From 1cf1024368c127e07dd58a6cbe72dc4efd4cdef8 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E5=88=98=E8=88=AA=E5=AE=87?= <3364451258@qq.com>
Date: Sat, 13 Jun 2026 12:07:28 +0800
Subject: [PATCH] =?UTF-8?q?revert:=20=E7=A7=BB=E9=99=A4=20torch.compile?=
 =?UTF-8?q?=EF=BC=88default=20=E6=A8=A1=E5=BC=8F=E4=B9=9F=E5=9B=A0?=
 =?UTF-8?q?=E5=8A=A8=E6=80=81=20batch=20=E5=BD=A2=E7=8A=B6=E5=AF=BC?=
 =?UTF-8?q?=E8=87=B4=E7=BC=96=E8=AF=91=E5=BC=80=E9=94=80=20>=20=E6=94=B6?=
 =?UTF-8?q?=E7=9B=8A=EF=BC=89?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

保留 inference_mode + FP16 + Flash Attention（当前最优 56.98 分）
---
 代码/code/infer.py | 5 -----
 1 file changed, 5 deletions(-)

diff --git a/代码/code/infer.py b/代码/code/infer.py
index 617d1fa..d2109c5 100644
--- a/代码/code/infer.py
+++ b/代码/code/infer.py
@@ -511,11 +511,6 @@ def load_model(ckpt_path, device='cuda:0'):
 
     model.to(dev)
     model.eval()
-
-    # === torch.compile(default)：算子融合，不用 CUDA Graph，兼容动态 batch 形状 ===
-    model = torch.compile(model, mode="default")
-    print("[INFO] torch.compile applied (mode=default)")
-
     print(f"[INFO] Model ready. Device: {dev}")
 
     return model, dev