weight matching complete

Author: heyyanshuman

Diff for: keras_hub/src/models/qwen_moe/qwen_moe_decoder.py

@@ -63,12 +63,6 @@ def build(self, decoder_sequence_shape):
             )
         )
 
-        self._feedforward_layernorm = QwenLayerNorm(
-            epsilon=self.layer_norm_epsilon,
-            dtype=self.dtype_policy,
-            name="feedforward_layernorm",
-        )
-        self._feedforward_layernorm.build(decoder_sequence_shape)
         self.activation = keras.activations.get(self.activation_fn)
         self.built = True
 
@@ -201,7 +195,7 @@ def call(self, hidden_states):
                 final_hidden_states, top_x[:, None], updated_values
             )
 
-        shared_expert_output = self.shared_expert(hidden_states)
+        shared_expert_output = self.shared_expert_dense(hidden_states)
         shared_expert_output = (
             ops.sigmoid(self.shared_expert_gate_dense(hidden_states))
             * shared_expert_output

Diff for: tools/checkpoint_conversion/convert_qwen_moe_checkpoints.py

@@ -15,7 +15,8 @@
 torch.set_default_device(device)
 
 from keras import ops  # noqa: E402
-from transformers import AutoModelForCausalLM, AutoTokenizer  # noqa: E402
+from transformers import AutoModelForCausalLM  # noqa: E402
+from transformers import AutoTokenizer  # noqa: E402
 
 import keras_hub  # noqa: E402