parallel mode must be set before create an initializer

mindspore/ccsrc/frontend/parallel/context.cc

@@ -81,6 +81,8 @@ void ParallelContext::set_mirror_mean(bool mirror_mean) { mirror_mean_ = mirror_
 
 void ParallelContext::set_full_batch(bool full_batch) { full_batch_ = full_batch; }
 
+void ParallelContext::set_has_initializer(bool has_initializer) { has_initializer_ = has_initializer; }
+
 void ParallelContext::set_cast_before_mirror(bool cast_before_mirror) { cast_before_mirror_ = cast_before_mirror; }
 
 void ParallelContext::set_loss_repeated_mean(bool loss_repeated_mean) { loss_repeated_mean_ = loss_repeated_mean; }

mindspore/ccsrc/frontend/parallel/context.h

@@ -58,6 +58,9 @@ class ParallelContext {
   void set_full_batch(bool full_batch);
   bool full_batch() const { return full_batch_; }
 
+  void set_has_initializer(bool has_initializer);
+  bool has_initializer() const { return has_initializer_; }
+
   void set_cast_before_mirror(bool cast_before_mirror);
   bool cast_before_mirror() const { return cast_before_mirror_; }
 
@@ -112,6 +115,7 @@ class ParallelContext {
   static std::shared_ptr<ParallelContext> inst_context_;
   bool mirror_mean_;
   bool full_batch_;
+  bool has_initializer_ = false;
   bool cast_before_mirror_;
   bool loss_repeated_mean_;
   int32_t device_num_;

mindspore/ccsrc/pipeline/jit/init.cc

@@ -193,6 +193,8 @@ PYBIND11_MODULE(_c_expression, m) {
     .def("get_strategy_ckpt_save_file", &ParallelContext::strategy_ckpt_save_file, "Get strategy checkpoint save file.")
     .def("set_full_batch", &ParallelContext::set_full_batch, "Set whether load full batch on each device.")
     .def("get_full_batch", &ParallelContext::full_batch, "Get whether load full batch on each device.")
+    .def("set_has_initializer", &ParallelContext::set_has_initializer, "Set whether any Initializer has been created.")
+    .def("get_has_initializer", &ParallelContext::has_initializer, "Get whether any Initializer has been created.")
     .def("set_enable_parallel_optimizer", &ParallelContext::set_enable_parallel_optimizer,
          "Set enable/disable parallel optimizer.")
     .def("get_enable_parallel_optimizer", &ParallelContext::enable_parallel_optimizer,

mindspore/common/api.py

@@ -24,7 +24,7 @@ from mindspore import log as logger
 from .._c_expression import generate_key, Executor_, Tensor, MetaTensor, PynativeExecutor_
 from .._c_expression import verify_inputs_signature, init_exec_dataset, _set_dataset_mode_config, init_backend
 from .tensor import Tensor as MsTensor
-from ..parallel._utils import _get_device_num, _get_global_rank, _need_to_full, _to_full_tensor
+from ..parallel._utils import _get_device_num, _get_global_rank, _need_to_full, _to_full_tensor, _set_has_initializer
 # store ms_function class compiled pipeline cache
 ms_compile_cache = {}
 
@@ -383,6 +383,7 @@ class _Executor:
             Str, the full phase of the cell.
             Bool, if the graph has been compiled before, return False, else return True.
         """
+        _set_has_initializer(False)
         obj.check_names()
         args_names, args_list = _generate_pip_args(obj, *args)
         dic = dict(zip(args_names, args_list))

mindspore/common/initializer.py

@@ -24,6 +24,7 @@ from mindspore import log as logger
 from . import dtype as mstype
 from .tensor import Tensor
 from .._c_expression import random_normal
+from ..parallel._utils import _set_has_initializer
 
 _INITIALIZER_ALIAS = dict()
 
@@ -42,6 +43,7 @@ class Initializer:
         self._kwargs = kwargs
         self.shape = None
         self.dtype = None
+        _set_has_initializer(True)
 
     def _initialize(self, *kwargs):
         raise NotImplementedError('Must be overridden!')

mindspore/context.py

@@ -437,6 +437,8 @@ def set_auto_parallel_context(**kwargs):
         If a program has tasks with different parallel modes, then before setting new parallel mode for
         next task, interface mindspore.context.reset_auto_parallel_context() needs to be called to reset
         the configuration.
+        Setting or changing parallel modes must be called before any Initializer created, or RuntimeError
+        will be raised.
 
     Args:
         device_num (int): Available device number, the value must be in [1, 4096]. Default: 1.
@@ -477,6 +479,7 @@ def set_auto_parallel_context(**kwargs):
 
     Raises:
         ValueError: If input key is not attribute in auto parallel context.
+        RuntimeError: If there is any Initializer created before setting or changing parallel_mode.
 
     Examples:
         >>> context.set_auto_parallel_context(device_num=8)

mindspore/parallel/_auto_parallel_context.py

@@ -176,8 +176,12 @@ class _AutoParallelContext:
 
         Raises:
             ValueError: If parallel mode is not supported.
+            RuntimeError: If there is any Initializer created before setting or changing parallel_mode.
         """
         self.check_context_handle()
+        if self.get_has_initializer():
+            self.set_has_initializer(False)
+            raise RuntimeError("Must set or change parallel mode before any Initializer created.")
         ret = self._context_handle.set_parallel_mode(parallel_mode)
         if ret is False:
             raise ValueError("Parallel mode does not support {}".format(parallel_mode))
@@ -249,6 +253,21 @@ class _AutoParallelContext:
         self.check_context_handle()
         return self._context_handle.get_full_batch()
 
+    def set_has_initializer(self, has_initializer):
+        """
+        Set whether any Initializer has been created.
+
+        Args:
+            has_initializer (bool): True if a Initializer created.
+        """
+        self.check_context_handle()
+        self._context_handle.set_has_initializer(has_initializer)
+
+    def get_has_initializer(self):
+        """Get whether any Initializer has been created."""
+        self.check_context_handle()
+        return self._context_handle.get_has_initializer()
+
     def set_strategy_ckpt_save_file(self, strategy_ckpt_save_file):
         """
         Set strategy checkpoint save path.
@@ -543,6 +562,7 @@ def _set_auto_parallel_context(**kwargs):
 
     Raises:
         ValueError: If input key is not attribute in auto parallel context.
+        RuntimeError: If there is any Initializer created before setting or changing parallel_mode.
     """
     for key, value in kwargs.items():
         if key not in _set_auto_parallel_context_func_map:

mindspore/parallel/_utils.py

@@ -32,6 +32,19 @@ def _get_full_batch():
     """Get whether to use full_batch."""
     return auto_parallel_context().get_full_batch()
 
+def _get_has_initializer():
+    """Get whether any Initializer has been created."""
+    return auto_parallel_context().get_has_initializer()
+
+def _set_has_initializer(has_initializer):
+    """
+    Set whether any Initializer has been created.
+
+    Args:
+        has_initializer (bool): True if a Initializer created.
+    """
+    auto_parallel_context().set_has_initializer(has_initializer)
+
 
 def _need_to_full():
     """Check whether to convert input to full shape or tensor."""

tests/st/nccl/test_nccl_lenet.py

@@ -78,6 +78,7 @@ def multisteplr(total_steps, gap, base_lr=0.9, gamma=0.1, dtype=mstype.float32):
 
 
 def test_lenet_nccl():
+    context.set_auto_parallel_context(parallel_mode="data_parallel", mirror_mean=True, device_num=get_group_size())
     net = LeNet()
     net.set_train()
 
@@ -86,7 +87,6 @@ def test_lenet_nccl():
     mom_optimizer = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), learning_rate, momentum)
     criterion = nn.SoftmaxCrossEntropyWithLogits(is_grad=False, sparse=True)
     net_with_criterion = WithLossCell(net, criterion)
-    context.set_auto_parallel_context(parallel_mode="data_parallel", mirror_mean=True, device_num=get_group_size())
     train_network = TrainOneStepCell(net_with_criterion, mom_optimizer)
     train_network.set_train()
     losses = []

tests/ut/python/communication/test_data_parallel_lenet.py

@@ -24,6 +24,7 @@ import mindspore.nn as nn
 from mindspore import Tensor, Model, ParallelMode
 from mindspore.nn.optim import Momentum
 from mindspore.ops import operations as P
+from mindspore.parallel._utils import _set_has_initializer
 
 _current_dir = os.path.dirname(os.path.realpath(__file__)) + "/../test_data"
 
@@ -89,3 +90,4 @@ def test_lenet5_train_step_training_pynative():
     Model(network=network, loss_fn=loss_fn, optimizer=optimizer)
     context.set_context(mode=context.GRAPH_MODE)
     context.reset_auto_parallel_context()
+    _set_has_initializer(False)

tests/ut/python/model/test_mix_precision.py

@@ -96,6 +96,8 @@ def test_on_momentum():
 
 def test_data_parallel_with_cast():
     """test_data_parallel_with_cast"""
+    context.reset_auto_parallel_context()
+    context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL, mirror_mean=True, device_num=8)
     predict = Tensor(np.ones([1, 1, 32, 32]).astype(np.float32) * 0.01)
     label = Tensor(np.zeros([1, 10]).astype(np.float32))
     net = LeNet5()
@@ -107,8 +109,6 @@ def test_data_parallel_with_cast():
                          learning_rate=0.1,
                          momentum=0.9)
     net = WithLossCell(net, loss_fn)
-    context.reset_auto_parallel_context()
-    context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL, mirror_mean=True, device_num=8)
     net = TrainOneStepCell(net, optimizer)
 
     _executor.compile(net, predict, label)

tests/ut/python/nn/test_parameter.py

@@ -21,7 +21,7 @@ from mindspore import context, Tensor, Parameter, ParameterTuple
 from mindspore._checkparam import _check_str_by_regular
 from mindspore.common import dtype as mstype
 from mindspore.common.initializer import initializer
-
+from mindspore.parallel._utils import _set_has_initializer
 
 def test_parameter_init():
     dat = np.array([[1, 2, 3], [2, 3, 4]])
@@ -170,6 +170,7 @@ def test_scalar_parameter_update():
 
 
 def test_parameter_lazy_init():
+    _set_has_initializer(False)
     # support lazy init in SEMI_AUTO_PARALLEL mode
     context.reset_auto_parallel_context()
     context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8)

tests/ut/python/parallel/test_add_relu_redistribution.py

@@ -20,6 +20,7 @@ from mindspore import context
 from mindspore.common.api import _executor
 from mindspore.ops import composite as C
 from mindspore.ops import operations as P
+from mindspore.parallel._utils import _set_has_initializer
 from tests.ut.python.ops.test_math_ops import VirtualLoss
 
 
@@ -60,12 +61,13 @@ def compile_net(net, x, y):
 
 
 def test_add_relu_stride_slice():
+    _set_has_initializer(False)
     context.set_auto_parallel_context(device_num=8, global_rank=7)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     strategy0 = ((1, 1), (1, 1))
     strategy1 = ((8, 1),)
     net = Grad(NetWithLoss(AddRelu(strategy0, strategy1)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([128, 32]), dtype=ms.float32)
     y = Tensor(np.ones([128, 32]), dtype=ms.float32)
@@ -73,12 +75,13 @@ def test_add_relu_stride_slice():
 
 
 def test_add_relu_all_gather():
+    _set_has_initializer(False)
     context.set_auto_parallel_context(device_num=8, global_rank=7)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     strategy0 = ((8, 1), (8, 1))
     strategy1 = ((1, 1),)
     net = Grad(NetWithLoss(AddRelu(strategy0, strategy1)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([128, 32]), dtype=ms.float32)
     y = Tensor(np.ones([128, 32]), dtype=ms.float32)

tests/ut/python/parallel/test_allreduce_fusion.py

@@ -23,6 +23,7 @@ from mindspore.nn.optim.momentum import Momentum
 from mindspore.parallel import _cost_model_context as cost_model_context
 from mindspore.parallel._auto_parallel_context import auto_parallel_context
 from mindspore.train import Model, ParallelMode
+from mindspore.parallel._utils import _set_has_initializer
 from tests.dataset_mock import MindData
 
 
@@ -105,10 +106,8 @@ def train_common(net):
     momentum = 0.9
     epoch_size = 2
     device_num = 4
-    context.reset_auto_parallel_context()
     auto_parallel_context().set_enable_all_reduce_fusion(enable_all_reduce_fusion=True)
-    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL, device_num=device_num,
-                                      parameter_broadcast=False)
+    context.set_auto_parallel_context(device_num=device_num, parameter_broadcast=False)
     context.set_context(mode=context.GRAPH_MODE)
 
     predict = Tensor(np.ones([batch_size, 128]), dtype=ms.float32)
@@ -183,9 +182,12 @@ def test_allreduce_fusion_parameters():
 
 
 def test_allreduce_fusion1():
+    _set_has_initializer(False)
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_algorithm=1)
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_times=2)
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_tail_percent=0.5)
+    context.reset_auto_parallel_context()
+    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL)
     net = SimpleDMLNet(DenseNet1(has_bias=False, activation=None), DenseNet2(has_bias=False, activation=None))
     allreduce_fusion_dict = train_common(net)
     expect_dict = {'backbone2.fc8.weight': 2,
@@ -210,6 +212,8 @@ def test_allreduce_fusion2():
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_times=2)
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_tail_percent=0.5)
     cost_model_context.reset_cost_model_context()
+    context.reset_auto_parallel_context()
+    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL)
     net = SimpleDMLNet(DenseNet1(has_bias=False, activation=None), DenseNet2(has_bias=False, activation=None))
     allreduce_fusion_dict = train_common(net)
     expect_dict = {}
@@ -221,6 +225,8 @@ def test_allreduce_fusion3():
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_algorithm=1)
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_times=3)
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_tail_percent=0.3333333)
+    context.reset_auto_parallel_context()
+    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL)
     net = SimpleDMLNet(DenseNet1(has_bias=True, activation='relu'), DenseNet2(has_bias=False, activation='relu'))
     allreduce_fusion_dict = train_common(net)
     expect_dict = {'backbone2.fc8.weight': 3,
@@ -247,6 +253,8 @@ def test_allreduce_fusion4():
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_algorithm=1)
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_times=2)
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_tail_percent=0.5)
+    context.reset_auto_parallel_context()
+    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL)
     net = SimpleDMLNet(DenseNet2(has_bias=False, activation=None), DenseNet2(has_bias=False, activation=None))
     allreduce_fusion_dict = train_common(net)
     expect_dict = {'backbone2.fc8.weight': 2,
@@ -276,6 +284,8 @@ def test_allreduce_fusion5():
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_allreduce_inherent_time=0.05)
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_allreduce_bandwidth=0.000001)
     cost_model_context.set_cost_model_context(costmodel_allreduce_fusion_computation_time_parameter=0.0000015)
+    context.reset_auto_parallel_context()
+    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL)
     net = SimpleDMLNet(DenseNet2(has_bias=False, activation=None), DenseNet2(has_bias=False, activation=None))
     allreduce_fusion_dict = train_common(net)
 

tests/ut/python/parallel/test_alltoall.py

@@ -23,7 +23,7 @@ from mindspore.common.parameter import Parameter
 from mindspore.nn.loss import SoftmaxCrossEntropyWithLogits
 from mindspore.nn.optim.momentum import Momentum
 from mindspore.ops import operations as P
-from mindspore.parallel._utils import _reset_op_id
+from mindspore.parallel._utils import _reset_op_id, _set_has_initializer
 from mindspore.train import Model, ParallelMode
 from tests.dataset_mock import MindData
 
@@ -90,6 +90,7 @@ def all_to_all_common(strategy1):
 
 
 def test_all_to_all():
+    _set_has_initializer(False)
     strategy1 = ((8, 1),)
     context.set_context(mode=context.GRAPH_MODE, save_graphs=False)
     _reset_op_id()

tests/ut/python/parallel/test_arithmetic.py

@@ -20,6 +20,7 @@ from mindspore import Parameter, Tensor, context
 from mindspore.common.api import _executor
 from mindspore.ops import composite as C
 from mindspore.ops import operations as P
+from mindspore.parallel._utils import _set_has_initializer
 from tests.ut.python.ops.test_math_ops import VirtualLoss
 
 
@@ -60,11 +61,12 @@ def test_matmul_sub():
             out = self.sub(out, b)
             return out
 
+    _set_has_initializer(False)
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((4, 2), (4, 2))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
@@ -85,10 +87,10 @@ def test_matmul_add():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((4, 2), (4, 2))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
@@ -109,10 +111,10 @@ def test_matmul_mul():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((4, 2), (4, 2))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
@@ -133,10 +135,10 @@ def test_matmul_div():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((4, 2), (4, 2))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
@@ -157,10 +159,10 @@ def test_matmul_greater():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((4, 2), (4, 2))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
@@ -181,10 +183,10 @@ def test_matmul_add_broadcast():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((4, 2), (2,))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
@@ -205,10 +207,10 @@ def test_matmul_add_broadcast2():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 4), (4, 1))
     strategy2 = ((4, 1), (1, 2))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 1]), dtype=ms.float32)
@@ -229,10 +231,10 @@ def test_matmul_sub_broadcast():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((4, 2), (2,))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
@@ -253,10 +255,10 @@ def test_matmul_sub_broadcast2():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 4), (4, 1))
     strategy2 = ((4, 1), (1, 2))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 1]), dtype=ms.float32)
@@ -277,10 +279,10 @@ def test_matmul_mul_broadcast():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((4, 2), (2,))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
@@ -301,10 +303,10 @@ def test_matmul_mul_broadcast2():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 4), (4, 1))
     strategy2 = ((4, 1), (1, 2))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 1]), dtype=ms.float32)
@@ -325,10 +327,10 @@ def test_matmul_div_broadcast():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((4, 2), (2,))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
@@ -349,10 +351,10 @@ def test_matmul_div_broadcast2():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 4), (4, 1))
     strategy2 = ((4, 1), (1, 2))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 1]), dtype=ms.float32)
@@ -373,10 +375,10 @@ def test_matmul_greater_broadcast():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((4, 2), (2,))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
@@ -397,10 +399,10 @@ def test_matmul_greater_broadcast2():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 4), (4, 1))
     strategy2 = ((4, 1), (1, 2))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 1]), dtype=ms.float32)
@@ -421,10 +423,10 @@ def test_matmul_floordiv():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((4, 2), (4, 2))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
@@ -445,10 +447,10 @@ def test_matmul_floordiv_broadcast():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((4, 2), (2,))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
@@ -469,10 +471,10 @@ def test_matmul_floordiv_broadcast2():
             return out
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 4), (4, 1))
     strategy2 = ((4, 1), (1, 2))
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 1]), dtype=ms.float32)

tests/ut/python/parallel/test_auto_parallel_BN_PReLU.py

@@ -64,10 +64,10 @@ def test_auto_parallel_bn_with_prelu():
 
     size = 8
     context.set_auto_parallel_context(device_num=size, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="auto_parallel")
 
     x = Tensor(np.random.rand(16, 16, 32, 64), dtype=ms.float32)
 
     net = GradWrap(NetWithLoss(Net()))
-    context.set_auto_parallel_context(parallel_mode="auto_parallel")
     net.set_auto_parallel()
     _executor.compile(net, x)

tests/ut/python/parallel/test_auto_parallel_double_subgraphs.py

@@ -106,8 +106,8 @@ def test_double_subgraphs():
     cost_model_context.set_cost_model_context(multi_subgraphs=True)
     context.set_context(save_graphs=True)
     context.set_auto_parallel_context(device_num=8, global_rank=0)
-    net = TrainStepWarp(NetWithLoss(Net()))
     context.set_auto_parallel_context(parallel_mode="auto_parallel")
+    net = TrainStepWarp(NetWithLoss(Net()))
     net.set_auto_parallel()
 
     x = Tensor(np.ones([8, 8, 8, 8]), dtype=ms.float32)

tests/ut/python/parallel/test_auto_parallel_tuple_depend.py

@@ -68,9 +68,9 @@ def test_virtual_dataset_3_input():
             out = self.matmul2(out, b)
             return out
 
-    net = GradWrap(NetWithLoss(Net()))
     context.set_auto_parallel_context(parallel_mode="auto_parallel")
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    net = GradWrap(NetWithLoss(Net()))
     net.set_auto_parallel()
     x = Tensor(np.ones([128, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)

tests/ut/python/parallel/test_auto_parallel_two_bn.py

@@ -68,11 +68,11 @@ def test_two_bn():
             out = self.block2(out)
             return out
 
-    net = NetWithLoss(Net())
-    x = Tensor(np.ones([64, 64]), dtype=ms.float32)
     context.set_context(save_graphs=True)
     context.set_auto_parallel_context(device_num=8, global_rank=0)
     context.set_auto_parallel_context(parallel_mode="auto_parallel")
+    net = NetWithLoss(Net())
+    x = Tensor(np.ones([64, 64]), dtype=ms.float32)
     net.set_auto_parallel()
     set_algo_parameters(elementwise_op_strategy_follow=True)
     reset_op_id()

tests/ut/python/parallel/test_batch_parallel.py

@@ -94,12 +94,12 @@ def test_batch():
             return out4
 
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((8, 1, 1, 1), (1, 1, 1, 1))
     strategy2 = ((1, 1, 1, 8), (1, 1, 1, 8))
     strategy3 = ((4, 1, 1, 2), (4, 1, 1, 2))
 
     net = GradWrap(NetWithLoss(Net(strategy1, strategy2, strategy3)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     net.set_auto_parallel()
 
     x = Tensor(np.ones([128, 16, 34, 34]), dtype=ms.float32)

tests/ut/python/parallel/test_batchnorm_batch_parallel.py

@@ -118,6 +118,9 @@ def batchnorm_net(num_classes):
 
 
 def test_batchnorm_batch_parallel():
+    context.reset_auto_parallel_context()
+    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL, device_num=dev_num)
+    context.set_context(mode=context.GRAPH_MODE)
     num_classes = 1001
     batch_size = 32
     learning_rate = 0.1
@@ -134,9 +137,6 @@ def test_batchnorm_batch_parallel():
     loss.softmax_cross_entropy.set_strategy(((dev_num, 1), (dev_num, 1)))
     opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), learning_rate, momentum)
 
-    context.reset_auto_parallel_context()
-    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL, device_num=dev_num)
-    context.set_context(mode=context.GRAPH_MODE)
     model = Model(net, loss, opt)
     model.train(epoch_size, dataset, dataset_sink_mode=False)
 

tests/ut/python/parallel/test_bn_prelu_cell.py

@@ -198,6 +198,7 @@ def bn_net():
 
 def bn_common(parallel_mode, train_flag, strategy_loss=None):
     context.set_context(mode=context.GRAPH_MODE)
+    context.set_auto_parallel_context(parallel_mode=parallel_mode, device_num=8)
     learning_rate = 0.1
     momentum = 0.9
     epoch_size = 2
@@ -218,7 +219,6 @@ def bn_common(parallel_mode, train_flag, strategy_loss=None):
 
     if parallel_mode == ParallelMode.DATA_PARALLEL:
         context.set_auto_parallel_context(parameter_broadcast=True)
-    context.set_auto_parallel_context(parallel_mode=parallel_mode, device_num=8)
     model = Model(net, loss, opt)
     if train_flag:
         model.train(epoch_size, dataset, dataset_sink_mode=False)

tests/ut/python/parallel/test_get_next.py

@@ -88,13 +88,13 @@ def test_get_next_semi_auto_parallel():
             return x
 
     context.set_auto_parallel_context(device_num=4, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     network = Net(strategy1=((1, 4),), strategy2=((4, 1), (1,)))
     strategy3 = ((4, 1), (), ())
     strategy4 = ((4, 1), (4, 1))
     net_with_loss = NetWithLoss(network, [ms.float32, ms.int32], [[32, 64], [32]], 2, strategy3=strategy3,
                                 strategy4=strategy4)
     net = GradWrap(net_with_loss)
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     compile_net(net)
 
 
@@ -112,13 +112,13 @@ def test_get_next_semi_auto_parallel1():
             return x
 
     context.set_auto_parallel_context(device_num=4, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     network = Net(strategy1=((1, 4),), strategy2=((4, 1), (1,)))
     strategy3 = ((1, 4), (), ())
     strategy4 = ((4, 1), (4, 1))
     net_with_loss = NetWithLoss(network, [ms.float32, ms.int32], [[32, 64], [32]], 2, strategy3=strategy3,
                                 strategy4=strategy4)
     net = GradWrap(net_with_loss)
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     compile_net(net)
 
 
@@ -136,10 +136,10 @@ def test_get_next_auto_parallel():
             return x
 
     context.set_auto_parallel_context(device_num=4, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="auto_parallel")
     network = Net()
     net_with_loss = NetWithLoss(network, [ms.float32, ms.int32], [[32, 64], [32]], 2)
     net = GradWrap(net_with_loss)
-    context.set_auto_parallel_context(parallel_mode="auto_parallel")
     compile_net(net)
 
 
@@ -153,6 +153,6 @@ def test_only_one_get_next():
             return self.get_next()
 
     context.set_auto_parallel_context(device_num=4, global_rank=0)
-    net = Net()
     context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
+    net = Net()
     compile_net(net)

tests/ut/python/parallel/test_initializer_weight_slice.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import numpy as np
+import pytest
 from mindspore import context
 import mindspore.nn as nn
 from mindspore.ops import operations as P
@@ -22,20 +23,19 @@ import mindspore.common.api as me
 from mindspore.common.initializer import initializer
 from hccl_test.manage.api import Hccl
 
+class Net(nn.Cell):
+    def __init__(self, strategy1, strategy2, weight):
+        super().__init__()
+        self.weight = Parameter(weight, "w1")
+        self.matmul = P.MatMul(transpose_a=False, transpose_b=True).set_strategy(strategy1)
+        self.relu = P.ReLU().set_strategy(strategy2)
 
-def check_initializer_weight_slice(init_name="Uniform"):
-    class Net(nn.Cell):
-        def __init__(self, strategy1, strategy2, weight):
-            super().__init__()
-            self.weight = Parameter(weight, "w1")
-            self.matmul = P.MatMul(transpose_a=False, transpose_b=True).set_strategy(strategy1)
-            self.relu = P.ReLU().set_strategy(strategy2)
-
-        def construct(self, x):
-            out = self.matmul(x, self.weight)
-            out = self.relu(out)
-            return out
+    def construct(self, x):
+        out = self.matmul(x, self.weight)
+        out = self.relu(out)
+        return out
 
+def check_initializer_weight_slice(init_name="Uniform"):
     def get_slice(rank):
         hccl = Hccl()
         rank_save = hccl.rank_id
@@ -77,5 +77,28 @@ def test_initializer_weight_slice():
     for init_name in initializers:
         check_initializer_weight_slice(init_name)
 
+def test_wrong_order_set_parallel_mode_with_initializer():
+    weight = initializer("Normal", [64, 32], ms.float32)
+    strategy1 = ((2, 1), (4, 1))
+    strategy2 = ((2, 4),)
+    net = Net(strategy1, strategy2, weight)
+    exe = me._executor
+    x = Tensor(np.ones([32, 32]), dtype=ms.float32)
+    with pytest.raises(RuntimeError):
+        context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
+        net.set_auto_parallel()
+        exe.compile(net, x, auto_parallel_mode=True, phase='train')
+
+def test_wrong_order_set_parallel_mode_without_initializer():
+    weight = Tensor(np.ones([64, 32]), ms.float32)
+    strategy1 = ((2, 1), (4, 1))
+    strategy2 = ((2, 4),)
+    net = Net(strategy1, strategy2, weight)
+    exe = me._executor
+    x = Tensor(np.ones([32, 32]), dtype=ms.float32)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=8, global_rank=0)
+    net.set_auto_parallel()
+    exe.compile(net, x, auto_parallel_mode=True, phase='train')
+
 if __name__ == '__main__':
     test_initializer_weight_slice()

tests/ut/python/parallel/test_linear.py

@@ -58,12 +58,12 @@ def test_linear():
             return out
 
     context.set_auto_parallel_context(device_num=16, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy0 = ((2, 4), (2, 4))
     strategy1 = ((2, 4), (4,))
     strategy2 = ((2, 8),)
     strategy3 = ((16, 1), (16, 1))
     net = GradWrap(NetWithLoss(Net(strategy0, strategy1, strategy2), strategy3))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     net.set_auto_parallel()
 
     x = Tensor(np.ones([64, 32]), dtype=ms.float32)

tests/ut/python/parallel/test_loss_and_optimizer.py

@@ -54,6 +54,7 @@ def test_momentum():
             return out
 
     context.set_auto_parallel_context(device_num=4, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 1), (2, 1))
     strategy2 = ((4, 1),)
     strategy3 = ((4, 1), (4, 1))
@@ -69,7 +70,6 @@ def test_momentum():
     net_with_loss = NetWithLoss(net, strategy3)
 
     train_net = TrainOneStepCell(net_with_loss, optimizer)
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     compile_net(train_net, x, b)
 
@@ -88,6 +88,7 @@ def test_momentum_with_loss_scale():
             return out
 
     context.set_auto_parallel_context(device_num=4, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 1), (2, 1))
     strategy2 = ((4, 1),)
     strategy3 = ((4, 1), (4, 1))
@@ -103,7 +104,6 @@ def test_momentum_with_loss_scale():
     net_with_loss = NetWithLoss(net, strategy3)
 
     train_net = TrainOneStepCell(net_with_loss, optimizer)
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     compile_net(train_net, x, b)
 
@@ -122,6 +122,7 @@ def test_momentum_with_dynamic_lr():
             return out
 
     context.set_auto_parallel_context(device_num=4, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 1), (2, 1))
     strategy2 = ((4, 1),)
     strategy3 = ((4, 1), (4, 1))
@@ -138,7 +139,6 @@ def test_momentum_with_dynamic_lr():
     net_with_loss = NetWithLoss(net, strategy3)
 
     train_net = TrainOneStepCell(net_with_loss, optimizer)
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     compile_net(train_net, x, b)
 
@@ -157,6 +157,7 @@ def test_momentum_with_loss_scale_and_dynamic_lr():
             return out
 
     context.set_auto_parallel_context(device_num=4, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     strategy1 = ((2, 1), (2, 1))
     strategy2 = ((4, 1),)
@@ -174,7 +175,6 @@ def test_momentum_with_loss_scale_and_dynamic_lr():
     net_with_loss = NetWithLoss(net, strategy3)
 
     train_net = TrainOneStepCell(net_with_loss, optimizer)
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     compile_net(train_net, x, b)
 
@@ -193,6 +193,7 @@ def test_lars():
             return out
 
     context.set_auto_parallel_context(device_num=4, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     strategy1 = ((2, 1), (2, 1))
     strategy2 = ((4, 1),)
     strategy3 = ((4, 1), (4, 1))
@@ -209,6 +210,5 @@ def test_lars():
                      lars_filter=lambda x: 'bn' not in x.name)
     net_with_loss = NetWithLoss(net, strategy3)
     train_net = TrainOneStepCell(net_with_loss, optimizer)
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     compile_net(train_net, x, b)

tests/ut/python/parallel/test_one_hot_net.py

@@ -266,11 +266,11 @@ class BNReshapeDenseBNNet(nn.Cell):
 def test_bn_reshape_dense_bn_train_loss():
     batch_size = 16
     context.set_auto_parallel_context(device_num=device_num, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     input_ = Tensor(np.ones([batch_size, 2, 32, 32]).astype(np.float32) * 0.01)
     label = Tensor(np.ones([batch_size]), dtype=ms.int32)
 
     net = GradWrap(NetWithLoss(BNReshapeDenseBNNet()))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     net.set_auto_parallel()
 
     _executor.compile(net, input_, label)
@@ -279,12 +279,12 @@ def test_bn_reshape_dense_bn_train_loss():
 def test_semi_one_hot_net_batch():
     batch_size = 16
     context.set_auto_parallel_context(device_num=device_num, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     input_ = Tensor(np.ones([batch_size * 1, 512]).astype(np.float32) * 0.01)
     label = Tensor(np.ones([batch_size]), dtype=ms.int32)
 
     net = SemiAutoOneHotNet(args=Args(), strategy=StrategyBatch())
     net = GradWrap(NetWithLoss(net))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     net.set_auto_parallel()
 
     _executor.compile(net, input_, label)
@@ -300,10 +300,10 @@ def test_semi_one_hot_net_model():
     label = Tensor(np.ones([batch_size]), dtype=ms.int32)
     dataset = Dataset(predict, label, 2, input_num=2)
 
-    net = SemiAutoOneHotNet(args=Args(), strategy=StrategyModel())
-    opt = Momentum(net.trainable_params(), learning_rate, momentum)
     context.reset_auto_parallel_context()
     context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL, device_num=16)
     context.set_context(mode=context.GRAPH_MODE)
+    net = SemiAutoOneHotNet(args=Args(), strategy=StrategyModel())
+    opt = Momentum(net.trainable_params(), learning_rate, momentum)
     model = Model(net, optimizer=opt)
     model.train(epoch_size, dataset, dataset_sink_mode=False)

tests/ut/python/parallel/test_operator_model_parallel.py

@@ -353,6 +353,8 @@ def test_resnet_operator_batch_parallel():
 
     context.reset_auto_parallel_context()
     context.set_auto_parallel_context(device_num=dev_num, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL, device_num=dev_num)
+    context.set_context(mode=context.GRAPH_MODE)
     predict = Tensor(np.ones([batch_size, 3, 224, 224]), dtype=ms.float32)
     label = Tensor(np.ones([batch_size]), dtype=ms.int32)
 
@@ -363,9 +365,6 @@ def test_resnet_operator_batch_parallel():
     loss.softmax_cross_entropy.set_strategy(((dev_num, 1), (dev_num, 1)))
     opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), learning_rate, momentum)
 
-    context.reset_auto_parallel_context()
-    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL, device_num=dev_num)
-    context.set_context(mode=context.GRAPH_MODE)
     model = Model(net, loss, opt)
     model.train(epoch_size, dataset, dataset_sink_mode=False)
 
@@ -379,6 +378,8 @@ def test_resnet_model_parallel():
 
     context.reset_auto_parallel_context()
     context.set_auto_parallel_context(device_num=dev_num, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL, device_num=dev_num)
+    context.set_context(mode=context.GRAPH_MODE)
     predict = Tensor(np.ones([batch_size, 64, 112, 112]), dtype=ms.float32)
     label = Tensor(np.ones([batch_size]), dtype=ms.int32)
 
@@ -389,9 +390,6 @@ def test_resnet_model_parallel():
     loss.softmax_cross_entropy.set_strategy(((dev_num, 1), (dev_num, 1)))
     opt = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), learning_rate, momentum)
 
-    context.reset_auto_parallel_context()
-    context.set_auto_parallel_context(parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL, device_num=dev_num)
-    context.set_context(mode=context.GRAPH_MODE)
     model = Model(net, loss, opt)
     model.train(epoch_size, dataset, dataset_sink_mode=False)
 

tests/ut/python/parallel/test_optimizer.py

@@ -45,6 +45,8 @@ class Net(nn.Cell):
 
 def test_dense_gen_graph():
     context.set_context(mode=context.GRAPH_MODE)
+    context.reset_auto_parallel_context()
+    context.set_auto_parallel_context(parallel_mode=ParallelMode.HYBRID_PARALLEL, mirror_mean=True, device_num=8)
     init()
     network = Net(512, 128)
 
@@ -53,8 +55,6 @@ def test_dense_gen_graph():
                          learning_rate=0.1,
                          momentum=0.9)
     network = WithLossCell(network, loss_fn)
-    context.reset_auto_parallel_context()
-    context.set_auto_parallel_context(parallel_mode=ParallelMode.HYBRID_PARALLEL, mirror_mean=True, device_num=8)
     network = TrainOneStepCell(network, optimizer)
 
     predict = Tensor(np.ones([64, 512]).astype(np.float32) * 0.01)

tests/ut/python/parallel/test_optimizer_clone_weight.py

@@ -54,6 +54,7 @@ def test_optimizer_clone_weight():
             return out
 
     context.set_auto_parallel_context(device_num=4, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     strategy1 = ((2, 1), (2, 1))
     strategy2 = ((4, 1),)
@@ -70,7 +71,6 @@ def test_optimizer_clone_weight():
     net_with_loss = NetWithLoss(net, strategy3)
 
     train_net = TrainOneStepCell(net_with_loss, optimizer)
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     compile_net(train_net, x, b)
 
@@ -89,6 +89,7 @@ def test_optimizer_clone_weight2():
             return out
 
     context.set_auto_parallel_context(device_num=4, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     strategy1 = ((2, 1), (2, 1))
     strategy2 = ((4, 1),)
@@ -105,6 +106,5 @@ def test_optimizer_clone_weight2():
     net_with_loss = NetWithLoss(net, strategy3)
 
     train_net = TrainOneStepCell(net_with_loss, optimizer)
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     compile_net(train_net, x, b)

tests/ut/python/parallel/test_reshape.py

@@ -320,10 +320,10 @@ def reshape_net2(backbone):
     batch_size = 16
     device_num = 16
     context.set_auto_parallel_context(device_num=device_num, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     input_ = Tensor(np.ones([batch_size * device_num, 512, 7, 7]).astype(np.float32) * 0.01)
 
     net = GradWrap(NetWithLoss(backbone))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     compile_net(net, input_)
 
@@ -530,10 +530,10 @@ def test_bn_reshape_dense_bn_train():
     batch_size = 16
     device_num = 16
     context.set_auto_parallel_context(device_num=device_num, global_rank=0)
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
     input_ = Tensor(np.ones([batch_size, 2, 32, 32]).astype(np.float32) * 0.01)
 
     net = GradWrap(NetWithLoss(BNReshapeDenseBNNet()))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
 
     compile_net(net, input_)
 

tests/ut/python/parallel/test_using_seed_for_initializer.py

@@ -18,6 +18,7 @@ from numpy import allclose
 import mindspore.common.initializer as init
 import mindspore.nn as nn
 from mindspore import Parameter
+from mindspore.parallel._utils import _set_has_initializer
 
 parameter_shape = [16, 4]
 
@@ -46,6 +47,7 @@ def test_using_same_seed_for_initializer():
     np.random.seed(0)
     net2 = ParameterNet()
     net2.init_parameters_data()
+    _set_has_initializer(False)
     for key in net1.parameters_dict():
         if key not in net2.parameters_dict():
             assert False
@@ -60,6 +62,7 @@ def test_using_diffserent_seed_for_initializer():
     np.random.seed(1)
     net2 = ParameterNet()
     net2.init_parameters_data()
+    _set_has_initializer(False)
     for key in net1.parameters_dict():
         if key not in net2.parameters_dict():
             assert False

tests/ut/python/parallel/test_virtual_dataset_3_input.py

@@ -62,13 +62,13 @@ def test_virtual_dataset_3_input():
             out = self.matmul2(out, b)
             return out
 
+    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
+    context.set_auto_parallel_context(device_num=8, global_rank=0)
     strategy0 = ((2, 1), (2, 1), (2, 1))
     strategy1 = ((2, 2), (2, 2))
     strategy2 = ((2, 2), (2, 2))
     strategy3 = ((2, 4),)
     net = GradWrap(NetWithLoss(Net(strategy0, strategy1, strategy2, strategy3)))
-    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
-    context.set_auto_parallel_context(device_num=8, global_rank=0)
     x = Tensor(np.ones([128, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
     b = Tensor(np.ones([64, 2048]), dtype=ms.float32)
@@ -89,10 +89,10 @@ def test_virtualdataset_cell_3_inputs():
             out = self.matmul2(out, b)
             return out
 
-    net = GradWrap(VirtualDatasetCellTriple(NetWithLoss(Net(None, None, None))))
     context.set_context(save_graphs=True)
     context.set_auto_parallel_context(parallel_mode="auto_parallel")
     context.set_auto_parallel_context(device_num=8, global_rank=0)
+    net = GradWrap(VirtualDatasetCellTriple(NetWithLoss(Net(None, None, None))))
     x = Tensor(np.ones([128, 32]), dtype=ms.float32)
     y = Tensor(np.ones([32, 64]), dtype=ms.float32)
     b = Tensor(np.ones([64, 2048]), dtype=ms.float32)

tests/ut/python/train/test_amp.py

@@ -146,6 +146,10 @@ def test_compile_model_train_O2():
 def test_compile_model_train_O2_parallel():
     dataset_types = (np.float32, np.float32)
     dataset_shapes = ((16, 16), (16, 16))
+    context.set_auto_parallel_context(
+        global_rank=0, device_num=8,
+        mirror_mean=True, parameter_broadcast=True,
+        parallel_mode=ParallelMode.DATA_PARALLEL)
 
     dataset = MindDataSet(dataset_types, dataset_shapes)
 
@@ -153,10 +157,6 @@ def test_compile_model_train_O2_parallel():
     loss = nn.MSELoss()
     optimizer = nn.Momentum(net.trainable_params(), 0.1, 0.9, 0.00004, 1024.0)
 
-    context.set_auto_parallel_context(
-        global_rank=0, device_num=8,
-        mirror_mean=True, parameter_broadcast=True,
-        parallel_mode=ParallelMode.DATA_PARALLEL)
     init()
 
     model = Model(net, loss_fn=loss, optimizer=optimizer, metrics={"acc"}, amp_level="O2")