Use correct master weights in AdamW. (#30895) (#31142)

* Use correct master weights in AdamW.

* Just modify the master weight.

* Update for CI Coverage.

Author: wzzju

python/paddle/fluid/contrib/tests/test_multi_precision_fp16_train.py

@@ -97,7 +97,7 @@ def train(use_pure_fp16=True, use_nesterov=False, use_adam=False):
         test_program = train_program.clone(for_test=True)
 
         if use_adam:
-            optimizer = paddle.optimizer.Adam(
+            optimizer = paddle.optimizer.AdamW(
                 learning_rate=0.001,
                 epsilon=1e-8,
                 weight_decay=0.0,

python/paddle/optimizer/adamw.py

@@ -14,6 +14,7 @@
 
 from .optimizer import Optimizer
 from .adam import Adam
+from ..fluid import core
 from ..fluid import framework
 from ..fluid.dygraph import base as imperative_base
 import paddle
@@ -182,8 +183,16 @@ def _append_decoupled_weight_decay(self, block, param_and_grad):
                 decay_coeff = 1.0 - learning_rate * self._coeff
                 self._lr_to_coeff[learning_rate] = decay_coeff
 
-            scaled_param = param * decay_coeff
-            paddle.fluid.layers.assign(input=scaled_param, output=param)
+            find_master = (self._multi_precision and
+                           param.dtype == core.VarDesc.VarType.FP16)
+            if find_master:
+                master_weight = self._master_weights[param.name]
+                scaled_param = master_weight * decay_coeff
+                paddle.fluid.layers.assign(
+                    input=scaled_param, output=master_weight)
+            else:
+                scaled_param = param * decay_coeff
+                paddle.fluid.layers.assign(input=scaled_param, output=param)
 
     def _append_optimize_op(self, block, param_and_grad):
         self._append_decoupled_weight_decay(block, param_and_grad)