Ifeval: Dowload punkt_tab on rank 0

[baberabb]

closes #2266. Also removed the pkg_resources dependency as that's depreciated.

[baberabb]

This isn't really necessary as the code runs at the beginning (before generations). But couldn't hurt.

[al093]

suggestion: if sleep is not needed we should not add it.

[al093]

Sharing a slightly more verbose version.
I would do this:

def download_nltk_resources_guarded() -> None:
    """Download 'punkt_tab' tokenizer.

    Downloading nltk with distributed barrier otherwise race condition can occur
    when multiple processes try to download the same resource later.
    """
    local_rank = os.environ("LOCAL_RANK", 0)

    if local_rank == 0:
        try:
            nltk.data.find("tokenizers/punkt_tab")
        except LookupError:
            logger.info(f"Local rank {local_rank}: Downloading NLTK 'punkt_tab' resource.")
            nltk.download("punkt_tab")
            logger.info(f"Local rank {local_rank}: Downloaded NLTK 'punkt_tab' resource.")

    if torch.distributed.is_initialized():
         torch.distributed.barrier()
    try:
        nltk.data.find("tokenizers/punkt_tab")
    except LookupError:
        logger.error(
            f"Local rank {local_rank}: NLTK 'punkt' resource not found."
            f"This should have been downloaded by local rank 0."
        )
        raise

[al093]

suggestion: would rethink the current download on import behaviour.

[baberabb]

@al093 . Thanks very much for your suggestion. I'm hesitant in add a torch dependency here, as none of the tasks currently require it. I removed time.sleep() as you suggested and we do have a barrier later in the evaluation loop. To add an extra layer we could add a condition to have the user manually download and cache the data: please run python -c "import nltk; nltk.download('punkt'). Thoughts @haileyschoelkopf ?