- Handle error metadata properly in c_api_tfrt.

- Update is_async_ based on the EagerExecutor.
- In async mode, don't report error synchronously

PiperOrigin-RevId: 337548544
Change-Id: Ib71900724ac4bd64da95d167a60545989da03c85

tensorflow/python/eager/context.py

@@ -414,7 +414,7 @@ class Context(object):
       raise ValueError(
           "execution_mode should be None/SYNC/ASYNC. Got %s" % execution_mode)
     if execution_mode is None:
-      execution_mode = ASYNC if tfrt_utils.enabled() else SYNC
+      execution_mode = SYNC
     self._default_is_async = execution_mode == ASYNC
     self._lazy_remote_inputs_copy = None
     self._use_tfrt = tfrt_utils.enabled()

tensorflow/python/eager/ops_test.py

@@ -481,8 +481,8 @@ class OpsTest(test_util.TensorFlowTestCase, parameterized.TestCase):
     self.assertIs(weak_x(), None)
     self.assertIs(weak_y(), None)
 
-  @test_util.disable_tfrt('TFE_ContextGetExecutorForThread not implemented '
-                          'b/156188669')
+  @test_util.disable_tfrt(
+      'b/153697193: tfrt cannot decode python stacktrace yet')
   def testAsyncExceptionStackTrace(self):
     config.set_synchronous_execution(False)