app.py

import gradio as gr
import os
from threading import Thread
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer, pipeline

# --- 1. Load a Standard CPU-Friendly Model ---
# No PEFT model needed. We are loading a pre-trained chat model directly.
model_name = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"

# Load the model and tokenizer
# No need for tokens if it's a public model. No special settings for CPU.
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

print(f"✅ Model '{model_name}' loaded successfully on CPU!")

# --- 2. Create a Pipeline for Easy Inference ---
pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
)

# --- 3. Define the Respond Function for the Chatbot ---
# This function takes the user message and history, and generates a response using the pipeline.
def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    # Build the prompt using the specific chat template for TinyLlama
    messages = [{"role": "system", "content": system_message}]
    for user_msg, assistant_msg in history:
        messages.append({"role": "user", "content": user_msg})
        messages.append({"role": "assistant", "content": assistant_msg})
    messages.append({"role": "user", "content": message})
    
    prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

    # Generate the response
    # This will be slow on a CPU and will wait for the full response.
    outputs = pipe(
        prompt,
        max_new_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p,
        do_sample=True,
    )
    
    # Extract only the generated text from the full output
    full_response = outputs[0]['generated_text']
    # The response includes the prompt, so we split it to get only the new part
    new_response = full_response.split(prompt)[1]
    
    return new_response

# --- 4. Launch the Gradio Interface ---
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(value="You are a friendly and helpful chatbot.", label="System message"),
        gr.Slider(minimum=10, maximum=512, value=128, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=0.95,
            step=0.05,
            label="Top-p (nucleus sampling)",
        ),
    ],
    title="TinyLlama 1.1B Chat",
    description="A simple chatbot running on a CPU-friendly model from Hugging Face."
)

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