From 982d0521d515423a328b7de5865c01b7a4fd3437 Mon Sep 17 00:00:00 2001 From: songsenand Date: Fri, 13 Feb 2026 01:44:30 +0800 Subject: [PATCH] =?UTF-8?q?=E6=B7=BB=E5=8A=A0=E6=97=A5=E5=BF=97=E8=AE=B0?= =?UTF-8?q?=E5=BD=95=E5=92=8C=E7=A1=AE=E4=BF=9D=E6=A8=A1=E5=9E=8B=E5=A4=84?= =?UTF-8?q?=E4=BA=8E=E8=AE=AD=E7=BB=83=E6=A8=A1=E5=BC=8F?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- src/trainer/model.py | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/src/trainer/model.py b/src/trainer/model.py index 8183dcc..0852548 100644 --- a/src/trainer/model.py +++ b/src/trainer/model.py @@ -7,6 +7,8 @@ from modelscope import AutoModel import pickle from importlib.resources import files +from loguru import logger + from tqdm import tqdm from .monitor import TrainingMonitor @@ -426,16 +428,20 @@ class MoEModel(nn.Module): and global_step % eval_frequency == 0 ): acc, _ = self.model_eval(eval_dataloader, criterion) + super().train() if monitor is not None: monitor.add_step( global_step, {"loss": loss.item() * grad_accum_steps, "acc": acc}, ) + logger.info({"loss": loss.item() * grad_accum_steps, "acc": acc}) + elif monitor is not None: # 仅记录训练损失 monitor.add_step( global_step, {"loss": loss.item() * grad_accum_steps} ) + logger.info({"loss": loss.item() * grad_accum_steps}) # ============================ 使用示例 ============================