Type casting error on compile

[fonnesbeck]

Despite having set floatX = pytensor.config.floatX = "float32", I'm getting type conversion errors when compiling some models. I have also checked to ensure that all data are float32 or int32

File ~/mambaforge/envs/pymc/lib/python3.10/site-packages/pytensor/tensor/type.py:272, in TensorType.filter_variable(self, other, allow_convert)
    269     if other2 is not None:
    270         return other2
--> 272 raise TypeError(
    273     f"Cannot convert Type {other.type} "
    274     f"(of Variable {other}) into Type {self}. "
    275     f"You can try to manually convert {other} into a {self}."
    276 )

TypeError: Cannot convert Type TensorType(float64, ()) (of Variable Reshape{0}.0) into Type TensorType(float32, ()). You can try to manually convert Reshape{0}.0 into a TensorType(float32, ()).
``

Unfortunately, the traceback does not provide sufficient feedback to isolate which line of model code is responsible for the error.

[aseyboldt]

I didn't really test float32 things at all yet. Sounds like what's happening is that nutpie itself currently only works with float64, and passes those into the logp function. I think numba is smart enough to handle that (although I woudn't be surprised if there are corner cases where things won't compile). But the aesara function that computes deterministics now receives float64 values, even though it expects float32.

[aseyboldt]

Closing in favor of #7