clear fluid apis: fix apis in fleet and passes (#48021)

* clear fluid apis in fleet and passes

* fix model.py

* fix model.py

* fix cpp_pass.py

python/paddle/distributed/fleet/fleet.py

@@ -15,7 +15,7 @@
 import copy
 import paddle
 import os
-from paddle.fluid.framework import _global_flags
+from paddle.framework import _global_flags
 from paddle.fluid import compiler
 from .base.role_maker import PaddleCloudRoleMaker, RoleMakerBase
 from .base.strategy_compiler import StrategyCompiler
@@ -271,14 +271,14 @@ class Fleet:
         self.strategy_compiler = StrategyCompiler()
 
         if self._role_maker._is_non_distributed() and self._is_collective:
-            if paddle.fluid.core.is_compiled_with_cuda():
-                gpus_num = paddle.fluid.core.get_cuda_device_count()
+            if paddle.framework.core.is_compiled_with_cuda():
+                gpus_num = paddle.framework.core.get_cuda_device_count()
                 if gpus_num != 1:
                     raise ValueError(
                         "CUDA_VISIBLE_DEVICES shoule be set only 1 card if you use `python` to launch fleet program."
                     )
 
-        if paddle.fluid.framework._non_static_mode():
+        if paddle.framework._non_static_mode():
             if self.worker_num() == 1:
                 # if worker_num is 1, should construct default topology & hcg
                 self._topology = tp.CommunicateTopology()
@@ -1011,8 +1011,8 @@ class Fleet:
                 import paddle.distributed.fleet as fleet
                 fleet.init()
                 import paddle
-                place = paddle.fluid.CPUPlace()
-                exe = paddle.fluid.Executor(place)
+                place = paddle.CPUPlace()
+                exe =  paddle.static.Executor(place)
 
                 # build net
                 # fleet.distributed_optimizer(...)
@@ -1242,7 +1242,7 @@ class Fleet:
             )
         else:
             if (
-                paddle.fluid.framework._non_static_mode()
+                paddle.framework._non_static_mode()
                 or self._role_maker._is_non_distributed()
                 or self._is_collective
             ):
@@ -1258,7 +1258,7 @@ class Fleet:
         context["user_defined_strategy"] = copy.deepcopy(
             self._user_defined_strategy
         )
-        if paddle.fluid.framework._non_static_mode():
+        if paddle.framework._non_static_mode():
             # imitate target optimizer retrieval
             target_opt = self.user_defined_optimizer
             self._context = context
@@ -1418,7 +1418,7 @@ class Fleet:
             logger.debug("default program id: " + str(id(default_program)))
 
             if id(default_program) != id(loss.block.program):
-                paddle.fluid.framework.switch_main_program(loss.block.program)
+                paddle.framework.switch_main_program(loss.block.program)
             logger.debug(
                 "default program id after switch: " + str(id(default_program))
             )
@@ -1532,7 +1532,7 @@ class Fleet:
         # default_program = paddle.static.default_main_program()
 
         # if id(default_program) != id(losses[0].block.program):
-        #     paddle.fluid.framework.switch_main_program(losses[0].block.program)
+        #     paddle.framework.switch_main_program(losses[0].block.program)
 
         context["program_optimize_ops"] = optimize_ops
         context["program_params_grads"] = params_grads

python/paddle/distributed/fleet/fleet_executor_utils.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from paddle.distributed.fleet.meta_optimizers.common import OpRole, OP_ROLE_KEY
-from paddle.fluid import core
+from paddle.framework import core
 from paddle.static import Program
 
 

python/paddle/distributed/fleet/launch.py

@@ -62,7 +62,7 @@ import time
 import copy
 import pathlib
 from argparse import ArgumentParser, REMAINDER
-import paddle.fluid as fluid
+import paddle.framework as framework
 from paddle.distributed.fleet import launch_utils
 from paddle.distributed.fleet.launch_utils import (
     get_host_name_ip,
@@ -136,7 +136,7 @@ see: http://www.paddlepaddle.org/documentation/docs/zh/1.6/user_guides/howto/tra
         help="run mode of job, can be:collective/ps/ps-heter",
     )
 
-    if fluid.core.is_compiled_with_cuda():
+    if framework.core.is_compiled_with_cuda():
         base_group.add_argument(
             "--gpus",
             type=str,
@@ -147,7 +147,7 @@ see: http://www.paddlepaddle.org/documentation/docs/zh/1.6/user_guides/howto/tra
         )
         base_group.add_argument("--selected_gpus", dest="gpus")
 
-    if fluid.core.is_compiled_with_xpu():
+    if framework.core.is_compiled_with_xpu():
         base_group.add_argument(
             "--xpus",
             type=str,
@@ -157,7 +157,7 @@ see: http://www.paddlepaddle.org/documentation/docs/zh/1.6/user_guides/howto/tra
         )
         base_group.add_argument("--selected_xpus", dest="xpus")
 
-    if fluid.core.is_compiled_with_npu():
+    if framework.core.is_compiled_with_npu():
         base_group.add_argument(
             "--npus",
             type=str,
@@ -167,7 +167,7 @@ see: http://www.paddlepaddle.org/documentation/docs/zh/1.6/user_guides/howto/tra
         )
         base_group.add_argument("--selected_npus", dest="npus")
 
-    if fluid.core.is_compiled_with_mlu():
+    if framework.core.is_compiled_with_mlu():
         base_group.add_argument(
             "--mlus",
             type=str,
@@ -505,13 +505,13 @@ def launch_ps(args, distribute_mode):
 def infer_backend(args):
     if args.backend != "auto":
         return
-    if fluid.core.is_compiled_with_cuda():
+    if framework.core.is_compiled_with_cuda():
         args.backend = 'nccl'
-    elif fluid.core.is_compiled_with_npu():
+    elif framework.core.is_compiled_with_npu():
         args.backend = 'unknown'
-    elif fluid.core.is_compiled_with_xpu():
+    elif framework.core.is_compiled_with_xpu():
         args.backend = 'bkcl'
-    elif fluid.core.is_compiled_with_mlu():
+    elif framework.core.is_compiled_with_mlu():
         args.backend = 'cncl'
     else:
         args.backend = 'gloo'
@@ -559,14 +559,14 @@ def which_distributed_mode(args):
             "Only one mode(Collective or Parameter-Server) can be selected at the same time, but more than one configuration was received."
         )
 
-    if fluid.core.is_compiled_with_cuda():
-        accelerators = fluid.core.get_cuda_device_count()
-    elif fluid.core.is_compiled_with_npu():
-        accelerators = fluid.core.get_npu_device_count()
-    elif fluid.core.is_compiled_with_xpu():
-        accelerators = fluid.core.get_xpu_device_count()
-    elif fluid.core.is_compiled_with_mlu():
-        accelerators = fluid.core.get_mlu_device_count()
+    if framework.core.is_compiled_with_cuda():
+        accelerators = framework.core.get_cuda_device_count()
+    elif framework.core.is_compiled_with_npu():
+        accelerators = framework.core.get_npu_device_count()
+    elif framework.core.is_compiled_with_xpu():
+        accelerators = framework.core.get_xpu_device_count()
+    elif framework.core.is_compiled_with_mlu():
+        accelerators = framework.core.get_mlu_device_count()
     else:
         accelerators = 0
 
@@ -591,9 +591,9 @@ def which_distributed_mode(args):
         return DistributeMode.COLLECTIVE
     else:
         if (
-            not fluid.core.is_compiled_with_cuda()
-            and not fluid.core.is_compiled_with_xpu()
-            and not fluid.core.is_compiled_with_mlu()
+            not framework.core.is_compiled_with_cuda()
+            and not framework.core.is_compiled_with_xpu()
+            and not framework.core.is_compiled_with_mlu()
         ):
             if args.servers:
                 logger.warning(

python/paddle/distributed/fleet/launch_utils.py

@@ -27,7 +27,7 @@ import socket
 import struct
 import json
 
-import paddle.fluid as fluid
+import paddle.framework as framework
 from distutils.util import strtobool
 import paddle.utils.cpp_extension.extension_utils as utils
 
@@ -572,7 +572,7 @@ def start_local_trainers(
                 [str(g) for g in t.accelerators]
             )
         # to do: same code style in future
-        if fluid.core.is_compiled_with_xpu() and len(t.accelerators) > 0:
+        if framework.core.is_compiled_with_xpu() and len(t.accelerators) > 0:
             proc_env["FLAGS_selected_xpus"] = "%s" % ",".join(
                 [str(g) for g in t.accelerators]
             )
@@ -706,7 +706,7 @@ def watch_local_trainers(procs, nranks):
 
 def get_gpus(gpus):
     if gpus is None:
-        gpus_num = fluid.core.get_cuda_device_count()
+        gpus_num = framework.core.get_cuda_device_count()
         res_gpus = [str(x) for x in range(0, gpus_num)]
     else:
         cuda_visible_devices = os.getenv("CUDA_VISIBLE_DEVICES")
@@ -740,7 +740,7 @@ def get_gpus(gpus):
 
 def get_xpus(xpus):
     if xpus is None:
-        xpus_num = fluid.core.get_xpu_device_count()
+        xpus_num = framework.core.get_xpu_device_count()
         res_xpus = [str(x) for x in range(0, xpus_num)]
     else:
         xpu_visible_devices = os.getenv("XPU_VISIBLE_DEVICES")
@@ -775,7 +775,7 @@ def get_xpus(xpus):
 
 def get_npus(npus):
     if npus is None:
-        npus_num = fluid.core.get_npu_device_count()
+        npus_num = framework.core.get_npu_device_count()
         res_npus = [str(x) for x in range(0, npus_num)]
     else:
         npu_visible_devices = os.getenv("ASCEND_VISIBLE_DEVICES")
@@ -809,7 +809,7 @@ def get_npus(npus):
 
 def get_mlus(mlus):
     if mlus is None:
-        mlus_num = fluid.core.get_mlu_device_count()
+        mlus_num = framework.core.get_mlu_device_count()
         res_mlus = [str(x) for x in range(0, mlus_num)]
     else:
         mlu_visible_devices = os.getenv("MLU_VISIBLE_DEVICES")
@@ -845,37 +845,37 @@ def get_mlus(mlus):
 def get_device_mode(backend):
     if backend == 'heter':
         if (
-            fluid.core.is_compiled_with_cuda()
-            and fluid.core.get_cuda_device_count() > 0
+            framework.core.is_compiled_with_cuda()
+            and framework.core.get_cuda_device_count() > 0
         ):
             print("launch train in heter mode with GPU device.")
             return DeviceMode.GPU
         if (
-            fluid.core.is_compiled_with_xpu()
-            and fluid.core.get_xpu_device_count() > 0
+            framework.core.is_compiled_with_xpu()
+            and framework.core.get_xpu_device_count() > 0
         ):
             print("launch train in heter mode with XPU device.")
             return DeviceMode.XPU
         if (
-            fluid.core.is_compiled_with_npu()
-            and fluid.core.get_npu_device_count() > 0
+            framework.core.is_compiled_with_npu()
+            and framework.core.get_npu_device_count() > 0
         ):
             print("launch train in heter mode with NPU device.")
             return DeviceMode.ASCEND_NPU
 
-    if backend == 'hccl' and fluid.core.get_npu_device_count() > 0:
+    if backend == 'hccl' and framework.core.get_npu_device_count() > 0:
         print("launch train in ascend npu mode!")
         return DeviceMode.ASCEND_NPU
 
-    if backend == 'nccl' and fluid.core.get_cuda_device_count() > 0:
+    if backend == 'nccl' and framework.core.get_cuda_device_count() > 0:
         print("launch train in GPU mode!")
         return DeviceMode.GPU
 
-    if backend == 'bkcl' and fluid.core.get_xpu_device_count() > 0:
+    if backend == 'bkcl' and framework.core.get_xpu_device_count() > 0:
         print("launch train in XPU mode")
         return DeviceMode.XPU
 
-    if backend == 'cncl' and fluid.core.get_mlu_device_count() > 0:
+    if backend == 'cncl' and framework.core.get_mlu_device_count() > 0:
         print("launch train in MLU mode")
         return DeviceMode.MLU
 
@@ -1063,7 +1063,7 @@ def get_mapped_cluster_from_args_without_rank_mapping(args, device_mode):
     assert (
         device_mode == DeviceMode.GPU
     ), "Only support get mapped cluster for gpu now."
-    gpus_num = fluid.core.get_cuda_device_count()
+    gpus_num = framework.core.get_cuda_device_count()
 
     # parse ip-ranks json file
     cluster_topo = None
@@ -1192,7 +1192,7 @@ def get_mapped_cluster_from_args_with_rank_mapping(args, device_mode):
     assert (
         device_mode == DeviceMode.GPU
     ), "Only support get mapped cluster for gpu now."
-    gpus_num = fluid.core.get_cuda_device_count()
+    gpus_num = framework.core.get_cuda_device_count()
 
     # parse ip-ranks json file
     rank_mapping_path = args.rank_mapping_path or os.getenv(
@@ -1866,11 +1866,11 @@ class ParameterServerLauncher:
 
         heter_device_num = 0
         device_list = []
-        if fluid.core.is_compiled_with_cuda():
+        if framework.core.is_compiled_with_cuda():
             device_list = get_gpus(args.gpus)
             heter_device_num = len(device_list)
-        elif fluid.core.is_compiled_with_xpu():
-            heter_device_num = fluid.core.get_xpu_device_count()
+        elif framework.core.is_compiled_with_xpu():
+            heter_device_num = framework.core.get_xpu_device_count()
             device_list = [str(x) for x in range(0, heter_device_num)]
 
         for idx, cur_worker in enumerate(pod.workers):
@@ -2042,11 +2042,11 @@ class ParameterServerLauncher:
 
         heter_device_num = 0
         device_list = []
-        if fluid.core.is_compiled_with_cuda():
+        if framework.core.is_compiled_with_cuda():
             device_list = get_gpus(args.gpus)
             heter_device_num = len(device_list)
-        elif fluid.core.is_compiled_with_xpu():
-            heter_device_num = fluid.core.get_xpu_device_count()
+        elif framework.core.is_compiled_with_xpu():
+            heter_device_num = framework.core.get_xpu_device_count()
             device_list = [str(x) for x in range(0, heter_device_num)]
 
         for idx, cur_heter_worker in enumerate(pod.heter_workers):
@@ -2144,25 +2144,25 @@ def check_backend(backend):
             "but got %s" % backend
         )
 
-    if backend == 'nccl' and not fluid.core.is_compiled_with_cuda():
+    if backend == 'nccl' and not framework.core.is_compiled_with_cuda():
         raise ValueError(
             "paddle.distributed initialize error, "
             "your paddle is not compiled with cuda but you assign 'nccl' as backend."
         )
 
-    if backend == 'bkcl' and not fluid.core.is_compiled_with_xpu():
+    if backend == 'bkcl' and not framework.core.is_compiled_with_xpu():
         raise ValueError(
             "paddle.distributed initialize error, "
             "your paddle is not compiled with xpu but you assign 'bkcl' as backend."
         )
 
-    if backend == 'hccl' and not fluid.core.is_compiled_with_npu():
+    if backend == 'hccl' and not framework.core.is_compiled_with_npu():
         raise ValueError(
             "paddle.distributed initialize error, "
             "your paddle is not compiled with npu but you assign 'hccl' as backend."
         )
 
-    if backend == 'cncl' and not fluid.core.is_compiled_with_mlu():
+    if backend == 'cncl' and not framework.core.is_compiled_with_mlu():
         raise ValueError(
             "paddle.distributed initialize error, "
             "your paddle is not compiled with mlu but you assign 'cncl' as backend."
@@ -2183,16 +2183,16 @@ def block_windows_and_macos(backend):
 
 
 def get_backend_by_compile_flag():
-    if fluid.core.is_compiled_with_cuda():
+    if framework.core.is_compiled_with_cuda():
         return 'nccl'
 
-    if fluid.core.is_compiled_with_xpu():
+    if framework.core.is_compiled_with_xpu():
         return 'bkcl'
 
-    if fluid.core.is_compiled_with_npu():
+    if framework.core.is_compiled_with_npu():
         return 'hccl'
 
-    if fluid.core.is_compiled_with_mlu():
+    if framework.core.is_compiled_with_mlu():
         return 'cncl'
 
     return 'gloo'

python/paddle/distributed/fleet/optimizer.py

@@ -72,7 +72,7 @@ def _dygraph_distributed_optimizer(optimizer, strategy=None):
 
 
 def distributed_optimizer(*args, **kwargs):
-    if paddle.fluid.framework._non_static_mode():
+    if paddle.framework._non_static_mode():
         return _dygraph_distributed_optimizer(*args, **kwargs)
     else:
         return fleet.fleet.distributed_optimizer(*args, **kwargs)

python/paddle/distributed/fleet/scaler.py

@@ -16,7 +16,7 @@ import paddle
 from .base.topology import ParallelMode
 from paddle.distributed import fleet
 from types import MethodType
-from paddle.fluid import core
+from paddle.framework import core
 from paddle.fluid.dygraph import to_variable
 import numpy as np
 from paddle import _legacy_C_ops

python/paddle/distributed/passes/cpp_pass.py

@@ -14,7 +14,7 @@
 
 from paddle.static import Executor
 from .pass_base import PassType, CPPPassWrapper, register_pass
-from paddle.fluid.framework import core, _apply_pass as _apply_cpp_pass
+from paddle.framework import core, _apply_pass as _apply_cpp_pass
 
 
 @register_pass("fuse_elewise_add_act")

python/paddle/distributed/passes/fuse_all_reduce.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from paddle.framework import core
-from paddle.fluid import unique_name
+from paddle.utils import unique_name
 from .pass_base import PassBase, PassType, register_pass
 import numpy as np
 

python/paddle/distributed/passes/pass_base.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from abc import ABC, abstractmethod
-from paddle.fluid.framework import _apply_pass as _apply_cpp_pass
+from paddle.framework import _apply_pass as _apply_cpp_pass
 
 
 class PassContext:

python/paddle/framework/__init__.py

@@ -66,5 +66,8 @@ from ..fluid.framework import _dygraph_tracer  # noqa: F401
 from ..fluid.layer_helper import LayerHelper  # noqa: F401
 from ..fluid.framework import in_dygraph_mode  # noqa: F401
 from ..fluid.framework import _in_legacy_dygraph  # noqa: F401
+from ..fluid.framework import _global_flags  # noqa: F401
+from ..fluid.framework import _apply_pass  # noqa: F401
+from ..fluid.framework import switch_main_program
 
 __all__ = []