Skip to content

README.md

Latest commit

 

History

History
66 lines (47 loc) · 1.61 KB

README.md

File metadata and controls

66 lines (47 loc) · 1.61 KB

pickett Python package

Python bindings for SPFIT/SPCAT via nanobind.

Prerequisites

  • CMake ≥ 3.15
  • A C++17 compiler
  • Python ≥ 3.8 with pip / uv

Build and install

# From the project root, using uv
uv venv && source .venv/bin/activate
uv pip install ".[test]" ./python

Or with plain pip:

pip install "./python[test]"

The build compiles the bundled dblas.c with -ffp-contract=off to preserve bit-identical numerical results with the v2008 reference baseline. Do not override these flags or link an external BLAS.

Usage

import pickett

# Fit spectroscopic parameters
fit = pickett.fit_files("path/to/molecule")  # reads .par + .lin
print(f"RMS = {fit.xsqbest:.6f}, B = {fit.par[0]:.4f} MHz")

# Generate catalog (no disk output)
cat = pickett.cat_files("path/to/molecule")  # reads .var + .int
for line in cat.cat_lines[:3]:
    print(line)
print(f"Q(300 K) ≈ {dict(zip(cat.temp, cat.qsum)).get(300.0, '?')}")

For the low-level path (access to raw session objects and inputs):

session = pickett.FitSession("mol.par", "mol.lin", engine="spinv")
out = session.run()  # may only be called once

session = pickett.CatSession("mol.int", "mol.var", engine="dpi")
out = session.run()

Running tests

pytest python/tests

Numerical reproducibility

-ffp-contract=off disables FMA and dblas.c replaces OpenBLAS. Together they prevent sign-flipped Cholesky columns and ensure the Python API returns results byte-identical to the CLI on the same machine. See CLAUDE.md for the full explanation.