diff --git a/test_tipc/prepare_npu_test.sh b/test_tipc/prepare_npu_test.sh
new file mode 100644
index 000000000..ed050987f
--- /dev/null
+++ b/test_tipc/prepare_npu_test.sh
@@ -0,0 +1,26 @@
+#!/bin/bash
+
+BASEDIR=$(dirname "$0")
+
+function readlinkf() {
+    perl -MCwd -e 'print Cwd::abs_path shift' "$1";
+}
+
+REPO_ROOT_PATH=$(readlinkf ${BASEDIR}/../)
+
+config_files=$(find ${REPO_ROOT_PATH}/test_tipc/configs -name "train_infer_python.txt")
+for file in ${config_files}; do
+   echo $file
+   sed -i "s/Global.use_gpu/Global.use_npu/g" $file
+   sed -i '16s/$/ -o use_npu=True/' $file
+   sed -i '24s/$/ -o use_npu=True/' $file
+   sed -i '40s/$/ --use_gpu=False/' $file
+   sed -i "s/--use_gpu:True|False/--use_npu:True|False/g" $file
+done
+
+yaml_files=$(find ${REPO_ROOT_PATH}/configs/recognition/ -name "*.yaml")
+for file in ${yaml_files}; do
+   echo $file
+   sed -i "s/num_workers: 4/num_workers: 0/g" $file
+   sed -i "s/num_workers: 8/num_workers: 0/g" $file
+done
diff --git a/test_tipc/test_train_inference_python.sh b/test_tipc/test_train_inference_python.sh
index ca7a97350..8b932a2ec 100644
--- a/test_tipc/test_train_inference_python.sh
+++ b/test_tipc/test_train_inference_python.sh
@@ -363,9 +363,9 @@ else
                 if [ ${#gpu} -le 2 ];then  # train with cpu or single gpu
                     cmd="${python} ${run_train} ${set_amp_config} ${set_use_gpu}  ${set_save_model} ${set_epoch} ${set_pretrain} ${set_batchsize} ${set_train_params1} ${set_train_params2} > ${LOG_PATH}/train.log 2>&1"
                 elif [ ${#ips} -le 15 ];then  # train with multi-gpu
-                    cmd="${python} -B -m paddle.distributed.launch --gpus=\"${gpu}\" ${run_train} ${set_amp_config} ${set_use_gpu} ${set_save_model} ${set_epoch} ${set_pretrain} ${set_batchsize} ${set_train_params1} ${set_train_params2} > ${LOG_PATH}/train.log 2>&1"
+                    cmd="${python} -B -m paddle.distributed.launch --devices=\"${gpu}\" ${run_train} ${set_amp_config} ${set_use_gpu} ${set_save_model} ${set_epoch} ${set_pretrain} ${set_batchsize} ${set_train_params1} ${set_train_params2} > ${LOG_PATH}/train.log 2>&1"
                 else     # train with multi-machine
-                    cmd="${python} -B -m paddle.distributed.launch --ips=${ips} --gpus=\"${gpu}\" ${run_train} ${set_amp_config} ${set_use_gpu} ${set_save_model} ${set_pretrain} ${set_epoch} ${set_batchsize} ${set_train_params1} ${set_train_params2} > ${LOG_PATH}/train.log 2>&1"
+                    cmd="${python} -B -m paddle.distributed.launch --ips=${ips} --devices=\"${gpu}\" ${run_train} ${set_amp_config} ${set_use_gpu} ${set_save_model} ${set_pretrain} ${set_epoch} ${set_batchsize} ${set_train_params1} ${set_train_params2} > ${LOG_PATH}/train.log 2>&1"
                 fi
 
                 # run train
diff --git a/tools/predict.py b/tools/predict.py
index 57a27cf42..bdf29e9f7 100644
--- a/tools/predict.py
+++ b/tools/predict.py
@@ -24,6 +24,7 @@
 
 
 def parse_args():
+
     def str2bool(v):
         return v.lower() in ("true", "t", "1")
 
@@ -50,6 +51,7 @@ def str2bool(v):
     # params for paddle predict
     parser.add_argument("-b", "--batch_size", type=int, default=1)
     parser.add_argument("--use_gpu", type=str2bool, default=True)
+    parser.add_argument("--use_npu", type=str2bool, default=False)
     parser.add_argument("--precision", type=str, default="fp32")
     parser.add_argument("--ir_optim", type=str2bool, default=True)
     parser.add_argument("--use_tensorrt", type=str2bool, default=False)
@@ -67,6 +69,8 @@ def create_paddle_predictor(args, cfg):
     config = Config(args.model_file, args.params_file)
     if args.use_gpu:
         config.enable_use_gpu(args.gpu_mem, 0)
+    elif args.use_npu:
+        config.enable_npu()
     else:
         config.disable_gpu()
         if args.cpu_threads: