app.py

import os
import torch
from huggingface_hub import login
from transformers import AutoTokenizer, AutoModelForCausalLM
import gradio as gr

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

model_id = "MindVR/JohnTran_Fine-tune"
tokenizer = AutoTokenizer.from_pretrained(model_id, token=HF_TOKEN)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    low_cpu_mem_usage=True,
    token=HF_TOKEN
)
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

def build_prompt(history, new_message):
    prompt = ""
    if history:
        prompt += "\n".join(history) + "\n"
    prompt += f"User: {new_message}\nAI:"
    return prompt

def chat(history, new_message):
    # Đảm bảo history là list (khi nhập trực tiếp trên UI đôi khi là str)
    if isinstance(history, str):
        import ast
        try:
            history = ast.literal_eval(history)
        except:
            history = [history]
    prompt = build_prompt(history, new_message)
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
    with torch.no_grad():
        output = model.generate(
            input_ids,
            max_new_tokens=200,
            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()
    else:
        response = output_text.strip()
    return response

iface = gr.Interface(
    fn=chat,
    inputs=[
        gr.Textbox(lines=8, label="History (JSON list, ví dụ: [\"User: Xin chào\"] )"),
        gr.Textbox(label="New message")
    ],
    outputs=gr.Textbox(label="AI Response"),
    title="MindVR Therapy Chatbot",
    allow_flagging="never"
)

iface.launch()