Benchmarks: Fix Bug - Fix OOM issue when run pytorch models sequentially. (#93)

* Clean up the cache.

superbench/benchmarks/base.py

@@ -128,24 +128,25 @@ def run(self):
         Return:
             True if run benchmark successfully.
         """
-        if not self._preprocess():
-            return False
-
+        ret = True
+        try:
+            ret &= self._preprocess()
+            if ret:
                 self._start_time = datetime.utcnow().strftime('%Y-%m-%d %H:%M:%S')
                 for self._curr_run_index in range(self._args.run_count):
-            if not self._benchmark():
-                return False
-
+                    ret &= self._benchmark()
                 self._end_time = datetime.utcnow().strftime('%Y-%m-%d %H:%M:%S')
                 self._result.set_timestamp(self._start_time, self._end_time)
 
-        if not self.__check_result_format():
-            return False
-
-        if not self._postprocess():
-            return False
+                if ret:
+                    ret &= self.__check_result_format()
+        except BaseException as e:
+            self._result.set_return_code(ReturnCode.RUNTIME_EXCEPTION_ERROR)
+            logger.error('Run benchmark failed - benchmark: {}, message: {}'.format(self._name, str(e)))
+        finally:
+            ret &= self._postprocess()
 
-        return True
+        return ret
 
     def __check_result_format(self):
         """Check the validation of result object.

superbench/benchmarks/model_benchmarks/pytorch_base.py

@@ -183,6 +183,12 @@ def _postprocess(self):
             )
             return False
 
+        del self._model
+        del self._optimizer
+        del self._target
+
+        torch.cuda.empty_cache()
+
         return True
 
     def _cal_params_count(self):

superbench/benchmarks/return_code.py

@@ -13,6 +13,7 @@ class ReturnCode(Enum):
     INVALID_ARGUMENT = 1
     INVALID_BENCHMARK_TYPE = 2
     INVALID_BENCHMARK_RESULT = 3
+    RUNTIME_EXCEPTION_ERROR = 4
     # Return codes related with model benchmarks.
     NO_SUPPORTED_PRECISION = 10
     DISTRIBUTED_SETTING_INIT_FAILURE = 13

tests/benchmarks/model_benchmarks/test_pytorch_base.py

@@ -173,17 +173,15 @@ def _inference_step(self, precision):
 @decorator.pytorch_test
 def test_pytorch_base():
     """Test PytorchBase class."""
-    # Register BERT Base benchmark.
+    # Register mnist benchmark.
     BenchmarkRegistry.register_benchmark('pytorch-mnist', PytorchMNIST)
 
     # Launch benchmark with --no_gpu for testing.
-    context = BenchmarkRegistry.create_benchmark_context(
-        'pytorch-mnist',
-        parameters='--batch_size 32 --num_warmup 8 --num_steps 64 --model_action train inference --no_gpu'
-    )
-
-    benchmark = BenchmarkRegistry.launch_benchmark(context)
+    parameters = '--batch_size 32 --num_warmup 8 --num_steps 64 --model_action train inference --no_gpu'
+    benchmark = PytorchMNIST('pytorch-mnist', parameters=parameters)
     assert (benchmark)
+    assert (benchmark._preprocess())
+    assert (benchmark._benchmark())
     assert (benchmark.name == 'pytorch-mnist')
     assert (benchmark.return_code == ReturnCode.SUCCESS)
 
@@ -231,3 +229,34 @@ def test_pytorch_base():
     assert (isinstance(benchmark._optimizer, torch.optim.SGD))
     benchmark._optimizer_type = None
     assert (benchmark._create_optimizer() is False)
+
+    # Test _postprocess().
+    assert (benchmark._postprocess())
+
+
+@decorator.cuda_test
+@decorator.pytorch_test
+def test_pytorch_empty_cache():
+    """Test PytorchBase class."""
+    # Register mnist benchmark.
+    BenchmarkRegistry.register_benchmark('pytorch-mnist', PytorchMNIST)
+
+    # Test cache empty by manually calling torch.cuda.empty_cache().
+    parameters = '--batch_size 32 --num_warmup 8 --num_steps 64 --model_action train'
+    benchmark = PytorchMNIST('pytorch-mnist', parameters=parameters)
+    assert (benchmark)
+    assert (benchmark._preprocess())
+    assert (benchmark._benchmark())
+    del benchmark
+    assert (torch.cuda.memory_stats()['reserved_bytes.all.current'] > 0)
+    torch.cuda.empty_cache()
+    assert (torch.cuda.memory_stats()['reserved_bytes.all.current'] == 0)
+
+    # Test automatic cache empty.
+    context = BenchmarkRegistry.create_benchmark_context(
+        'pytorch-mnist', parameters='--batch_size 32 --num_warmup 8 --num_steps 64 --model_action train'
+    )
+
+    benchmark = BenchmarkRegistry.launch_benchmark(context)
+    assert (benchmark)
+    assert (torch.cuda.memory_stats()['reserved_bytes.all.current'] == 0)