diff --git a/test/cpp/jit/test.cpp b/test/cpp/jit/test.cpp
index 20cab91f2827..cc5b26fecb07 100644
--- a/test/cpp/jit/test.cpp
+++ b/test/cpp/jit/test.cpp
@@ -90,9 +90,11 @@ TH_FORALL_TESTS_CUDA(JIT_GTEST_CUDA)
 #endif
 
 #define JIT_TEST(name) test##name();
-void runJITCPPTests() {
+void runJITCPPTests(bool runCuda) {
   TH_FORALL_TESTS(JIT_TEST)
-  TH_FORALL_TESTS_CUDA(JIT_TEST)
+  if (runCuda) {
+    TH_FORALL_TESTS_CUDA(JIT_TEST)
+  }
 
   // This test is special since it requires prior setup in python.
   // So it's included here but not in the pure cpp gtest suite
diff --git a/test/test_jit.py b/test/test_jit.py
index 5763645ae41f..98d8bd4a9036 100644
--- a/test/test_jit.py
+++ b/test/test_jit.py
@@ -1484,13 +1484,23 @@ class TestJit(JitTestCase):
         for node in g.nodes():
             self.assertTrue(g2.findNode(node.kind()) is not None)
 
+    @unittest.skipIf(IS_WINDOWS, "NYI: JIT tests not yet supported on windows")
+    @unittest.skipIf(IS_SANDCASTLE, "gtest runs these in sandcastle")
+    @unittest.skipIf(RUN_CUDA, "covered by test_cpp_cuda")
+    @skipIfRocm
+    def test_cpp(self):
+        from cpp.jit import tests_setup
+        tests_setup.setup()
+        torch._C._jit_run_cpp_tests(run_cuda=False)
+        tests_setup.shutdown()
+
     @unittest.skipIf(IS_WINDOWS, "NYI: fuser support for Windows")
     @unittest.skipIf(not RUN_CUDA, "cpp tests require CUDA")
     @skipIfRocm
     def test_cpp_cuda(self):
         from cpp.jit import tests_setup
         tests_setup.setup()
-        torch._C._jit_run_cpp_tests()
+        torch._C._jit_run_cpp_tests(run_cuda=True)
         tests_setup.shutdown()
 
     def test_batchnorm(self):
diff --git a/torch/csrc/jit/init.cpp b/torch/csrc/jit/init.cpp
index e693a1584fe3..922873fee58d 100644
--- a/torch/csrc/jit/init.cpp
+++ b/torch/csrc/jit/init.cpp
@@ -79,11 +79,11 @@ bool loadPythonClasses() {
 } // anonymous namespace
 
 #if defined(_WIN32)
-void runJITCPPTests() {
+void runJITCPPTests(bool runCuda) {
   AT_ERROR("JIT tests not yet supported on Windows");
 }
 #else
-void runJITCPPTests();
+void runJITCPPTests(bool runCuda);
 #endif
 
 void initJITBindings(PyObject* module) {
@@ -182,14 +182,15 @@ void initJITBindings(PyObject* module) {
           [](std::shared_ptr<Graph> graph) { CreateAutodiffSubgraphs(graph); })
       .def(
           "_jit_run_cpp_tests",
-          [] {
+          [](bool runCuda) {
             // We have to release the GIL inside this method, because if we
             // happen to initialize the autograd engine in these tests, the
             // newly spawned worker threads will try to initialize their
             // PyThreadState*, and they need the GIL for this.
             AutoNoGIL _no_gil;
-            return runJITCPPTests();
-          })
+            return runJITCPPTests(runCuda);
+          },
+          py::arg("run_cuda"))
       .def(
           "_jit_flatten",
           [](py::handle& obj) {