app.py

import gradio as gr
from huggingface_hub import InferenceClient
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load Inference Client for the response model
client = InferenceClient("Qwen/Qwen2.5-3B-Instruct")

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

# Function to compute EOU probability
def compute_eou_probability(chat_ctx: list[dict[str, str]], max_tokens: int = 512) -> float:
    # Extract only the 'content' from the chat context (messages) and use a list of strings for tokenization
    conversation = ["Assistant ready to help."]  # Add system message directly as a string
    
    # Only append the 'content' of each message to the conversation list
    for msg in chat_ctx:
        content = msg.get("content", "")
        if content:
            conversation.append(content)  # Only append the content (string)

    # Tokenize the conversation content (just the text) as a list of strings
    inputs = tokenizer(
        conversation, padding=True, truncation=True, max_length=max_tokens, return_tensors="pt"
    )
    
    # Get model logits
    with torch.no_grad():
        outputs = model(**inputs)
    
    # Get the logits for the last token in the sequence
    logits = outputs.logits[0, -1, :]
    
    # Apply softmax to get probabilities
    probabilities = torch.nn.functional.softmax(logits, dim=-1)
    
    # Get the EOU token index (typically "<|im_end|>" token in the model)
    eou_token_id = tokenizer.encode("<|im_end|>")[0]
    eou_probability = probabilities[eou_token_id].item()
    
    return eou_probability

# Respond function with EOU checking logic
def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
    eou_threshold: float = 0.2,  # Default EOU threshold
):
    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]})
    
    # Compute EOU probability before responding
    eou_probability = compute_eou_probability(messages, max_tokens=max_tokens)
    print(eou_probability)
    # Only respond if EOU probability exceeds threshold
    if eou_probability >= eou_threshold:
        # Prepare message for assistant response
        messages.append({"role": "user", "content": message})

        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
    else:
        # Let the user continue typing if the EOU probability is low
        yield "Waiting for user to finish... Please continue."
        print("Waiting for user to finish... Please continue.")

# Gradio UI
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)",
        ),
        gr.Slider(
            minimum=0.0, maximum=1.0, value=0.7, step=0.05, label="EOU Threshold"
        ),  # Add EOU threshold slider
    ],
)

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