[matthiaskoenig]

Problem

compile_formula imported petab.v2.math for its parser sympify_petab. That import pulls in:

• the package petab, with the problems of PEtab v1, matplotlib.pyplot and scipy.stats;
• petab.v2, with its SciML extension, petab_sciml and torch.

Every process and every worker of a pool paid this on its first Formula observable or Data function: 0.7 s on an idle machine, 1.4 s under load. Measured with the sampling profiler of python 3.15 on a scan with observables.

Fix

The parser of petab, petab.v1.math, is self-contained: it uses relative imports and needs only sympy, numpy and the runtime of ANTLR.

• parse_math(formula) imports that directory as the package sbmlsim._petab_math. It is the same code as sympify_petab, so PEtab math keeps exactly petab's semantics.
• A process which imported petab already uses petab.v2.math.
• If a future petab moves its parser, parse_math falls back to importing petab.v2.math as before.

The first formula now takes 0.04 s instead of 0.67 s, and torch, petab, matplotlib.pyplot and scipy.stats are no longer imported.

Tests

• test_compiling_a_formula_does_not_import_petab: in a fresh python, none of these modules is imported. On develop all four are.
• test_the_parser_is_the_one_of_petab: ten formulas give the same expression as sympify_petab. They cover the precedence of ^ and unary minus, the booleans, piecewise, the logarithms, comparisons, the constants, min/max and numbers.
• pytest --skip-slow tests/simulator tests/simulation tests/data tests/uncertainty: 878 passed.
• tests/test_data*.py tests/test_imports.py tests/fit tests/sciml: 2313 passed.

After review

test_the_math_of_petab_v2_is_the_parser_of_petab_v1 fails if petab.v2.math stops re-exporting the parser of v1, which parse_math reads.