support load hf ckpt (#10976)

Author: zhangbo9674

paddlenlp/trainer/trainer.py

@@ -170,6 +170,7 @@
     nested_numpify,
     nested_truncate,
 )
+from .utils.load_hf_ckpt import load_huggingface_ckpt
 from .utils.sharding_io import ShardingIO
 
 DEFAULT_CALLBACKS = [DefaultFlowCallback]
@@ -1009,6 +1010,9 @@ def _inner_training_loop(
         if self.args.ignore_data_skip:
             self.timers and self.timers("read-data").start()
 
+        if self.args.resume_from_huggingface_ckpt is not None:
+            load_huggingface_ckpt(model, self.args.resume_from_huggingface_ckpt)
+
         for epoch in range(epochs_trained, num_train_epochs):
             if isinstance(train_dataloader, paddle.io.DataLoader) and isinstance(
                 train_dataloader.batch_sampler, DistributedBatchSampler

paddlenlp/trainer/training_args.py

@@ -885,6 +885,10 @@ class TrainingArguments:
         default=None,
         metadata={"help": "The path to a folder with a valid checkpoint for your model."},
     )
+    resume_from_huggingface_ckpt: Optional[str] = field(
+        default=None,
+        metadata={"help": "The path to a folder with a valid huggingface checkpoint for your model."},
+    )
     auto_parallel_resume_form_hybrid_parallel: Optional[bool] = field(
         default=False,
         metadata={"help": "Wether hybrid paralle checkpoints be loaded in auto parallel mode."},