Keras 3 vs tf.keras import breakage in CI

TensorFlow 2.16 switched its default tf.keras to Keras 3, which changed imports and behavior. Libraries built for Keras 2 (or that import from removed paths) fail at import time when CI installs the newer stack.

What this error means

An import fails with "Your currently installed version of Keras is Keras 3, but this is not yet supported" or "cannot import name ... from keras", after a TensorFlow upgrade pulled in Keras 3.

python

RuntimeError: Your currently installed version of Keras is Keras 3, but this is not yet
supported. Please install the backwards-compatible tf-keras package with
`pip install tf-keras`.

Common causes

TensorFlow 2.16+ defaults to Keras 3

Newer TensorFlow bundles Keras 3, which moves APIs and is not a drop-in replacement for the Keras 2 surface many libraries still expect.

A dependency requires the legacy tf.keras

Transformers, TF-Hub, or other libraries detect Keras 3 and demand the backwards-compatible tf-keras package.

How to fix it

Install the backwards-compatible tf-keras

Add tf-keras and point TensorFlow at it so the legacy API is available.

Terminal

pip install tf-keras
# in the environment for code that needs the legacy API:
export TF_USE_LEGACY_KERAS=1

Pin the stack to a compatible version set

Pin TensorFlow and Keras together to a combination your code supports rather than letting CI pull Keras 3.