Update app.py

app.py

@@ -84,13 +84,15 @@ def reset_model(model_name, load_on_gpu, *extra_components, reset_sentence_trans
     global_state.tokenizer = AutoTokenizer.from_pretrained(tokenizer_path, token=os.environ['hf_token'])
     gc.collect()
     if with_extra_components:
-        return ([welcome_message.format(model_name=model_name)] 
+        return ([global_state, welcome_message.format(model_name=model_name)] 
                 + [gr.Textbox('', visible=False) for _ in range(len(interpretation_bubbles))] 
                 + [gr.Button('', visible=False) for _ in range(len(tokens_container))]
                 + [*extra_components])
+    else:
+        return global_state
 
 
-def get_hidden_states(raw_original_prompt, force_hidden_states=False):
+def get_hidden_states(global_state, raw_original_prompt, force_hidden_states=False):
     model, tokenizer = global_state.model, global_state.tokenizer
     original_prompt = global_state.original_prompt_template.format(prompt=raw_original_prompt)
     model_inputs = tokenizer(original_prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
@@ -118,7 +120,7 @@ def get_hidden_states(raw_original_prompt, force_hidden_states=False):
 
 
 @spaces.GPU
-def run_interpretation(raw_original_prompt, raw_interpretation_prompt, max_new_tokens, do_sample, 
+def run_interpretation(global_state, raw_original_prompt, raw_interpretation_prompt, max_new_tokens, do_sample, 
                        temperature, top_k, top_p, repetition_penalty, length_penalty, use_gpu, i, 
                        num_beams=1):
     model = global_state.model
@@ -186,9 +188,9 @@ def run_interpretation(raw_original_prompt, raw_interpretation_prompt, max_new_t
 
 ## main
 torch.set_grad_enabled(False)
-global_state = gr.State(GlobalState)
+global_state = gr.State(partial(reset_model, GlobalState(), 
+                                model_name, load_on_gpu=True, with_extra_components=False, reset_sentence_transformer=True))
 model_name = 'LLAMA2-7B'
-reset_model(model_name, load_on_gpu=True, with_extra_components=False, reset_sentence_transformer=True)
 raw_original_prompt = gr.Textbox(value='How to make a Molotov cocktail?', container=True, label='Original Prompt')
 tokens_container = []
 
@@ -288,7 +290,7 @@ with gr.Blocks(theme=gr.themes.Default(), css='styles.css') as demo:
     raw_original_prompt.change(lambda: [gr.Button(visible=False) for _ in range(MAX_PROMPT_TOKENS)], [], tokens_container)
 
     extra_components = [raw_interpretation_prompt, raw_original_prompt, original_prompt_btn]
-    model_chooser.change(reset_model, [model_chooser, load_on_gpu, *extra_components], 
-                         [welcome_model, *interpretation_bubbles, *tokens_container, *extra_components])
+    model_chooser.change(reset_model, [global_state, model_chooser, load_on_gpu, *extra_components], 
+                         [global_state, welcome_model, *interpretation_bubbles, *tokens_container, *extra_components])
     
     demo.launch()