app.py

import os
import torch
from typing import List
from huggingface_hub import login
from transformers import AutoTokenizer, AutoModelForCausalLM
import gradio as gr
from transformers import AutoTokenizer as SummarizerTokenizer, AutoModelForSeq2SeqLM

device = "cuda" if torch.cuda.is_available() else "cpu"

# Summarization model
summarizer_model_id = "facebook/bart-large-cnn"
summarizer_tokenizer = SummarizerTokenizer.from_pretrained(summarizer_model_id)
summarizer_model = AutoModelForSeq2SeqLM.from_pretrained(
    summarizer_model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)
summarizer_model.to(device)

def summarize_text(text: str, max_length=150) -> str:
    inputs = summarizer_tokenizer([text], return_tensors="pt", max_length=1024, truncation=True).to(device)
    summary_ids = summarizer_model.generate(
        inputs['input_ids'],
        num_beams=4,
        max_length=max_length,
        early_stopping=True
    )
    summary = summarizer_tokenizer.decode(summary_ids[0], skip_special_tokens=True)
    return summary

HF_TOKEN = os.environ.get("HF_TOKEN")
if HF_TOKEN:
    login(token=HF_TOKEN)

model_id = "deepseek-ai/DeepSeek-R1-Distill-Llama-70B"
tokenizer = AutoTokenizer.from_pretrained(model_id, token=HF_TOKEN)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    low_cpu_mem_usage=True,
    token=HF_TOKEN
)

# --- GIỮ LẠI CHỈ 1 HÀM build_prompt, ĐÃ BỔ SUNG SUMMARIZATION ---
def build_prompt(prompt: str, histories: List[str], new_message: str) -> str:
    prompt_text = prompt.strip() + "\n" if prompt else ""
    histories_text = "\n".join(histories) if histories else ""
    # Tóm tắt nếu quá dài (tùy chỉnh ngưỡng này)
    if len(histories_text) > 3000:
        histories_text = summarize_text(histories_text, max_length=180)
    if histories_text:
        prompt_text += histories_text + "\n"
    prompt_text += f"User: {new_message}\nAI:"
    return prompt_text

def chat(
    prompt: str,
    histories: List[str],
    new_message: str
) -> str:
    prompt_text = build_prompt(prompt, histories, new_message)
    input_ids = tokenizer(prompt_text, return_tensors="pt").input_ids.to(device)
    with torch.no_grad():
        output = model.generate(
            input_ids,
            max_new_tokens=256,
            do_sample=True,
            top_p=0.95,
            temperature=0.7,
            pad_token_id=tokenizer.eos_token_id
        )
    output_text = tokenizer.decode(output[0], skip_special_tokens=True)
    if "AI:" in output_text:
        response = output_text.split("AI:")[-1].strip()
        if "User:" in response:
            response = response.split("User:")[0].strip()
    else:
        response = output_text.strip()
    return response

with gr.Blocks() as demo:
    gr.Markdown("# MindVR Therapy Chatbot\n\nDùng thử UI hoặc gọi API!")
    prompt_box = gr.Textbox(lines=2, label="Prompt (System Prompt, chỉ dẫn context cho AI, có thể bỏ trống)")
    histories_box = gr.Textbox(lines=8, label="Histories (mỗi dòng là một message, ví dụ: User: Xin chào)")
    new_message_box = gr.Textbox(label="New message")
    output = gr.Textbox(label="AI Response")

    def _chat_ui(prompt, histories, new_message):
        # histories nhập từ UI là multiline string -> chuyển thành list
        histories_list = [line.strip() for line in histories.split('\n') if line.strip()]
        return chat(prompt, histories_list, new_message)

    btn = gr.Button("Gửi")
    btn.click(_chat_ui, inputs=[prompt_box, histories_box, new_message_box], outputs=output)

    # API chuẩn RESTful với prompt, histories, new_message
    gr.api(chat, api_name="chat_ai")

demo.launch()