app.py

import gradio as gr
import subprocess
import os
import time
from transformers import AutoTokenizer, AutoModelForCausalLM
import logging

# Configure logging
logging.basicConfig(level=logging.INFO)

# Path to the cloned repository
BITNET_REPO_PATH = "/home/user/app/BitNet"
SETUP_SCRIPT = os.path.join(BITNET_REPO_PATH, "setup_env.py")
INFERENCE_SCRIPT = os.path.join(BITNET_REPO_PATH, "run_inference.py")

# Function to set up the environment by running setup.py
def setup_bitnet(model_name):
    try:
        result = subprocess.run(
            f"python {SETUP_SCRIPT} --hf-repo {model_name} -q i2_s",
            shell=True,
            cwd=BITNET_REPO_PATH,
            capture_output=True,
            text=True
        )
        if result.returncode == 0:
            return "Setup completed successfully!"
        else:
            return f"Error in setup: {result.stderr}"
    except Exception as e:
        return str(e)

# Function to run inference using the `run_inference.py` file
def run_inference(model_name, input_text, num_tokens=6):
    try:
        # Call the `run_inference.py` script with the model and input
        start_time = time.time()
        result = subprocess.run(
            f"python run_inference.py -m models/Llama3-8B-1.58-100B-tokens/ggml-model-i2_s.gguf -p \"{input_text}\" -n {num_tokens} -temp 0",
            shell=True,
            cwd=BITNET_REPO_PATH,
            capture_output=True,
            text=True
        )
        end_time = time.time()
        
        if result.returncode == 0:
            inference_time = round(end_time - start_time, 2)
            return result.stdout, f"Inference took {inference_time} seconds."
        else:
            return f"Error during inference: {result.stderr}", None
    except Exception as e:
        return str(e), None

def run_transformers(profile: gr.OAuthProfile | None, oauth_token: gr.OAuthToken | None, model_name, input_text, num_tokens):

    if oauth_token is None : 
        return "Error : To Compare please login to your HF account and make sure you have access to the used Llama models"
    # Load the model and tokenizer dynamically if needed (commented out for performance)
    tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token=oauth_token.token)
    model = AutoModelForCausalLM.from_pretrained(model_name, use_auth_token=oauth_token.token)

    # Encode the input text
    input_ids = tokenizer.encode(input_text, return_tensors="pt")

    # Start time for inference
    start_time = time.time()

    # Generate output with the specified number of tokens
    output = model.generate(input_ids, max_length=len(input_ids[0]) + num_tokens, num_return_sequences=1)

    # Calculate inference time
    inference_time = time.time() - start_time

    # Decode the generated output
    generated_text = tokenizer.decode(output[0], skip_special_tokens=True)

    return generated_text, f"{inference_time:.2f} seconds"

# Gradio Interface
def interface():
    with gr.Blocks(css=".gr-button {background-color: #5C6BC0; color: white;} .gr-button:hover {background-color: #3F51B5;}") as demo:

        gr.LoginButton(elem_id="login-button", elem_classes="center-button")

        gr.Markdown(
            """
            <h1 style="text-align: center; color: #4A148C;">BitNet.cpp Speed Demonstration</h1>
            <p style="text-align: center; color: #6A1B9A;">Compare the speed and performance of BitNet with Transformers!</p>
            """, 
            elem_id="header"
        )

        # Model selection and setup row
        with gr.Row():
            model_dropdown = gr.Dropdown(
                label="Select Model",
                choices=["HF1BitLLM/Llama3-8B-1.58-100B-tokens", "1bitLLM/bitnet_b1_58-3B", "1bitLLM/bitnet_b1_58-large"],  # Replace with available models
                value="HF1BitLLM/Llama3-8B-1.58-100B-tokens",
                interactive=True,
                elem_id="model-dropdown"
            )
            setup_button = gr.Button("Run Setup", elem_id="setup-button")
            setup_status = gr.Textbox(label="Setup Status", interactive=False, placeholder="Setup status will appear here...")

        # Inference row
        with gr.Row():
            num_tokens = gr.Slider(minimum=1, maximum=100, label="Number of Tokens to Generate", value=50, step=1)
            input_text = gr.Textbox(label="Input Text", placeholder="Enter your input text here...")
            infer_button = gr.Button("Run Inference", elem_id="infer-button")
            result_output = gr.Textbox(label="Output", interactive=False, placeholder="Inference output will appear here...")
            time_output = gr.Textbox(label="Inference Time", interactive=False, placeholder="Inference time will appear here...")

        # Comparison with Transformers
        with gr.Row():
            transformer_model_dropdown = gr.Dropdown(
                label="Select Transformers Model",
                choices=["meta-llama/Llama-3.1-8B", "meta-llama/Llama-3.2-3B", "meta-llama/Llama-3.2-1B"],  # Replace with actual models
                value="meta-llama/Llama-3.1-8B",
                interactive=True
            )
            compare_button = gr.Button("Run Transformers Inference", elem_id="compare-button")
            transformer_result_output = gr.Textbox(label="Transformers Output", interactive=False, placeholder="Transformers output will appear here...")
            transformer_time_output = gr.Textbox(label="Transformers Inference Time", interactive=False, placeholder="Transformers inference time will appear here...")

        # Actions
        setup_button.click(setup_bitnet, inputs=model_dropdown, outputs=setup_status)
        infer_button.click(run_inference, inputs=[model_dropdown, input_text, num_tokens], outputs=[result_output, time_output])
        compare_button.click(run_transformers, inputs=[transformer_model_dropdown, input_text, num_tokens], outputs=[transformer_result_output, transformer_time_output])

# Launch the Gradio app    
    return demo

demo = interface()
demo.launch()