algorithmicsuperintelligence/OptiLLM

1import os2import gradio as gr3 4from openai import OpenAI5 6from optillm.cot_reflection import cot_reflection7from optillm.rto import round_trip_optimization8from optillm.z3_solver import Z3SymPySolverSystem9from optillm.self_consistency import advanced_self_consistency_approach10from optillm.plansearch import plansearch11from optillm.leap import leap12from optillm.reread import re2_approach13 14 15API_KEY = os.environ.get("OPENROUTER_API_KEY")16 17def compare_responses(message, model1, approach1, model2, approach2, system_message, max_tokens, temperature, top_p):18    response1 = respond(message, [], model1, approach1, system_message, max_tokens, temperature, top_p)19    response2 = respond(message, [], model2, approach2, system_message, max_tokens, temperature, top_p)20    return response1, response221 22def parse_conversation(messages):23    system_prompt = ""24    conversation = []25    26    for message in messages:27        role = message['role']28        content = message['content']29        30        if role == 'system':31            system_prompt = content32        elif role in ['user', 'assistant']:33            conversation.append(f"{role.capitalize()}: {content}")34    35    initial_query = "\n".join(conversation)36    return system_prompt, initial_query37 38def respond(message, history, model, approach, system_message, max_tokens, temperature, top_p):39    try:40        client = OpenAI(api_key=API_KEY, base_url="https://openrouter.ai/api/v1")41        messages = [{"role": "system", "content": system_message}]42        43        for val in history:44            if val[0]:45                messages.append({"role": "user", "content": val[0]})46            if val[1]:47                messages.append({"role": "assistant", "content": val[1]})48        49        messages.append({"role": "user", "content": message})50        51        if approach == "none":52            response = client.chat.completions.create(53                extra_headers={54                    "HTTP-Referer": "https://github.com/codelion/optillm",55                    "X-Title": "optillm"56                },57                model=model,58                messages=messages,59                max_tokens=max_tokens,60                temperature=temperature,61                top_p=top_p,62            )63            return response.choices[0].message.content64        else:65            system_prompt, initial_query = parse_conversation(messages)66            67            if approach == 'rto':68                final_response, _ = round_trip_optimization(system_prompt, initial_query, client, model)69            elif approach == 'z3':70                z3_solver = Z3SymPySolverSystem(system_prompt, client, model)71                final_response, _ = z3_solver.process_query(initial_query)72            elif approach == "self_consistency":73                final_response, _ = advanced_self_consistency_approach(system_prompt, initial_query, client, model)74            elif approach == "cot_reflection":75                final_response, _ = cot_reflection(system_prompt, initial_query, client, model)76            elif approach == 'plansearch':77                response, _ = plansearch(system_prompt, initial_query, client, model)78                final_response = response[0]79            elif approach == 'leap':80                final_response, _ = leap(system_prompt, initial_query, client, model)81            elif approach == 're2':82                final_response, _ = re2_approach(system_prompt, initial_query, client, model)83            84            return final_response85            86    except Exception as e:87        error_message = f"Error in respond function: {str(e)}\nType: {type(e).__name__}"88        print(error_message)89 90def create_model_dropdown():91    return gr.Dropdown(92        [ "meta-llama/llama-3.1-8b-instruct:free", "nousresearch/hermes-3-llama-3.1-405b:free","meta-llama/llama-3.2-1b-instruct:free",93         "mistralai/mistral-7b-instruct:free","mistralai/pixtral-12b:free","meta-llama/llama-3.1-70b-instruct:free",94         "qwen/qwen-2-7b-instruct:free", "qwen/qwen-2-vl-7b-instruct:free", "google/gemma-2-9b-it:free", "liquid/lfm-40b:free", "meta-llama/llama-3.1-405b-instruct:free",95         "openchat/openchat-7b:free", "meta-llama/llama-3.2-90b-vision-instruct:free", "meta-llama/llama-3.2-11b-vision-instruct:free",96         "meta-llama/llama-3-8b-instruct:free", "meta-llama/llama-3.2-3b-instruct:free", "microsoft/phi-3-medium-128k-instruct:free",97         "microsoft/phi-3-mini-128k-instruct:free", "huggingfaceh4/zephyr-7b-beta:free"],98        value="meta-llama/llama-3.2-1b-instruct:free", label="Model"99    )100 101def create_approach_dropdown():102    return gr.Dropdown(103        ["none", "leap", "plansearch", "cot_reflection", "rto", "self_consistency", "z3", "re2"],104        value="none", label="Approach"105    )106 107html = """<iframe src="https://ghbtns.com/github-btn.html?user=codelion&repo=optillm&type=star&count=true&size=large" frameborder="0" scrolling="0" width="170" height="30" title="GitHub"></iframe>108"""109 110with gr.Blocks() as demo:111    gr.Markdown("# OptiLLM - Optimizing LLM Inference")112    gr.HTML(html)113    114    with gr.Row():115        system_message = gr.Textbox(value="", label="System message")116        max_tokens = gr.Slider(minimum=1, maximum=4096, value=1024, step=1, label="Max new tokens")117        temperature = gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature")118        top_p = gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)")119    120    with gr.Tabs():121        with gr.TabItem("Chat"):122            model = create_model_dropdown()123            approach = create_approach_dropdown()124            chatbot = gr.Chatbot()125            msg = gr.Textbox()126            with gr.Row():127                submit = gr.Button("Submit")128                clear = gr.Button("Clear")129 130            def user(user_message, history):131                return "", history + [[user_message, None]]132 133            def bot(history, model, approach, system_message, max_tokens, temperature, top_p):134                user_message = history[-1][0]135                bot_message = respond(user_message, history[:-1], model, approach, system_message, max_tokens, temperature, top_p)136                history[-1][1] = bot_message137                return history138 139            msg.submit(user, [msg, chatbot], [msg, chatbot]).then(140                bot, [chatbot, model, approach, system_message, max_tokens, temperature, top_p], chatbot141            )142            submit.click(user, [msg, chatbot], [msg, chatbot]).then(143                bot, [chatbot, model, approach, system_message, max_tokens, temperature, top_p], chatbot144            )145            clear.click(lambda: None, None, chatbot, queue=False)146 147        with gr.TabItem("Compare"):148            with gr.Row():149                model1 = create_model_dropdown()150                approach1 = create_approach_dropdown()151                model2 = create_model_dropdown()152                approach2 = create_approach_dropdown()153            154            compare_input = gr.Textbox(label="Enter your message for comparison")155            compare_button = gr.Button("Compare")156            157            with gr.Row():158                output1 = gr.Textbox(label="Response 1")159                output2 = gr.Textbox(label="Response 2")160            161            compare_button.click(162                compare_responses,163                inputs=[compare_input, model1, approach1, model2, approach2, system_message, max_tokens, temperature, top_p],164                outputs=[output1, output2]165            )166 167if __name__ == "__main__":168    demo.launch()