app.py

import gradio as gr
import models
import pandas as pd
import theme
import matplotlib.pyplot as plt
    
text = "<h1 style='text-align: center; color: #333333; font-size: 40px;'>TCO Comparison Calculator"
text2 = "Please note that the cost/request only defines the infrastructure cost for deployment. The labor cost must be added for the whole AI model service deployment TCO."
description=f"""
<p>In this demo application, we help you compare different AI model services, such as Open source or SaaS solutions, based on the Total Cost of Ownership for their deployment. 😊</p> 
<p>Please note that we focus on getting the service up and running, but not the maintenance that follows.🚀</p>
<p>If you want to <strong>contribute to the calculator</strong> by adding your own AI service option, follow this <a href="https://huggingface.co/spaces/mithril-security/TCO_calculator/blob/main/How_to_contribute.md">tutorial</a> 👈. </p>
"""
formula = r"""
$CR = \frac{CIT_{1K} \times IT + COT_{1K} \times OT}{1000}$  <br>
with: <br>
$CR$ = Cost per Request <br>
$CIT_{1K}$ = Cost per 1000 Input Tokens <br>
$COT_{1K}$ = Cost per 1000 Output Tokens <br>
$IT$ = Input Tokens <br>
$OT$ = Output Tokens
"""

def on_use_case_change(use_case):
    if use_case == "Summarize":
        return gr.update(value=500), gr.update(value=200)
    elif use_case == "Question-Answering":
        return gr.update(value=300), gr.update(value=300)
    else:
        return gr.update(value=50), gr.update(value=10)
    
def compare_info(tco1, tco2, dropdown, dropdown2):
    #Compute the cost/request ratio
    r = tco1 / tco2   
    if r < 1:
        comparison_result = f"""The cost/request of the second {dropdown2} service is <b>{1/r:.5f} times more expensive</b> than the one of the first {dropdown} service."""   
    elif r > 1:
        comparison_result = f"""The cost/request of the second {dropdown2} service is <b>{r:.5f} times cheaper</b> than the one of the first {dropdown} service."""
    else:
        comparison_result = f"""Both solutions have the <b>same cost/request</b>."""
        
    # Create a bar chart
    services = [dropdown, dropdown2]
    costs_to_compare = [tco1, tco2]
    
    plt.figure(figsize=(6, 4))
    plt.bar(services, costs_to_compare, color=['red', 'green'])
    plt.xlabel('AI option services', fontsize=10)
    plt.ylabel('($) Cost/Request', fontsize=10)
    plt.title('Comparison of Cost/Request', fontsize=14)

    plt.tight_layout()
    plt.savefig('cost_comparison.png')  # Save to a file

    return gr.update(value='cost_comparison.png', visible=True), comparison_result

def create_table(tco1, tco2, labor_cost1, labor_cost2, dropdown, dropdown2, latency, latency2):
    list_values = []
    first_sol = [tco1, labor_cost1, latency]
    second_sol = [tco2, labor_cost2, latency2]
    list_values.append(first_sol)
    list_values.append(second_sol)

    data = pd.DataFrame(list_values, index=[dropdown, dropdown2], columns=["Cost/request ($) ", "Labor Cost ($/month)", "Average latency (s)"])
    
    formatted_data = data.copy()
    formatted_data["Cost/request ($) "] = formatted_data["Cost/request ($) "].apply('{:.5f}'.format)
    formatted_data["Labor Cost ($/month)"] = formatted_data["Labor Cost ($/month)"].apply('{:.0f}'.format)

    styled_data = formatted_data.style\
        .set_properties(**{'background-color': '#ffffff', 'color': '#000000', 'border-color': '#e0e0e0', 'border-width': '1px', 'border-style': 'solid'})\
        .to_html()
    centered_styled_data = f"<center>{styled_data}</center>"
    
    return gr.update(value=centered_styled_data)

def compute_cost_per_request(*args):
    dropdown_id = args[-2]
    dropdown_id2 = args[-1]
    if dropdown_id!=None and dropdown_id2!=None:
        # Separate the arguments for page1 and page2
        args_list = list(args)
        args_page1 = args_list[:len(page1.get_all_components_for_cost_computing())] + [dropdown_id, input_tokens, output_tokens]
        args_page2 = args_list[len(page1.get_all_components_for_cost_computing()):] + [dropdown_id2, input_tokens, output_tokens]
        # Compute and compare using both pages
        result_page1 = page1.compute_cost_per_token(*args_page1)
        result_page2 = page2.compute_cost_per_token(*args_page2)
        # Unpack the results from the functions
        tco1, latency, labor_cost1 = result_page1
        tco2, latency2, labor_cost2 = result_page2
        
        return tco1, latency, labor_cost1, tco2, latency2, labor_cost2
    else: 
        raise gr.Error("Please select two AI service options.")
    
    

def update_plot(tco1, tco2, dropdown, dropdown2, labour_cost1, labour_cost2):
    
    request_ranges = list(range(0, 1001, 100)) + list(range(1000, 10001, 500)) + list(range(10000, 100001, 1000)) + list(range(100000, 2000001, 100000))
    costs_tco1 = [(tco1 * req + labour_cost1) for req in request_ranges]
    costs_tco2 = [(tco2 * req + labour_cost2) for req in request_ranges]

    data = pd.DataFrame({
        "Number of requests": request_ranges * 2,
        "Cost ($)": costs_tco1 + costs_tco2,
        "AI model service": ["1)" + " " + dropdown] * len(request_ranges) + ["2)" + " " + dropdown2] * len(request_ranges)
        }
    )
    return gr.LinePlot.update(data, visible=True, x="Number of requests", y="Cost ($)",color="AI model service",color_legend_position="bottom", title="Set-up TCO for one month", height=300, width=500, tooltip=["Number of requests", "Cost ($)", "AI model service"])

style = theme.Style()

with gr.Blocks(theme=style) as demo:
    Models: list[models.BaseTCOModel] = [models.OpenAIModelGPT4, models.OpenAIModelGPT3_5, models.CohereModel, models.DIYLlama2Model]
    model_names = [Model().get_name() for Model in Models]
    gr.Markdown(value=text)
    gr.Markdown(value=description)
    
    with gr.Row():
        with gr.Column():
            with gr.Row():
                use_case = gr.Dropdown(["Summarize", "Question-Answering", "Classification"], value="Question-Answering", label=" Describe your use case ")
            with gr.Accordion("Click here if you want to customize the number of input and output tokens per request", open=False):    
                with gr.Row():
                    input_tokens = gr.Slider(minimum=1, maximum=1000, value=300, step=1, label=" Input tokens per request", info="We suggest a value that we believe best suit your use case choice but feel free to adjust", interactive=True)
                    output_tokens = gr.Slider(minimum=1, maximum=1000, value=300, step=1, label=" Output tokens per request", info="We suggest a value that we believe best suit your use case choice but feel free to adjust", interactive=True)
                with gr.Row(visible=False):    
                    num_users = gr.Number(value="1000", interactive = True, label=" Number of users for your service ")
    
    use_case.change(on_use_case_change, inputs=use_case, outputs=[input_tokens, output_tokens])
    
    with gr.Row():
        with gr.Column():
            page1 = models.ModelPage(Models)
            dropdown = gr.Dropdown(model_names, interactive=True, label=" First AI service option ")
            with gr.Accordion("Click here for more information on the computation parameters for your first AI service option", open=False):    
                page1.render()

        with gr.Column():
            page2 = models.ModelPage(Models)
            dropdown2 = gr.Dropdown(model_names, interactive=True, label=" Second AI service option ")
            with gr.Accordion("Click here for more information on the computation parameters for your second AI service option", open=False):        
                page2.render()
            
    dropdown.change(page1.make_model_visible, inputs=[dropdown, use_case], outputs=page1.get_all_components())
    dropdown2.change(page2.make_model_visible, inputs=[dropdown2, use_case], outputs=page2.get_all_components())
    
    compute_tco_btn = gr.Button("Compute & Compare", size="lg", variant="primary", scale=1) 
    tco1 = gr.State()
    tco2 = gr.State()
    labor_cost1 = gr.State()
    labor_cost2 = gr.State()
    latency = gr.State()
    latency2 = gr.State()
    
    with gr.Row():
        with gr.Accordion("Click here to see the cost/request computation formula", open=False):
            tco_formula = gr.Markdown(formula)
        
    with gr.Row(variant='panel'):
        with gr.Column():
            with gr.Row():
                table = gr.Markdown()
            with gr.Row():
                info = gr.Markdown(text2)
            with gr.Row():
                with gr.Column(scale=1):
                    image = gr.Image(visible=False)
                    ratio = gr.Markdown()
                with gr.Column(scale=2):
                    plot = gr.LinePlot(visible=False)
    
    compute_tco_btn.click(compute_cost_per_request, inputs=page1.get_all_components_for_cost_computing() + page2.get_all_components_for_cost_computing() + [dropdown, dropdown2], outputs=[tco1, latency, labor_cost1, tco2, latency2, labor_cost2]).then(create_table, inputs=[tco1, tco2, labor_cost1, labor_cost2, dropdown, dropdown2, latency, latency2], outputs=table).then(compare_info, inputs=[tco1, tco2, dropdown, dropdown2], outputs=[image, ratio]).then(update_plot, inputs=[tco1, tco2, dropdown, dropdown2, labor_cost1, labor_cost2], outputs=plot)

demo.launch(debug=True)