GAIA-ML

GAIA-ML is an experimental compiler that turns ONNX models into optimized C++ kernels and graphs for AMD AI Engine v1, targeting the Versal Premium VP2802-2M.

Features

• Dense/MLP networks, supported CNN patterns and bounded tree ensembles.
• Float32, int32, int16 and int8 neural-network kernels.
• Custom model-local ONNX functions using supported float32 arithmetic and tensor operations.
• Operation fusion, weight packing and implementation selection across AIE tiles.
• Scheduling for first-output latency, steady-state output interval or event bursts.
• New opt-in surrogate selection for supported FP32 CNN row kernels and tensor reductions, plus vectorized reductions of up to 4096 elements.

GAIA generates an ADF graph, C++ kernels, weights, test vectors and build files. Generating a project does not require Vitis. Building and simulating the emitted AIE code requires the AMD toolchain separately.

Install

From a checkout, with Python 3.9 or newer:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
gaia_ml version

Compile a model

Inspect the model and target constraints, then generate a project:

gaia_ml inspect model.onnx --stage constraints

gaia_ml compile model.onnx \
  --output gaia_project \
  --schedule-objective latency \
  --data-file input.csv \
  --simulation-events 10

Each CSV row is one flattened input event; provide at least as many rows as --simulation-events. For dynamic inputs, supply a static --input-shape, for example --input-shape 1,3,18,18. Use a new output directory or --force to regenerate an existing project.

Useful compilation options:

Surrogate selection requires a trained bundle for the supported compute family. Use gaia_ml compile --help for all options, including tensor microkernel selection. The generated manifest.json records resource use, the selected schedule and estimated latency.

Build and validate

After loading your AMD Vitis environment:

cd gaia_project
make aie
make aiesim
cd ..
gaia_ml verify gaia_project

Latency estimates guide optimization; validate numerical accuracy and measured timing for your model with the external AMD toolchain.

Developed by the NGT WP2.1 Group. Core developer: Ioannis Xiotidis. Distributed under the MIT license.