pm_fast

pm_fast

Rust hot-path kernels for the gnn process-mining pipeline. Built with PyO3 + maturin.

Two kernels are exposed today:

function	replaces
build_task_adjacency	modules.process_mining.build_task_adjacency (per-case loop)
build_padded_prefixes	_build_prefixes + make_padded_dataset (fused, one pass)

Both keep the original Python contracts; the parent package transparently prefers the Rust kernels and falls back to pure-Python when the extension isn't installed (so you can still pip install -r requirements.txt and work without a Rust toolchain).

Build & install (from the repo root)

pip install maturin
cd pm_fast
maturin develop --release       # builds + installs into the active env
cd .. && python bench/bench_hotpaths.py

maturin develop is the equivalent of pip install -e . for a Rust extension — fastest dev loop. For wheels (CI / publishing) use maturin build --release.

Why these two

A cProfile run of gnn run on BPI2020 puts the per-case Python loops at >40% of pre-training wall time. The two functions above are the densest of those loops; both translate to tight integer arithmetic in Rust with zero Python-level allocations. Numbers in the parent README's "Performance" section.