@@ -0,0 +1,46 @@ |
| 1 | +"""Minimal `slow` + `online` sanity check on the tiny-model fixture. |
| 2 | + |
| 3 | +Placeholder until Sprint 09 lands the real LoRA-loop assertions (audit 02 |
| 4 | +M3). What this guards today: |
| 5 | + |
| 6 | +- The cache-and-pre-warm CI step actually downloads SmolLM2-135M. |
| 7 | +- The downloaded cache contains a tokenizer with a chat template. |
| 8 | +- Subsequent `tiny_model_dir` invocations are a no-op (cache hit). |
| 9 | + |
| 10 | +Scope is deliberately narrow so the slow-tests CI job has *something* |
| 11 | +non-trivial to exercise. |
| 12 | +""" |
| 13 | + |
| 14 | +from __future__ import annotations |
| 15 | + |
| 16 | +from pathlib import Path |
| 17 | + |
| 18 | +import pytest |
| 19 | + |
| 20 | + |
| 21 | +@pytest.mark.slow |
| 22 | +@pytest.mark.online |
| 23 | +def test_tiny_model_dir_yields_usable_tokenizer(tiny_model_dir: Path) -> None: |
| 24 | + # Import inside the test so collection stays cheap for the fast |
| 25 | + # subset (transformers import is ~seconds). |
| 26 | + from transformers import AutoTokenizer |
| 27 | + |
| 28 | + tokenizer = AutoTokenizer.from_pretrained(str(tiny_model_dir)) |
| 29 | + assert tokenizer.chat_template, ( |
| 30 | + "SmolLM2-135M-Instruct must ship a chat template out of the box; " |
| 31 | + "absence here signals a broken cache or an upstream revision drift" |
| 32 | + ) |
| 33 | + # Basic encode roundtrip (no chat template applied — just vocab sanity). |
| 34 | + ids = tokenizer.encode("hello world", add_special_tokens=False) |
| 35 | + assert ids |
| 36 | + assert tokenizer.decode(ids).strip() == "hello world" |
| 37 | + |
| 38 | + |
| 39 | +@pytest.mark.slow |
| 40 | +@pytest.mark.online |
| 41 | +def test_tiny_model_dir_cache_hit_is_fast(tiny_model_dir: Path) -> None: |
| 42 | + """Second call should resolve from the session cache.""" |
| 43 | + from tests.fixtures.tiny_model import tiny_model_path |
| 44 | + |
| 45 | + again = tiny_model_path() |
| 46 | + assert again == tiny_model_dir |
