app.py

import time
import logging
from fastapi import FastAPI, Request
from pydantic import BaseModel
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

app = FastAPI()

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("VietAI/vit5-large-vietnews-summarization")
model = AutoModelForSeq2SeqLM.from_pretrained("VietAI/vit5-large-vietnews-summarization")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# Warm-up model to reduce first-request latency
dummy_input = tokenizer("Tin nhanh: Đây là văn bản mẫu để warmup mô hình.", return_tensors="pt").to(device)
with torch.no_grad():
    _ = model.generate(**dummy_input, max_length=32)

class SummarizeRequest(BaseModel):
    text: str

@app.get("/")
async def root():
    return {"message": "Model is ready."}

@app.post("/summarize")
async def summarize(req: Request, body: SummarizeRequest):
    start_time = time.time()
    client_ip = req.client.host
    logger.info(f"[{time.strftime('%Y-%m-%d %H:%M:%S')}] 🔵 Received request from {client_ip}")

    text = body.text.strip()

    # Tiền xử lý: nếu không giống tin tức thì thêm "Tin nhanh:"
    if not text.lower().startswith(("theo", "trong khi", "bộ", "ngày", "việt nam", "công an")):
        text = "Tin nhanh: " + text
    else:
        text = "Vietnews: " + text

    input_text = text + " </s>"
    encoding = tokenizer(input_text, return_tensors="pt")
    input_ids = encoding["input_ids"].to(device)
    attention_mask = encoding["attention_mask"].to(device)

    # Sinh tóm tắt với cấu hình ổn định (loại bỏ early_stopping và dùng greedy decoding)
    outputs = model.generate(
        input_ids=input_ids,
        attention_mask=attention_mask,
        max_length=256,
        num_beams=1,  # greedy decoding
        no_repeat_ngram_size=2
    )

    summary = tokenizer.decode(outputs[0], skip_special_tokens=True, clean_up_tokenization_spaces=True)

    end_time = time.time()
    logger.info(f"[{time.strftime('%Y-%m-%d %H:%M:%S')}] ✅ Response sent — total time: {end_time - start_time:.2f}s")

    return {"summary": summary}