work on dialogue models

model.py

@@ -13,7 +13,7 @@ model = AutoModelForCausalLM.from_pretrained(
     config=config,
     #torch_dtype=torch.float16,
     #load_in_4bit=True,
-    device_map='cpu'
+    device_map='auto'
     )
 #else:
 #    model = None
@@ -55,7 +55,7 @@ def run(message: str,
     inputs = tokenizer([prompt], return_tensors='pt', add_special_tokens=False)
 
     streamer = TextIteratorStreamer(tokenizer,
-                                    timeout=40.,
+                                    timeout=60.,
                                     skip_prompt=True,
                                     skip_special_tokens=True)
     generate_kwargs = dict(