app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import numpy as np
import string

# Load tokenizer and model for EOU detection
tokenizer = AutoTokenizer.from_pretrained("livekit/turn-detector")
model = AutoModelForCausalLM.from_pretrained("livekit/turn-detector")

# Define function to calculate softmax
def _softmax(logits: np.ndarray) -> np.ndarray:
    exp_logits = np.exp(logits - np.max(logits))
    return exp_logits / np.sum(exp_logits)

# Define the EOU probability calculation
def get_eou_probability(chat_ctx: list) -> float:
    """Calculate the probability of End of Utterance (EOU)"""
    # Normalize and prepare the chat context
    text = " ".join([msg["content"] for msg in chat_ctx])
    inputs = tokenizer(text, return_tensors="pt", max_length=512, truncation=True)

    # Run the model and get the logits
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits[0, -1, :]  # Get logits of the last token
        probs = _softmax(logits.numpy())   # Convert logits to probabilities

    # Assuming <|im_end|> token corresponds to EOU, get the probability of that token
    eou_token_id = tokenizer.encode("<|im_end|>")[-1]
    return probs[eou_token_id]

# Define the main response function for Gradio
def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    messages = [{"role": "system", "content": system_message}]
    
    for val in history:
        if val[0]:
            messages.append({"role": "user", "content": val[0]})
        if val[1]:
            messages.append({"role": "assistant", "content": val[1]})

    messages.append({"role": "user", "content": message})

    # Get the response from the Qwen model (e.g., for conversation generation)
    response = ""
    for message in client.chat_completion(
        messages,
        max_tokens=max_tokens,
        stream=True,
        temperature=temperature,
        top_p=top_p,
    ):
        token = message.choices[0].delta.content
        response += token
        yield response

    # After generating the response, get the EOU probability
    eou_probability = get_eou_probability(messages)  # Get EOU prediction
    print(f"EOU Probability: {eou_probability}")

    # Include the EOU probability in the output
    yield f"\nEOU Probability: {eou_probability:.2f}"

# Gradio interface setup
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(value="Bạn là một trợ lý ảo", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, 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)",
        ),
    ],
)

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