fix typo

Author: Difers

paddleformers/trainer/trainer_callback.py

@@ -31,6 +31,7 @@
 from tqdm.auto import tqdm
 
 from ..transformers.moe_gate import PretrainedMoEGate
+from ..transformers.moe_utils import offload, reload
 from ..utils.log import logger
 from .trainer_utils import IntervalStrategy, has_length
 from .training_args import TrainingArguments
@@ -688,8 +689,6 @@ def on_optimizer_begin(self, args, state, control, **kwargs):
                 reload(optimizer._master_weights[name])
 
 
-class MoECorrectionBiasAdjustCallback(TrainerCallback):
-    """used for moe aux loss free balance"""
 class MoECorrectionBiasAdjustCallback(TrainerCallback):
     """
     used for moe aux loss free balance

paddleformers/trainer/training_args.py

@@ -1100,7 +1100,7 @@ class TrainingArguments:
     reorder_pipeline_priority: Optional[bool] = field(
         default=False,
         metadata={"help": "Controls the parallel execution order. False (pp first), True (sharding first)."},
-    pp_seg_method: Optional[str] = field(default=None, metadata={"help": "PP Segmentation Method"})
+    )
     nccl_comm_group_config: Optional[str] = field(
         default=None,
         metadata={
@@ -2233,6 +2233,7 @@ def moe_sharding_parallel_rank(self):
             hcg = fleet.get_hybrid_communicate_group()
             if hasattr(hcg, "get_moe_sharding_parallel_group"):
                 return max(hcg.get_moe_sharding_parallel_group().rank, 0)
+
     def context_parallel_rank(self):
         if self.use_hybrid_parallel:
             hcg = fleet.get_hybrid_communicate_group()