app.py

import gradio as gr
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, pipeline, AutoConfig
from huggingface_hub import cached_download, hf_hub_url, list_models
import requests
import json
import os
import matplotlib.pyplot as plt
from io import BytesIO
import base64
from transformers.models.auto import AutoModel
from transformers.modeling_utils import PreTrainedModel
import torch
from torch.nn.utils import prune

# Function to fetch open-weight LLM models
def fetch_open_weight_models():
    models = list_models(filter="open-weight", sort="downloads", limit=12)
    return [model["id"] for model in models]

# Function to prune a model using the "merge-kit" approach
def prune_model(llm_model_name, target_size, output_dir):
    # Load the LLM model and tokenizer
    llm_tokenizer = AutoTokenizer.from_pretrained(llm_model_name)
    llm_model = AutoModelForSeq2SeqLM.from_pretrained(llm_model_name)

    # Get the model config
    config = AutoConfig.from_pretrained(llm_model_name)
    # Calculate the target number of parameters
    target_num_parameters = int(config.num_parameters * (target_size / 100))

    # Use merge-kit to prune the model
    pruned_model = merge_kit_prune(llm_model, target_num_parameters)

    # Save the pruned model
    pruned_model.save_pretrained(output_dir)

    # Create a visualization
    fig, ax = plt.subplots(figsize=(10, 5))
    ax.bar(["Original", "Pruned"], [config.num_parameters, pruned_model.num_parameters])
    ax.set_ylabel("Number of Parameters")
    ax.set_title("Model Size Comparison")
    buf = BytesIO()
    fig.savefig(buf, format="png")
    buf.seek(0)
    image_base64 = base64.b64encode(buf.read()).decode("utf-8")
    return f"Pruned model saved to {output_dir}", f"data:image/png;base64,{image_base64}"

# Merge-kit Pruning Function
def merge_kit_prune(model: PreTrainedModel, target_num_parameters: int) -> PreTrainedModel:
    """Prunes a model using a merge-kit approach.

    Args:
        model (PreTrainedModel): The model to be pruned.
        target_num_parameters (int): The target number of parameters after pruning.

    Returns:
        PreTrainedModel: The pruned model.
    """

    # Define the pruning method
    pruning_method = "unstructured"

    # Calculate the pruning amount
    amount = 1 - (target_num_parameters / model.num_parameters)

    # Prune the model using the selected method
    for name, module in model.named_modules():
        if isinstance(module, (torch.nn.Linear, torch.nn.Conv2d)):
            prune.random_unstructured(module, name="weight", amount=amount)

    # Remove the pruned weights
    for name, module in model.named_modules():
        if isinstance(module, (torch.nn.Linear, torch.nn.Conv2d)):
            prune.remove(module, name="weight")

    return model

# Function to create a Gradio interface
def create_interface():
    with gr.Blocks() as demo:
        gr.Markdown("## Create a Smaller LLM")

        # Fetch open-weight models from Hugging Face
        available_models = gr.Dropdown(
            label="Choose a Large Language Model",
            choices=fetch_open_weight_models(),
            interactive=True,
        )

        # Input for target model size
        target_size = gr.Slider(
            label="Target Model Size (%)",
            minimum=1,
            maximum=100,
            step=1,
            value=50,
            interactive=True,
        )

        # Output for pruning status
        pruning_status = gr.Textbox(label="Pruning Status")

        # Output for saving the model
        save_model_path = gr.Textbox(label="Save Model Path", placeholder="Path to save the pruned model", interactive=True)

        # Button to start pruning
        prune_button = gr.Button("Prune Model")

        # Output for visualization
        visualization = gr.Image(label="Model Size Comparison")

        # Connect components
        prune_button.click(
            fn=prune_model,
            inputs=[available_models, target_size, save_model_path],
            outputs=[pruning_status, visualization],
        )

        # Example usage of the pruned model (optional)
        text_input = gr.Textbox(label="Input Text")
        text_output = gr.Textbox(label="Generated Text")

        # Generate text button
        generate_button = gr.Button("Generate Text")

        def generate_text(text, model_path):
            # Load the pruned model and tokenizer
            tokenizer = AutoTokenizer.from_pretrained(model_path)
            model = AutoModelForSeq2SeqLM.from_pretrained(model_path)

            # Use the pipeline for text generation
            generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
            generated_text = generator(text, max_length=50, num_beams=5, num_return_sequences=1)[0]["generated_text"]
            return generated_text

        generate_button.click(fn=generate_text, inputs=[text_input, save_model_path], outputs=text_output)

    return demo

# Create and launch the Gradio interface
demo = create_interface()
demo.launch(share=True)