[Bug]: RangeParameterConfig error when using some float stepsize

[AlbanMor]

What happened?

Hi,
First of all, Thank you for this amazing package!

Here is the issue. When trying to use a step size of 0.1 in the range of [8 to 9.8] in a RangeParameterConfig, when trying to configure the experiment, it throws the error "The range of the parameter must be evenly divisible by the step size." when trying to configure the experiment.

The issue is that in that case for some reason (upper - lower) % step_size != 0. It is equal to 6.106226635438361e-16...

Please provide a minimal, reproducible example of the unexpected behavior.

from ax.api.configs import RangeParameterConfig

test = RangeParameterConfig(name="test", parameter_type="float", bounds=(8, 9.8), step_size=0.1)

print((9.8 - 8) % 0.1)

client = Client()
client.configure_experiment(parameters=[test])

Please paste any relevant traceback/logs produced by the example provided.

Ax Version

1.2.1

Python Version

3.13.3

Operating System

MS

(Optional) Describe any potential fixes you've considered to the issue outlined above.

Might be worth changing the code for if (upper - lower) % step_size > 1e-14. But I don<t know what would be the implications further down the road in your package.

In the mean time I'll just be changing my units ;-).

Pull Request

None

Code of Conduct

• I agree to follow Ax's Code of Conduct

[mgrange1998]

Hi @AlbanMor, thank you for using Ax and for opening this issue!

This is a known bug (#4599) and a PR was landed to fix this issue #4604. I reproduced w/ your code on the latest build to ensure it would be fixed.

This change was not picked up in the latest Ax Version 1.2.1, but the next Ax release will include this change.
For the time being, thank you for your flexibility in changing to work around this issue!

[wizardlyrocket8]

Hello @mgrange1998 I am the author of the original bug report that you cited here. And actually I changed the code chunk that was provided, but it still does not work, and I am not sure, but it seems that the logic of the function is wrong, because by default if the step size is provided the function return ChoiceParameter instead of Range.

[VMLC-PV]

I don't know if this helps track things down, but it seems that this problem also raises issues when trying to attach_trial()
Even if the attach trials are compatible with the bounds and stepsize, it raises:
Parameterization {Some_dic} is in out-of-design. Ax will still attach the trial for use in candidate generation.

[VMLC-PV]

@wizardlyrocket8 or @AlbanMor
could you try to modify the ChoiceParameter class (ax/core/parameters)

change the def validate like so:

def validate(self, value: TParamValue, raises: bool = False) -> bool:
        """Checks that the input is in the list of allowed values.

        Args:
            value: Value being checked.
            raises: If true, and validation fails, raises a UserInputError.

        Raises:
            UserInputError: If validation fails and raises is True.

        Returns:
            True if valid, False otherwise.
        """
        is_valid = value in self._values
        # check if value is close to any of the allowed values for numeric parameters
        if not is_valid and self.parameter_type.is_numeric:
            for allowed_value in self._values:
                if isinstance(value, (int, float)) and isinstance(allowed_value, (int, float)):
                    if abs(float(value) - float(allowed_value)) < EPS:
                        is_valid = True
                        break
        if raises and not is_valid:
            raise UserInputError(
                f"Value {value} is not in the list of allowed values: {self._values}."
            )
        return is_valid

It worked to fix my issue when using the last version on git

[VMLC-PV]

My previous comment also seems to work in 1.2.4 that just got released