Update generate.py

generate.py

@@ -172,7 +172,7 @@ def generate(
 
     generated_token_ids = custom_generate(
         self,
-        input_ids,
+        input_ids=input_ids,  # Pass input_ids explicitly
         attention_mask=attention_mask,
         max_new_tokens=max_new_tokens,
         min_length=min_length,
@@ -207,4 +207,4 @@ def generate(
         **model_kwargs,
     )
 
-    return generated_token_ids
+    return generated_token_ids