SOLAR

PyTorch Model Analysis Toolkit — graph extraction, einsum conversion, and hardware-aware SOL performance analysis.

SOLAR = Speed of Light Analysis for Runtime

SOLAR is a toolkit for analyzing PyTorch model graphs, converting them to einsum representations, and performing SOL performance analysis.

It is used by SOL-ExecBench for deriving SOL performance metrics that serve as ground-truth references for evaluating LLM-generated GPU kernels.

Quick links: Quickstart · Documentation · License

Related: SOL-ExecBench GitHub · Dataset (HuggingFace) · Website

Quickstart

# From repo root (installs SOLAR + patched torchview)
bash install.sh  # or: bash install_uv.sh && source .venv/bin/activate

cd examples/Attention
bash run_solar.sh

Note: SOLAR depends on a patched torchview for parameter-tensor support. The install scripts apply patches/torchview-parameter-tensors.patch by default.

For PDF graph rendering, install Graphviz (dot).

Documentation

• docs/USAGE.md: End-to-end pipeline + CLI + Python API overview
• docs/SOL_GUIDE.md: SOL (Speed of Light Analysis for Runtime) methodology + metrics
• docs/EINSUM_GUIDE.md: Einsum conversion details and conventions

Contributing

See CONTRIBUTING.

License

Apache 2.0 License