Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Python cliconsumer-gpudpofine-tuningggufhuggingfacellmllmopslocal-ailocal-llmloralow-vrammachine-learningollamapeftpythonpytorchqlorasfttransformers
"""#1529 — unlearn's trainer result must carry the panel's 'duration' key.
``soup train --task unlearn`` crashed with ``KeyError: 'duration'`` in the
"Training Complete!" panel — AFTER the adapter was saved and the tracker row
was written — because ``UnlearnTrainerWrapper`` was the one transformers-backend
wrapper that returned only ``duration_secs``. These tests pin both halves of
the fix: the wrapper now formats the string like every other wrapper, and the
panel tolerates a result dict that lacks the key anyway.
"""
from __future__ import annotations
from pathlib import Path
from unittest.mock import MagicMock
from unittest.mock import patch as mock_patch
import pytest
from soup_cli.commands.train import _format_duration_display
from soup_cli.trainer.unlearn import UnlearnTrainerWrapper
TRAINER_DIR = Path(__file__).parent.parent / "src" / "soup_cli" / "trainer"
def _run_ready_unlearn_wrapper(tmp_path, monkeypatch) -> UnlearnTrainerWrapper:
"""A wrapper past setup() with an empty forget set (a zero-step run).
setup() itself refuses an empty forget set, so this bypasses it and
arranges exactly what a real call leaves behind: ``_setup_called`` set,
mock model/tokenizer, and an output dir under the (tmp) cwd that the
``_validated_output_dir`` gate accepts. ``train()`` then clears the
optimizer's gradients, finds no forget row to step over, and exercises
the save + result path.
"""
monkeypatch.chdir(tmp_path)
cfg = MagicMock()
cfg.training.epochs = 1
cfg.output = "unlearn-out"
wrapper = UnlearnTrainerWrapper(cfg)
wrapper._setup_called = True
wrapper.model = MagicMock()
wrapper.tokenizer = MagicMock()
wrapper._optimizer = MagicMock() # #1564: train() touches it even at zero steps
return wrapper
@pytest.mark.parametrize(("elapsed", "expected"), [(90, "1m"), (3720, "1h 2m")])
def test_unlearn_train_returns_formatted_duration(
tmp_path, monkeypatch, elapsed, expected
):
"""train() returns the pre-formatted 'duration' alongside duration_secs."""
wrapper = _run_ready_unlearn_wrapper(tmp_path, monkeypatch)
ticks = iter([0, elapsed])
with mock_patch(
"soup_cli.trainer.unlearn.time.monotonic", side_effect=lambda: next(ticks)
):
result = wrapper.train()
assert result["duration"] == expected
assert result["duration_secs"] == elapsed
@pytest.mark.parametrize(
("result", "expected"),
[
({"duration": "3m", "duration_secs": 185.0}, "3m"),
# The #1529 shape: seconds only, no pre-formatted string.
({"duration_secs": 90.0}, "1m"),
({"duration_secs": 3720.0}, "1h 2m"),
# A measured zero stays 0m; an absent measurement is not a zero.
({"duration_secs": 0.0}, "0m"),
# Neither key: a finished run still gets its summary panel.
({}, "unknown"),
],
)
def test_completion_panel_duration_falls_back_to_seconds(result, expected):
"""The panel reads 'duration' when present, else formats duration_secs."""
assert _format_duration_display(result) == expected
def test_completion_panel_never_indexes_result_duration_directly():
"""#1529 surfaced in the completion panel, after finish_run() had already run.
The helper tests above never reach that call site, so reverting the panel
line to ``result['duration']`` would leave them green. Pin the site itself.
"""
import ast
train_py = TRAINER_DIR.parent / "commands" / "train.py"
tree = ast.parse(train_py.read_text(encoding="utf-8"))
offenders = [
node.lineno
for node in ast.walk(tree)
if isinstance(node, ast.Subscript)
and isinstance(node.value, ast.Name)
and node.value.id == "result"
and isinstance(node.slice, ast.Constant)
and node.slice.value == "duration"
]
assert not offenders, (
f"commands/train.py indexes result['duration'] at line(s) {offenders}; "
"go through _format_duration_display(result) so a wrapper that omits the key "
"cannot crash a finished run"
)
def test_every_trainer_reporting_duration_secs_also_formats_duration():
"""Guard the whole wrapper family, not just unlearn (#1529).
Any trainer module that reports ``duration_secs`` but never emits the
``"duration"`` string key crashes the train completion panel after the
run is otherwise done. unlearn.py was the original offender; this scan
catches the next wrapper that repeats the mistake.
"""
offenders = []
for path in sorted(TRAINER_DIR.glob("*.py")):
src = path.read_text(encoding="utf-8")
if "duration_secs" not in src:
continue
if '"duration"' not in src and "'duration'" not in src:
offenders.append(path.name)
assert not offenders, f"trainer module(s) missing the duration key: {offenders}"