Skip to content

Development

Working from a checkout

uv venv
uv pip install mypy numpy pytest ty==0.0.97

PYTHONPATH=. uv run mypy demo.py      # expect 2 errors: rank, size
uvx ty@0.0.97 check numty/__init__.py demo.py  # the same 2 errors, none in numty/
uv run pytest                         # full suite: runtime, mypy- and ty-driven

The ty run ends in Found 2 diagnostics, both invalid-argument-type on demo.py:30 and demo.py:34; uvx pyright@1.1.408 --pythonpath .venv/bin/python numty/__init__.py demo.py reports the same two (without --pythonpath it cannot find numpy). ty 0.0.97 is pinned in the dev group, and tests/test_ty.py holds the ty line to exactly those two.

Repository layout

  • numty — rank aliases, generic shape bases, FlexArray rank-polymorphic base, narrow witness helper, shape-preserving ident, loose matmul fallback, familiar-style heads (F64 and friends) with ready-made dims (batch, features, …) plus the PEP 561 py.typed marker.
  • numty_plugin — the solver.
  • demo.py — good and bad usages, clean under both checkers except the two intentional errors.
  • bench/ — realistic benchmark (MLP training, math library, 17 of 18 seeded mistakes caught), with its own scorecard in bench/README.md.
  • tests/test_runtime.py — shapes, values, narrowing accept/reject (also under python -O), isinstance against heads.
  • tests/test_static.py — runs real mypy over fixtures: each bad one must fail with our message, and reveal_type pins must show precise shapes on the good paths. The f64_ok + bad_f64_* fixtures do the same for the familiar-style heads (clean end-to-end MLP included).
  • tests/test_differential.py + tests/differential.py — the differential oracle: generated numpy expressions, numty's type against numpy's result (see Checked against numpy).
  • tests/test_dim_algebra.py — properties of the symbolic dim algebra, generated with Hypothesis and checked at every assignment of small sizes, 0 and 1 included: form arithmetic matches integer arithmetic, a size has one spelling however it was built, "provably unequal" is never claimed for forms some sizes make equal, forms survive the cache round trip, and a printed formula evaluates to its form. String dims are checked through mypy against Python's own arithmetic.
  • tests/test_ty.py — runs the pinned ty over numty/__init__.py demo.py (exactly the 2 demo errors) and over flex_portable.py (FlexArray keeps its rank checks without the plugin).
  • numty_convert + tests/test_convert.py — the jaxtyping bridge: one-time annotation rewrite with goldens for every dim kind and an end-to-end MLP (clean + seeded-bug twins in fixtures).