app.py

from adapters import AutoAdapterModel  # Ensure this library is correctly installed
from transformers import AutoTokenizer
import gradio as gr
import onnxruntime as ort
import numpy as np
import string
from huggingface_hub import InferenceClient
import os

# Load Base Model and Adapter
BASE_MODEL = "Qwen/Qwen2.5-1.5B-Instruct"  # Replace with the actual base model ID
ADAPTER_NAME = "ystemsrx/Qwen2.5-Sex"  # Replace with the correct adapter name

model = AutoAdapterModel.from_pretrained(BASE_MODEL)
model.load_adapter(ADAPTER_NAME, set_active=True)
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

# ONNX setup
ONNX_FILENAME = "model_quantized.onnx"
onnx_session = ort.InferenceSession(ONNX_FILENAME, providers=["CPUExecutionProvider"])

PUNCS = string.punctuation.replace("'", "")
MAX_HISTORY = 4
MAX_HISTORY_TOKENS = 512
EOU_THRESHOLD = 0.5

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

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

# Format chat context
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]

# Calculate EOU probability
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]

# Respond function
def respond(
    message,
    history: list[tuple[str, str]],
    max_tokens,
    temperature,
    top_p,
):
    messages = [{"role": "system", "content": os.environ.get("CHARACTER_DESC")}]
    for val in history[-10:]:
        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)
    print(f"EOU Probability: {eou_prob}")
    if eou_prob < EOU_THRESHOLD:
        yield "[Waiting for user to continue input...]"
        return
    response = ""
    for message in qwen_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
    print(f"Generated response: {response}")

# Gradio interface
demo = gr.ChatInterface(
    respond,
)

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