[Auto Parallel] Remove some deprecated fluid APIs (#49099)

* [Auto Parallel] Remove some fluid APIs

* [Auto Parallel] Fix the wrong import

* [Auto Parallel] Remove unnecessary comments

* [Auto Parallel] Fix the importing bug

python/paddle/distributed/auto_parallel/cluster_v2.py

@@ -16,9 +16,7 @@ from enum import IntEnum, unique
 
 import numpy as np
 
-from paddle.fluid import core
-from paddle.fluid.core import Device  # noqa: F401
-from paddle.fluid.core import Link  # noqa: F401
+from paddle.framework import core
 
 
 @unique

python/paddle/distributed/auto_parallel/completion.py

@@ -16,7 +16,7 @@ import copy
 import logging
 
 from paddle.distributed.fleet.meta_optimizers.common import OpRole
-from paddle.fluid import core
+from paddle.framework import core
 
 from .dist_attribute import OperatorDistAttr, TensorDistAttr
 from .dist_context import _node_id

python/paddle/distributed/auto_parallel/cost/tensor_cost.py

@@ -16,7 +16,7 @@ from functools import reduce
 
 import paddle
 from paddle.distributed.auto_parallel.dist_tensor import DistributedTensor
-from paddle.fluid.framework import Variable
+from paddle.static import Variable
 
 from .base_cost import Cost
 

python/paddle/distributed/auto_parallel/cost_model.py

@@ -20,7 +20,7 @@ import numpy as np
 
 import paddle
 from paddle.distributed.fleet.meta_optimizers.common import OpRole
-from paddle.fluid import core
+from paddle.framework import core
 
 SUCC = 0  # successor
 PRED = 1  # predecessor

python/paddle/distributed/auto_parallel/dist_context.py

@@ -16,8 +16,7 @@ import copy
 from collections import defaultdict
 
 from paddle.distributed.passes import PassContext
-from paddle.fluid import core, framework
-from paddle.fluid.framework import set_flags
+from paddle.framework import IrGraph, core, set_flags
 
 from .dist_op import DistributedOperator
 from .dist_tensor import DistributedTensor
@@ -437,7 +436,7 @@ class DistributedContext:
 
             if with_graph:
                 set_flags({"FLAGS_convert_all_blocks": True})
-                self._serial_graph = framework.IrGraph(
+                self._serial_graph = IrGraph(
                     core.Graph(self._serial_main_program.desc)
                 )
                 self._init_dist_attr_for_graph()

python/paddle/distributed/auto_parallel/dist_op.py

@@ -15,7 +15,7 @@
 import copy
 
 import paddle
-from paddle.fluid.framework import Variable
+from paddle.static import Variable
 
 from .dist_attribute import OperatorDistAttr
 from .utils import (
@@ -303,7 +303,7 @@ class DistributedOperatorHelper:
                 tensor_to_dims_mapping[arg.name] = self._in_dims_mappings[index]
             index += 1
 
-        default_prog = paddle.fluid.default_main_program()
+        default_prog = paddle.static.default_main_program()
         cur_block = default_prog.current_block()
         op_size = len(cur_block.ops)
         output = self._serial_op(*args, **kwargs)

python/paddle/distributed/auto_parallel/dist_saver.py

@@ -21,8 +21,7 @@ import re
 import numpy as np
 
 import paddle
-from paddle import fluid
-from paddle.fluid import core
+from paddle.framework import core
 
 from ..utils.log_utils import get_logger
 from .process_group import _g_process_group_map
@@ -167,7 +166,7 @@ class DistributedSaver:
 
         dist_main_prog = kwargs.get('program', None)
         if not dist_main_prog:
-            dist_main_prog = fluid.default_main_program()
+            dist_main_prog = paddle.static.default_main_program()
         global_block = dist_main_prog.global_block()
 
         ops = global_block.ops

python/paddle/distributed/auto_parallel/dist_tensor.py

@@ -16,7 +16,8 @@ import copy
 import inspect
 
 import paddle
-from paddle.fluid.framework import Block, Parameter, Variable
+from paddle.framework import Block
+from paddle.static import Parameter, Variable
 
 from .dist_attribute import TensorDistAttr
 from .utils import __no_shape_var_type__, _linear_idx2coordinate

python/paddle/distributed/auto_parallel/engine.py

@@ -24,17 +24,16 @@ import numpy as np
 import paddle
 import paddle.distributed.auto_parallel.utils as auto_utils
 import paddle.utils as utils
-from paddle import fluid, static
+from paddle import static
 from paddle.distributed import fleet
-from paddle.fluid import Variable, core
 from paddle.fluid.dygraph.parallel import ParallelEnv
-from paddle.fluid.executor import _to_name_str, global_scope
-from paddle.fluid.framework import IrGraph, Operator
-from paddle.fluid.framework import _current_expected_place as _get_device
-from paddle.fluid.framework import in_dygraph_mode
+from paddle.fluid.executor import _to_name_str
 from paddle.fluid.layers.utils import flatten
+from paddle.framework import IrGraph
+from paddle.framework import _current_expected_place as _get_device
+from paddle.framework import core, in_dygraph_mode
 from paddle.metric import Metric
-from paddle.static import InputSpec
+from paddle.static import InputSpec, Operator, Variable, global_scope
 
 from ..utils.log_utils import get_logger
 from .callbacks import config_callbacks
@@ -151,11 +150,11 @@ class Engine:
 
         if optimizer and not isinstance(
             optimizer,
-            (paddle.optimizer.Optimizer, paddle.fluid.optimizer.Optimizer),
+            (paddle.optimizer.Optimizer, paddle.static.Optimizer),
         ):
             raise TypeError(
                 "'optimizer' must be object of class `paddle.optimizer.Optimizer`"
-                " or `paddle.fluid.optimizer.Optimizer`."
+                " or `paddle.static.Optimizer`."
             )
         self._optimizer = auto_utils.validate_opt(optimizer)
         self._orig_optimizer = copy.deepcopy(self._optimizer)
@@ -769,8 +768,8 @@ class Engine:
                     process_group.instantiate()
 
         self._place = _get_device()
-        if isinstance(self._place, fluid.CUDAPlace):
-            self._place = fluid.CUDAPlace(ParallelEnv().dev_id)
+        if isinstance(self._place, paddle.framework.CUDAPlace):
+            self._place = paddle.framework.CUDAPlace(ParallelEnv().dev_id)
 
         if self._strategy.seed:
             paddle.seed(self._strategy.seed + self._dp_ranks[0])

python/paddle/distributed/auto_parallel/helper.py

@@ -15,11 +15,10 @@
 import logging
 from collections import defaultdict
 
-from paddle.fluid.executor import global_scope
-from paddle.fluid.framework import Parameter, program_guard
 from paddle.jit import not_to_static, to_static
 from paddle.jit.dy2static.program_translator import StaticFunction
 from paddle.nn import Layer
+from paddle.static import Parameter, global_scope, program_guard
 
 from .converter import Converter
 from .utils import get_logger, to_list

python/paddle/distributed/auto_parallel/interface.py

@@ -207,7 +207,7 @@ def recompute(op):
             self._op = op
 
         def __call__(self, *args, **kwargs):
-            default_prog = paddle.fluid.default_main_program()
+            default_prog = paddle.static.default_main_program()
             cur_block = default_prog.current_block()
             op_size = len(cur_block.ops)
             output = self._op(*args, **kwargs)

python/paddle/distributed/auto_parallel/operators/dist_check_finite_and_unscale.py

@@ -16,7 +16,7 @@ from paddle.distributed.auto_parallel.process_group import (
     get_world_process_group,
 )
 from paddle.distributed.fleet.meta_optimizers.common import OP_ROLE_KEY, OpRole
-from paddle.fluid import core
+from paddle.framework import core
 
 from ..dist_attribute import OperatorDistAttr
 from ..process_group import new_process_group

python/paddle/distributed/auto_parallel/operators/dist_embedding.py

@@ -17,8 +17,9 @@ from paddle.distributed.auto_parallel.cost.comm_op_cost import (
     IdentityOpCost,
 )
 from paddle.distributed.fleet.meta_optimizers.common import OP_ROLE_KEY, OpRole
-from paddle.fluid import core, unique_name
 from paddle.fluid.data_feeder import check_dtype, check_variable_and_dtype
+from paddle.framework import core
+from paddle.utils import unique_name
 
 from ..cost import (
     EmbeddingGradOpCost,

python/paddle/distributed/auto_parallel/operators/dist_matmul.py

@@ -19,8 +19,9 @@ from paddle.distributed.auto_parallel.cost.comm_op_cost import (
     IdentityOpCost,
 )
 from paddle.distributed.fleet.meta_optimizers.common import OP_ROLE_KEY, OpRole
-from paddle.fluid import core, unique_name
 from paddle.fluid.data_feeder import check_dtype, check_variable_and_dtype
+from paddle.framework import core
+from paddle.utils import unique_name
 
 from ..cost import (
     MatmulGradOpCost,

python/paddle/distributed/auto_parallel/operators/dist_pnorm.py

@@ -14,9 +14,9 @@
 
 import copy
 
-from paddle.fluid import core
 from paddle.fluid.data_feeder import check_dtype, check_variable_and_dtype
-from paddle.fluid.framework import Operator
+from paddle.framework import core
+from paddle.static import Operator
 
 from ..dist_attribute import OperatorDistAttr, TensorDistAttr
 from ..process_group import new_process_group

python/paddle/distributed/auto_parallel/parallelizer.py

@@ -24,11 +24,10 @@ import sys
 import time
 
 import paddle
-import paddle.fluid.core as core
 from paddle.distributed.passes import PassContext, new_pass
 from paddle.distributed.utils.log_utils import get_logger
-from paddle.fluid import program_guard
-from paddle.fluid.backward import append_backward
+from paddle.framework import core
+from paddle.static import append_backward, program_guard
 
 from .cluster import Cluster
 from .completion import Completer

python/paddle/distributed/auto_parallel/parallelizer_v2.py

@@ -17,9 +17,8 @@ import logging
 import time
 
 from paddle.distributed.passes import new_pass
-from paddle.fluid import program_guard
-from paddle.fluid.backward import append_backward
-from paddle.fluid.framework import unique_name
+from paddle.static import append_backward, program_guard
+from paddle.utils import unique_name
 
 from ..utils.log_utils import get_logger
 from .partitioner import Partitioner

python/paddle/distributed/auto_parallel/partitioner.py

@@ -14,13 +14,13 @@
 
 import copy
 
-import paddle.fluid as fluid
+import paddle
 from paddle.distributed.auto_parallel.dist_context import DistributedContext
 from paddle.distributed.auto_parallel.operators.common import (
     get_distributed_operator_impl_container,
 )
-from paddle.fluid import core
-from paddle.fluid.framework import Parameter, Program
+from paddle.framework import Program, core
+from paddle.static import Parameter
 
 from .dist_attribute import OperatorDistAttr
 from .operators.common import BACKWARD_ONLY_DIST_OPS
@@ -52,12 +52,12 @@ class Partitioner:
     def __init__(self, dist_context, rank_id=0):
         """
         Args:
-            dist_context (paddle.fluid.DistributedContext): used to access the distributed_attr of var & op, every Partitioner object could maintain its own DistributedContext member, and partition program base on that shard scenario.
+            dist_context (DistributedContext): used to access the distributed_attr of var & op, every Partitioner object could maintain its own DistributedContext member, and partition program base on that shard scenario.
             rank_id (int): global rank id to which the partitioned distributed program belong.
         """
         if not isinstance(dist_context, DistributedContext):
             raise TypeError(
-                "dist_context be paddle.fluid.DistributedContext, got %s here"
+                "dist_context be DistributedContext, got %s here"
                 % type(dist_context)
             )
 
@@ -71,7 +71,7 @@ class Partitioner:
     ):
         if not isinstance(serial_main_program, (Program)):
             raise TypeError(
-                "main_program be paddle.fluid.framework.program, got %s here"
+                "main_program be paddle.framework.Program, got %s here"
                 % type(serial_main_program)
             )
 
@@ -113,11 +113,11 @@ class Partitioner:
 
         if not isinstance(serial_startup_program, (Program)):
             raise TypeError(
-                "dist_context be paddle.fluid.framework.program, got %s here"
+                "dist_context be paddle.framework.Program, got %s here"
                 % type(serial_startup_program)
             )
 
-        partitioned_startup_prog = fluid.Program()
+        partitioned_startup_prog = paddle.framework.Program()
         ref_block = serial_main_program.global_block()
         target_block = partitioned_startup_prog.global_block()
         var2shape = {}
@@ -183,7 +183,7 @@ class Partitioner:
         2. replace local op with corresponding dist op
         """
 
-        partitioned_main_prog = fluid.Program()
+        partitioned_main_prog = paddle.framework.Program()
         dist_op_context = self._dist_context.dist_op_context
         dist_op_context.dst_main_program = partitioned_main_prog
 

python/paddle/distributed/auto_parallel/process_group.py

@@ -15,9 +15,8 @@
 from collections import OrderedDict
 
 import paddle
-import paddle.fluid.core as core
 from paddle import _legacy_C_ops
-from paddle.fluid.framework import in_dygraph_mode
+from paddle.framework import core, in_dygraph_mode
 
 from ...fluid.layers.tensor import fill_constant
 from ..collective import _get_global_env, _new_ring_id

python/paddle/distributed/auto_parallel/process_mesh.py

@@ -157,7 +157,7 @@ class ProcessMesh(core.ProcessMesh):
 
     def __enter__(self):
         set_current_process_mesh(self)
-        default_prog = paddle.fluid.default_main_program()
+        default_prog = paddle.static.default_main_program()
         cur_block = default_prog.current_block()
         self._old_var_names = list(cur_block.vars.keys())
         self._old_op_size = len(cur_block.ops)
@@ -166,7 +166,7 @@ class ProcessMesh(core.ProcessMesh):
         from .dist_op import DistributedOperator
         from .dist_tensor import DistributedTensor
 
-        default_prog = paddle.fluid.default_main_program()
+        default_prog = paddle.static.default_main_program()
         cur_block = default_prog.current_block()
         new_var_names = list(cur_block.vars.keys())
         new_op_size = len(cur_block.ops)

python/paddle/distributed/auto_parallel/process_mesh_v2.py

@@ -14,7 +14,7 @@
 
 import numpy as np
 
-from paddle.fluid import core
+from paddle.framework import core
 
 
 class ProcessMesh(core.ProcessMesh):

python/paddle/distributed/auto_parallel/reshard.py

@@ -15,11 +15,9 @@
 from functools import reduce
 
 import paddle
-import paddle.fluid.core as core
 import paddle.fluid.layers.utils as utils
 from paddle.distributed.fleet.meta_optimizers.common import OpRole
-from paddle.fluid.framework import OpProtoHolder, Program
-from paddle.fluid.layer_helper import LayerHelper
+from paddle.framework import LayerHelper, OpProtoHolder, Program, core
 from paddle.utils import unique_name
 
 from .cost import (
@@ -310,7 +308,7 @@ class Inserter:
     @staticmethod
     def insert_cast_op(block, idx, tensor, op_role, tensor_type):
         # to avoid name conflict with framework
-        new_var_name = paddle.fluid.unique_name.generate_with_ignorable_key(
+        new_var_name = paddle.utils.unique_name.generate_with_ignorable_key(
             ".".join(["cast@RESHARD", 'tmp'])
         )
         out = block.create_var(
@@ -380,7 +378,7 @@ class Inserter:
     def insert_reset_lod_op(block, idx, X, Y, op_role):
         """Insert reset_lod op into block at the given index."""
 
-        new_var_name = paddle.fluid.unique_name.generate_with_ignorable_key(
+        new_var_name = paddle.utils.unique_name.generate_with_ignorable_key(
             ".".join(["reset_lod@RESHARD", 'tmp'])
         )
         reset_lod_out = block.create_var(
@@ -412,7 +410,7 @@ class Inserter:
         helper = LayerHelper('concat@RESHARD', **locals())
         with paddle.static.program_guard(block.program):
             out = block.create_var(
-                name=paddle.fluid.unique_name.generate_with_ignorable_key(
+                name=paddle.utils.unique_name.generate_with_ignorable_key(
                     ".".join([helper.name, 'tmp'])
                 ),
                 dtype=tensors[0].dtype,
@@ -484,7 +482,7 @@ class Inserter:
             with paddle.static.program_guard(block.program):
                 outs = [
                     block.create_var(
-                        name=paddle.fluid.unique_name.generate_with_ignorable_key(
+                        name=paddle.utils.unique_name.generate_with_ignorable_key(
                             ".".join(['split@RESHARD', 'tmp'])
                         ),
                         dtype=tensor.dtype,
@@ -550,7 +548,7 @@ class Inserter:
         with paddle.static.program_guard(block.program):
             outs = [
                 block.create_var(
-                    name=paddle.fluid.unique_name.generate_with_ignorable_key(
+                    name=paddle.utils.unique_name.generate_with_ignorable_key(
                         ".".join([helper.name, 'tmp'])
                     ),
                     dtype=tensor.dtype,
@@ -576,7 +574,7 @@ class Inserter:
         # use paddle.int64 as dtype
         with paddle.static.program_guard(block.program):
             out = block.create_var(
-                name=paddle.fluid.unique_name.generate_with_ignorable_key(
+                name=paddle.utils.unique_name.generate_with_ignorable_key(
                     ".".join([helper.name, 'tmp'])
                 ),
                 dtype=paddle.int64,
@@ -650,7 +648,7 @@ class Inserter:
         helper = LayerHelper(op_type + "@RESHARD", **locals())
         with paddle.static.program_guard(block.program):
             allgather_out = block.create_var(
-                name=paddle.fluid.unique_name.generate_with_ignorable_key(
+                name=paddle.utils.unique_name.generate_with_ignorable_key(
                     ".".join([helper.name, 'tmp'])
                 ),
                 dtype=tensor.dtype,
@@ -695,7 +693,7 @@ class Inserter:
         helper = LayerHelper(op_type + "@RESHARD", **locals())
         with paddle.static.program_guard(block.program):
             c_concat_out = block.create_var(
-                name=paddle.fluid.unique_name.generate_with_ignorable_key(
+                name=paddle.utils.unique_name.generate_with_ignorable_key(
                     ".".join([helper.name, 'tmp'])
                 ),
                 dtype=tensor.dtype,

python/paddle/distributed/auto_parallel/tuner/optimization_tuner.py

@@ -41,8 +41,8 @@ from paddle.distributed.auto_parallel.utils import (
     set_grad_var_shape,
 )
 from paddle.distributed.passes import PassContext, new_pass
-from paddle.fluid import program_guard, unique_name
-from paddle.fluid.backward import append_backward
+from paddle.static import append_backward, program_guard
+from paddle.utils import unique_name
 
 from ..utils import get_logger
 from .algorithms import new_algorithm

python/paddle/distributed/auto_parallel/tuner/profiler.py

@@ -28,7 +28,8 @@ from paddle.distributed.auto_parallel.process_group import (
     new_process_group,
 )
 from paddle.distributed.collective import _get_global_env
-from paddle.fluid.framework import Operator, Program, _current_expected_place
+from paddle.framework import Program, _current_expected_place
+from paddle.static import Operator
 
 paddle.enable_static()
 

python/paddle/distributed/auto_parallel/utils.py

@@ -22,9 +22,9 @@ from functools import reduce
 import numpy as np
 
 import paddle
-from paddle.fluid.framework import Variable
 from paddle.fluid.io import is_belong_to_optimizer, is_parameter
 from paddle.framework import core
+from paddle.static import Variable
 
 from .dist_attribute import OperatorDistAttr, TensorDistAttr
 from .process_group import get_all_process_groups
@@ -619,7 +619,7 @@ def save_distributed_checkpoint(
     """
     from .dist_context import get_default_distributed_context
 
-    assert isinstance(program, paddle.fluid.framework.Program)
+    assert isinstance(program, paddle.static.Program)
     assert isinstance(is_integrated, bool)
     if dist_context is None:
         dist_context = get_default_distributed_context()
@@ -702,7 +702,7 @@ def load_checkpoint_into_program(
     """
     from .dist_context import get_default_distributed_context
 
-    assert isinstance(program, paddle.fluid.framework.Program)
+    assert isinstance(program, paddle.static.Program)
     assert _check_valid_path(
         checkpoint_path
     ), "'checkpoint_path' cannot be None."
@@ -731,7 +731,7 @@ def load_parameter_into_program(param_dict, program):
         program(Program): the program to be updated
     """
     assert isinstance(param_dict, dict)
-    assert program and isinstance(program, paddle.fluid.framework.Program)
+    assert program and isinstance(program, paddle.static.Program)
     if not param_dict:
         return
     program.set_state_dict(param_dict)
@@ -818,7 +818,7 @@ def get_dist_attr(program, dist_context=None):
     """
     from .dist_context import get_default_distributed_context
 
-    assert isinstance(program, paddle.fluid.framework.Program)
+    assert isinstance(program, paddle.static.Program)
     if dist_context is None:
         dist_context = get_default_distributed_context()
     dist_attr = {}
@@ -1845,7 +1845,7 @@ def get_var_numel(var):
 def get_lr(optimizer):
     if isinstance(optimizer, paddle.optimizer.Optimizer):
         return optimizer.get_lr()
-    elif isinstance(optimizer, paddle.fluid.optimizer.Optimizer):
+    elif isinstance(optimizer, paddle.static.Optimizer):
         if isinstance(optimizer._learning_rate, float):
             return optimizer._learning_rate
         else:
@@ -1853,9 +1853,7 @@ def get_lr(optimizer):
     else:
         raise TypeError(
             "'optimizer' must be object of class `paddle.optimizer.Optimizer`"
-            " or `paddle.fluid.optimizer.Optimizer`, but got {}.".format(
-                type(optimizer)
-            )
+            " or `paddle.static.Optimizer`, but got {}.".format(type(optimizer))
         )
 
 

python/paddle/distributed/passes/auto_parallel_amp.py

@@ -23,7 +23,6 @@ from paddle.distributed.auto_parallel.utils import (
     set_var_dist_attr,
 )
 from paddle.distributed.fleet.meta_optimizers.common import OpRole
-from paddle.fluid import unique_name
 from paddle.fluid.contrib.mixed_precision.fp16_utils import (
     AutoMixedPrecisionLists,
     _dtype_to_str,
@@ -38,6 +37,7 @@ from paddle.fluid.contrib.mixed_precision.fp16_utils import (
 )
 from paddle.fluid.data_feeder import check_type, check_variable_and_dtype
 from paddle.framework import core
+from paddle.utils import unique_name
 
 from ..auto_parallel.process_mesh import ProcessMesh
 from ..auto_parallel.utils import is_backward_op, is_forward_op, is_loss_op
@@ -523,7 +523,7 @@ def _update_backward_cast_ops(params_grads, dist_context):
             # add new op in the python and cpp at the same time
             new_op_desc = main_block.desc.append_op()
             new_op_desc.copy_from(op.desc)
-            new_op = paddle.fluid.framework.Operator(
+            new_op = paddle.static.Operator(
                 block=main_block,
                 desc=new_op_desc,
                 type=None,
@@ -898,7 +898,7 @@ class AMPPass(PassBase):
                 OP_ROLE_KEY, core.op_proto_and_checker_maker.OpRole.Backward
             )
             elementwise_mul_grad_op_desc._set_attr('axis', -1)
-            elementwise_mul_grad_op = paddle.fluid.framework.Operator(
+            elementwise_mul_grad_op = paddle.static.Operator(
                 main_block, elementwise_mul_grad_op_desc
             )
             main_block.ops.insert(loss_op_idx + 3, elementwise_mul_grad_op)

python/paddle/distributed/passes/auto_parallel_data_parallel_optimization.py

@@ -29,9 +29,9 @@ from paddle.distributed.auto_parallel.utils import (
     ring_id_to_process_group,
 )
 from paddle.distributed.fleet.meta_optimizers.common import OP_ROLE_KEY, OpRole
-from paddle.fluid import unique_name
 from paddle.fluid.executor import _is_enable_standalone_executor
-from paddle.fluid.framework import default_main_program
+from paddle.static import default_main_program
+from paddle.utils import unique_name
 
 from .pass_base import PassBase, PassType, register_pass
 

python/paddle/distributed/passes/auto_parallel_fp16.py

@@ -26,7 +26,6 @@ from paddle.distributed.auto_parallel.utils import (
     set_var_dist_attr,
 )
 from paddle.distributed.fleet.meta_optimizers.common import OP_ROLE_KEY, OpRole
-from paddle.fluid import unique_name
 from paddle.fluid.contrib.mixed_precision.fp16_utils import (
     AutoMixedPrecisionLists,
     _dtype_to_str,
@@ -35,8 +34,9 @@ from paddle.fluid.contrib.mixed_precision.fp16_utils import (
     _valid_types,
 )
 from paddle.fluid.data_feeder import check_type, check_variable_and_dtype
-from paddle.fluid.framework import default_main_program, default_startup_program
 from paddle.framework import core
+from paddle.static import default_main_program, default_startup_program
+from paddle.utils import unique_name
 
 from ..auto_parallel.process_mesh import ProcessMesh
 from .auto_parallel_amp import AMPPass
@@ -790,7 +790,7 @@ class FP16Pass(AMPPass):
 
                         # all_infs = paddle.fluid.layers.concat(found_infs)
                         all_infs = block.create_var(
-                            name=paddle.fluid.unique_name.generate_with_ignorable_key(
+                            name=paddle.utils.unique_name.generate_with_ignorable_key(
                                 ".".join(['concat', 'tmp'])
                             ),
                             dtype=found_infs[0].dtype,
@@ -821,7 +821,7 @@ class FP16Pass(AMPPass):
 
                         # found_inf = paddle.fluid.layers.reduce_any(all_infs)
                         found_inf = block.create_var(
-                            name=paddle.fluid.unique_name.generate_with_ignorable_key(
+                            name=paddle.utils.unique_name.generate_with_ignorable_key(
                                 ".".join(['reduce_any', 'tmp'])
                             ),
                             dtype=all_infs.dtype,
@@ -867,7 +867,8 @@ class FP16Pass(AMPPass):
             if self.get_attr("use_optimizer_fp16"):
                 base_opt._multi_precision = False
             if isinstance(
-                base_opt, (paddle.fluid.optimizer.Adam, paddle.optimizer.AdamW)
+                base_opt,
+                (paddle.static.Adam, paddle.optimizer.AdamW),
             ):
                 with main_program._optimized_guard([]):
                     # found_inf = paddle.tensor.creation._memcpy(

python/paddle/distributed/passes/auto_parallel_gradient_merge.py

@@ -26,8 +26,8 @@ from paddle.distributed.auto_parallel.utils import (
 )
 from paddle.distributed.fleet.meta_optimizers.common import OP_ROLE_KEY, OpRole
 from paddle.fluid import layers
-from paddle.fluid.framework import device_guard
 from paddle.framework import core
+from paddle.static import device_guard
 
 from .pass_base import PassBase, PassType, register_pass
 

python/paddle/distributed/passes/auto_parallel_quantization.py

@@ -17,8 +17,8 @@ import logging
 import numpy as np
 
 import paddle
-from paddle.fluid import core, framework
 from paddle.fluid.dygraph.parallel import ParallelEnv
+from paddle.framework import IrGraph, core
 from paddle.static.quantization import (
     AddQuantDequantForInferencePass,
     AddQuantDequantPassV2,
@@ -72,7 +72,7 @@ class QuantizationPass(PassBase):
         # TODO: scope and place will be removed,
         # cause params should be initialized by engine module.
         scope = paddle.static.global_scope()
-        place = paddle.fluid.CUDAPlace(ParallelEnv().dev_id)
+        place = paddle.framework.CUDAPlace(ParallelEnv().dev_id)
 
         # 0. record the relation among blocks
         parent_idx_dict = dict()
@@ -81,7 +81,7 @@ class QuantizationPass(PassBase):
 
         is_test = True if mode != "train" else False
         # 1. Program convert to Graph, and this pass is only for train mode
-        main_graph = framework.IrGraph(
+        main_graph = IrGraph(
             core.Graph(main_program.desc), for_test=mode != "train"
         )
 

python/paddle/distributed/passes/auto_parallel_recompute.py

@@ -14,8 +14,8 @@
 
 import logging
 
+import paddle
 from paddle.distributed.fleet.meta_optimizers.common import OP_ROLE_KEY, OpRole
-from paddle.fluid import core, framework, unique_name
 from paddle.fluid.backward import (
     ProgramStats,
     _append_grad_suffix_,
@@ -23,6 +23,8 @@ from paddle.fluid.backward import (
     _get_no_grad_set_name,
     _rename_arg_,
 )
+from paddle.framework import core
+from paddle.utils import unique_name
 
 from ..auto_parallel.dist_attribute import OperatorDistAttr
 from ..auto_parallel.utils import (
@@ -221,7 +223,8 @@ def _add_needed_descs_to_block(
 
     result_descs = []
     for desc in descs:
-        if isinstance(desc, framework.Operator):
+        # if isinstance(desc, framework.Operator):
+        if isinstance(desc, paddle.static.Operator):
             desc = desc.desc
         if isinstance(desc, tuple):
             desc = desc[0]

python/paddle/distributed/passes/auto_parallel_sharding.py

@@ -35,10 +35,10 @@ from paddle.distributed.auto_parallel.utils import (
     set_var_dist_attr,
 )
 from paddle.distributed.fleet.meta_optimizers.sharding.utils import get_var_size
-from paddle.fluid import unique_name
 from paddle.fluid.executor import _is_enable_standalone_executor
-from paddle.fluid.framework import default_main_program, default_startup_program
 from paddle.framework import core
+from paddle.static import default_main_program, default_startup_program
+from paddle.utils import unique_name
 
 from .pass_base import PassBase, register_pass
 

python/paddle/fluid/tests/unittests/auto_parallel/test_cluster_v2.py

@@ -14,7 +14,8 @@
 
 import unittest
 
-from paddle.distributed.auto_parallel.cluster_v2 import Device, DeviceMesh, Link
+from paddle.distributed.auto_parallel.cluster_v2 import DeviceMesh
+from paddle.framework import core
 
 
 class TestDeviceMesh(unittest.TestCase):
@@ -38,12 +39,12 @@ class TestDeviceMesh(unittest.TestCase):
         self.assertEqual(device_mesh.contains(0), True)
         self.assertEqual(device_mesh.contains(6), False)
 
-        dev0 = Device(global_id=0, local_id=0, machine_id=0, type="GPU")
-        dev1 = Device(global_id=1, local_id=1, machine_id=0, type="GPU")
-        dev2 = Device(global_id=2, local_id=2, machine_id=0, type="GPU")
-        dev3 = Device(global_id=3, local_id=0, machine_id=1, type="GPU")
-        dev4 = Device(global_id=4, local_id=1, machine_id=1, type="GPU")
-        dev5 = Device(global_id=5, local_id=2, machine_id=1, type="GPU")
+        dev0 = core.Device(global_id=0, local_id=0, machine_id=0, type="GPU")
+        dev1 = core.Device(global_id=1, local_id=1, machine_id=0, type="GPU")
+        dev2 = core.Device(global_id=2, local_id=2, machine_id=0, type="GPU")
+        dev3 = core.Device(global_id=3, local_id=0, machine_id=1, type="GPU")
+        dev4 = core.Device(global_id=4, local_id=1, machine_id=1, type="GPU")
+        dev5 = core.Device(global_id=5, local_id=2, machine_id=1, type="GPU")
         device_mesh.add_device(dev0)
         device_mesh.add_device(dev1)
         device_mesh.add_device(dev2)
@@ -57,10 +58,10 @@ class TestDeviceMesh(unittest.TestCase):
         self.assertEqual(device_mesh.device(4), dev4)
         self.assertEqual(device_mesh.device(5), dev5)
 
-        link0 = Link(source_id=0, target_id=1, type="NVL")
-        link1 = Link(source_id=1, target_id=0, type="NVL")
-        link2 = Link(source_id=3, target_id=4, type="NVL")
-        link3 = Link(source_id=4, target_id=5, type="NVL")
+        link0 = core.Link(source_id=0, target_id=1, type="NVL")
+        link1 = core.Link(source_id=1, target_id=0, type="NVL")
+        link2 = core.Link(source_id=3, target_id=4, type="NVL")
+        link3 = core.Link(source_id=4, target_id=5, type="NVL")
         device_mesh.add_link(link0)
         device_mesh.add_link(link1)
         device_mesh.add_link(link2)
@@ -90,7 +91,7 @@ class TestDeviceMesh(unittest.TestCase):
         self.assertEqual(str(device_mesh), str(device_mesh))
 
     def test_device(self):
-        device = Device(global_id=0, local_id=1, machine_id=2, type="GPU")
+        device = core.Device(global_id=0, local_id=1, machine_id=2, type="GPU")
         device.capability.sflops = 100
         device.capability.dflops = 200
         device.capability.memory = 32
@@ -107,7 +108,7 @@ class TestDeviceMesh(unittest.TestCase):
         self.assertEqual(str(device), str(device))
 
     def test_link(self):
-        link = Link(source_id=0, target_id=1, type="NVL")
+        link = core.Link(source_id=0, target_id=1, type="NVL")
         link.capability.bandwidth = 100
         link.capability.latency = 1
         self.assertEqual(link.source_id, 0)

python/paddle/framework/__init__.py

@@ -69,6 +69,7 @@ from ..fluid.framework import _apply_pass  # noqa: F401
 from ..fluid.framework import switch_main_program
 from ..fluid.framework import _set_expected_place  # noqa: F401
 from ..fluid.framework import Block, Program  # noqa: F401
+from ..fluid.framework import IrGraph  # noqa: F401
 from ..fluid.dygraph import parallel_helper  # noqa: F401
 from ..fluid.dygraph.parallel import (
     _split_tensors,

python/paddle/static/__init__.py

@@ -56,11 +56,15 @@ from ..fluid.framework import xpu_places  # noqa: F401
 from ..fluid.framework import mlu_places  # noqa: F401
 from ..fluid.framework import npu_places  # noqa: F401
 from ..fluid.framework import Variable  # noqa: F401
+from ..fluid.framework import Operator  # noqa: F401
+from ..fluid.framework import Parameter  # noqa: F401
 from ..fluid.framework import ipu_shard_guard  # noqa: F401
 from ..fluid.framework import set_ipu_shard  # noqa: F401
 from ..fluid.layers.control_flow import Print  # noqa: F401
 from ..fluid.parallel_executor import ParallelExecutor  # noqa: F401
 from ..fluid.param_attr import WeightNormParamAttr  # noqa: F401
+from ..fluid.optimizer import Optimizer  # noqa: F401
+from ..fluid.optimizer import Adam  # noqa: F401
 from ..fluid.optimizer import ExponentialMovingAverage  # noqa: F401
 from ..fluid.io import save  # noqa: F401
 from ..fluid.io import load  # noqa: F401

python/paddle/utils/unique_name.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 
 from ..fluid.unique_name import generate  # noqa: F401
+from ..fluid.unique_name import generate_with_ignorable_key  # noqa: F401
 from ..fluid.unique_name import guard  # noqa: F401
 from ..fluid.unique_name import switch  # noqa: F401