adapt fc

examples/pre-training/ernie/pretrain.py

@@ -296,16 +296,16 @@ def formatv(v):
         and not args.overwrite_output_dir
     ):
         last_checkpoint = get_last_checkpoint(args.output_dir)
-        if last_checkpoint is None and len(os.listdir(args.output_dir)) > 0:
-            raise ValueError(
-                f"Output directory ({args.output_dir}) already exists and is not empty. "
-                "Use --overwrite_output_dir to overcome."
-            )
-        elif last_checkpoint is not None and args.resume_from_checkpoint is None:
-            logger.info(
-                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
-                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
-            )
+        # if last_checkpoint is None and len(os.listdir(args.output_dir)) > 0:
+        #     raise ValueError(
+        #         f"Output directory ({args.output_dir}) already exists and is not empty. "
+        #         "Use --overwrite_output_dir to overcome."
+        #     )
+        # elif last_checkpoint is not None and args.resume_from_checkpoint is None:
+        #     logger.info(
+        #         f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+        #         "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+        #     )
 
     def compute_metrics(p):
         preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions
@@ -439,6 +439,7 @@ def sname_to_tname(pp_model):
     cfg.token_balance_seqlen = args.max_seq_length * args.per_device_train_batch_size
     cfg.fp16_opt_level = args.fp16_opt_level
     cfg.moe_group = args.moe_group
+    cfg.moe_group_name = args.moe_group
     cfg.dtype = dtype
     cfg.use_fp8 = args.use_fp8
     cfg.enable_mtp_magic_send = args.enable_mtp_magic_send
@@ -502,7 +503,7 @@ def sname_to_tname(pp_model):
     logger.info(f"using model type:{type(model)}")
     paddle.set_default_dtype("float32")
 
-    logger.info(f"using model={type(model)}, cfg={cfg}")
+    # logger.info(f"using model={type(model)}, cfg={cfg}")
 
     train_dataset, eval_dataset, test_dataset, data_collator = (
         create_pretrained_dataset(args)

examples/pre-training/ernie/src/trainers/pretraining_trainer.py

@@ -1260,7 +1260,7 @@ def _maybe_log_save_evaluate(
             )
             logs["learning_rate"] = float(self._get_learning_rate())
             logs["global_step"] = int(self.state.global_step)
-
+            logs["loss_md5"] = paddle.to_tensor(logs["loss"])._md5sum()
             divisor = 2**30
 
             current_device = framework._current_expected_place_()

examples/pre-training/models/ernie/configuration.py

@@ -149,6 +149,7 @@ def __init__(
         global_aux_loss=False,
         moe_dropout_prob=0.0,
         moe_group="world",
+        moe_group_name="world",
         num_experts_per_tok: int = 8,
         moe_intermediate_size: Union[int, list] = 0,
         moe_num_shared_experts: int = 0,
@@ -356,6 +357,7 @@ def update_nested_dict(default_dict, update_dict):
         self.moe_layer_interval = moe_layer_interval
         self.moe_dropout_prob = moe_dropout_prob
         self.moe_group = moe_group
+        self.moe_group_name = moe_group_name
         self.num_experts_per_tok = num_experts_per_tok
         self.moe_num_shared_experts = moe_num_shared_experts
         self.moe_num_dense_experts = moe_num_dense_experts
@@ -395,7 +397,6 @@ def update_nested_dict(default_dict, update_dict):
 
         self.use_linear_residual_norm_recompute = use_linear_residual_norm_recompute
         self.use_rms_qkv_recompute = use_rms_qkv_recompute
-
         assert aux_loss_type in ["", "default", "seq_aux_loss", "switch_aux_loss"]
         self.aux_loss_type = aux_loss_type