Update configs.py

configs.py

@@ -35,7 +35,7 @@ model_info = {
                                  ),
     'CodeLLAMA 70B Instruct (Experimental)': dict(model_path='TheBloke/CodeLlama-70B-Instruct-GPTQ',
                                    token=os.environ['hf_token'],
-                                   wait_with_hidden_states=True, dont_cuda=True, device_map='cuda', disable_exllama=True,
+                                   wait_with_hidden_states=True, dont_cuda=True, device_map='cuda', # disable_exllama=True,
                                    original_prompt_template='<s>{prompt}',
                                    interpretation_prompt_template='<s>[INST] [X] [/INST] {prompt}',
                                    layers_format=llama_layers_format