How does batch option work in pipeline transformers library

Question

I have a collection of news articles and I want to produce some new (unbiased) news articles using meta-llama/Meta-Llama-3-8B-Instruct. The articles are in a huggingface Dataset and to feed the transformers library pipeline I am using a KeyDataset like this:

key_dataset = KeyDataset(content, "prompt")

where prompt = "orders for LLM + article_content"

I want to produce new articles in batches to utilize better the GPU like this:

outputs = list(tqdm(pipeline(key_dataset, 
 batch_size=4, 
 max_new_tokens = 2*2024,
 eos_token_id=terminators,
 do_sample=True,
 temperature=1,
 top_p=0.9,), 
 total=len(key_dataset)))

The problem is that sometimes there is not generated text at all for some articles. Why is that? Doesn't batching work for different length inputs?