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
from pathlib import Path
import pytest
import yaml
from soup_cli.utils.ra_dit_run import run_ra_dit
@pytest.mark.parametrize(
"fail_generator",
[False, True],
ids=["success", "generator-failure"],
)
def test_ra_dit_preserves_input_configs(
tmp_path: Path,
monkeypatch: pytest.MonkeyPatch,
fail_generator: bool,
) -> None:
monkeypatch.chdir(tmp_path)
retriever = Path("retriever.yaml")
generator = Path("generator.yaml")
retriever.write_bytes(
b"# Original retriever config\n"
b"output: ./out/retriever\n"
b"training:\n"
b" ra_dit_stage: retriever\n"
)
generator.write_bytes(
b"# Original generator config\n"
b"output: ./out/generator\n"
b"training:\n"
b" ra_dit_stage: generator\n"
)
original_retriever = retriever.read_bytes()
original_generator = generator.read_bytes()
calls: list[Path] = []
def fake_runner(config_path: str) -> None:
path = Path(config_path)
calls.append(path)
if len(calls) == 1:
assert path.resolve() == retriever.resolve()
return
assert path.resolve() != generator.resolve()
assert path.resolve() != retriever.resolve()
config = yaml.safe_load(path.read_text(encoding="utf-8"))
assert (
config["training"]["ra_dit_retriever_model"]
== "./out/retriever"
)
if fail_generator:
raise RuntimeError("Simulated generator failure")
try:
if fail_generator:
with pytest.raises(RuntimeError, match="Simulated generator failure"):
run_ra_dit(
str(retriever),
str(generator),
_runner=fake_runner,
)
else:
result = run_ra_dit(
str(retriever),
str(generator),
_runner=fake_runner,
)
assert result.autolinked is True
finally:
assert retriever.read_bytes() == original_retriever
assert generator.read_bytes() == original_generator
assert len(calls) == 2
assert not calls[1].exists()
assert not calls[1].parent.exists()