[BUG] `BARTModel.sample` fails on float32 inputs

[Gattocrucco]

From the Python interface, BARTModel fails when the data is float32. Easy to sidestep by casting, weird though. Pops up for me because I want to replace BART3 with stochtree as reference implementation in my unit tests of bartz, and all my arrays are float32 by default because jax.

import numpy as np
import stochtree

rng = np.random.default_rng(0)
X = rng.normal(size=(100, 3)).astype(np.float32)
y = rng.normal(size=100).astype(np.float32)

m = stochtree.BARTModel()
m.sample(X_train=X, y_train=y)

Traceback (most recent call last):
  File "<python-input-0>", line 9, in <module>
    m.sample(X_train=X, y_train=y)
    ~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/giacomo/Library/Caches/uv/archive-v0/ewzn4Tz9IlUrLkmx/lib/python3.14/site-packages/stochtree/bart.py", line 1421, in sample
    global_model_config = GlobalModelConfig(global_error_variance=current_sigma2)
  File "/Users/giacomo/Library/Caches/uv/archive-v0/ewzn4Tz9IlUrLkmx/lib/python3.14/site-packages/stochtree/config.py", line 677, in __init__
    raise ValueError("`global_error_variance` must be a positive scalar")
ValueError: `global_error_variance` must be a positive scalar

Versions: stochtree 0.4.2, numpy 2.4.6, python 3.14, macOS 26.4.1

[andrewherren]

Hi Giacomo! Thanks for raising this!

I suspect that, because we use double under the hood in C++ (something we will likely change in the future), there might be some parts in the codebase where we aren't being careful about the size of the underlying floating point type when copying from Python / R to C++. I'll investigate and get back to you.

We've got a 0.4.3 release going out this week, so if I can debug this quickly, you should have an updated version on PyPI in a few days' time. If not, the fix will be installable from github ahead of the next release.

[Gattocrucco]

Thanks

[andrewherren]

A fix is now available on main (i.e. pip install git+https://github.com/StochasticTree/stochtree.git). It will be on PyPI either later today or tomorrow as version 0.4.3.

Thanks for catching this!