[LLM] Fix Qwen2 (#8584)

* fix output_router_logits

* fix with __future__

paddlenlp/transformers/qwen2/modeling.py

@@ -18,6 +18,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Paddle Qwen2 model."""
+from __future__ import annotations
 
 import math
 import warnings
@@ -187,11 +188,11 @@ def scaled_dot_product_attention(
     else:
         #  [ bz, seqlen, nhead, head_dim] -> [bs, nhead, seq_len, head_dim]
         query_states = paddle.transpose(query_states, [0, 2, 1, 3])
-        # merge with the next tranpose
+        # merge with the next transpose
         key_states = paddle.transpose(key_states, [0, 2, 1, 3])
         value_states = paddle.transpose(value_states, [0, 2, 1, 3])
 
-        # matmul and devide by sqrt(head_dim)
+        # matmul and divide by sqrt(head_dim)
         attn_weights = paddle.matmul(query_states / math.sqrt(head_dim), key_states.transpose([0, 1, 3, 2]))
 
         if attn_weights.shape != [bsz, num_heads, q_len, kv_seq_len]:
@@ -1127,7 +1128,7 @@ def forward(self, prediction_scores, masked_lm_labels):
         if self.enable_parallel_cross_entropy:
             if prediction_scores.shape[-1] == self.config.vocab_size:
                 warnings.warn(
-                    f"enable_parallel_cross_entropy, the vocab_size should be splited: {prediction_scores.shape[-1]}, {self.config.vocab_size}"
+                    f"enable_parallel_cross_entropy, the vocab_size should be splitted: {prediction_scores.shape[-1]}, {self.config.vocab_size}"
                 )
                 self.loss_func = paddle.nn.CrossEntropyLoss(reduction="none", ignore_index=self.ignore_index)
 
@@ -1202,14 +1203,7 @@ def get_decoder(self):
         return self.qwen2
 
     def prepare_inputs_for_generation(
-        self,
-        input_ids,
-        use_cache=False,
-        past_key_values=None,
-        attention_mask=None,
-        inputs_embeds=None,
-        output_router_logits=False,
-        **kwargs
+        self, input_ids, use_cache=False, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
     ):
         batch_size, seq_length = input_ids.shape
         position_ids = kwargs.get("position_ids", paddle.arange(seq_length).expand((batch_size, seq_length)))
@@ -1230,7 +1224,6 @@ def prepare_inputs_for_generation(
                 "past_key_values": past_key_values,
                 "use_cache": use_cache,
                 "attention_mask": attention_mask,
-                "output_router_logits": output_router_logits,
             }
         )
         return model_inputs
@@ -1325,7 +1318,7 @@ def forward(
         hidden_states = outputs[0]
 
         # if labels is None，means we need full output, instead of tensor_parallel_output
-        # tensor_parallel_output is togather with ParallelCrossEntropy
+        # tensor_parallel_output is together with ParallelCrossEntropy
         tensor_parallel_output = (
             self.config.tensor_parallel_output and labels is not None and self.config.tensor_parallel_degree > 1
         )