRAPHX_BENCHMARKING_BUNDLE
Sets the number of configurations to run in a bundle during benchmarking.

Takes a positive integer.
MIGRAPHX_BENCHMARKING_NRUNS
Sets the number of timing runs for each configuration bundle being benchmarked.

Takes a positive integer.