Spaces:
Running
Running
| 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) |