diff --git a/python_module/megengine/_internal/__init__.py b/python_module/megengine/_internal/__init__.py
index e46ec7cd4f6f67ae7bbbad04b75c0f509e99f951..684aa51e94e40c7fa318e466315dd51da92f3764 100644
--- a/python_module/megengine/_internal/__init__.py
+++ b/python_module/megengine/_internal/__init__.py
@@ -25,7 +25,7 @@ from . import config, craniotome, dtype
 from . import global_init as _global_init
 from . import helper as _helper
 from . import mgb as _detail
-from . import opr, opr_param_defs, plugin
+from . import opr, opr_extra, opr_param_defs, plugin
 from .exc import MegBrainError
 from .logconf import get_logger
 from .mgb import (
diff --git a/python_module/megengine/_internal/opr_extra.py b/python_module/megengine/_internal/opr_extra.py
new file mode 100644
index 0000000000000000000000000000000000000000..7a59a2cfc0e933c1c744874791387748bbda5467
--- /dev/null
+++ b/python_module/megengine/_internal/opr_extra.py
@@ -0,0 +1,3 @@
+# -*- coding: utf-8 -*-
+# Copyright (c) 2015-2019 Megvii Inc. All rights reserved.
+
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 20ddcbbcb338119e58236882920493ee308dfe93..0a9263ac8f8a2bc1c689778b9a24b2771fa5f4b3 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -34,6 +34,11 @@ if(MGE_WITH_CUDA AND MGE_WITH_TRT)
 endif()
 
 
+if(MGE_WITH_CUDA)
+    file(GLOB_RECURSE SOURCES_ opr/impl/standalone/*.cu)
+    list(APPEND SOURCES ${SOURCES_})
+endif()
+
 add_library(megbrain OBJECT EXCLUDE_FROM_ALL ${SOURCES})
 target_link_libraries(megbrain PUBLIC mgb_opr_param_defs)
 target_include_directories(megbrain
diff --git a/src/opr/impl/mgb_cpp_opr.fbs b/src/opr/impl/mgb_cpp_opr.fbs
index 75b76795c285e4d6b286e7824e836d7422e0d6d2..48dfe44c99d3c21e35b94cdb42d9fdd90f67900e 100644
--- a/src/opr/impl/mgb_cpp_opr.fbs
+++ b/src/opr/impl/mgb_cpp_opr.fbs
@@ -55,3 +55,8 @@ struct IndexDescMaskItem {
 table IndexDescMaskDump {
     items:[IndexDescMaskItem];
 }
+
+table NMSKeep {
+    iou_thresh:float;
+    max_output:uint;
+}
diff --git a/src/serialization/impl/schema.fbs b/src/serialization/impl/schema.fbs
index d0a85bfdbf259810b335a1b54ab177d1a9b00ec7..623f5a704c0ff006c55f70b5949cce959af8d972 100644
--- a/src/serialization/impl/schema.fbs
+++ b/src/serialization/impl/schema.fbs
@@ -98,6 +98,7 @@ union OperatorParam {
     param.IndexDescMaskDump = 66,
     DType = 67,
     param.Remap = 68,
+    param.NMSKeep = 69,
 }
 
 table Operator {