def calculate_perplexity(model, text):
    # Encode the text
    encodings = tokenizer(text, return_tensors='pt').to(device)

    # Define input_ids and target_ids
    input_ids = encodings.input_ids
    target_ids = input_ids.clone()

    with torch.no_grad():
        outputs = model(input_ids, labels=target_ids)

    # Loss calculation
    neg_log_likelihood = outputs.loss

    # Perplexity calculation
    ppl = torch.exp(neg_log_likelihood)

    return ppl

ppl     = calculate_perplexity(model, original_text)
ppl_abs = calculate_perplexity(model_abs, absmax_text)
ppl_zp  = calculate_perplexity(model_zp, absmax_text)

print(f"Original perplexity:  {ppl.item():.2f}")
print(f"Absmax perplexity:    {ppl_abs.item():.2f}")
print(f"Zeropoint perplexity: {ppl_zp.item():.2f}")