Fix: Keep use_cudnn in nested RNNs in Bidirectional layer (#21534)

- the use_cudnn attribute of RNNs such as GRU/LSTM is not serialized
  and Bidirectional lost it when creating nested copies for forward and
  backward RNN layers
- let's pass it explicitly to the nested layers as it should not be
  serialized
- do nothing for other RNN layers where it's not defined

Author: bzamecnik

keras/src/layers/rnn/bidirectional.py

@@ -125,6 +125,10 @@ def __init__(
             )
         else:
             self.backward_layer = backward_layer
+        # Keep the use_cudnn attribute if defined (not serialized).
+        if hasattr(layer, "use_cudnn"):
+            self.forward_layer.use_cudnn = layer.use_cudnn
+            self.backward_layer.use_cudnn = layer.use_cudnn
         self._verify_layer_config()
 
         def force_zero_output_for_mask(layer):

keras/src/layers/rnn/bidirectional_test.py

@@ -260,3 +260,18 @@ def test_output_shape(self):
         output_shape = layer.compute_output_shape(x.shape)
         for out, shape in zip(output, output_shape):
             self.assertEqual(out.shape, shape)
+
+    def test_keeps_use_cudnn(self):
+        # keep use_cudnn if the layer has it
+        for rnn_class in [layers.GRU, layers.LSTM]:
+            for use_cudnn in [True, False, "auto"]:
+                rnn = rnn_class(1, use_cudnn=use_cudnn)
+                bidi = layers.Bidirectional(rnn)
+                self.assertEqual(bidi.forward_layer.use_cudnn, use_cudnn)
+                self.assertEqual(bidi.backward_layer.use_cudnn, use_cudnn)
+
+        # otherwise ignore it
+        rnn = layers.SimpleRNN(1)
+        bidi = layers.Bidirectional(rnn)
+        self.assertFalse(hasattr(bidi.forward_layer, "use_cudnn"))
+        self.assertFalse(hasattr(bidi.backward_layer, "use_cudnn"))