code

app.py

@@ -3,15 +3,6 @@ import gradio as gr
 from llama_cpp import Llama
 import os
 
-# 初始化LLM
-llm = Llama.from_pretrained(
-    repo_id="matteogeniaccio/phi-4",
-    filename="phi-4-Q4_K_M.gguf",
-    verbose=True,
-    main_gpu=0,
-    n_gpu_layers=-1
-)
-
 # 响应函数
 @spaces.GPU
 def respond(
@@ -31,6 +22,13 @@ def respond(
             messages.append({"role": "assistant", "content": assistant_msg})
     messages.append({"role": "user", "content": message})
 
+    llm = Llama.from_pretrained(
+        repo_id="matteogeniaccio/phi-4",
+        filename="phi-4-Q4_K_M.gguf",
+        verbose=True,
+        main_gpu=0,
+        n_gpu_layers=-1
+    )
     # 使用llama-cpp-python的方式生成响应
     response = llm.create_chat_completion(
         messages=messages,