app.py

import gradio as gr
from transformers import AutoTokenizer
import onnxruntime as ort
import numpy as np
import string
from huggingface_hub import InferenceClient
import os

# Initialize Qwen client
qwen_client = InferenceClient(os.environ.get("HF_TOKEN"))

# Model and ONNX setup
HG_MODEL = "livekit/turn-detector"
ONNX_FILENAME = "model_quantized.onnx"
PUNCS = string.punctuation.replace("'", "")
MAX_HISTORY = 4
MAX_HISTORY_TOKENS = 512
EOU_THRESHOLD = 0.5

# Initialize ONNX model
tokenizer = AutoTokenizer.from_pretrained(HG_MODEL)
onnx_session = ort.InferenceSession(ONNX_FILENAME, providers=["CPUExecutionProvider"])

def softmax(logits):
    exp_logits = np.exp(logits - np.max(logits))
    return exp_logits / np.sum(exp_logits)

def normalize_text(text):
    def strip_puncs(text):
        return text.translate(str.maketrans("", "", PUNCS))
    return " ".join(strip_puncs(text).lower().split())

def format_chat_ctx(chat_ctx):
    new_chat_ctx = []
    for msg in chat_ctx:
        if msg["role"] in ("user", "assistant"):
            content = normalize_text(msg["content"])
            if content:
                msg["content"] = content
                new_chat_ctx.append(msg)
    convo_text = tokenizer.apply_chat_template(
        new_chat_ctx, add_generation_prompt=False, add_special_tokens=False, tokenize=False
    )
    ix = convo_text.rfind("<|im_end|>")
    return convo_text[:ix]

def calculate_eou(chat_ctx, session):
    formatted_text = format_chat_ctx(chat_ctx[-MAX_HISTORY:])
    inputs = tokenizer(
        formatted_text,
        return_tensors="np",
        truncation=True,
        max_length=MAX_HISTORY_TOKENS,
    )
    input_ids = np.array(inputs["input_ids"], dtype=np.int64)
    outputs = session.run(["logits"], {"input_ids": input_ids})
    logits = outputs[0][0, -1, :]
    probs = softmax(logits)
    eou_token_id = tokenizer.encode("<|im_end|>")[-1]
    return probs[eou_token_id]

def respond(
    message,
    history: list[tuple[str, str]],
	max_tokens=2048,
    temperature=0.6,
    top_p=0.95,
):
    messages = [{"role": "system", "content": os.environ.get("CHARACTER_DESC", "You are a helpful assistant.")}]
    
    for val in history[-MAX_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})

    eou_prob = calculate_eou(messages, onnx_session)
    if eou_prob < EOU_THRESHOLD:
        yield "[Wait... Keep typing...]"
        return

    # Generate raw response without any processing
    full_response = ""
    stream = qwen_client.chat.completions.create(
        model="deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B", 
        messages=messages,
        max_tokens=max_tokens,
        stream=True,
        temperature=temperature,
        top_p=top_p,
        stream=True
    )

    for chunk in stream:
        full_response += chunk.choices[0].delta.content
        yield chunk.choices[0].delta.content  # Send raw unmodified response to Gradio

    # This will match both console and Gradio output
# Create Gradio interface
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Slider(1, 4096, value=256, label="Max Tokens"),
        gr.Slider(0.1, 4.0, value=0.7, label="Temperature"),
        gr.Slider(0.1, 1.0, value=0.95, label="Top-p"),
    ]
)

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