* loader and prevent it from being garbage collected. This poses a problem for environments where
 * you want to throw away the class loader. For example, dynamically reloading a web application or
 * unloading an OSGi bundle.
 *
 * <p>{@code com.google.common.base.FinalizableReferenceQueue} loads this class in its own class

   * thread keeps an indirect strong reference to the queue in ReferenceMap, which keeps the
   * Finalizer running, and as a result, the application class loader can never be reclaimed.
   *
   * This means that dynamically loaded web applications and OSGi bundles can't be unloaded.
   *
   * If the library is loaded in an application class loader, we try to break the cycle by loading
   * Finalizer in its own independent class loader:
   *

bool_setting(
    name = "oss_setting",
    build_setting_default = if_oss(True, False),
    visibility = ["//visibility:private"],
)

# Setting to use when loading kernels dynamically
config_setting(
    name = "dynamic_loaded_kernels",
    define_values = {
        "dynamic_loaded_kernels": "true",
        "framework_shared_object": "true",
    },
    visibility = ["//visibility:public"],

    List<List<Integer>> input = newArrayList(list1, list2);

    Iterable<Integer> result = Iterables.concat(input);
    assertEquals(asList(1, 4), newArrayList(result));

    // Now change the inputs and see result dynamically change as well

    list1.add(2);
    List<Integer> list3 = newArrayList(3);
    input.add(1, list3);

    assertEquals(asList(1, 2, 3, 4), newArrayList(result));

        the input layer.

## Bug Fixes and Other Changes

### TF Core:

*   Set `tf2_behavior` to 1 to enable V2 for early loading cases.
*   Add `execute_fn_for_device function` to dynamically choose the
    implementation based on underlying device placement.
*   Eager:
    *   Add `reduce_logsumexp` benchmark with experiment compile.
    *   Give `EagerTensor`s a meaningful `__array__` implementation.

higher-dimensional objects.
With arrays of arrays, the inner arrays are, by construction, always the same length;
however with slices of slices (or arrays of slices), the inner lengths may vary dynamically.
Moreover, the inner slices must be initialized individually.
</p>

<h3 id="Struct_types">Struct types</h3>

<p>
A struct is a sequence of named elements, called fields, each of which has a

    * The same new ID generation applies to response models.
    * This also changes the generated title for those models.
    * Only composite bodies and response models are affected because those are generated dynamically, they don't have a module (a Python file).