@@ -0,0 +1,106 @@ |
| 1 | +"""Pack → unpack → `dlm prompt` produces byte-identical output (Sprint 14). |
| 2 | + |
| 3 | +Sprint 14 DoD §1: "pack a tiny-model store → delete original → unpack → |
| 4 | +`dlm prompt` works and produces the same output as before pack". The |
| 5 | +structural round-trip check lives in `test_round_trip.py`; this test |
| 6 | +owns the prompt-output invariant at `temperature=0` (greedy decoding, |
| 7 | +deterministic when weights are bit-identical). |
| 8 | + |
| 9 | +Doesn't destroy the shared `trained_store.home` — other session tests |
| 10 | +still need it. Instead, unpacks into a fresh home and compares |
| 11 | +generations from the two live stores side by side. |
| 12 | +""" |
| 13 | + |
| 14 | +from __future__ import annotations |
| 15 | + |
| 16 | +import os |
| 17 | +from pathlib import Path |
| 18 | + |
| 19 | +import pytest |
| 20 | +from typer.testing import CliRunner |
| 21 | + |
| 22 | +pytestmark = pytest.mark.slow |
| 23 | + |
| 24 | + |
| 25 | +_PROMPT_QUERY = "hello" |
| 26 | +_PROMPT_MAX_TOKENS = 32 |
| 27 | + |
| 28 | + |
| 29 | +def _run_prompt(home: Path, doc: Path) -> str: |
| 30 | + """Invoke `dlm prompt --temp 0` under the given home. Return stdout.""" |
| 31 | + from dlm.cli.app import app |
| 32 | + |
| 33 | + os.environ["DLM_HOME"] = str(home) |
| 34 | + runner = CliRunner(mix_stderr=False) |
| 35 | + result = runner.invoke( |
| 36 | + app, |
| 37 | + [ |
| 38 | + "prompt", |
| 39 | + str(doc), |
| 40 | + _PROMPT_QUERY, |
| 41 | + "--temp", |
| 42 | + "0", |
| 43 | + "--max-tokens", |
| 44 | + str(_PROMPT_MAX_TOKENS), |
| 45 | + ], |
| 46 | + ) |
| 47 | + assert result.exit_code == 0, f"prompt failed: {result.stderr}" |
| 48 | + return result.stdout |
| 49 | + |
| 50 | + |
| 51 | +@pytest.mark.slow |
| 52 | +def test_pack_unpack_prompt_is_byte_identical(trained_store, tmp_path: Path) -> None: |
| 53 | + """Pre-pack prompt output == post-unpack prompt output (greedy decode).""" |
| 54 | + from dlm.cli.app import app |
| 55 | + |
| 56 | + offline_vars = ("HF_HUB_OFFLINE", "TRANSFORMERS_OFFLINE", "HF_DATASETS_OFFLINE") |
| 57 | + saved = {k: os.environ.pop(k, None) for k in offline_vars} |
| 58 | + |
| 59 | + try: |
| 60 | + pre = _run_prompt(trained_store.home, trained_store.doc) |
| 61 | + |
| 62 | + # Pack under the original home. |
| 63 | + pack_path = tmp_path / "round-trip.pack" |
| 64 | + runner = CliRunner() |
| 65 | + pack_result = runner.invoke( |
| 66 | + app, |
| 67 | + [ |
| 68 | + "pack", |
| 69 | + str(trained_store.doc), |
| 70 | + "--out", |
| 71 | + str(pack_path), |
| 72 | + ], |
| 73 | + ) |
| 74 | + assert pack_result.exit_code == 0, pack_result.output |
| 75 | + assert pack_path.is_file() |
| 76 | + |
| 77 | + # Unpack into a fresh home. The original store stays intact so |
| 78 | + # other session-scoped tests aren't affected. |
| 79 | + fresh_home = tmp_path / "restored-home" |
| 80 | + fresh_home.mkdir() |
| 81 | + restored_dir = tmp_path / "restored-doc" |
| 82 | + restored_dir.mkdir() |
| 83 | + os.environ["DLM_HOME"] = str(fresh_home) |
| 84 | + unpack_result = runner.invoke( |
| 85 | + app, |
| 86 | + [ |
| 87 | + "unpack", |
| 88 | + str(pack_path), |
| 89 | + "--out", |
| 90 | + str(restored_dir), |
| 91 | + ], |
| 92 | + ) |
| 93 | + assert unpack_result.exit_code == 0, unpack_result.output |
| 94 | + |
| 95 | + restored_doc = restored_dir / trained_store.doc.name |
| 96 | + assert restored_doc.is_file(), f"no restored doc at {restored_doc}" |
| 97 | + post = _run_prompt(fresh_home, restored_doc) |
| 98 | + |
| 99 | + assert pre == post, ( |
| 100 | + "pack round-trip broke prompt output at temp=0.\n" |
| 101 | + f"pre: {pre!r}\npost: {post!r}" |
| 102 | + ) |
| 103 | + finally: |
| 104 | + for k, v in saved.items(): |
| 105 | + if v is not None: |
| 106 | + os.environ[k] = v |
