Update configs.py

configs.py

@@ -34,7 +34,8 @@ model_info = {
                                   layers_format=llama_layers_format
                                  ),
     'CodeLLAMA 70B Instruct (Experimental)': dict(model_path='TheBloke/CodeLlama-70B-Instruct-GPTQ',
-                                   token=os.environ['hf_token'], wait_with_hidden_states=True,
+                                   token=os.environ['hf_token'],
+                                   wait_with_hidden_states=True, dont_cuda=True, device_map='cuda',
                                    original_prompt_template='<s>{prompt}',
                                    interpretation_prompt_template='<s>[INST] [X] [/INST] {prompt}',
                                    layers_format=llama_layers_format