Add a hacky fix for TF 2.13 and 2.14 weights.h5 loading

We have a bug where weights.h5 for a functional model will read and
write to the wrong paths in TF 2.13 and 2.14. We can work around this
for these versions (while thankfully needing none of this for Keras 3).

keras_nlp/utils/preset_utils.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import datetime
+import inspect
 import json
 import os
 
@@ -159,6 +160,21 @@ def save_to_preset(
             metadata_file.write(json.dumps(metadata, indent=4))
 
 
+def legacy_load_weights(layer, weights_path):
+    # Hacky fix for TensorFlow 2.13 and 2.14 when loading a `.weights.h5` file.
+    # We find the `Functional` class, and temporarily remove the
+    # `_layer_checkpoint_dependencies` property, which on older version of
+    # TensorFlow complete broke the variable paths for functional models.
+    functional_cls = None
+    for cls in inspect.getmro(layer.__class__):
+        if cls.__name__ == "Functional":
+            functional_cls = cls
+    property = functional_cls._layer_checkpoint_dependencies
+    functional_cls._layer_checkpoint_dependencies = None
+    layer.load_weights(weights_path)
+    functional_cls._layer_checkpoint_dependencies = property
+
+
 def load_from_preset(
     preset,
     load_weights=True,
@@ -186,7 +202,10 @@ def load_from_preset(
     load_weights = load_weights and config["weights"]
     if load_weights:
         weights_path = get_file(preset, config["weights"])
-        layer.load_weights(weights_path)
+        if hasattr(layer, "_layer_checkpoint_dependencies"):
+            legacy_load_weights(layer, weights_path)
+        else:
+            layer.load_weights(weights_path)
 
     return layer