[remove fluid] under fleet meta_optimizers (#47864)

* [remove fluid] under fleet meta_optimizers

* [remove fluid] under fleet meta_optimizers

* [remove fluid] under fleet meta_optimizers

* [remove fluid] under fleet meta_optimizers

* [remove fluid] under fleet meta_optimizers

* [remove fluid] under fleet meta_optimizers

* [remove fluid] under fleet meta_optimizers

* [remove fluid] under fleet meta_optimizers

* [remove fluid] under fleet meta_optimizers

* [remove fluid] under fleet meta_optimizers

* [remove fluid] under fleet meta_optimizers

* [remove fluid] under fleet meta_optimizers

python/paddle/distributed/__init__.py

@@ -12,9 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from . import io
 from .spawn import spawn  # noqa: F401
 from .launch.main import launch  # noqa: F401
-
 from .parallel import init_parallel_env  # noqa: F401
 from .parallel import get_rank  # noqa: F401
 from .parallel import get_world_size  # noqa: F401
@@ -74,6 +74,7 @@
 from . import rpc
 
 __all__ = [  # noqa
+    "io",
     "spawn",
     "launch",
     "scatter",

python/paddle/distributed/fleet/meta_optimizers/ascend/ascend_optimizer.py

@@ -12,8 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from paddle.fluid.optimizer import Optimizer
-import paddle.fluid.core as core
+from paddle.optimizer import Optimizer
+import paddle.framework.core as core
 from . import ascend_parser
 from paddle.distributed import fleet
 import hccl.manage.api as hccl

python/paddle/distributed/fleet/meta_optimizers/ascend/ascend_parser.py

@@ -11,7 +11,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-import paddle.fluid.core as core
+import paddle.framework.core as core
 import numpy as np
 from functools import reduce
 

python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/heter_parallel_optimizer.py

@@ -12,8 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from paddle.fluid.dygraph import base as imperative_base
-from paddle.fluid import framework
+import paddle.autograd as imperative_base
+from paddle import framework
 
 __all__ = []
 
@@ -41,13 +41,13 @@ def __init__(self, optimizer, strategy):
         # NOTE(liubo48): In pure DataParallel mode,
         # the gradient synchronization is achieved through reducer.
 
-    @imperative_base.no_grad
+    @imperative_base.no_grad()
     @framework.dygraph_only
     def step(self):
         parameters_list = _obtain_optimizer_parameters_list(self._inner_opt)
         self._inner_opt.step()
 
-    @imperative_base.no_grad
+    @imperative_base.no_grad()
     def minimize(
         self, loss, startup_program=None, parameters=None, no_grad_set=None
     ):

python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/hybrid_parallel_gradscaler.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from ...base.topology import ParallelMode
-from paddle.fluid.dygraph import base as imperative_base
+import paddle.autograd as imperative_base
 import paddle
 from paddle import _legacy_C_ops
 
@@ -51,7 +51,7 @@ def minimize(self, optimizer, *args, **kwargs):
 
         return optimize_ops, params_grads
 
-    @imperative_base.no_grad
+    @imperative_base.no_grad()
     def _unscale(self, optimizer):
         if not self._enable:
             return

python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/hybrid_parallel_optimizer.py

@@ -19,10 +19,10 @@
     sharding_reduce_gradients,
 )
 from ...base.topology import ParallelMode
-from paddle.fluid.dygraph import base as imperative_base
-from paddle.fluid import framework
+from paddle.autograd import no_grad
+from paddle import framework
 from ...utils.log_util import logger
-from paddle.fluid import core
+from paddle.framework import core
 from paddle.fluid import layers
 
 __all__ = []
@@ -47,7 +47,7 @@ def __init__(self, clip, hcg):
         self._clip = clip
         self._hcg = hcg
 
-    @imperative_base.no_grad
+    @no_grad()
     def _dygraph_clip(self, params_grads):
         sum_square_dist_fp16 = []
         sum_square_dist_fp32 = []
@@ -63,8 +63,8 @@ def _dygraph_clip(self, params_grads):
             if g.type == core.VarDesc.VarType.SELECTED_ROWS:
                 merge_grad = layers.merge_selected_rows(g)
                 merge_grad = layers.get_tensor_from_selected_rows(merge_grad)
-            square = layers.square(merge_grad)
-            sum_square = layers.reduce_sum(square)
+            square = paddle.square(merge_grad)
+            sum_square = paddle.sum(square)
 
             not_shared_enable = (not hasattr(p, 'is_firstly_shared')) or (
                 hasattr(p, 'is_firstly_shared')
@@ -89,8 +89,8 @@ def _dygraph_clip(self, params_grads):
                 [0.0], dtype=paddle.float32
             )
         else:
-            global_norm_dist_fp16 = layers.concat(sum_square_dist_fp16)
-            global_norm_dist_fp16 = layers.reduce_sum(global_norm_dist_fp16)
+            global_norm_dist_fp16 = paddle.concat(sum_square_dist_fp16)
+            global_norm_dist_fp16 = paddle.sum(global_norm_dist_fp16)
             global_norm_dist_fp16 = paddle.cast(
                 global_norm_dist_fp16, dtype=paddle.float32
             )
@@ -101,29 +101,27 @@ def _dygraph_clip(self, params_grads):
                 [0.0], dtype=paddle.float32
             )
         else:
-            global_norm_not_dist_fp16 = layers.concat(sum_square_not_dist_fp16)
-            global_norm_not_dist_fp16 = layers.reduce_sum(
-                global_norm_not_dist_fp16
-            )
+            global_norm_not_dist_fp16 = paddle.concat(sum_square_not_dist_fp16)
+            global_norm_not_dist_fp16 = paddle.sum(global_norm_not_dist_fp16)
             global_norm_not_dist_fp16 = paddle.cast(
                 global_norm_not_dist_fp16, dtype=paddle.float32
             )
 
         # global norm of distributed FP32 params_and_grads
         global_norm_dist_fp32 = (
-            layers.concat(sum_square_dist_fp32)
+            paddle.concat(sum_square_dist_fp32)
             if len(sum_square_dist_fp32) != 0
             else paddle.to_tensor([0.0], dtype=paddle.float32)
         )
-        global_norm_dist_fp32 = layers.reduce_sum(global_norm_dist_fp32)
+        global_norm_dist_fp32 = paddle.sum(global_norm_dist_fp32)
 
         # global norm of non-distributed FP32 params_and_grads
         global_norm_not_dist_fp32 = (
-            layers.concat(sum_square_not_dist_fp32)
+            paddle.concat(sum_square_not_dist_fp32)
             if len(sum_square_not_dist_fp32) != 0
             else paddle.to_tensor([0.0], dtype=paddle.float32)
         )
-        global_norm_not_dist_fp32 = layers.reduce_sum(global_norm_not_dist_fp32)
+        global_norm_not_dist_fp32 = paddle.sum(global_norm_not_dist_fp32)
 
         global_norm_var_dist = global_norm_dist_fp16 + global_norm_dist_fp32
         global_norm_var_not_dist = (
@@ -151,14 +149,16 @@ def _dygraph_clip(self, params_grads):
                 group=self._hcg.get_sharding_parallel_group(),
             )
 
-        global_norm_var_fp32 = layers.sqrt(
+        global_norm_var_fp32 = paddle.sqrt(
             global_norm_var_dist + global_norm_var_not_dist
         )
 
-        max_global_norm = layers.fill_constant(
-            shape=[1], dtype=global_norm_var_fp32.dtype, value=self.clip_norm
+        max_global_norm = paddle.full(
+            shape=[1],
+            dtype=global_norm_var_fp32.dtype,
+            fill_value=self.clip_norm,
         )
-        clip_var = layers.elementwise_div(
+        clip_var = paddle.divide(
             x=max_global_norm,
             y=paddle.maximum(x=global_norm_var_fp32, y=max_global_norm),
         )
@@ -229,7 +229,7 @@ def __init__(self, optimizer, hcg, strategy):
                                 self._inner_opt._grad_clip, hcg
                             )
 
-    @imperative_base.no_grad
+    @no_grad()
     @framework.dygraph_only
     def step(self):
         parameters_list = _obtain_optimizer_parameters_list(self._inner_opt)
@@ -241,7 +241,7 @@ def step(self):
 
         self._inner_opt.step()
 
-    @imperative_base.no_grad
+    @no_grad()
     def minimize(
         self, loss, startup_program=None, parameters=None, no_grad_set=None
     ):

python/paddle/distributed/fleet/meta_optimizers/dygraph_optimizer/sharding_optimizer_stage2.py

@@ -28,7 +28,7 @@
 
 import paddle
 import paddle.distributed as dist
-from paddle.fluid import core
+from paddle.framework import core
 from paddle.optimizer import Optimizer
 from paddle.fluid.clip import ClipGradByGlobalNorm
 from paddle.distributed.collective import (

python/paddle/distributed/fleet/meta_optimizers/sharding/fp16_helper.py

@@ -18,7 +18,7 @@
     OpRole,
 )
 
-from paddle.fluid import core
+from paddle.framework import core
 
 __all__ = []
 

python/paddle/distributed/fleet/meta_optimizers/sharding/offload_helper.py

@@ -13,7 +13,8 @@
 # limitations under the License.
 
 from ..common import is_optimizer_op, OP_ROLE_KEY, OpRole, is_update_op
-from paddle.fluid import core, unique_name
+from paddle.framework import core
+from paddle.utils import unique_name
 
 __all__ = []
 

python/paddle/distributed/fleet/meta_optimizers/sharding/utils.py

@@ -12,7 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import paddle
-from paddle.fluid import core, unique_name
+from paddle.framework import core
+from paddle.utils import unique_name
 from functools import reduce
 from paddle.distributed.fleet.meta_optimizers.common import (
     is_loss_grad_op,
@@ -1046,11 +1047,11 @@ def sharding_predicate(var):
         )
 
     if int(os.environ.get('PADDLE_TRAINER_ID', 0)) == 0:
-        paddle.fluid.io.save_persistables(
-            exe, dirname, main_program=main_program, filename=None
+        paddle.distributed.io.save_persistables(
+            exe, dirname, main_program=main_program, filename=filename
         )
     else:
-        paddle.fluid.io.save_vars(
+        paddle.static.save_vars(
             exe,
             dirname,
             main_program=main_program,