fix fp16 for paddle version (#4283)

Author: ZHUI

paddlenlp/trainer/trainer.py

@@ -111,6 +111,28 @@ def is_datasets_available():
 if is_datasets_available():
     import datasets
 
+
+@contextlib.contextmanager
+def device_guard(device="cpu", dev_id=0):
+    origin_device = paddle.device.get_device()
+    if device == "cpu":
+        paddle.set_device(device)
+    elif device in ["gpu", "xpu", "npu"]:
+        paddle.set_device("{}:{}".format(device, dev_id))
+    try:
+        yield
+    finally:
+        paddle.set_device(origin_device)
+
+
+def paddlenlp_load(path, return_numpy=False):
+    if return_numpy:
+        with device_guard():
+            return paddle.load(path)
+    else:
+        return paddle.load(path, return_numpy=return_numpy)
+
+
 __all__ = ["Trainer"]
 
 
@@ -267,7 +289,11 @@ def __init__(
             self.amp_dtype = "float16" if args.fp16 else "bfloat16"
             # fix for load saved fp16 or bf16 ckpt, decorate model first.
             if self.args.fp16_opt_level == "O2":
-                paddle.amp.decorate(models=model, level=self.args.fp16_opt_level, dtype=self.amp_dtype)
+                if self.amp_dtype == "bfloat16":
+                    # fix for paddlepaddle < 2.4.1, not support for bf16
+                    paddle.amp.decorate(models=model, level=self.args.fp16_opt_level, dtype=self.amp_dtype)
+                else:
+                    paddle.amp.decorate(models=model, level=self.args.fp16_opt_level)
 
             if self.sharding is not None:
                 self.scaler = paddle.amp.GradScaler(init_loss_scaling=self.args.scale_loss)
@@ -1130,9 +1156,16 @@ def _wrap_model(self, model, training=True):
         # Mixed precision training
         if training and self.do_grad_scaling:  # self.args.fp16_opt_level=="O2":
             # model, self.optimizer
-            decorated = paddle.amp.decorate(
-                models=model, optimizers=self.optimizer, level=self.args.fp16_opt_level, dtype=self.amp_dtype
-            )
+            if self.amp_dtype == "bfloat16":
+                # fix for paddlepaddle < 2.4.1, not support for bf16
+                decorated = paddle.amp.decorate(
+                    models=model, optimizers=self.optimizer, level=self.args.fp16_opt_level, dtype=self.amp_dtype
+                )
+            else:
+                decorated = paddle.amp.decorate(
+                    models=model, optimizers=self.optimizer, level=self.args.fp16_opt_level
+                )
+
             if self.optimizer is None:
                 model = decorated
             else:
@@ -1459,15 +1492,18 @@ def _load_optimizer_and_scheduler(self, checkpoint):
             # Load in optimizer and scheduler states
             if self.sharding is not None:
                 self.optimizer.set_state_dict(
-                    paddle.load(
+                    paddlenlp_load(
                         os.path.join(checkpoint, OPTIMIZER_NAME + f"_shard{self.sharding_group.rank}"),
                         return_numpy=True,
                     )
                 )
                 empty_dict = paddle.load(os.path.join(checkpoint, OPTIMIZER_NAME), return_numpy=True)
                 assert len(empty_dict) == 0, "Optimizer file of sharding, should be empty!"
             else:
-                self.optimizer.set_state_dict(paddle.load(os.path.join(checkpoint, OPTIMIZER_NAME), return_numpy=True))
+                self.optimizer.set_state_dict(
+                    paddlenlp_load(os.path.join(checkpoint, OPTIMIZER_NAME), return_numpy=True)
+                )
+
             self.lr_scheduler.set_state_dict(paddle.load(os.path.join(checkpoint, SCHEDULER_NAME)))
             if self.do_grad_scaling and os.path.isfile(os.path.join(checkpoint, SCALER_NAME)):
                 self.scaler.load_state_dict(paddle.load(os.path.join(checkpoint, SCALER_NAME), return_numpy=True))