app.py

import gradio as gr
import pandas as pd
import numpy as np
import plotly.express as px
import plotly.graph_objects as go

# Updated pricing data - restructured for better comparison
aws_instances = {
    # T4 GPU Instances (entry level)
    "g4dn.xlarge": {"vcpus": 4, "memory": 16, "gpu": "1x NVIDIA T4", "hourly_rate": 0.526, "gpu_memory": "16GB", "tier": "Entry"},
    "g4dn.2xlarge": {"vcpus": 8, "memory": 32, "gpu": "1x NVIDIA T4", "hourly_rate": 0.752, "gpu_memory": "16GB", "tier": "Entry"},
    
    # A10G GPU Instances (mid-tier)
    "g5.xlarge": {"vcpus": 4, "memory": 16, "gpu": "1x NVIDIA A10G", "hourly_rate": 0.65, "gpu_memory": "24GB", "tier": "Mid"},
    "g5.2xlarge": {"vcpus": 8, "memory": 32, "gpu": "1x NVIDIA A10G", "hourly_rate": 1.006, "gpu_memory": "24GB", "tier": "Mid"},
    
    # V100 GPU Instances (high-tier)
    "p3.2xlarge": {"vcpus": 8, "memory": 61, "gpu": "1x NVIDIA V100", "hourly_rate": 3.06, "gpu_memory": "16GB", "tier": "High"},
    
    
    # Added comparable instances to match GCP
    "p4d.xlarge": {"vcpus": 12, "memory": 85, "gpu": "1x NVIDIA A100", "hourly_rate": 4.10, "gpu_memory": "40GB", "tier": "Premium"},
    "p4d.2xlarge": {"vcpus": 24, "memory": 170, "gpu": "2x NVIDIA A100", "hourly_rate": 8.20, "gpu_memory": "2x40GB", "tier": "Premium"},
    "p4d.4xlarge": {"vcpus": 48, "memory": 340, "gpu": "4x NVIDIA A100", "hourly_rate": 16.40, "gpu_memory": "4x40GB", "tier": "Premium"},
}

gcp_instances = {
    # T4 GPU Instances (entry level)
    "n1-standard-4-t4": {"vcpus": 4, "memory": 15, "gpu": "1x NVIDIA T4", "hourly_rate": 0.49, "gpu_memory": "16GB", "tier": "Entry"},
    "n1-standard-8-t4": {"vcpus": 8, "memory": 30, "gpu": "1x NVIDIA T4", "hourly_rate": 0.69, "gpu_memory": "16GB", "tier": "Entry"},
    
    # L4 GPU Instances (mid-tier)
    "g2-standard-4": {"vcpus": 4, "memory": 16, "gpu": "1x NVIDIA L4", "hourly_rate": 0.59, "gpu_memory": "24GB", "tier": "Mid"},
    "g2-standard-8": {"vcpus": 8, "memory": 32, "gpu": "1x NVIDIA L4", "hourly_rate": 0.89, "gpu_memory": "24GB", "tier": "Mid"},
    
    # Added comparable V100 instance
    "n1-standard-8-v100": {"vcpus": 8, "memory": 60, "gpu": "1x NVIDIA V100", "hourly_rate": 2.95, "gpu_memory": "16GB", "tier": "High"},
    
    # A100 GPU Instances (premium)
    "a2-highgpu-1g": {"vcpus": 12, "memory": 85, "gpu": "1x NVIDIA A100", "hourly_rate": 1.46, "gpu_memory": "40GB", "tier": "Premium"},
    "a2-highgpu-2g": {"vcpus": 24, "memory": 170, "gpu": "2x NVIDIA A100", "hourly_rate": 2.93, "gpu_memory": "2x40GB", "tier": "Premium"},
    "a2-highgpu-4g": {"vcpus": 48, "memory": 340, "gpu": "4x NVIDIA A100", "hourly_rate": 5.86, "gpu_memory": "4x40GB", "tier": "Premium"},
}

api_pricing = {
    "OpenAI": {
        "GPT-3.5-Turbo": {"input_per_1M": 0.5, "output_per_1M": 1.5, "token_context": 16385},
        "GPT-4o": {"input_per_1M": 5.0, "output_per_1M": 15.0, "token_context": 32768},
        "GPT-4o-mini": {"input_per_1M": 2.5, "output_per_1M": 7.5, "token_context": 32768},
    },
    "TogetherAI": {
        "Llama-3-8B": {"input_per_1M": 0.15, "output_per_1M": 0.15, "token_context": 8192},
        "Llama-3-70B": {"input_per_1M": 0.9, "output_per_1M": 0.9, "token_context": 8192},
        "Llama-2-13B": {"input_per_1M": 0.6, "output_per_1M": 0.6, "token_context": 4096},
        "Llama-2-70B": {"input_per_1M": 2.5, "output_per_1M": 2.5, "token_context": 4096},
        "DeepSeek-Coder-33B": {"input_per_1M": 2.0, "output_per_1M": 2.0, "token_context": 16384},
    },
    "Anthropic": {
        "Claude-3-Opus": {"input_per_1M": 15.0, "output_per_1M": 75.0, "token_context": 200000},
        "Claude-3-Sonnet": {"input_per_1M": 3.0, "output_per_1M": 15.0, "token_context": 200000},
        "Claude-3-Haiku": {"input_per_1M": 0.25, "output_per_1M": 1.25, "token_context": 200000},
    }
}

model_sizes = {
    "Small (7B parameters)": {"memory_required": 14},
    "Medium (13B parameters)": {"memory_required": 26},
    "Large (70B parameters)": {"memory_required": 140},
    "XL (180B parameters)": {"memory_required": 360},
}

def calculate_aws_cost(instance, hours, storage, reserved=False, spot=False, years=1):
    data = aws_instances[instance]
    rate = data['hourly_rate']
    if spot:
        rate *= 0.3
    elif reserved:
        factors = {1: 0.6, 3: 0.4}
        rate *= factors.get(years, 0.6)
    compute = rate * hours
    storage_cost = storage * 0.10
    return {'total_cost': compute + storage_cost, 'details': data}

def calculate_gcp_cost(instance, hours, storage, reserved=False, spot=False, years=1):
    data = gcp_instances[instance]
    rate = data['hourly_rate']
    if spot:
        rate *= 0.2
    elif reserved:
        factors = {1: 0.7, 3: 0.5}
        rate *= factors.get(years, 0.7)
    compute = rate * hours
    storage_cost = storage * 0.04
    return {'total_cost': compute + storage_cost, 'details': data}

def calculate_api_cost(provider, model, in_tokens, out_tokens, calls):
    m = api_pricing[provider][model]
    input_cost = in_tokens * m['input_per_1M'] / 1000000
    output_cost = out_tokens * m['output_per_1M'] / 1000000
    call_cost = calls * 0.0001 if provider == 'TogetherAI' else 0
    return {'total_cost': input_cost + output_cost + call_cost, 'details': m}

def filter_compatible(instances, min_mem):
    res = {}
    for name, data in instances.items():
        mem_str = data['gpu_memory']
        if 'x' in mem_str and not mem_str.startswith(('1x','2x','4x','8x')):
            val = int(mem_str.replace('GB',''))
        elif 'x' in mem_str:
            parts = mem_str.split('x')
            val = int(parts[0]) * int(parts[1].replace('GB',''))
        else:
            val = int(mem_str.replace('GB',''))
        if val >= min_mem:
            res[name] = data
    return res

def generate_cost_comparison(
    compute_hours, tokens_per_month, input_ratio, api_calls,
    model_size, storage_gb, reserved_instances, spot_instances, multi_year_commitment,
    comparison_tier
):
    years = int(multi_year_commitment)
    in_tokens = tokens_per_month * 1000000 * (input_ratio/100)
    out_tokens = tokens_per_month * 1000000 - in_tokens
    min_mem = model_sizes[model_size]['memory_required']

    # Filter by both memory requirements and tier if a tier is selected
    aws_comp = filter_compatible(aws_instances, min_mem)
    gcp_comp = filter_compatible(gcp_instances, min_mem)
    
    if comparison_tier != "All":
        aws_comp = {k: v for k, v in aws_comp.items() if v.get('tier', '') == comparison_tier}
        gcp_comp = {k: v for k, v in gcp_comp.items() if v.get('tier', '') == comparison_tier}

    results = []

    # AWS table
    aws_html = '<h3>AWS Instances</h3>'
    aws_html += '<table width="100%"><tr><th>Instance</th><th>vCPUs</th><th>Memory</th><th>GPU</th><th>Tier</th><th>Monthly Cost ($)</th></tr>'
    if aws_comp:
        for inst in aws_comp:
            res = calculate_aws_cost(inst, compute_hours, storage_gb, reserved_instances, spot_instances, years)
            aws_html += f'<tr><td>{inst}</td><td>{res["details"]["vcpus"]}</td><td>{res["details"]["memory"]}GB</td><td>{res["details"]["gpu"]}</td><td>{res["details"].get("tier", "")}</td><td>${res["total_cost"]:.2f}</td></tr>'
        # best AWS
        best_aws = min(aws_comp, key=lambda x: calculate_aws_cost(x, compute_hours, storage_gb, reserved_instances, spot_instances, years)['total_cost'])
        best_aws_cost = calculate_aws_cost(best_aws, compute_hours, storage_gb, reserved_instances, spot_instances, years)['total_cost']
        best_aws_tier = aws_instances[best_aws].get('tier', '')
        results.append({'provider': f'AWS ({best_aws})', 'cost': best_aws_cost, 'type': 'Cloud', 'tier': best_aws_tier})
    else:
        aws_html += '<tr><td colspan="6">No compatible instances</td></tr>'
    aws_html += '</table>'

    # GCP table
    gcp_html = '<h3>GCP Instances</h3>'
    gcp_html += '<table width="100%"><tr><th>Instance</th><th>vCPUs</th><th>Memory</th><th>GPU</th><th>Tier</th><th>Monthly Cost ($)</th></tr>'
    if gcp_comp:
        for inst in gcp_comp:
            res = calculate_gcp_cost(inst, compute_hours, storage_gb, reserved_instances, spot_instances, years)
            gcp_html += f'<tr><td>{inst}</td><td>{res["details"]["vcpus"]}</td><td>{res["details"]["memory"]}GB</td><td>{res["details"]["gpu"]}</td><td>{res["details"].get("tier", "")}</td><td>${res["total_cost"]:.2f}</td></tr>'
        best_gcp = min(gcp_comp, key=lambda x: calculate_gcp_cost(x, compute_hours, storage_gb, reserved_instances, spot_instances, years)['total_cost'])
        best_gcp_cost = calculate_gcp_cost(best_gcp, compute_hours, storage_gb, reserved_instances, spot_instances, years)['total_cost']
        best_gcp_tier = gcp_instances[best_gcp].get('tier', '')
        results.append({'provider': f'GCP ({best_gcp})', 'cost': best_gcp_cost, 'type': 'Cloud', 'tier': best_gcp_tier})
    else:
        gcp_html += '<tr><td colspan="6">No compatible instances</td></tr>'
    gcp_html += '</table>'

    # API table
    api_html = '<h3>API Options</h3>'
    api_html += '<table width="100%"><tr><th>Provider</th><th>Model</th><th>Input Cost</th><th>Output Cost</th><th>Total Cost ($)</th><th>Context</th></tr>'
    api_costs = {}
    for prov in api_pricing:
        for mdl in api_pricing[prov]:
            res = calculate_api_cost(prov, mdl, in_tokens, out_tokens, api_calls)
            details = api_pricing[prov][mdl]
            api_html += f'<tr><td>{prov}</td><td>{mdl}</td><td>${in_tokens * details["input_per_1M"] / 1000000:.2f}</td><td>${out_tokens * details["output_per_1M"] / 1000000:.2f}</td><td>${res["total_cost"]:.2f}</td><td>{details["token_context"]:,}</td></tr>'
            api_costs[(prov, mdl)] = res['total_cost']
    api_html += '</table>'
    
    if api_costs:
        best_api = min(api_costs, key=api_costs.get)
        results.append({'provider': f'{best_api[0]} ({best_api[1]})', 'cost': api_costs[best_api], 'type': 'API', 'tier': 'API'})

    # Direct comparison tables for similar instances
    direct_comparison_html = ""
    if comparison_tier != "All" and comparison_tier != "API":
        direct_comparison_html = f'<h3>Direct {comparison_tier} Tier Comparison</h3>'
        direct_comparison_html += '<table width="100%"><tr><th>Provider</th><th>Instance</th><th>vCPUs</th><th>Memory</th><th>GPU</th><th>Monthly Cost ($)</th></tr>'
        
        aws_filtered = {k: v for k, v in aws_instances.items() if v.get('tier', '') == comparison_tier}
        gcp_filtered = {k: v for k, v in gcp_instances.items() if v.get('tier', '') == comparison_tier}
        
        # Group by vCPU for comparison
        vcpu_groups = {}
        for inst, data in aws_filtered.items():
            vcpu = data['vcpus']
            if vcpu not in vcpu_groups:
                vcpu_groups[vcpu] = {'aws': [], 'gcp': []}
            vcpu_groups[vcpu]['aws'].append(inst)
        
        for inst, data in gcp_filtered.items():
            vcpu = data['vcpus']
            if vcpu not in vcpu_groups:
                vcpu_groups[vcpu] = {'aws': [], 'gcp': []}
            vcpu_groups[vcpu]['gcp'].append(inst)
        
        # Display direct comparisons
        for vcpu in sorted(vcpu_groups.keys()):
            group = vcpu_groups[vcpu]
            for aws_inst in group['aws']:
                aws_cost = calculate_aws_cost(aws_inst, compute_hours, storage_gb, reserved_instances, spot_instances, years)
                aws_data = aws_cost['details']
                direct_comparison_html += f'<tr><td>AWS</td><td>{aws_inst}</td><td>{aws_data["vcpus"]}</td><td>{aws_data["memory"]}GB</td><td>{aws_data["gpu"]}</td><td>${aws_cost["total_cost"]:.2f}</td></tr>'
            
            for gcp_inst in group['gcp']:
                gcp_cost = calculate_gcp_cost(gcp_inst, compute_hours, storage_gb, reserved_instances, spot_instances, years)
                gcp_data = gcp_cost['details']
                direct_comparison_html += f'<tr><td>GCP</td><td>{gcp_inst}</td><td>{gcp_data["vcpus"]}</td><td>{gcp_data["memory"]}GB</td><td>{gcp_data["gpu"]}</td><td>${gcp_cost["total_cost"]:.2f}</td></tr>'
            
            # Add separator between different vCPU groups
            if vcpu != sorted(vcpu_groups.keys())[-1]:
                direct_comparison_html += '<tr><td colspan="6" style="border-bottom: 1px solid #ccc; height: 10px;"></td></tr>'
        
        direct_comparison_html += '</table>'

    # Chart with annotations
    df = pd.DataFrame(results)
    colors = {'Entry': '#66BB6A', 'Mid': '#42A5F5', 'High': '#FFA726', 'Premium': '#EF5350', 'API': '#AB47BC'}
    
    # Create figure using plotly graph objects for more control
    fig = go.Figure()
    
    # Add bars
    for i, row in df.iterrows():
        tier_color = colors.get(row.get('tier', 'API'), '#9E9E9E')
        fig.add_trace(go.Bar(
            x=[row['provider']], 
            y=[row['cost']], 
            name=row['provider'],
            marker_color=tier_color
        ))
    
    # Add annotations on top of each bar
    for i, row in df.iterrows():
        fig.add_annotation(
            x=row['provider'],
            y=row['cost'],
            text=f"${row['cost']:.2f}",
            showarrow=False,
            yshift=10,  # Position above the bar
            font=dict(size=14)
        )
    
    # Update layout
    fig.update_layout(
        showlegend=False,
        height=500,
        yaxis=dict(title='Monthly Cost ($)', tickprefix='$'),
        xaxis=dict(title=''),
        title='Cost Comparison'
    )

    html = f"""
    <div style='padding:20px;font-family:Arial;'>
      {direct_comparison_html}
      {aws_html}
      {gcp_html}
      {api_html}
    </div>
    """
    return html, fig

# UI setup
with gr.Blocks(title="Cloud Cost Estimator", theme=gr.themes.Soft(primary_hue="indigo")) as demo:
    gr.HTML('<h1 style="text-align:center;">Cloud Cost Estimator</h1>')
    with gr.Row():
        with gr.Column(scale=1):
            compute_hours = gr.Slider(label="Compute Hours per Month", minimum=1, maximum=300, value=50)
            tokens_per_month = gr.Slider(label="Tokens per Month (M)", minimum=1, maximum=200, value=5)
            input_ratio = gr.Slider(label="Input Ratio (%)", minimum=10, maximum=70, value=25)
            api_calls = gr.Slider(label="API Calls per Month", minimum=100, maximum=100000, value=5000, step=100)
            model_size = gr.Dropdown(label="Model Size", choices=list(model_sizes.keys()), value="Medium (13B parameters)")
            storage_gb = gr.Slider(label="Storage (GB)", minimum=10, maximum=1000, value=100)
            comparison_tier = gr.Radio(label="Comparison Tier", choices=["All", "Entry", "Mid", "High", "Premium", "API"], value="All")
            reserved_instances = gr.Checkbox(label="Reserved Instances", value=False)
            spot_instances = gr.Checkbox(label="Spot Instances", value=False)
            multi_year_commitment = gr.Radio(label="Commitment Period (years)", choices=["1","3"], value="1")
        with gr.Column(scale=2):
            out_html = gr.HTML()
            out_plot = gr.Plot()
    
    # Create inputs list for the function
    inputs = [compute_hours, tokens_per_month, input_ratio, api_calls,
             model_size, storage_gb, reserved_instances, spot_instances, multi_year_commitment, comparison_tier]
    outputs = [out_html, out_plot]
    
    # Initial calculation on load
    demo.load(generate_cost_comparison, inputs, outputs)
    
    # Update on each input change
    for input_component in inputs:
        input_component.change(generate_cost_comparison, inputs, outputs)

demo.launch()