Bypass ThreadedEngine in test_operator_gpu.py:test_convolution_multip…

…le_streams. (apache#14338)

tests/python/gpu/test_operator_gpu.py

@@ -564,8 +564,18 @@ def _conv_with_num_streams(seed):
 
 @with_seed()
 def test_convolution_multiple_streams():
+    engines = ['NaiveEngine', 'ThreadedEngine', 'ThreadedEnginePerDevice']
+
+    if os.getenv('MXNET_ENGINE_TYPE') is not None:
+        engines = [os.getenv('MXNET_ENGINE_TYPE'),]
+        print("Only running against '%s'" % engines[0], file=sys.stderr, end='')
+    # Remove this else clause when the ThreadedEngine can handle this test
+    else:
+        engines.remove('ThreadedEngine')
+        print("SKIP: 'ThreadedEngine', only running against %s" % engines, file=sys.stderr, end='')
+
     for num_streams in [1, 2]:
-        for engine in ['NaiveEngine', 'ThreadedEngine', 'ThreadedEnginePerDevice']:
+        for engine in engines:
             _test_in_separate_process(_conv_with_num_streams,
                 {'MXNET_GPU_WORKER_NSTREAMS' : num_streams, 'MXNET_ENGINE_TYPE' : engine})