From 02f851205f4b05a5c62bbcc3ef3028f55f54f580 Mon Sep 17 00:00:00 2001
From: songsenand <songsenand@163.com>
Date: Fri, 13 Feb 2026 11:29:22 +0800
Subject: [PATCH] =?UTF-8?q?=E4=BF=AE=E5=A4=8D=E5=91=A8=E6=9C=9F=E6=80=A7?=
 =?UTF-8?q?=E8=AF=84=E4=BC=B0=E6=97=B6=E5=B9=B3=E5=9D=87=E6=8D=9F=E5=A4=B1?=
 =?UTF-8?q?=E8=AE=A1=E7=AE=97=E9=94=99=E8=AF=AF?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 src/trainer/model.py | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/src/trainer/model.py b/src/trainer/model.py
index e78c4ec..dbb4f05 100644
--- a/src/trainer/model.py
+++ b/src/trainer/model.py
@@ -437,13 +437,12 @@ class MoEModel(nn.Module):
                     global_step += 1
                     original_loss = loss.item() * grad_accum_steps
                     batch_loss_sum += original_loss
-
                     # 周期性评估（与原代码相同）
                     if (
                         eval_dataloader is not None
                         and global_step % eval_frequency == 0
                     ):
-                        avg_loss = batch_loss_sum / global_step
+                        avg_loss = batch_loss_sum / eval_frequency
                         acc, _ = self.model_eval(eval_dataloader, criterion)
                         super().train()
                         if monitor is not None: