[Multi-worer training with Keras](https://www.tensorflow.org/tutorials/distribute/multi_worker_with_keras).
*   Adds `tf.distribute.Strategy.gather` and
    `tf.distribute.ReplicaContext.all_gather` APIs to support gathering dense
    distributed values.
*   Fixes various issues with saving a distributed model.

### `tf.keras`:

*   Improvements from the Functional API refactoring:

## Changelog since v1.15.5

### Other notable changes

* Fix kubelet metrics gathering on non-English Windows hosts ([#84156](https://github.com/kubernetes/kubernetes/pull/84156), [@wawa0210](https://github.com/wawa0210))

- Refactored `test/e2e/framework` so that the core framework is smaller. Optional functionality like resource monitoring, log size monitoring, metrics gathering and debug information dumping must be imported by specific e2e test suites. Init packages are provided which can be imported to re-enable the functionality that traditionally was in the core framework. If you have code that no longer compiles because...

## Changelog since v1.14.8

### Other notable changes

* Fix kubelet metrics gathering on non-English Windows hosts ([#84156](https://github.com/kubernetes/kubernetes/pull/84156), [@wawa0210](https://github.com/wawa0210))

* Update kubernetes-e2e charm to use snaps ([#45044](https://github.com/kubernetes/kubernetes/pull/45044), [@Cynerva](https://github.com/Cynerva))
* Log the error (if any) in e2e metrics gathering step ([#45039](https://github.com/kubernetes/kubernetes/pull/45039), [@shyamjvs](https://github.com/shyamjvs))