diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt
index de23d3e477ecf1810df19d6c1c54767003ce6775..b5e8450f35f788dcb14d3b3e132f33a275719869 100644
--- a/paddle/fluid/inference/CMakeLists.txt
+++ b/paddle/fluid/inference/CMakeLists.txt
@@ -109,6 +109,8 @@ set(SHARED_INFERENCE_SRCS
     ${PADDLE_CUSTOM_OP_SRCS})
 
 # shared inference library deps
+list(REMOVE_ITEM fluid_modules standalone_executor
+     interpretercore_garbage_collector)
 set(SHARED_INFERENCE_DEPS ${fluid_modules} phi analysis_predictor
                           ${utils_modules})
 
diff --git a/paddle/fluid/platform/device/gpu/CMakeLists.txt b/paddle/fluid/platform/device/gpu/CMakeLists.txt
index dd2a7d1eb88306d8ecf61c3f62464a0c7c5c1726..a6f6bc232e69420e6c6839a12e3ff1394032c10f 100644
--- a/paddle/fluid/platform/device/gpu/CMakeLists.txt
+++ b/paddle/fluid/platform/device/gpu/CMakeLists.txt
@@ -3,7 +3,13 @@ if(WITH_GPU)
   nv_library(
     gpu_info
     SRCS gpu_info.cc
-    DEPS phi_backends gflags glog enforce monitor dynload_cuda)
+    DEPS phi_backends
+         gflags
+         glog
+         enforce
+         monitor
+         dynload_cuda
+         malloc)
 
   nv_test(cuda_helper_test SRCS cuda_helper_test.cu)
   nv_test(
diff --git a/tools/get_op_list.py b/tools/get_op_list.py
index aa1fd4dbfb353ba221b8fba999e02108daea4fcc..8d71af4a6e7e76c9180308575dbc5b5798e3e5d5 100644
--- a/tools/get_op_list.py
+++ b/tools/get_op_list.py
@@ -30,15 +30,20 @@ def parse_args():
         '--model_dir',
         type=str,
         default="",
-        help='Directory of the inference models.',
+        help='Directory of the inference models that named with pdmodel.',
+    )
+    parser.add_argument(
+        '--op_list',
+        type=str,
+        default="",
+        help='List of ops like "conv2d;pool2d;relu".',
     )
     return parser.parse_args()
 
 
-def get_model_ops(model_file):
+def get_model_ops(model_file, ops_set):
     model_bytes = paddle.static.load_from_file(model_file)
     pg = paddle.static.deserialize_program(model_bytes)
-    ops_set = set()
 
     for i in range(0, pg.desc.num_blocks()):
         block = pg.desc.block(i)
@@ -47,12 +52,12 @@ def get_model_ops(model_file):
         for j in range(0, size):
             ops_set.add(block.op(j).type())
 
-    return ops_set
-
 
 def get_model_phi_kernels(ops_set):
     phi_set = set()
     for op in ops_set:
+        print(op)
+        print(_get_phi_kernel_name(op))
         phi_set.add(_get_phi_kernel_name(op))
 
     return phi_set
@@ -60,10 +65,17 @@ def get_model_phi_kernels(ops_set):
 
 if __name__ == '__main__':
     args = parse_args()
-    for root, dirs, files in os.walk(args.model_dir, topdown=True):
-        for name in files:
-            if re.match(r'.*pdmodel', name):
-                ops_set = get_model_ops(os.path.join(root, name))
+    ops_set = set()
+    if args.op_list != "":
+        op_list = args.op_list.split(";")
+        for op in op_list:
+            ops_set.add(op)
+
+    if args.model_dir != "":
+        for root, dirs, files in os.walk(args.model_dir, topdown=True):
+            for name in files:
+                if re.match(r'.*pdmodel', name):
+                    get_model_ops(os.path.join(root, name), ops_set)
     phi_set = get_model_phi_kernels(ops_set)
     ops = ";".join(ops_set)
     kernels = ";".join(phi_set)