OpenACM 16-bit PEA-GNN Surrogate

This repository packages the 16-bit OpenACM PEA-GNN surrogate model files for Hugging Face.

The model predicts five approximate multiplier metrics from graph-structured multiplier configurations:

• mred
• nmed
• delay
• area
• power

Provenance

• OpenACM source project: https://github.com/ShenShan123/OpenACM
• OpenACM source directory: openacm/multiplier_gnn/16-bitGNN
• Helper code, FEATURE.csv, and label_minmax_16.txt come from the OpenACM model files.
• Weight file: best_model_weights_16.pth
• Weight source: OpenACM GitHub 16-bit final model weight

All helper files and the 16-bit weight follow the OpenACM GitHub 16-bit model files.

Files

• best_model_weights_16.pth: final model parameters.
• gnn_predictor.py: GraphSAGE model definition and prediction wrapper.
• my_io.py: helper code for building PyTorch Geometric graph inputs from multiplier configurations.
• FEATURE.csv: static feature table used by the input builder.
• label_minmax_16.txt: min/max values used to map normalized model outputs to raw target units.
• config.json: compact metadata and hash information.
• requirements.txt: Python package dependencies.

Associated Dataset

https://huggingface.co/datasets/xuzhuo0417/openacm-gnn-16bit-dataset

Download Example

from huggingface_hub import hf_hub_download

weights_path = hf_hub_download(
    repo_id="xuzhuo0417/openacm-gnn-16bit",
    filename="best_model_weights_16.pth"
)
print(weights_path)

Notes

These are final model parameters only, without optimizer or epoch state.