// Update name_map with newly-created ops.
  // TODO(skyewm): right now BuildWhileLoop() may alter the graph if it returns
  // a bad status. Once we fix this, we may want to return early instead of
  // executing the following code.
  for (int i = first_new_node_id; i < parent->graph.num_node_ids(); ++i) {
    Node* new_node = parent->graph.FindNodeId(i);
    if (new_node == nullptr) continue;

- `ValidatingAdmissionPolicy` now preserves types of composition variables, and raises type-related errors early. ([#121001](https://github.com/kubernetes/kubernetes/pull/121001), [@jiahuif](https://github.com/jiahuif))

    *   Keras-based LSTM models must be converted with an explicit batch size in
        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.

- [`KT-24901`](https://youtrack.jetbrains.com/issue/KT-24901) No smart cast for `when` with early return
- [`KT-53086`](https://youtrack.jetbrains.com/issue/KT-53086) "Cannot access '<init>' before superclass constructor has been called" with inner class secondary constructor

The focus of this release is **compatibility** with Pydantic v1 and v2, to make sure your current apps keep working. Later there will be more focus on refactors, correctness, code improvements, and then **performance** improvements. Some third-party early beta testers that ran benchmarks on the beta releases of FastAPI reported improvements of **2x - 3x**. Which is not bad for just doing `pip install --upgrade fastapi pydantic`. This was not an official benchmark and I didn't check it myself, but it's...