[LLM] Fix P-Tuning difference (#2240)

* Support multicards

* fix ptuning diff

* Update engine.py

llm/fastdeploy_llm/engine.py

@@ -257,6 +257,7 @@ def _init_predictor(self):
             dist_config.enable_dist_model(True)
             parent_dir = os.path.abspath(os.path.join(self.model_dir, ".."))
             mapping_file = os.path.join(parent_dir, "rank_mapping.csv")
+
             if self.mp_degree > 1 and not os.path.exists(mapping_file):
                 raise Exception("There's no file {}.".format(mapping_file))
             dist_config.set_comm_init_config(mapping_file)
@@ -383,6 +384,8 @@ def dy_input_preprocess(inputs):
 
     for i in range(bsz):
         if stop_flags[i] == 1:
+            attention_mask[i] = 0
+            tgt_generation_mask[i] = 0
             length = int(enc_length[i, 0])
             if args.is_ptuning:
                 model_id = inputs['model_id'][i]