from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("HoangQuocViet25/nlp-btl-xlsum-x2")
net = AutoModelForSeq2SeqLM.from_pretrained("HoangQuocViet25/nlp-btl-xlsum-x2").eval()
# ViT5's own preprocessing: no task prefix, EOS appended to the raw string.
enc = tok(article.strip() + "</s>", return_tensors="pt", truncation=True, max_length=1024)
out = net.generate(**enc, num_beams=4, length_penalty=1.0, max_new_tokens=256,
                   early_stopping=True)
print(tok.decode(out[0], skip_special_tokens=True))