[hotfix] fix aten default bug (#2158)

colossalai/fx/profiler/opcount.py

@@ -7,6 +7,7 @@ from numbers import Number
 from typing import Any, Callable, List
 
 import torch
+from packaging import version
 
 aten = torch.ops.aten
 
@@ -188,6 +189,7 @@ def zero_flop_jit(*args):
     return 0
 
 
+if version.parse(torch.__version__) >= version.parse('1.12.0'):
     flop_mapping = {
     # gemm
         aten.mm.default: matmul_flop_jit,
@@ -276,7 +278,6 @@ elementwise_flop_aten = [
         aten.native_dropout.default,
         aten.native_dropout_backward.default,
     ]
-
     for op in elementwise_flop_aten:
         flop_mapping[op] = elementwise_flop_counter(1, 0)
 
@@ -316,3 +317,8 @@ zero_flop_aten = [
 
     for op in zero_flop_aten:
         flop_mapping[op] = zero_flop_jit
+
+else:
+    flop_mapping = {}
+    elementwise_flop_aten = {}
+    zero_flop_aten = {}

tests/test_auto_parallel/test_tensor_shard/test_node_handler/test_binary_elementwise_handler.py

@@ -207,9 +207,9 @@ def check_binary_elementwise_handler_with_int(rank, op, other_dim, world_size, p
         assert input_sharding_spec.sharding_sequence == output_sharding_spec.sharding_sequence
 
 
+@run_on_environment_flag(name='AUTO_PARALLEL')
 @parameterize('op', [torch.add])
 @parameterize('other_dim', [1, 2])
-@run_on_environment_flag(name='AUTO_PARALLEL')
 @pytest.mark.dist
 @rerun_if_address_is_in_use()
 def test_binary_elementwise_handler(op, other_dim):

tests/test_auto_parallel/test_tensor_shard/test_node_handler/test_bmm_handler.py

@@ -203,8 +203,8 @@ def check_1d_device_mesh(rank, module, world_size, port):
         assert other_sharding_spec.sharding_sequence[-1] == output_sharding_spec.sharding_sequence[-1]
 
 
-@parameterize('module', [BMMTensorMethodModule, BMMTorchFunctionModule])
 @run_on_environment_flag(name='AUTO_PARALLEL')
+@parameterize('module', [BMMTensorMethodModule, BMMTorchFunctionModule])
 @pytest.mark.dist
 @rerun_if_address_is_in_use()
 def test_bmm_handler(module):

tests/test_auto_parallel/test_tensor_shard/test_node_handler/test_getitem_handler.py

@@ -23,6 +23,7 @@ class GetItemFromTensorModel(nn.Module):
         return x
 
 
+@run_on_environment_flag(name='AUTO_PARALLEL')
 def test_getitem_from_tensor_handler():
     model = GetItemFromTensorModel()
     tracer = ColoTracer()
@@ -96,6 +97,7 @@ class GetItemFromTupleModel(nn.Module):
         return x
 
 
+@run_on_environment_flag(name='AUTO_PARALLEL')
 def test_getitem_from_tuple_handler():
     model = GetItemFromTupleModel()
     tracer = ColoTracer()

tests/test_auto_parallel/test_tensor_shard/test_node_handler/test_linear_handler.py

@@ -308,8 +308,8 @@ def check_linear_function_handler(rank, bias, input_shape, world_size, port):
             assert bias_sharding_spec.sharding_sequence[-1] == output_sharding_spec.sharding_sequence[-1]
 
 
-@parameterize('input_shape', [(1, 4, 4, 16), (4, 4, 4, 16)])
 @run_on_environment_flag(name='AUTO_PARALLEL')
+@parameterize('input_shape', [(1, 4, 4, 16), (4, 4, 4, 16)])
 @pytest.mark.dist
 @rerun_if_address_is_in_use()
 def test_linear_handler(input_shape, bias=False):

tests/test_auto_parallel/test_tensor_shard/test_node_handler/test_norm_pooling_handler.py

@@ -2,15 +2,15 @@ import pytest
 import torch
 import torch.nn as nn
 
-from colossalai.auto_parallel.tensor_shard.node_handler.normal_pooling_handler import \
+from colossalai.auto_parallel.tensor_shard.node_handler.normal_pooling_handler import NormPoolingHandler
-    NormPoolingHandler
+from colossalai.auto_parallel.tensor_shard.sharding_strategy import OperationData, OperationDataType, StrategiesVector
-from colossalai.auto_parallel.tensor_shard.sharding_strategy import (OperationData, OperationDataType, StrategiesVector)
 from colossalai.device.device_mesh import DeviceMesh
 from colossalai.fx import ColoGraphModule, ColoTracer
 from colossalai.fx.tracer.meta_patch.patched_module import linear
 from colossalai.testing.pytest_wrapper import run_on_environment_flag
 
 
+@run_on_environment_flag(name='AUTO_PARALLEL')
 def test_norm_pool_handler():
     model = nn.Sequential(nn.MaxPool2d(4, padding=1).to('meta'))
     tracer = ColoTracer()

tests/test_auto_parallel/test_tensor_shard/test_node_handler/test_reshape_handler.py

@@ -20,6 +20,7 @@ class ReshapeModel(nn.Module):
         return reshape_node
 
 
+@run_on_environment_flag(name='AUTO_PARALLEL')
 def test_reshape_handler():
     model = ReshapeModel()
     tracer = ColoTracer()

tests/test_auto_parallel/test_tensor_shard/test_node_handler/test_tensor_constructor.py

@@ -5,6 +5,7 @@ from colossalai.auto_parallel.tensor_shard.node_handler.tensor_constructor_handl
 from colossalai.auto_parallel.tensor_shard.sharding_strategy import OperationData, OperationDataType, StrategiesVector
 from colossalai.device.device_mesh import DeviceMesh
 from colossalai.fx import ColoGraphModule, ColoTracer
+from colossalai.testing.pytest_wrapper import run_on_environment_flag
 
 
 class TensorConstructorModel(nn.Module):
@@ -18,6 +19,7 @@ class TensorConstructorModel(nn.Module):
         return x
 
 
+@run_on_environment_flag(name='AUTO_PARALLEL')
 def test_where_handler():
     model = TensorConstructorModel()
     tracer = ColoTracer()

tests/test_auto_parallel/test_tensor_shard/test_node_handler/test_unary_element_wise_handler.py

@@ -22,6 +22,7 @@ class ReLuModel(nn.Module):
         return relu_node
 
 
+@run_on_environment_flag(name='AUTO_PARALLEL')
 def test_elementwise_handler():
     model = ReLuModel()
     tracer = ColoTracer()

tests/test_auto_parallel/test_tensor_shard/test_param_resharding_cost.py

@@ -10,6 +10,7 @@ from colossalai.auto_parallel.tensor_shard.solver import (
 )
 from colossalai.device.device_mesh import DeviceMesh
 from colossalai.fx import ColoGraphModule, ColoTracer
+from colossalai.testing.pytest_wrapper import run_on_environment_flag
 
 
 def _param_resharding_cost_assertion(node):
@@ -51,6 +52,7 @@ class ConvModel(torch.nn.Module):
         return x
 
 
+@run_on_environment_flag(name='AUTO_PARALLEL')
 def test_linear_module():
     model = LinearModel(4, 8)
     physical_mesh_id = torch.arange(0, 4)
@@ -86,6 +88,7 @@ def test_linear_module():
     _param_resharding_cost_assertion(linear_node)
 
 
+@run_on_environment_flag(name='AUTO_PARALLEL')
 def test_conv_module():
     model = ConvModel(3, 6, 2)
     physical_mesh_id = torch.arange(0, 4)