Revert grad scale optimization pr (#50839)

* Revert "fixoptminizer _set_auxiliary_var bug (#50335)"

This reverts commit c44005f0.

* Revert "refine optimizer create accumulators (#50188)"

This reverts commit 244e7546.

* Revert "fix found_inf bug for custom optimizer (#50158)"

This reverts commit 64573f9f.

* Revert "refine amp scaler found_inf (#49864)"

This reverts commit 382e9a06.

* fix code format

* fix conflict

python/paddle/amp/grad_scaler.py

@@ -18,7 +18,7 @@ from enum import Enum
 
 import numpy as np
 
-from paddle import _C_ops, _legacy_C_ops
+from paddle import _legacy_C_ops
 from paddle.fluid import core, in_dygraph_mode
 from paddle.fluid.data_feeder import check_type
 from paddle.fluid.dygraph import to_variable
@@ -228,11 +228,6 @@ class AmpScaler:
 
         optimize_ops, params_grads = (None, None)
 
-        if hasattr(optimizer, "_set_auxiliary_var"):
-            optimizer._set_auxiliary_var('found_inf', self._found_inf)
-            optimize_ops, params_grads = optimizer.minimize(*args, **kwargs)
-            self._cache_founf_inf = optimizer._get_auxiliary_var('found_inf')
-        else:
         if self._found_inf:
             self._cache_founf_inf = True
         else:
@@ -335,9 +330,6 @@ class AmpScaler:
                     param_grads_fp16,
                     self._temp_found_inf_fp16,
                 )
-                self._found_inf = _C_ops.bitwise_or(
-                    self._found_inf, self._temp_found_inf_fp16
-                )
             if len(param_grads_bf16):
                 _legacy_C_ops.check_finite_and_unscale(
                     param_grads_bf16,
@@ -346,9 +338,6 @@ class AmpScaler:
                     param_grads_bf16,
                     self._temp_found_inf_bf16,
                 )
-                self._found_inf = _C_ops.bitwise_or(
-                    self._found_inf, self._temp_found_inf_bf16
-                )
             if len(param_grads_fp32):
                 _legacy_C_ops.check_finite_and_unscale(
                     param_grads_fp32,
@@ -357,9 +346,6 @@ class AmpScaler:
                     param_grads_fp32,
                     self._temp_found_inf_fp32,
                 )
-                self._found_inf = _C_ops.bitwise_or(
-                    self._found_inf, self._temp_found_inf_fp32
-                )
         else:
             if len(param_grads_fp16):
                 _legacy_C_ops.check_finite_and_unscale(
@@ -368,9 +354,6 @@ class AmpScaler:
                     param_grads_fp16,
                     self._temp_found_inf_fp16,
                 )
-                self._found_inf = _C_ops.bitwise_or(
-                    self._found_inf, self._temp_found_inf_fp16
-                )
             if len(param_grads_bf16):
                 _legacy_C_ops.check_finite_and_unscale(
                     param_grads_bf16,
@@ -378,9 +361,6 @@ class AmpScaler:
                     param_grads_bf16,
                     self._temp_found_inf_bf16,
                 )
-                self._found_inf = _C_ops.bitwise_or(
-                    self._found_inf, self._temp_found_inf_bf16
-                )
             if len(param_grads_fp32):
                 _legacy_C_ops.check_finite_and_unscale(
                     param_grads_fp32,
@@ -388,8 +368,11 @@ class AmpScaler:
                     param_grads_fp32,
                     self._temp_found_inf_fp32,
                 )
-                self._found_inf = _C_ops.bitwise_or(
-                    self._found_inf, self._temp_found_inf_fp32
+
+        self._found_inf = (
+            self._temp_found_inf_fp16
+            or self._temp_found_inf_bf16
+            or self._temp_found_inf_fp32
         )
 
         optimizer_state["state"] = OptimizerState.UNSCALED
@@ -778,11 +761,6 @@ class GradScaler(AmpScaler):
         if optimizer_state["state"] is OptimizerState.INIT:
             self._unscale(optimizer)
 
-        if hasattr(optimizer, "_set_auxiliary_var"):
-            optimizer._set_auxiliary_var('found_inf', self._found_inf)
-            optimizer.step()
-            self._cache_founf_inf = optimizer._get_auxiliary_var('found_inf')
-        else:
         if self._found_inf:
             self._cache_founf_inf = True
         else:

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

@@ -236,10 +236,6 @@ class AscendOptimizer(Optimizer):
                 ret_list.append(var)
         return ret_list
 
-    def _set_auxiliary_var(self, key, val):
-        super()._set_auxiliary_var(key, val)
-        self.inner_opt._set_auxiliary_var(key, val)
-
     def minimize(
         self,
         loss,

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

@@ -41,11 +41,6 @@ class HybridParallelGradScaler:
 
         optimize_ops, params_grads = (None, None)
 
-        if hasattr(optimizer, "_set_auxiliary_var"):
-            optimizer._set_auxiliary_var('found_inf', self._found_inf)
-            optimize_ops, params_grads = optimizer.minimize(*args, **kwargs)
-            self._cache_founf_inf = optimizer._get_auxiliary_var('found_inf')
-        else:
         if self._found_inf:
             self._cache_founf_inf = True
         else:

python/paddle/distributed/fleet/meta_optimizers/meta_optimizer_base.py

@@ -25,10 +25,6 @@ class MetaOptimizerBase(Optimizer):
         self.meta_optimizers_white_list = []
         self.meta_optimizers_black_list = []
 
-    def _set_auxiliary_var(self, key, val):
-        super()._set_auxiliary_var(key, val)
-        self.inner_opt._set_auxiliary_var(key, val)
-
     def _set_basic_info(
         self, loss, role_maker, user_defined_optimizer, user_defined_strategy
     ):

python/paddle/distributed/fleet/meta_parallel/sharding/group_sharded_optimizer_stage2.py

@@ -203,10 +203,6 @@ class GroupShardedOptimizerStage2(Optimizer):
         # Update optimizer parameters and adjust parameter storage and use according to rank.
         self._update_opt_status()
 
-    def _set_auxiliary_var(self, key, val):
-        super()._set_auxiliary_var(key, val)
-        self._optim._set_auxiliary_var(key, val)
-
     @paddle.autograd.no_grad()
     def _sync_params_and_buffers(self):
         """

python/paddle/distributed/fleet/meta_parallel/sharding/group_sharded_utils.py

@@ -19,10 +19,10 @@ from types import MethodType
 import numpy as np
 
 import paddle
-from paddle import _C_ops, _legacy_C_ops
+from paddle import _legacy_C_ops
 from paddle.common_ops_import import dygraph_only
-from paddle.fluid import core
 from paddle.fluid.dygraph import to_variable
+from paddle.framework import core
 from paddle.nn import clip
 
 
@@ -270,9 +270,6 @@ def GroupShardedScaler(scaler):
                     param_grads_bfp16,
                     temp_found_inf_bfp16,
                 )
-                self._found_inf = _C_ops.bitwise_or(
-                    self._found_inf, temp_found_inf_bfp16
-                )
             if len(param_grads_fp16):
                 _legacy_C_ops.check_finite_and_unscale(
                     param_grads_fp16,
@@ -280,9 +277,6 @@ def GroupShardedScaler(scaler):
                     param_grads_fp16,
                     temp_found_inf_fp16,
                 )
-                self._found_inf = _C_ops.bitwise_or(
-                    self._found_inf, temp_found_inf_fp16
-                )
             if len(param_grads_fp32):
                 _legacy_C_ops.check_finite_and_unscale(
                     param_grads_fp32,
@@ -290,17 +284,21 @@ def GroupShardedScaler(scaler):
                     param_grads_fp32,
                     temp_found_inf_fp32,
                 )
-                self._found_inf = _C_ops.bitwise_or(
-                    self._found_inf, temp_found_inf_fp32
-                )
 
-        self._found_inf = self._found_inf.cast("int32")
+        self._found_inf = (
+            1
+            if temp_found_inf_bfp16
+            or temp_found_inf_fp16
+            or temp_found_inf_fp32
+            else 0
+        )
+        is_found_inf = paddle.to_tensor([self._found_inf], dtype="int32")
 
         paddle.distributed.all_reduce(
-            self._found_inf, op=paddle.distributed.ReduceOp.MAX, group=None
+            is_found_inf, op=paddle.distributed.ReduceOp.SUM, group=None
         )
 
-        self._found_inf = self._found_inf.cast("bool")
+        self._found_inf = is_found_inf.numpy()[0]
 
     scaler._unscale = MethodType(unscale_method, scaler)
     return scaler

python/paddle/distributed/fleet/scaler.py

@@ -17,7 +17,7 @@ from types import MethodType
 import numpy as np
 
 import paddle
-from paddle import _C_ops, _legacy_C_ops
+from paddle import _legacy_C_ops
 from paddle.distributed import fleet
 from paddle.fluid.dygraph import to_variable
 from paddle.framework import core
@@ -73,9 +73,6 @@ def distributed_scaler(scaler):
                 param_grads_fp16,
                 temp_found_inf_fp16,
             )
-            self._found_inf = _C_ops.bitwise_or(
-                self._found_inf, temp_found_inf_fp16
-            )
         if len(param_grads_fp32):
             _legacy_C_ops.check_finite_and_unscale(
                 param_grads_fp32,
@@ -83,19 +80,17 @@ def distributed_scaler(scaler):
                 param_grads_fp32,
                 temp_found_inf_fp32,
             )
-            self._found_inf = _C_ops.bitwise_or(
-                self._found_inf, temp_found_inf_fp32
-            )
 
-        self._found_inf = self._found_inf.cast("int32")
+        self._found_inf = 1 if temp_found_inf_fp16 or temp_found_inf_fp32 else 0
+        is_found_inf = paddle.to_tensor([self._found_inf], dtype="int32")
 
         # TODO(shenliang03) Since dp allreduce in the optimizer is
         # after the gradscaler, check_finite needs to synchronize global
         # information. In the future, we should use check_group to speed.
         paddle.distributed.all_reduce(
-            self._found_inf, op=paddle.distributed.ReduceOp.MAX, group=None
+            is_found_inf, op=paddle.distributed.ReduceOp.MAX, group=None
         )
-        self._found_inf = self._found_inf.cast("bool")
+        self._found_inf = is_found_inf.numpy()[0]
 
     # Only data_parallel doesn't need to modify scaler
     fleet_env = fleet.fleet

python/paddle/fluid/optimizer.py

@@ -19,6 +19,7 @@ from collections import defaultdict
 
 import paddle
 
+
 from paddle.fluid.framework import (
     Program,
     Variable,
@@ -899,13 +900,6 @@ class Optimizer:
         self._create_global_learning_rate()
 
         if in_dygraph_mode():
-            found_inf = self._get_auxiliary_var('found_inf')
-            if found_inf:
-                if isinstance(found_inf, core.eager.Tensor):
-                    self._set_auxiliary_var('found_inf', True)
-            else:
-                if isinstance(found_inf, core.eager.Tensor):
-                    self._set_auxiliary_var('found_inf', False)
             for param_and_grad in parameters_and_grads:
                 if param_and_grad[1] is None:
                     continue

python/paddle/incubate/optimizer/lookahead.py

@@ -144,10 +144,6 @@ class LookAhead(Optimizer):
         self._global_step_var = None
         self._k_var = None
 
-    def _set_auxiliary_var(self, key, val):
-        super()._set_auxiliary_var(key, val)
-        self.inner_optimizer._set_auxiliary_var(key, val)
-
     @framework.dygraph_only
     @imperative_base.no_grad
     def step(self):

python/paddle/optimizer/adadelta.py

@@ -145,11 +145,8 @@ class Adadelta(Optimizer):
             parameters = parameters.get('params')
 
         for p in parameters:
-            if p.name in self._already_create_accumulater:
-                continue
             self._add_accumulator(self._avg_squared_grad_acc_str, p)
             self._add_accumulator(self._avg_squared_update_acc_str, p)
-            self._already_create_accumulater.add(p.name)
 
     def _append_optimize_op(self, block, param_and_grad):
         if isinstance(param_and_grad, dict):

python/paddle/optimizer/adagrad.py

@@ -139,14 +139,11 @@ class Adagrad(Optimizer):
             parameters = self._update_param_group(parameters)
 
         for p in parameters:
-            if p.name in self._already_create_accumulater:
-                continue
             self._add_accumulator(
                 self._moment_acc_str,
                 p,
                 fill_value=self.initial_accumulator_value,
             )
-            self._already_create_accumulater.add(p.name)
 
     def _append_optimize_op(self, block, param_and_grad):
         assert isinstance(block, framework.Block)

python/paddle/optimizer/adam.py

@@ -317,12 +317,9 @@ class Adam(Optimizer):
 
         # Create accumulator tensors for first and second moments
         for p in parameters:
-            if p.name in self._already_create_accumulater:
-                continue
             if self._multi_precision and self._is_dtype_fp16_or_bf16(p.dtype):
                 master_p = self._create_master_weight(p)
                 self._add_moments_pows(master_p)
-                self._already_create_accumulater.add(p.name)
                 continue
             if (
                 self._is_dtype_fp16_or_bf16(p.dtype)
@@ -333,7 +330,6 @@ class Adam(Optimizer):
                     "Consider using multi_precision=True option of the Adam optimizer."
                 )
             self._add_moments_pows(p)
-            self._already_create_accumulater.add(p.name)
 
     def _append_optimize_op(self, block, param_and_grad):
         assert isinstance(block, framework.Block)
@@ -364,6 +360,8 @@ class Adam(Optimizer):
         # create the adam optimize op
 
         if framework.in_dygraph_mode():
+            found_inf = self._get_auxiliary_var('found_inf')
+
             _beta1 = (
                 self._beta1
                 if not isinstance(self._beta1, Variable)
@@ -384,7 +382,7 @@ class Adam(Optimizer):
                 beta1_pow_acc,
                 beta2_pow_acc,
                 master_weight,
-                None,
+                found_inf,
                 _beta1,
                 _beta2,
                 self._epsilon,
@@ -695,13 +693,6 @@ class Adam(Optimizer):
                         if master_weight is not None
                         else None
                     )
-                    found_inf = self._get_auxiliary_var('found_inf')
-                    if found_inf:
-                        if isinstance(found_inf, core.eager.Tensor):
-                            self._set_auxiliary_var('found_inf', True)
-                    else:
-                        if isinstance(found_inf, core.eager.Tensor):
-                            self._set_auxiliary_var('found_inf', False)
                     _, _, _, _, _, _ = _C_ops.merged_adam_(
                         self._param_dict[key][param_group_idx],
                         grad_dict[key],

python/paddle/optimizer/adamax.py

@@ -176,8 +176,6 @@ class Adamax(Optimizer):
 
         # Create accumulator tensors for first moment and infinity norm
         for p in parameters:
-            if p.name in self._already_create_accumulater:
-                continue
             self._add_accumulator(self._moment_acc_str, p)
             self._add_accumulator(self._inf_norm_acc_str, p)
             self._add_accumulator(
@@ -186,7 +184,6 @@ class Adamax(Optimizer):
                 fill_value=self._beta1,
                 shape=[1],
             )
-            self._already_create_accumulater.add(p.name)
 
     def _append_optimize_op(self, block, param_and_grad):
         assert isinstance(block, framework.Block)

python/paddle/optimizer/adamw.py

@@ -281,7 +281,6 @@ class AdamW(Optimizer):
         self._use_multi_tensor = None
         self.regularization = None
         self._auxiliary_vars = {}
-        self._already_create_accumulater = set()
 
     def _set_auxiliary_var(self, key, val):
         self._auxiliary_vars[key] = val
@@ -423,12 +422,9 @@ class AdamW(Optimizer):
 
         # Create accumulator tensors for first and second moments
         for p in parameters:
-            if p.name in self._already_create_accumulater:
-                continue
             if self._multi_precision and self._is_dtype_fp16_or_bf16(p.dtype):
                 master_p = self._create_master_weight(p)
                 self._add_moments_pows(master_p)
-                self._already_create_accumulater.add(p.name)
                 continue
             if (
                 self._is_dtype_fp16_or_bf16(p.dtype)
@@ -439,7 +435,6 @@ class AdamW(Optimizer):
                     "Consider using multi_precision=True option of the Adam optimizer."
                 )
             self._add_moments_pows(p)
-            self._already_create_accumulater.add(p.name)
 
     def _append_optimize_op(self, block, param_and_grad):
         assert isinstance(block, framework.Block)
@@ -496,6 +491,7 @@ class AdamW(Optimizer):
                 else self._beta2.numpy().item(0)
             )
 
+            found_inf = self._get_auxiliary_var('found_inf')
             _, _, _, _, _, _ = _C_ops.adamw_(
                 param_and_grad[0],
                 param_and_grad[1],
@@ -505,7 +501,7 @@ class AdamW(Optimizer):
                 beta1_pow_acc,
                 beta2_pow_acc,
                 master_weight,
-                None,
+                found_inf,
                 _beta1,
                 _beta2,
                 self._epsilon,

python/paddle/optimizer/lamb.py

@@ -190,15 +190,11 @@ class Lamb(Optimizer):
 
         # Create accumulator tensors for first and second moments
         for p in parameters:
-            if p.name in self._already_create_accumulater:
-                continue
             if self._multi_precision and p.dtype == core.VarDesc.VarType.FP16:
                 master_p = self._create_master_weight(p)
                 self._add_moments_pows(master_p)
-                self._already_create_accumulater.add(p.name)
             else:
                 self._add_moments_pows(p)
-                self._already_create_accumulater.add(p.name)
 
     def _get_accumulator(self, name, param):
         """Utility function to fetch an accumulator for a parameter
@@ -297,6 +293,7 @@ class Lamb(Optimizer):
             self._used_master_weights[p_name] = master_weight.name
         else:
             master_weight = None
+        found_inf = self._get_auxiliary_var('found_inf')
 
         if framework.in_dygraph_mode():
             _C_ops.lamb_(
@@ -308,7 +305,7 @@ class Lamb(Optimizer):
                 beta1_pow_acc,
                 beta2_pow_acc,
                 master_weight,
-                None,
+                found_inf,
                 weight_decay,
                 self._beta1,
                 self._beta2,
@@ -346,7 +343,6 @@ class Lamb(Optimizer):
                 inputs["MasterParam"] = master_weight
                 outputs["MasterParamOut"] = master_weight
 
-            found_inf = self._get_auxiliary_var('found_inf')
             if found_inf:
                 inputs["SkipUpdate"] = found_inf
 

python/paddle/optimizer/momentum.py

@@ -270,12 +270,9 @@ class Momentum(Optimizer):
             parameters = self._update_param_group(parameters)
 
         for p in parameters:
-            if p.name in self._already_create_accumulater:
-                continue
             if self._multi_precision and p.dtype == core.VarDesc.VarType.FP16:
                 master_p = self._create_master_weight(p)
                 self._add_accumulator(self._velocity_acc_str, master_p)
-                self._already_create_accumulater.add(p.name)
                 continue
             if (
                 p.dtype == core.VarDesc.VarType.FP16
@@ -286,7 +283,6 @@ class Momentum(Optimizer):
                     "Consider using multi_precision=True option of the Momentum optimizer."
                 )
             self._add_accumulator(self._velocity_acc_str, p)
-            self._already_create_accumulater.add(p.name)
 
     def _create_regularization_of_grad(self, param, grad, regularization=None):
         """Create and add backward regularization Operators
@@ -534,13 +530,6 @@ class Momentum(Optimizer):
                 )
 
                 if in_dygraph_mode():
-                    found_inf = self._get_auxiliary_var('found_inf')
-                    if found_inf:
-                        if isinstance(found_inf, core.eager.Tensor):
-                            self._set_auxiliary_var('found_inf', True)
-                    else:
-                        if isinstance(found_inf, core.eager.Tensor):
-                            self._set_auxiliary_var('found_inf', False)
                     _, _, _ = _C_ops.merged_momentum_(
                         self._param_dict[key][param_group_idx],
                         grad_dict[key],
@@ -549,12 +538,8 @@ class Momentum(Optimizer):
                         master_weight,
                         self._momentum,
                         self._use_nesterov,
-                            self._regularization_method_dict[key][
-                                param_group_idx
-                            ],
-                            self._regularization_coeff_dict[key][
-                                param_group_idx
-                            ],
+                        self._regularization_method_dict[key][param_group_idx],
+                        self._regularization_coeff_dict[key][param_group_idx],
                         find_master,
                         self._rescale_grad,
                     )

python/paddle/optimizer/optimizer.py

@@ -275,7 +275,6 @@ class Optimizer:
 
         self._param_dict = self._create_multi_tensor_dict()
         self._auxiliary_vars = {}
-        self._already_create_accumulater = set()
 
     def _set_auxiliary_var(self, key, val):
         self._auxiliary_vars[key] = val
@@ -924,13 +923,6 @@ class Optimizer:
                 self._create_accumulators(target_block, params_acc_dict)
 
             if framework._non_static_mode():
-                found_inf = self._get_auxiliary_var('found_inf')
-                if found_inf:
-                    if isinstance(found_inf, core.eager.Tensor):
-                        self._set_auxiliary_var('found_inf', True)
-                else:
-                    if isinstance(found_inf, core.eager.Tensor):
-                        self._set_auxiliary_var('found_inf', False)
                 if isinstance(parameters_and_grads, list):
                     for param_and_grad in parameters_and_grads:
                         if param_and_grad[1] is None:

python/paddle/optimizer/rmsprop.py

@@ -199,12 +199,9 @@ class RMSProp(Optimizer):
             parameters = parameters.get('params')
 
         for p in parameters:
-            if p.name in self._already_create_accumulater:
-                continue
             self._add_accumulator(self._momentum_acc_str, p)
             self._add_accumulator(self._mean_square_acc_str, p)
             self._add_accumulator(self._mean_grad_acc_str, p)
-            self._already_create_accumulater.add(p.name)
 
     def _append_optimize_op(self, block, param_and_grad):
         if not isinstance(block, framework.Block):

python/paddle/optimizer/sgd.py

@@ -129,11 +129,8 @@ class SGD(Optimizer):
 
         # Create accumulator tensors for first and second moments
         for p in parameters:
-            if p.name in self._already_create_accumulater:
-                continue
             if self._multi_precision and p.dtype == core.VarDesc.VarType.FP16:
                 master_p = self._create_master_weight(p)
-                self._already_create_accumulater.add(p.name)
                 continue
             if (
                 p.dtype == core.VarDesc.VarType.FP16