app.py

import gradio as gr

def calculate_training_metrics(
    gpu_choice, precision, num_gpus, num_parameters, dataset_tokens, 
    num_epochs, utilization_rate=0.5, overhead=1.10, cost_per_gpu_hour=1.85
):
    """
    Calculates both the training time and cost for LLM training with parallel computing.

    Args:
    - gpu_choice (str): The choice of GPU model
    - precision (str): The precision level for training
    - num_gpus (int): Number of GPUs for parallel computing
    - num_parameters (float): Number of model parameters in billions
    - dataset_tokens (float): Number of tokens in the dataset
    - num_epochs (int): Number of training epochs
    - utilization_rate (float): GPU utilization rate (0 < rate ≤ 1)
    - overhead (float): Overhead multiplier for additional costs
    - cost_per_gpu_hour (float): Cost per GPU hour in dollars

    Returns:
    - tuple: (total_cost, training_days, training_hours)
    """
    
    # GPU throughputs in FLOPS (operations per second)
    gpu_throughputs = {
        'A100 80GB PCIe': {'bf16': 312e12, 'tf32': 156e12},
        'A100 80GB SXM': {'bf16': 624e12, 'tf32': 312e12},
        'V100': {'tensor': 130e12},
        'H100 SXM': {'bf16': 1979e12, 'tf32': 989e12},
        'H100 PCIe': {'bf16': 1513e12, 'tf32': 756e12}
    }

    # Get the base GPU throughput
    base_throughput = gpu_throughputs[gpu_choice][precision]
    
    # Calculate effective throughput with multiple GPUs
    # Assuming 90% scaling efficiency for parallel computing
    parallel_efficiency = 0.9
    effective_throughput = base_throughput * num_gpus * parallel_efficiency
    
    # Calculate total tokens processed (dataset_tokens * epochs)
    total_tokens = dataset_tokens * num_epochs
    
    # Calculate total FLOPS needed (6 operations per parameter per token)
    total_flops = 6 * num_parameters * total_tokens
    
    # Calculate raw computation hours needed
    compute_hours = total_flops / (effective_throughput * 3600)
    
    # Adjust for utilization rate and overhead
    actual_hours = (compute_hours / utilization_rate) * overhead
    
    # Calculate days and remaining hours
    training_days = int(actual_hours // 24)
    training_hours = actual_hours % 24
    
    # Calculate total cost (cost per GPU * number of GPUs * hours)
    total_cost = actual_hours * cost_per_gpu_hour * num_gpus
    
    return total_cost, training_days, training_hours

def gradio_interface(
    gpu_choice, precision, num_gpus, num_parameters, dataset_tokens,
    num_epochs, utilization_rate, overhead, cost_per_gpu_hour
):
    # Convert inputs to appropriate numeric types
    num_parameters = float(num_parameters) * 1e9  # Convert billions to actual number
    dataset_tokens = float(dataset_tokens) * 1e9  # Convert billions to actual number
    num_gpus = int(num_gpus)
    num_epochs = int(num_epochs)
    utilization_rate = float(utilization_rate)
    overhead = float(overhead)
    cost_per_gpu_hour = float(cost_per_gpu_hour)

    # Calculate metrics
    cost, days, hours = calculate_training_metrics(
        gpu_choice, precision, num_gpus, num_parameters, dataset_tokens,
        num_epochs, utilization_rate, overhead, cost_per_gpu_hour
    )
    
    # Format the output message
    time_msg = f"{days} days and {hours:.1f} hours"
    cost_msg = f"{cost:,.2f}$"
    
    return time_msg, cost_msg

# Define available GPU choices and their default precisions
gpu_choices = ["A100 80GB PCIe", "A100 80GB SXM", "V100", "H100 SXM", "H100 PCIe"]

# Create the Gradio interface
title = "<h2 style='text-align: center;'>LLM Training Time and Cost Calculator</h2>"
description = """
<p style='text-align: center;'>Calculate both the training time and cost for large language models (LLM) with parallel computing support.</p>
<p><strong>Input Parameters:</strong></p>
<ul>
    <li><strong>GPU Selection:</strong> Choose from various GPU models with different compute capabilities</li>
    <li><strong>Number of GPUs:</strong> Specify how many GPUs to use in parallel</li>
    <li><strong>Model Size:</strong> Number of parameters in billions</li>
    <li><strong>Dataset Size:</strong> Number of tokens in your dataset in billions</li>
    <li><strong>Training Epochs:</strong> Number of times to iterate over the dataset</li>
    <li><strong>Utilization Rate:</strong> Expected GPU utilization (typically 0.4-0.7)</li>
    <li><strong>Overhead:</strong> Additional time/cost factor for data loading, checkpointing, etc.</li>
</ul>
Ouputs:
<ul>
    <li><strong>Estimated Training Time:</strong> Total days and hours required for training</li>
    <li><strong>Estimated Training Cost:</strong> Total cost in dollars based on GPU hours</li>
</ul>
Modified from <a href="https://huggingface.co/spaces/Heng666/LLM-Training-Cost-Calculator">this Hf Space</a>.
"""

iface = gr.Interface(
    fn=gradio_interface,
    inputs=[
        gr.Dropdown(choices=gpu_choices, label="Select GPU", value='A100 80GB PCIe'),
        gr.Dropdown(choices=['bf16', 'tf32', 'tensor'], label="Select Precision", value='bf16'),
        gr.Number(label="Number of GPUs", value=1, minimum=1, maximum=1024),
        gr.Number(label="Number of Parameters (billions)", value=70),
        gr.Number(label="Dataset Tokens (billions)", value=1),
        gr.Number(label="Number of Epochs", value=3, minimum=1),
        gr.Slider(minimum=0.1, maximum=1.0, step=0.1, value=0.5, label="GPU Utilization Rate"),
        gr.Slider(minimum=1.0, maximum=2.0, step=0.01, value=1.10, label="Overhead Factor"),
        gr.Number(label="Cost per GPU Hour ($)", value=1.85)
    ],
    outputs=[gr.Textbox(label="Estimated Training Time:"),
             gr.Textbox(label="Estimated Training Cost:")],
    title=title,
    description=description,
    article="<p style='text-align: center;'>Improved with good intentions by ghost.</p>"
)

if __name__ == "__main__":
    iface.launch()