* Fix GPU resource validation ([#28743](https://github.com/kubernetes/kubernetes/pull/28743), [@therc](https://github.com/therc))
* Scale kube-proxy conntrack limits by cores (new default behavior) ([#28876](https://github.com/kubernetes/kubernetes/pull/28876), [@thockin](https://github.com/thockin))

- Fixed a suboptimal scheduler preemption behavior where potential preemption victims were violating Pod Disruption Budgets. ([#128434](https://github.com/kubernetes/kubernetes/pull/128434), [@NoicFank](https://github.com/NoicFank)) [SIG Scheduling]

## Dependencies

- Fixed #118052: nodeAffinity on pods can change prior to scheduling gates being removed even when podSpec.affinity is nil in the initial spec, this matches the 1.28 behavior to allow consistent integrators to be written. ([#118199](https://github.com/kubernetes/kubernetes/pull/118199), [@aleskandro](https://github.com/aleskandro)) [SIG Apps, Scheduling and Testing]

    4.2.2.
 *  Fix: Drop partial support for HTTP/1.0. Previously OkHttp would send
    `HTTP/1.0` on connections after seeing a response with `HTTP/1.0`. The fixed
    behavior is consistent with Firefox and Chrome.
 *  Fix: Allow a body in `OPTIONS` requests.
 *  Fix: Don't percent-encode non-ASCII characters in URL fragments.
 *  Fix: Handle null fragments.

   - In version 1.8, enable this behavior by setting the
     `--horizontal-pod-autoscaler-use-rest-clients` flag to `true`.

   - In version 1.9, this behavior will be enabled by default, and must be explicitly
     disabled by setting the `--horizontal-pod-autoscaler-use-rest-clients` flag to `false`.

### Cluster Lifecycle

- The `auto-detect` behavior of the kubelet's `--cloud-provider` flag is deprecated.

     */
    // NB: yes, this is surprisingly high, but that's what the experiments said was necessary
    // Raising this number slows the worst-case contains behavior, speeds up hashFloodingDetected,
    // and reduces the false-positive probability.
    static final int MAX_RUN_MULTIPLIER = 13;

    /**

   * process in any way, for example being sent over RPC, or saved to disk. For a general-purpose,
   * non-cryptographic hash function that will never change behavior, we suggest {@link
   * #murmur3_128}.
   *
   * <p>Repeated calls to this method on the same loaded {@code Hashing} class, using the same value

    visibility = ["//visibility:public"],
)

# By default, XLA GPU is compiled into tensorflow when building with
# --config=cuda even when `with_xla_support` is false. The config setting
# here allows us to override the behavior if needed.
config_setting(
    name = "no_xla_deps_in_cuda",
    define_values = {"no_xla_deps_in_cuda": "true"},
    visibility = ["//visibility:public"],
)

build:windows_x86_cpu_2022 --platforms="//tensorflow/tools/toolchains/win2022:windows_ltsc2022_clang"

# Options to build TensorFlow 1.x or 2.x.
# TODO(kanglan): Change v2's define to default behavior
build:v2 --define=tf_api_version=2 --action_env=TF2_BEHAVIOR=1

# Enable all targets in XLA
build:cpu_cross --define=with_cross_compiler_support=true

# Disable XLA on mobile.

 *  Fix: Don't crash with a `NoSuchMethodError` when using OkHttp with the Sentry SDK.
 *  Fix: Retain the query data in the old `okhttp3.mockwebserver.RecordedRequest.path` property. We
    inadvertently changed this behavior when we introduced the `mockwebserver3` API.


## Version 5.0.0-alpha.17

_2025-06-29_

This release stabilizes many APIs for the imminent OkHttp 5.0.0 release.