Update app.py

app.py

@@ -284,9 +284,11 @@ def simple_chat(message: dict, temperature: float = 0.8, max_length: int = 4096,
         print("  ")
         print("--------------")
 
-        with torch.no_grad():
-            generated_ids = model.generate(input_ids['input_ids'], **generate_kwargs)
-            generated_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
+        # with torch.no_grad():
+        #     generated_ids = model.generate(input_ids['input_ids'], **generate_kwargs)
+        #     generated_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
+
+        generated_text = buffer
 
 
         return PlainTextResponse(generated_text)