Update generate.py

generate.py

@@ -123,7 +123,8 @@ def generate(
     synced_gpus=None,
     **model_kwargs,
 ):
-    generation_kwargs = dict(
+    return custom_generate(
+        self,
         input_ids=input_ids,
         attention_mask=attention_mask,
         max_length=max_length,
@@ -147,7 +148,6 @@ def generate(
         num_beam_groups=num_beam_groups,
         diversity_penalty=diversity_penalty,
         prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
-        stopping_criteria=stopping_criteria,
         output_attentions=output_attentions,
         output_hidden_states=output_hidden_states,
         output_scores=output_scores,
@@ -157,9 +157,4 @@ def generate(
         remove_invalid_values=remove_invalid_values,
         synced_gpus=synced_gpus,
         **model_kwargs,
-    )
-
-    generation_output = GenerationMixin.generate(self, **generation_kwargs)
-
-    # Return the generated token IDs as a list
-    return generation_output.tolist()
+    )