try

app.py

@@ -66,6 +66,21 @@ def generate(
     top_k: int = 50,
     repetition_penalty: float = 1.2,
 ) -> Iterator[str]:
+
+    ## check inside function
+    # 检查CUDA是否可用
+    if torch.cuda.is_available():
+        print("check inside function: CUDA is available. Listing available GPUs:")
+        # 获取并打印GPU数量
+        num_gpus = torch.cuda.device_count()
+        for i in range(num_gpus):
+            print(f"check inside function: GPU {i}: {torch.cuda.get_device_name(i)}")
+            # 其他相关信息，例如内存
+            print(f"check inside function:   Memory Allocated: {torch.cuda.memory_allocated(i) / 1024 ** 2:.0f} MB")
+            print(f"check inside function:  Memory Reserved: {torch.cuda.memory_reserved(i) / 1024 ** 2:.0f} MB")
+    else:
+        print("check inside function: CUDA is not available.")
+
     conversation = []
     if system_prompt:
         conversation.append({"role": "system", "content": system_prompt})