"""Pinned, networked tiny-model load plus cache-only replay."""

from __future__ import annotations

import uuid

import pytest
import torch

from obliteratus.models.loader import load_model


pytestmark = [pytest.mark.network, pytest.mark.download]
MODEL = "hf-internal-testing/tiny-random-gpt2"
REVISION = "71034c5d8bde858ff824298bdedc65515b97d2b9"


def test_pinned_tiny_model_download_inference_and_offline_cache(monkeypatch):
    handle = load_model(
        MODEL,
        revision=REVISION,
        device="cpu",
        dtype="float32",
        trust_remote_code=False,
        skip_snapshot=True,
    )
    encoded = handle.tokenizer("conditional gate", return_tensors="pt")
    with torch.no_grad():
        output = handle.model(**encoded)
    assert output.logits.shape[:2] == encoded["input_ids"].shape
    assert next(handle.model.parameters()).device.type == "cpu"
    handle.cleanup()

    monkeypatch.setenv("HF_HUB_OFFLINE", "1")
    monkeypatch.setenv("TRANSFORMERS_OFFLINE", "1")
    cached = load_model(
        MODEL,
        revision=REVISION,
        device="cpu",
        local_files_only=True,
        trust_remote_code=False,
        skip_snapshot=True,
    )
    assert cached.model_name == MODEL
    cached.cleanup()

    missing = f"obliteratus/offline-missing-{uuid.uuid4().hex}"
    with pytest.raises(OSError):
        load_model(missing, revision=REVISION, device="cpu", local_files_only=True)