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
"""#944: diverged sweep arms must stay visible and rank after finite losses."""
from __future__ import annotations
from unittest.mock import patch
from rich.console import Console
from typer.testing import CliRunner
from soup_cli.cli import app
from soup_cli.commands import sweep as sweep_mod
from .conftest import strip_ansi
def _result(name: str, loss: float) -> dict:
return {
"name": name,
"params": {"lr": name},
"run_id": name,
"final_loss": loss,
"duration": "1s",
"status": "completed",
}
def test_mixed_nonfinite_losses_sort_last_in_full_order():
results = [
_result("one", 1.0),
_result("nan", float("nan")),
_result("half", 0.5),
_result("inf", float("inf")),
_result("eight", 0.8),
]
ordered = sorted(results, key=sweep_mod._loss_sort_key)
assert [result["name"] for result in ordered] == ["half", "eight", "one", "nan", "inf"]
def test_failed_zero_loss_does_not_outrank_a_completed_run():
failed = _result("failed", 0.0)
failed["status"] = "failed"
ordered = sorted([failed, _result("completed", 0.5)], key=sweep_mod._loss_sort_key)
assert [result["name"] for result in ordered] == ["completed", "failed"]
def test_summary_labels_diverged_arms_and_never_selects_them_as_best(monkeypatch):
console = Console(record=True, width=120)
monkeypatch.setattr(sweep_mod, "console", console)
results = [
_result("nan-run", float("nan")),
_result("finite-run", 0.5),
_result("inf-run", float("inf")),
]
sweep_mod._display_summary(results, {"lr": ["nan", "finite", "inf"]})
output = console.export_text()
assert output.count("diverged") == 2
assert output.index("finite-run") < output.index("nan-run") < output.index("inf-run")
assert "Best run: finite-run (loss: 0.5000)" in output
def test_all_diverged_summary_has_no_best_marker_or_best_run(monkeypatch):
console = Console(record=True, width=120)
monkeypatch.setattr(sweep_mod, "console", console)
sweep_mod._display_summary(
[_result("nan-run", float("nan")), _result("inf-run", float("inf"))],
{"lr": ["nan", "inf"]},
)
output = console.export_text()
assert output.count("diverged") == 2
assert "*" not in output
assert "Best run:" not in output
def test_early_stop_treats_a_diverged_recent_arm_as_worse(tmp_path, monkeypatch):
config_file = tmp_path / "soup.yaml"
config_file.write_text(
"base: test-model\n"
"data:\n"
" train: ./data.jsonl\n",
encoding="utf-8",
)
run_results = [
{"final_loss": 1.0, "run_id": "run-1", "duration": "1s"},
{"final_loss": float("nan"), "run_id": "run-2", "duration": "1s"},
{"final_loss": 0.5, "run_id": "run-3", "duration": "1s"},
{"final_loss": float("inf"), "run_id": "run-4", "duration": "1s"},
{"final_loss": 0.8, "run_id": "run-5", "duration": "1s"},
]
runner = CliRunner()
with patch.object(sweep_mod, "_run_single", side_effect=run_results) as run_single:
result = runner.invoke(
app,
[
"sweep",
"--config",
str(config_file),
"--param",
"lr=1,2,3,4,5",
"--early-stop",
"1.5",
"--yes",
],
)
assert result.exit_code == 0, (result.output, repr(result.exception))
assert run_single.call_count == 2
output = strip_ansi(result.output)
assert output.lower().count("diverged") == 3
assert "skipping 3 remaining run(s)" in output