pymc-labs/pathmc

Summary A small grab-bag of low-risk polish items for the falsification feature (good first issue). 1. FalsificationResult.plot(bins=...) non-integer validation pathmc/falsify.py:307 rejects bins < 1 but accepts a non-integer like 2.5, which then fails later inside matplotlib with a less clear error. The type hint already says int | None. if bins is not None and (not isinstance(bins, int) or bins < 1): raise ValueError(f"bins must be a positive integer or None, got {bins!r}.") 2. A few explicit edge-case tests worth locking in The suite (tests/test_falsify.py) is thorough, but these explicit cases are nice to pin: A 3-variable residual-covariance block (X ~~ Y, Y ~~ Z) is rejected (currently covered only transitively by the residual_blocks check). plot(bins=2.5) raises a clear ValueError (after fix #1). Acceptance criteria plot(bins=2.5) raises a clear pathmc ValueError. plot(bins=0), bins=-1 still raise (unchanged); bins=None, bins=10 still work. New tests added under tests/

Context Issue #4 introduced stochastic latent nodes and sparse measurement equations. These models use pytensor.scan for AR(1) dynamics and PyMC's masked array imputation for sparse observations. The default PyMC NUTS sampler works but is slow — the scan graph is expensive to compile and evaluate. Options to explore 1. nutpie JAX backend (quick win) The numba backend fails because masked array imputation creates unnamed shared variables (ValueError: Shared variables must have unique names: None). The JAX backend compiles the graph differently and may bypass this. JAX is also documented as giving substantial speedups for scan-heavy models. model.sample( ..., nuts_sampler="nutpie", nuts_sampler_kwargs={"backend": "jax"}, ) 2. NumPyro sampler nuts_sampler="numpyro" compiles the entire model to JAX via NumPyro's NUTS. Well-established fast path, especially with GPU. Needs testing with masked array imputation. 3. pymc-extras statespace with Kalman filter pymc_extras.statespace