Update app.py

app.py

@@ -98,8 +98,8 @@ with gr.Blocks() as demo:
                     inputs['input_ids'] = torch.cat([past_ids, inputs.input_ids], dim=1) 
                 outputs = infer.model.generate(
                     **inputs, past_key_values=state.get('past_key_values', None), 
-                    return_dict_in_generate=True, do_sample=do_sample, 
-                    repetition_penalty=repetition_penalty,
+                    return_dict_in_generate=True, do_sample=False, 
+                    repetition_penalty=1.05,
                     max_new_tokens=512,
                 )
                 state['past_key_values'] = outputs.past_key_values