romeokienzler/gridfm-pf-case57-tiny

GridFM PowerFlow Reconstruction — case57 (tiny)

A heterogeneous graph neural network (GNS_heterogeneous) trained for PowerFlow reconstruction on the IEEE 57-bus case (case57_ieee). Given a power-grid case, it returns per-node latent embeddings and predictions for bus voltage magnitude/angle (Vm, Va) and generator active power (Pg).

This is a tiny variant (hidden_size = 12, 12 layers, seed 0) intended for lightweight experimentation and serving demos.

Serving with vLLM

This directory is a vLLM-loadable model (architectures: ["GridFMGNS"]). Install gridfm-graphkit with its vllm extra, then serve on the /pooling endpoint:

pip install "gridfm-graphkit[vllm]"

vllm serve <this-repo-or-dir> \
  --runner pooling \
  --trust-remote-code \
  --skip-tokenizer-init \
  --enforce-eager \
  --io-processor-plugin gridfm_pf_reconstruction \
  --enable-mm-embeds

Model details

• Architecture: GNS_heterogeneous (heterogeneous message-passing GNN)
• Task: PowerFlow reconstruction
• Network: case57_ieee
• hidden_size: 12, num_layers: 12, attention_head: 8
• Normalizer: HeteroDataMVANormalizer (baseMVA_orig 100.0, baseMVA 111.83, vn_kv_max 1.0)
• License: Apache-2.0

Files

• config.json — vLLM/HF config carrying the full GridFM config and normalizer stats under pretrained_cfg.
• model.safetensors — trained weights (416 tensors).