refine amp scaler found_inf (#49864)

* refine _found_inf

python/paddle/amp/grad_scaler.py

@@ -18,7 +18,7 @@
 
 import numpy as np
 
-from paddle import _legacy_C_ops
+from paddle import _C_ops, _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 +228,9 @@ def minimize(self, optimizer, *args, **kwargs):
 
         optimize_ops, params_grads = (None, None)
 
-        if self._found_inf:
-            self._cache_founf_inf = True
-        else:
-            optimize_ops, params_grads = optimizer.minimize(*args, **kwargs)
-            self._cache_founf_inf = False
+        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')
 
         if self._use_dynamic_loss_scaling:
             # uopdate the scale
@@ -330,6 +328,9 @@ def _unscale(self, optimizer):
                     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,
@@ -338,6 +339,9 @@ def _unscale(self, optimizer):
                     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,
@@ -346,6 +350,9 @@ def _unscale(self, optimizer):
                     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(
@@ -354,26 +361,29 @@ def _unscale(self, optimizer):
                     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,
                     self._scale,
                     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,
                     self._scale,
                     param_grads_fp32,
                     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
-        )
+                self._found_inf = _C_ops.bitwise_or(
+                    self._found_inf, self._temp_found_inf_fp32
+                )
 
         optimizer_state["state"] = OptimizerState.UNSCALED
 
@@ -761,11 +771,9 @@ def step(self, optimizer):
         if optimizer_state["state"] is OptimizerState.INIT:
             self._unscale(optimizer)
 
-        if self._found_inf:
-            self._cache_founf_inf = True
-        else:
-            optimizer.step()
-            self._cache_founf_inf = False
+        optimizer._set_auxiliary_var('found_inf', self._found_inf)
+        optimizer.step()
+        self._cache_founf_inf = optimizer._get_auxiliary_var('found_inf')
 
         optimizer_state["state"] = OptimizerState.STEPPED
 

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

@@ -41,11 +41,9 @@ def minimize(self, optimizer, *args, **kwargs):
 
         optimize_ops, params_grads = (None, None)
 
-        if self._found_inf:
-            self._cache_founf_inf = True
-        else:
-            optimize_ops, params_grads = optimizer.minimize(*args, **kwargs)
-            self._cache_founf_inf = False
+        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')
 
         if self._use_dynamic_loss_scaling:
             self._update()

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

@@ -19,10 +19,10 @@
 import numpy as np
 
 import paddle
-from paddle import _legacy_C_ops
+from paddle import _C_ops, _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
 
 
@@ -231,22 +231,27 @@ def unscale_method(self, optimizer):
                     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,
                     self._scale,
                     param_grads_fp32,
                     temp_found_inf_fp32,
                 )
+                self._found_inf = _C_ops.bitwise_or(
+                    self._found_inf, temp_found_inf_fp32
+                )
 
-        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")
+        self._found_inf = self._found_inf.cast("int32")
 
         paddle.distributed.all_reduce(
-            is_found_inf, op=paddle.distributed.ReduceOp.SUM, group=None
+            self._found_inf, op=paddle.distributed.ReduceOp.MAX, group=None
         )
 
-        self._found_inf = is_found_inf.numpy()[0]
+        self._found_inf = self._found_inf.cast("bool")
 
     scaler._unscale = MethodType(unscale_method, scaler)
     return scaler

python/paddle/distributed/fleet/scaler.py

@@ -17,7 +17,7 @@
 import numpy as np
 
 import paddle
-from paddle import _legacy_C_ops
+from paddle import _C_ops, _legacy_C_ops
 from paddle.distributed import fleet
 from paddle.fluid.dygraph import to_variable
 from paddle.framework import core
@@ -73,24 +73,29 @@ def unscale_method(self, optimizer):
                 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,
                 self._scale,
                 param_grads_fp32,
                 temp_found_inf_fp32,
             )
+            self._found_inf = _C_ops.bitwise_or(
+                self._found_inf, temp_found_inf_fp32
+            )
 
-        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")
+        self._found_inf = self._found_inf.cast("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(
-            is_found_inf, op=paddle.distributed.ReduceOp.MAX, group=None
+            self._found_inf, op=paddle.distributed.ReduceOp.MAX, group=None
         )
-        self._found_inf = is_found_inf.numpy()[0]
+        self._found_inf = self._found_inf.cast("bool")
 
     # Only data_parallel doesn't need to modify scaler
     fleet_env = fleet.fleet

python/paddle/fluid/optimizer.py

@@ -893,11 +893,18 @@ def _create_optimization_pass(self, parameters_and_grads):
         self._create_global_learning_rate()
 
         if in_dygraph_mode():
-            for param_and_grad in parameters_and_grads:
-                if param_and_grad[1] is None:
-                    continue
-                if param_and_grad[0].trainable is True:
-                    self._append_optimize_op(target_block, param_and_grad)
+            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
+                    if param_and_grad[0].trainable is True:
+                        self._append_optimize_op(target_block, param_and_grad)
         else:
             for param_and_grad in parameters_and_grads:
                 if param_and_grad[1] is None:

python/paddle/optimizer/adam.py

@@ -360,8 +360,6 @@ def _append_optimize_op(self, block, param_and_grad):
         # 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)
@@ -382,7 +380,7 @@ def _append_optimize_op(self, block, param_and_grad):
                 beta1_pow_acc,
                 beta2_pow_acc,
                 master_weight,
-                found_inf,
+                None,
                 _beta1,
                 _beta2,
                 self._epsilon,
@@ -693,21 +691,28 @@ def _append_optimize_multi_tensor_op(
                         if master_weight is not None
                         else None
                     )
-                    _, _, _, _, _, _ = _C_ops.merged_adam_(
-                        self._param_dict[key][param_group_idx],
-                        grad_dict[key],
-                        lr_dict[key],
-                        self._moment1_dict[key][param_group_idx],
-                        self._moment2_dict[key][param_group_idx],
-                        self._beta1_pow_acc_dict[key][param_group_idx],
-                        self._beta2_pow_acc_dict[key][param_group_idx],
-                        master_weight,
-                        _beta1,
-                        _beta2,
-                        self._epsilon,
-                        find_master,
-                        False,
-                    )
+                    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],
+                            lr_dict[key],
+                            self._moment1_dict[key][param_group_idx],
+                            self._moment2_dict[key][param_group_idx],
+                            self._beta1_pow_acc_dict[key][param_group_idx],
+                            self._beta2_pow_acc_dict[key][param_group_idx],
+                            master_weight,
+                            _beta1,
+                            _beta2,
+                            self._epsilon,
+                            find_master,
+                            False,
+                        )
                 else:
                     inputs = {
                         "Param": self._param_dict[key][param_group_idx],

python/paddle/optimizer/adamw.py

@@ -491,7 +491,6 @@ def _append_optimize_op(self, block, param_and_grad):
                 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],
@@ -501,7 +500,7 @@ def _append_optimize_op(self, block, param_and_grad):
                 beta1_pow_acc,
                 beta2_pow_acc,
                 master_weight,
-                found_inf,
+                None,
                 _beta1,
                 _beta2,
                 self._epsilon,

python/paddle/optimizer/lamb.py

@@ -293,7 +293,6 @@ def _append_optimize_op(self, block, param_and_grad):
             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_(
@@ -305,7 +304,7 @@ def _append_optimize_op(self, block, param_and_grad):
                 beta1_pow_acc,
                 beta2_pow_acc,
                 master_weight,
-                found_inf,
+                None,
                 weight_decay,
                 self._beta1,
                 self._beta2,
@@ -343,6 +342,7 @@ def _append_optimize_op(self, block, param_and_grad):
                 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

@@ -530,19 +530,30 @@ def _append_optimize_multi_tensor_op(
                 )
 
                 if in_dygraph_mode():
-                    _, _, _ = _C_ops.merged_momentum_(
-                        self._param_dict[key][param_group_idx],
-                        grad_dict[key],
-                        self._velocity_dict[key][param_group_idx],
-                        lr_dict[key],
-                        master_weight,
-                        self._momentum,
-                        self._use_nesterov,
-                        self._regularization_method_dict[key][param_group_idx],
-                        self._regularization_coeff_dict[key][param_group_idx],
-                        find_master,
-                        self._rescale_grad,
-                    )
+                    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],
+                            self._velocity_dict[key][param_group_idx],
+                            lr_dict[key],
+                            master_weight,
+                            self._momentum,
+                            self._use_nesterov,
+                            self._regularization_method_dict[key][
+                                param_group_idx
+                            ],
+                            self._regularization_coeff_dict[key][
+                                param_group_idx
+                            ],
+                            find_master,
+                            self._rescale_grad,
+                        )
                 else:
                     inputs = {
                         "Param": self._param_dict[key][param_group_idx],