support AMP training (#5067)

* support AMP training

PaddleNLP/benchmark/transformer/configs/transformer.base.yaml

@@ -96,9 +96,10 @@ dropout: 0.1
 # Vocabularies in source and target should be same for weight sharing.
 weight_sharing: True
 
-# Use amp or not
+# Mixed precision training
 use_amp: False
-scale_loss: 1.0
+use_pure_fp16: False
+scale_loss: 128.0
 
 # Whether to use multi-card/multi-node distributed training.
 # Only works for static graph for now.

PaddleNLP/benchmark/transformer/configs/transformer.big.yaml

@@ -96,9 +96,10 @@ dropout: 0.1
 # Vocabularies in source and target should be same for weight sharing.
 weight_sharing: True
 
-# Use amp or not
+# Mixed precision training
 use_amp: False
-scale_loss: 1.0
+use_pure_fp16: False
+scale_loss: 128.0
 
 # Whether to use multi-card/multi-node distributed training.
 # Only works for static graph for now.

PaddleNLP/benchmark/transformer/static/predict.py

@@ -20,6 +20,18 @@ FORMAT = '%(asctime)s-%(levelname)s: %(message)s'
 logging.basicConfig(level=logging.INFO, format=FORMAT)
 logger = logging.getLogger(__name__)
 
+def cast_parameters_to_fp32(place, program, scope=None):
+    all_parameters = []
+    for block in program.blocks:
+        all_parameters.extend(block.all_parameters())
+
+    var_scope = scope if scope else paddle.static.global_scope()
+    for param in all_parameters:
+        tensor = var_scope.find_var(param.name).get_tensor()
+        if 'fp16' in str(tensor._dtype()).lower() and \
+            'fp32' in str(param.dtype).lower():
+            data = np.array(tensor)
+            tensor.set(np.float32(data), place)
 
 def parse_args():
     parser = argparse.ArgumentParser()
@@ -93,6 +105,10 @@ def do_predict(args):
                        os.path.join(args.init_from_params, "transformer"), exe)
     print("finish initing model from params from %s" % (args.init_from_params))
 
+    # cast weights from fp16 to fp32 after loading
+    if args.use_pure_fp16:
+        cast_parameters_to_fp32(place, test_program)
+
     f = open(args.output_file, "w")
     for data in test_loader:
         finished_sequence, = exe.run(test_program,

PaddleNLP/benchmark/transformer/static/train.py

@@ -114,6 +114,17 @@ def do_train(args):
 
             optimizer = fleet.distributed_optimizer(
                 optimizer, strategy=dist_strategy)
+        else:
+            if args.use_amp:
+                amp_list = paddle.static.amp.AutoMixedPrecisionLists(
+                    custom_white_list=['softmax', 'layer_norm'],
+                    custom_black_list=['lookup_table_v2'])
+                optimizer = paddle.static.amp.decorate(
+                    optimizer,
+                    amp_list,
+                    init_loss_scaling=args.scale_loss,
+                    use_dynamic_loss_scaling=True,
+                    use_pure_fp16=args.use_pure_fp16)
         optimizer.minimize(avg_cost)
 
     if args.is_distributed:
@@ -130,6 +141,9 @@ def do_train(args):
                 exec_strategy=exec_strategy)
     exe.run(startup_program)
 
+    if not args.is_distributed and args.use_amp:
+        optimizer.amp_init(places[0])
+
     # the best cross-entropy value with label smoothing
     loss_normalizer = -(
         (1. - args.label_smooth_eps) * np.log(

PaddleNLP/paddlenlp/transformers/transformer/modeling.py

@@ -287,7 +287,7 @@ class TransformerModel(nn.Layer):
         trg_pos = paddle.cast(
             trg_word != self.bos_id, dtype="int64") * paddle.arange(
                 start=0, end=trg_max_len)
-
+        with paddle.static.amp.fp16_guard():
             src_emb = self.src_word_embedding(src_word)
             src_pos_emb = self.src_pos_embedding(src_pos)
             src_emb = src_emb + src_pos_emb