app.py

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from huggingface_hub import HfApi
import gradio as gr
import os
from peft import PeftModel, PeftConfig

# Function to merge models using PEFT
def merge_models():
    base_model_name = "meta-llama/Meta-Llama-3-8B"
    finetuned_model_name = "NoaiGPT/autotrain-14mrs-fc44l"
    
    # Load the base model
    base_model = AutoModelForCausalLM.from_pretrained(base_model_name)
    base_tokenizer = AutoTokenizer.from_pretrained(base_model_name)
    
    # Load the fine-tuned model using PEFT
    peft_config = PeftConfig.from_pretrained(finetuned_model_name)
    finetuned_model = PeftModel.from_pretrained(base_model, peft_config)
    
    # Merge the models (simple weight averaging here for demonstration; adjust as needed)
    for param_base, param_finetuned in zip(base_model.parameters(), finetuned_model.parameters()):
        param_base.data = (param_base.data + param_finetuned.data) / 2
    
    # Save the merged model
    merged_model_name = "./merged_model"
    base_model.save_pretrained(merged_model_name)
    base_tokenizer.save_pretrained(merged_model_name)
    
    return merged_model_name

# Function to upload the merged model to Hugging Face Hub
def upload_to_hf(repo_id, merged_model_name):
    api = HfApi()
    model_files = [os.path.join(merged_model_name, f) for f in os.listdir(merged_model_name)]
    
    for file in model_files:
        api.upload_file(
            path_or_fileobj=file,
            path_in_repo=os.path.basename(file),
            repo_id=repo_id,
            repo_type="model"
        )

    return f"Model uploaded to Hugging Face Hub at {repo_id}."

# Gradio function to handle the merge button click
def merge_button_clicked():
    merged_model_name = merge_models()
    repo_id = "NoaiGPT/autotrain-14mrs-fc44l"
    return upload_to_hf(repo_id, merged_model_name)

# Create the Gradio interface
with gr.Blocks() as demo:
    with gr.Row():
        merge_button = gr.Button("Merge Models")
        output = gr.Textbox(label="Output")

    merge_button.click(merge_button_clicked, outputs=output)

# Launch the Gradio interface
demo.launch()