From 1d26aa2e936e49bb5faec358d3b3accbb2a2fd02 Mon Sep 17 00:00:00 2001
From: ViperEkura <3081035982@qq.com>
Date: Tue, 26 May 2026 15:08:01 +0800
Subject: [PATCH] =?UTF-8?q?fix:=20=E7=A6=81=E7=94=A8DDP=20static=5Fgraph?=
 =?UTF-8?q?=E9=81=BF=E5=85=8DPyTorch=202.7.1=E4=B8=8Bno=5Fsync=E4=B8=8Ebac?=
 =?UTF-8?q?kward=E5=86=B2=E7=AA=81?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- static_graph=True时DDP.no_sync() + loss.backward()触发expect_autograd_hooks_内部断言
- PyTorch 2.7.1中no_sync上下文切换与静态图hook状态管理存在兼容性bug
- 将static_graph设为False恢复梯度累积正常执行
- find_unused_parameters保持False（模型无不参与计算的参数）
---
 scripts/tools/train.py | 2 --
 1 file changed, 2 deletions(-)

diff --git a/scripts/tools/train.py b/scripts/tools/train.py
index 5acbfef..f305352 100644
--- a/scripts/tools/train.py
+++ b/scripts/tools/train.py
@@ -255,8 +255,6 @@ def train(
     }
 
     executor_kwargs = {
-        "static_graph": True,
-        "find_unused_parameters": False,
         "gradient_as_bucket_view": True,
         "broadcast_buffers": False,
     }