fix ernie output

Author: cheng221

examples/config/ernie4_5/sft_argument_ernie4_5_0p3b.json

@@ -42,5 +42,6 @@
     "sequence_parallel": false,
     "report_to": "none",
     "convert_from_hf": true,
+    "save_to_hf": true,
     "pp_seg_method": "layer:DecoderLayer|EmptyLayer"
   }

examples/config/ernie4_5_moe/sft_argument_ernie4_5_21b_a3b.json

@@ -41,5 +41,6 @@
     "sequence_parallel": true,
     "report_to": "none",
     "convert_from_hf": true,
+    "save_to_hf": true,
     "pp_seg_method": "layer:DecoderLayer|EmptyLayer"
   }

paddleformers/transformers/ernie4_5/modeling.py

@@ -834,7 +834,8 @@ def forward(
 
         # Pretrain & Eval must have labels
         assert labels is not None
-        return self.criterion(logits, labels, loss_mask)
+        loss, _ = self.criterion(logits, labels, loss_mask)
+        return loss, logits
 
 
 class Ernie4_5ForCausalLMPipe(GeneralModelForCausalLMPipe):

paddleformers/transformers/ernie4_5_moe/modeling.py

@@ -1157,7 +1157,8 @@ def forward(
         # Pretrain & Eval must have labels
         assert labels is not None
 
-        return self.criterion(logits, labels, loss_mask, router_loss=router_loss, mtp_logits=mtp_logits)
+        loss, _ = self.criterion(logits, labels, loss_mask, router_loss=router_loss, mtp_logits=mtp_logits)
+        return loss, logits
 
 
 class Ernie4_5_MoeForCausalLMPipe(GeneralModelForCausalLMPipe):