PaddlePaddle
diff --git a/‎python/paddle/base/framework.py
Copy file name to clipboard
+17Lines changed: 17 additions & 0 deletions b/‎python/paddle/base/framework.py
Copy file name to clipboard
+17Lines changed: 17 additions & 0 deletions
diff --git a/‎python/paddle/base/layer_helper.py
Copy file name to clipboard
+6-1Lines changed: 6 additions & 1 deletion b/‎python/paddle/base/layer_helper.py
Copy file name to clipboard
+6-1Lines changed: 6 additions & 1 deletion
diff --git a/‎python/paddle/base/layer_helper_base.py
Copy file name to clipboard
+16-11Lines changed: 16 additions & 11 deletions b/‎python/paddle/base/layer_helper_base.py
Copy file name to clipboard
+16-11Lines changed: 16 additions & 11 deletions
diff --git a/‎python/paddle/base/layers/math_op_patch.py
Copy file name to clipboard
+3-3Lines changed: 3 additions & 3 deletions b/‎python/paddle/base/layers/math_op_patch.py
Copy file name to clipboard
+3-3Lines changed: 3 additions & 3 deletions
diff --git a/‎python/paddle/base/unique_name.py
Copy file name to clipboard
+17Lines changed: 17 additions & 0 deletions b/‎python/paddle/base/unique_name.py
Copy file name to clipboard
+17Lines changed: 17 additions & 0 deletions
diff --git a/‎python/paddle/distributed/auto_parallel/static/parallelizer_v2.py
Copy file name to clipboard
+1-2Lines changed: 1 addition & 2 deletions b/‎python/paddle/distributed/auto_parallel/static/parallelizer_v2.py
Copy file name to clipboard
+1-2Lines changed: 1 addition & 2 deletions
diff --git a/‎python/paddle/distributed/auto_parallel/static/tuner/optimization_tuner.py
Copy file name to clipboard
+1-2Lines changed: 1 addition & 2 deletions b/‎python/paddle/distributed/auto_parallel/static/tuner/optimization_tuner.py
Copy file name to clipboard
+1-2Lines changed: 1 addition & 2 deletions
diff --git a/‎python/paddle/distributed/auto_parallel/static/tuner/rule_based_tuner.py
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions b/‎python/paddle/distributed/auto_parallel/static/tuner/rule_based_tuner.py
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions
diff --git a/‎python/paddle/jit/dy2static/partial_program.py
Copy file name to clipboard
+2-45Lines changed: 2 additions & 45 deletions b/‎python/paddle/jit/dy2static/partial_program.py
Copy file name to clipboard
+2-45Lines changed: 2 additions & 45 deletions
@@ -6390,8 +6390,25 @@ def clone(self, for_test=False):
         p._copy_data_info_from(self, pruned_origin_block_id_map)
         p._copy_dist_param_info_from(self)
         p._copy_operator_info_from(self)
+        p._name_generator = self._name_generator.clone()
         return p
 
+    @signature_safe_contextmanager
+    def switch_name_generator_guard(self, new_generator):
+        if isinstance(new_generator, str):
+            new_generator = unique_name.UniqueNameGenerator(new_generator)
+        elif isinstance(new_generator, bytes):
+            new_generator = unique_name.UniqueNameGenerator(
+                new_generator.decode()
+            )
+
+        old_generator = self._name_generator
+        self._name_generator = new_generator
+        try:
+            yield
+        finally:
+            self._name_generator = old_generator
+
     def _prune(self, targets):
         """
         Prune operators and variables which are not needed to generate
 
@@ -36,7 +36,12 @@ def __init__(self, layer_type, **kwargs):
         # can not use both `layer_type` and `name`. Deprecate LayerHelper
         # and write a Helper for dygraph mode.
         if name is None:
-            self.kwargs['name'] = unique_name.generate(layer_type)
+            if in_dygraph_mode():
+                self.kwargs['name'] = unique_name.generate(layer_type)
+            else:
+                self.kwargs[
+                    'name'
+                ] = self.main_program._name_generator.generate(layer_type)
 
         super().__init__(self.kwargs['name'], layer_type=layer_type)
 
 
@@ -120,14 +120,14 @@ def __norm_op(
         ):
             if out is None:
                 out = block.create_var(
-                    name=unique_name.generate_with_ignorable_key(
+                    name=self.main_program._name_generator.generate_with_ignorable_key(
                         ".".join([self.name, 'weight_norm_norm'])
                     ),
                     dtype=dtype,
                     persistable=False,
                 )
             abs_out = block.create_var(
-                name=unique_name.generate_with_ignorable_key(
+                name=self.main_program._name_generator.generate_with_ignorable_key(
                     ".".join([self.name, 'weight_norm_abs'])
                 ),
                 dtype=dtype,
@@ -137,7 +137,7 @@ def __norm_op(
                 type='abs', inputs={'X': x}, outputs={'Out': abs_out}
             )
             pow_out = block.create_var(
-                name=unique_name.generate_with_ignorable_key(
+                name=self.main_program._name_generator.generate_with_ignorable_key(
                     ".".join([self.name, 'weight_norm_pow'])
                 ),
                 dtype=dtype,
@@ -150,7 +150,7 @@ def __norm_op(
                 attrs={'factor': float(p)},
             )
             sum_out = block.create_var(
-                name=unique_name.generate_with_ignorable_key(
+                name=self.main_program._name_generator.generate_with_ignorable_key(
                     ".".join([self.name, 'weight_norm_sum'])
                 ),
                 dtype=dtype,
@@ -179,7 +179,7 @@ def __reshape_op(
         ):
             if out is None:
                 out = block.create_var(
-                    name=unique_name.generate_with_ignorable_key(
+                    name=self.main_program._name_generator.generate_with_ignorable_key(
                         ".".join([self.name, 'weight_norm_reshape'])
                     ),
                     dtype=dtype,
@@ -199,7 +199,7 @@ def __transpose_op(
         ):
             if out is None:
                 out = block.create_var(
-                    name=unique_name.generate_with_ignorable_key(
+                    name=self.main_program._name_generator.generate_with_ignorable_key(
                         ".".join([self.name, 'weight_norm_transpose'])
                     ),
                     dtype=dtype,
@@ -219,7 +219,7 @@ def __norm_except_dim(
             """Computes the norm over all dimensions except dim"""
             if out is None:
                 out = block.create_var(
-                    name=unique_name.generate_with_ignorable_key(
+                    name=self.main_program._name_generator.generate_with_ignorable_key(
                         ".".join([self.name, 'weight_norm_norm'])
                     ),
                     dtype=dtype,
@@ -376,7 +376,12 @@ def create_parameter(
                 else default_initializer
             )
         if attr.name is None:
-            attr.name = unique_name.generate(".".join([self.name, suffix]))
+            if in_dygraph_mode():
+                attr.name = unique_name.generate(".".join([self.name, suffix]))
+            else:
+                attr.name = self.main_program._name_generator.generate(
+                    ".".join([self.name, suffix])
+                )
 
         if default_initializer is None and attr.initializer is None:
             if isinstance(dtype, core.VarDesc.VarType):
@@ -466,7 +471,7 @@ def create_variable_for_type_inference(
         if not dtype:
             dtype = self.__dtype
         return self.main_program.current_block().create_var(
-            name=unique_name.generate_with_ignorable_key(
+            name=self.main_program._name_generator.generate_with_ignorable_key(
                 ".".join([self.name, 'tmp'])
             ),
             dtype=dtype,
@@ -491,7 +496,7 @@ def _create_global_variable_for_type_inference(
         if not dtype:
             dtype = self.__dtype
         output = self.main_program.global_block().create_var(
-            name=unique_name.generate_with_ignorable_key(
+            name=self.main_program._name_generator.generate_with_ignorable_key(
                 ".".join([self.name, 'tmp'])
             ),
             dtype=dtype,
@@ -524,7 +529,7 @@ def create_sparse_variable_for_type_inference(
         if not dtype:
             dtype = self.__dtype
         return self.main_program.current_block().create_var(
-            name=unique_name.generate_with_ignorable_key(
+            name=self.main_program._name_generator.generate_with_ignorable_key(
                 ".".join([self.name, 'tmp'])
             ),
             dtype=dtype,
 
@@ -17,8 +17,8 @@
 
 from paddle.base.dygraph.base import in_to_static_mode
 
-from .. import core
-from ..framework import Variable, static_only, unique_name
+from .. import core, default_main_program
+from ..framework import Variable, static_only
 from .layer_function_generator import OpProtoHolder
 
 _supported_int_dtype_ = [
@@ -70,7 +70,7 @@
 
 def monkey_patch_variable():
     def unique_tmp_name():
-        return unique_name.generate("tmp")
+        return default_main_program()._name_generator.generate("tmp")
 
     def safe_get_dtype(var):
         try:
 
@@ -13,6 +13,7 @@
 # limitations under the License.
 
 import collections
+from copy import deepcopy
 
 from .wrapped_decorator import signature_safe_contextmanager
 
@@ -35,6 +36,9 @@ def __init__(self, prefix=None):
         self.prefix = prefix
 
     def __call__(self, key):
+        return self.generate(key)
+
+    def generate(self, key):
         """
         Generate unique names with prefix
 
@@ -47,6 +51,19 @@ def __call__(self, key):
         self.ids[key] += 1
         return self.prefix + "_".join([key, str(tmp)])
 
+    def generate_with_ignorable_key(self, key):
+        from .framework import _dygraph_tracer, in_dygraph_mode
+
+        if in_dygraph_mode():
+            return _dygraph_tracer()._generate_unique_name()
+
+        return self.generate(key)
+
+    def clone(self):
+        ret = UniqueNameGenerator(self.prefix)
+        ret.ids = deepcopy(self.ids)
+        return ret
+
 
 class DygraphParameterNameChecker:
     """
 
@@ -20,7 +20,6 @@
 from paddle.distributed.passes import PassManager, new_pass
 from paddle.framework import get_flags
 from paddle.static import append_backward, program_guard
-from paddle.utils import unique_name
 
 from ...utils.log_utils import get_logger
 from ..random import init_auto_parallel_rng
@@ -257,7 +256,7 @@ def _generate_optimizer(
         optimizer._sorted = False
 
         with program_guard(main_program, startup_program):
-            with unique_name.guard("opt_"):
+            with main_program.switch_name_generator_guard("opt_"):
                 optimizer_ops = optimizer.apply_gradients(params_grads)
         self._completer.complete_update_annotation(main_program)
         return optimizer_ops
 
@@ -41,7 +41,6 @@
 from paddle.distributed.auto_parallel.static.utils import debug_program
 from paddle.distributed.passes import PassContext, new_pass
 from paddle.static import append_backward, program_guard
-from paddle.utils import unique_name
 
 from ..utils import get_logger
 from .algorithms import new_algorithm
@@ -344,7 +343,7 @@ def _apply_optimization(self, trial):
         # Generate optimizer
         # FIXME should be remove from apply pass after pass support optimizers
         with program_guard(dist_main_prog, dist_startup_prog):
-            with unique_name.guard("opt_"):
+            with dist_main_prog.switch_name_generator_guard("opt_"):
                 optimizer_ops = dist_context.serial_optimizer.apply_gradients(
                     dist_params_grads
                 )
 
@@ -28,7 +28,7 @@
 import paddle
 from paddle.base import program_guard
 from paddle.base.backward import append_backward
-from paddle.base.framework import Parameter, unique_name
+from paddle.base.framework import Parameter
 from paddle.distributed.auto_parallel.process_mesh import ProcessMesh
 from paddle.distributed.auto_parallel.static.cluster_v2 import DeviceMesh
 from paddle.distributed.auto_parallel.static.completion import Completer
@@ -1162,7 +1162,7 @@ def gen_full_program(self):
             with program_guard(
                 self.full_main_program, self.full_startup_program
             ):
-                with unique_name.guard("opt_"):
+                with self.full_main_program.switch_name_generator_guard("opt_"):
                     optimizer_ops = optimizer.apply_gradients(params_grads)
 
             # op original id to grad op id
 
@@ -25,7 +25,6 @@
 from paddle.base.data_feeder import check_type, convert_dtype
 from paddle.base.dygraph.base import switch_to_static_graph
 from paddle.base.framework import _apply_pass, get_flags
-from paddle.base.unique_name import switch
 from paddle.optimizer.lr import LRScheduler
 
 from . import logging_utils
@@ -172,19 +171,12 @@ class PartialProgramLayer:
     """
 
     def __init__(
-        self,
-        main_program,
-        inputs,
-        outputs,
-        name_generator,
-        parameters=None,
-        **kwargs
+        self, main_program, inputs, outputs, parameters=None, **kwargs
     ):
         super().__init__()
         self._inputs = NestSequence(inputs)
         self._outputs = NestSequence(outputs, need_check=True)
         self._params = parameters if parameters is not None else []
-        self._name_generator = name_generator
 
         self._build_strategy = kwargs.get('build_strategy', BuildStrategy())
         assert isinstance(self._build_strategy, BuildStrategy)
@@ -234,8 +226,6 @@ def __call__(self, inputs):
         """
         Execute static graph by Interpreter and Return dynamic Tensors.
         """
-        old_generator, old_para_name_checker = switch(self._name_generator)
-
         in_vars, in_var_names = self._prepare_inputs(inputs)
         out_vars = self._prepare_outputs()
         self._cast_fp16_if_pure_fp16(in_vars)
@@ -261,16 +251,12 @@ def __call__(self, inputs):
         restored_nest_out = self._restore_out(out_vars)
         restored_nest_out = self._remove_no_value(restored_nest_out)
 
-        switch(old_generator, old_para_name_checker)
         return restored_nest_out
 
     def sot_call(self, inputs):
         """
-        Same as __call__, but set force_not_use_pt to False.
-        Currently _sot_call will cause CUDA 700 error, so we disable it temporarily.
+        In sot, inputs and outputs of partial program only contain tensors, so we can skip some step to speed up
         """
-        old_generator, old_para_name_checker = switch(self._name_generator)
-
         in_vars, in_var_names = self._prepare_inputs(inputs)
         out_vars = self._prepare_outputs()
         self._cast_fp16_if_pure_fp16(in_vars)
@@ -292,36 +278,8 @@ def sot_call(self, inputs):
 
         restored_nest_out = self._restore_out(out_vars)
         restored_nest_out = self._remove_no_value(restored_nest_out)
-
-        switch(old_generator, old_para_name_checker)
         return restored_nest_out
 
-    def _sot_call(self, inputs):
-        """
-        In sot, inputs and outputs of partial program only contain tensors, so we can skip some step to speed up
-        """
-        old_generator, old_para_name_checker = switch(self._name_generator)
-
-        out_vars = self._prepare_outputs()
-        self._cast_fp16_if_pure_fp16(inputs)
-        attrs = self._prepare_attributes()
-        attrs.extend(["x_names", self._in_var_names])
-        self._sync_lr_value_with_scheduler()
-
-        _legacy_C_ops.run_program(
-            self._valid_vars(inputs),
-            self._valid_vars(self._params),
-            self._valid_vars(out_vars),
-            self._create_scope_vec(
-                program_id=self.program_id, use_scope_cache=True
-            ),
-            self._cuda_graph_vec,
-            *attrs
-        )
-
-        switch(old_generator, old_para_name_checker)
-        return out_vars
-
     def _sync_lr_value_with_scheduler(self):
         """Update lr_var value with calculated by lr_scheduler."""
         main_program = self._origin_main_program
@@ -1207,7 +1165,6 @@ def partial_program_from(concrete_program, from_method=False):
         concrete_program.main_program,
         inputs,
         concrete_program.outputs,
-        concrete_program.name_generator,
         concrete_program.parameters,
         **concrete_program.kwargs
     )