From 6db70df2a7eacd0df619b79ae31ee6f3f0becc90 Mon Sep 17 00:00:00 2001
From: YWMditto <ditto@YWMdittodeMacBook-Pro.local>
Date: Wed, 31 Aug 2022 20:37:28 +0800
Subject: [PATCH] =?UTF-8?q?=E6=B7=BB=E5=8A=A0=E4=BA=86=20torch=5Fdriver.lo?=
 =?UTF-8?q?ad=5Fmodel=20=E5=9C=A8ddp=20=E6=97=B6=E5=87=BA=E7=8E=B0?=
 =?UTF-8?q?=E6=98=BE=E5=AD=98=E7=88=86=E7=82=B8=E7=9A=84todo=E4=BA=8B?=
 =?UTF-8?q?=E9=A1=B9?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 fastNLP/core/drivers/torch_driver/torch_driver.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/fastNLP/core/drivers/torch_driver/torch_driver.py b/fastNLP/core/drivers/torch_driver/torch_driver.py
index db011403..6ca33476 100644
--- a/fastNLP/core/drivers/torch_driver/torch_driver.py
+++ b/fastNLP/core/drivers/torch_driver/torch_driver.py
@@ -190,6 +190,7 @@ class TorchDriver(Driver):
         :param load_state_dict: 保存的内容是否只是权重
         """
         model = self.unwrap_model()
+        # todo torch.load 在加载时会使得卡 0 多出一个（甚至多个）model 的显存；因此在多卡断点重训时可能会出现错误；
         res = torch.load(filepath, map_location='cpu')
         if isinstance(res, dict) and only_state_dict is False:
             logger.rank_zero_warning(f"It seems like that {filepath} only contains state, you may need to use "