amdshark Tuner

amdshark Tuner automates the dispatch-level and model-level tuning for the IREE (Intermediate Representation Execution Environment) ML Compiler on AMD GPUs.

Installation

Install amdshark Tuner from PyPI:

pip install amdsharktuner

This will install all required dependencies including IREE compiler and runtime.

You can use the latest nightly IREE python bindings:

pip install --upgrade --pre \
    iree-base-compiler \
    iree-base-runtime \
    --find-links https://iree.dev/pip-release-links.html

Verify that the environment is set up correctly:

python -m model_tuner --help

or

python -m dispatch_tuner --help

Model Tuner

Use the Model Tuner to tune a dispatch and a model:

python -m model_tuner double_mmt.mlir mmt_benchmark.mlir \
    --compile-flags-file=compile_flags.txt \
    --model-benchmark-flags-file=model_benchmark_flags.txt \
    --devices=hip://0 \
    --num-candidates=30 \
    --model-tuner-num-dispatch-candidates=5 \
    --model-tuner-num-model-candidates=3`

Refer to Mode Tuner README for detailed information on flags and MLIR files.

Dispatch Tuner

Use the Dispatch Tuner to tune a dispatch:

python -m dispatch_tuner dispatch_sample.mlir dispatch_sample_benchmark.mlir \
    --compile-flags-file=compile_flags.txt
    --devices=hip://0 \
    --num-candidates=30

Refer to Dispatch Tuner README for detailed information on flags and MLIR files.