import io, os
import torch
from huggingface_hub import hf_hub_download

_CACHE = {}


def _wikitext(split):
    if split in _CACHE:
        return _CACHE[split]
    import pandas as pd
    fn = hf_hub_download("Salesforce/wikitext", repo_type="dataset",
                         filename=f"wikitext-2-raw-v1/{split}-00000-of-00001.parquet")
    txt = "\n\n".join(pd.read_parquet(fn)["text"].tolist())
    _CACHE[split] = txt
    return txt


def calib_batches(tok, nsamples, seqlen, seed=0):
    ids = tok(_wikitext("train"), return_tensors="pt").input_ids
    g = torch.Generator().manual_seed(seed)
    out = []
    for _ in range(nsamples):
        i = torch.randint(0, ids.shape[1] - seqlen - 1, (1,), generator=g).item()
        out.append(ids[:, i:i + seqlen])
    return out


def test_tokens(tok, seqlen):
    ids = tok(_wikitext("test"), return_tensors="pt").input_ids
    n = ids.shape[1] // seqlen
    return [ids[:, i * seqlen:(i + 1) * seqlen] for i in range(n)]