Environment:

• Keras version: 3.14.0
• Backend: JAX
• Python: 3.12

When using TextVectorization with a custom standardize callable and a non-TensorFlow backend (e.g. JAX), calling .adapt() raises an AttributeError because the callable receives a tf.EagerTensor instead of a plain Python string.

import os
os.environ["KERAS_BACKEND"] = "jax"

import re
import string
import keras

strip_chars = string.punctuation
def my_standardize(input_string):
    input_string = input_string.lower()          # AttributeError here
    return re.sub(f"[{re.escape(strip_chars)}]", "", input_string)

layer = keras.layers.TextVectorization(standardize=my_standardize)
layer.adapt(["Hello, world."])

AttributeError: 'tensorflow.python.framework.ops.EagerTensor' object has no attribute 'lower'

TextVectorization should be independent of TensorFlow.