[Feature][MTP]support new speculative decoding method named hybrid mtp with ngram (#3610)

fastdeploy/model_executor/models/ernie4_5_mtp.py

@@ -248,6 +248,7 @@ class Ernie4_5_MTPModel(nn.Layer):
 
         self.num_layers = fd_config.model_config.num_hidden_layers
         self.embed_tokens = fd_config.speculative_config.sharing_model.ernie.embed_tokens
+        self.norm = fd_config.speculative_config.sharing_model.ernie.norm
 
         self.layers = nn.LayerList(
             [
@@ -318,6 +319,8 @@ class Ernie4_5_MTPModel(nn.Layer):
 
         hidden_states = hidden_states + residual
 
+        hidden_states = self.norm(hidden_states)
+
         return hidden_states