@@ -1,3 +1,10 @@
+import os
+
+# 禁用 FlashAttention,解决沐曦显卡共享内存不足问题
+os.environ["FLASH_ATTENTION_ENABLED"] = "0"
+os.environ["PYTORCH_NO_FLASH"] = "1"
+os.environ["VLLM_ATTENTION_BACKEND"] = "NO_ATTENTION"
import asyncio
import json
from pathlib import Path