修复损失计算方式，使用NLLLoss替代原始criterion

src/trainer/model.py

@@ -349,7 +349,8 @@ class MoEModel(nn.Module):
 
                 # 前向传播
                 logits = self(input_ids, attention_mask, pg)
-                loss = criterion(logits, labels)
+                log_probs = torch.log(probs + 1e-12)
+                loss = nn.NLLLoss()(log_probs, labels)
                 total_loss += loss.item() * labels.size(0)
 
                 # 计算准确率