Update app.py

app.py

@@ -11,17 +11,16 @@ MODEL_NAMES = {
 
 HF_TOKEN = os.getenv("HF_TOKEN")
 
-
 def load_model(model_path):
-    tokenizer = AutoTokenizer.from_pretrained(
-        model_path, trust_remote_code=True, token=HF_TOKEN
-    )
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True, token=HF_TOKEN)
     model = AutoModelForCausalLM.from_pretrained(
-        model_path, torch_dtype=torch.float16, trust_remote_code=True, token=HF_TOKEN
-    ).cuda()
+        model_path,
+        trust_remote_code=True,
+        token=HF_TOKEN,
+        device_map="auto",  # 讓 transformers 自動決定使用 CPU/GPU
+    )
     return model, tokenizer
 
-
 # 預設載入 DeepSeek-V3
 current_model, current_tokenizer = load_model("deepseek-ai/DeepSeek-V3")