- Fixed a bug where the `scheduler_unschedulable_pods` metric could be artificially inflated (leak) when a pod fails `PreEnqueue` plugins after being previously marked unschedulable. ([#135981](https://github.com/kubernetes/kubernetes/pull/135981), [@vs...

- Improved Node's QueueingHint in the `NodeAffinity` plugin by ignoring unrelated changes that keep pods unschedulable. ([#127444](https://github.com/kubernetes/kubernetes/pull/127444), [@dom4ha](https://github.com/dom4ha)) [SIG Scheduling and Testing]

- DRA: the explanation for why a pod which wasn't using ResourceClaims was unscheduleable included a useless "no new claims to deallocate" when it was unscheduleable for some other reasons. ([#129823](https://github.com/kubernetes/kubernetes/pull/129823), [@googs1025](https://github.com/googs1025)) [SIG Node and Scheduling]

- Removed the deprecated `--register-schedulable` command line argument from the kubelet. ([#122384](https://github.com/kubernetes/kubernetes/pull/122384), [@carlory](https://github.com/carlory)) [SIG Cloud Provider, Node and Scalability]

    *   Converts hardswish subgraphs into atomic ops.
*   Add support for defaulting the value of `cycle_length` argument of
    `tf.data.Dataset.interleave` to the number of schedulable CPU cores.
*   `parallel_for`: Add converter for `MatrixDiag`.
*   Add `narrow_range` attribute to `QuantizeAndDequantizeV2` and V3.
*   Added new op: `tf.strings.unsorted_segment_join`.