app.py

import gradio as gr
import torch
from transformers import pipeline

# Model ID for Llama 3 8B instruct (replace with the exact model you want)
MODEL_ID = "manycore-research/SpatialLM-Llama-1B"

# Load the text-generation pipeline with device_map="auto" to use GPU if available
generator = pipeline(
    "text-generation",
    model=MODEL_ID,
    torch_dtype=torch.float16,
    device_map="auto",
)

def generate_response(prompt, max_length=512, temperature=0.7):
    # Format prompt for Llama 3 instruct style
    formatted_prompt = f"<s>[INST] {prompt} [/INST]"
    output = generator(
        formatted_prompt,
        max_length=max_length,
        temperature=temperature,
        do_sample=True,
        top_p=0.95,
        num_return_sequences=1,
    )
    generated_text = output[0]["generated_text"]
    # Extract the response after the [/INST] token
    response = generated_text.split("[/INST]")[-1].strip()
    return response

with gr.Blocks() as demo:
    gr.Markdown("# Chat with Llama 3 (8B Instruct)")
    with gr.Row():
        with gr.Column():
            user_input = gr.Textbox(lines=3, placeholder="Type your message here...", label="Your Message")
            submit_btn = gr.Button("Submit")
        with gr.Column():
            output = gr.Textbox(lines=10, label="Llama 3 Response")
    submit_btn.click(fn=generate_response, inputs=user_input, outputs=output)

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