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
"""Continuing a QuEST artifact must preserve its measured clipping table."""
import copy
from types import SimpleNamespace
import pytest
from soup_cli.trainer.sft import SFTTrainerWrapper
from soup_cli.utils import quest
def wrapper_for(path, monkeypatch, declaration):
from soup_cli.trainer import stream_setup
wrapper = object.__new__(SFTTrainerWrapper)
wrapper.deepspeed_config = None
wrapper.fsdp_config = None
wrapper.model = SimpleNamespace(config=SimpleNamespace(soup_quest=declaration))
wrapper.config = SimpleNamespace(base=str(path))
wrapper._quest_base_identity_before = "local-sha256:" + "a" * 64
monkeypatch.setattr(stream_setup, "_distributed_launch", lambda: False)
monkeypatch.setattr(quest, "validate_cuda_hardware", lambda: ("RTX 3090", (8, 6)))
monkeypatch.setattr(quest, "resolve_base_model_identity", lambda _: "local-sha256:" + "a" * 64)
monkeypatch.setattr(
quest,
"calibrate_activation_scales",
lambda *_: pytest.fail("existing artifact was recalibrated"),
)
return wrapper
def metadata():
return quest.build_metadata(
activation_scales=dict.fromkeys(quest.EXPECTED_MODULES, 3.0),
base_model="local-sha256:" + "b" * 64,
calibration_sha256="c" * 64,
)
def test_warm_start_restores_the_declared_route_without_new_calibration(tmp_path, monkeypatch):
original = metadata()
quest.write_metadata(tmp_path, original)
wrapper = wrapper_for(tmp_path, monkeypatch, copy.deepcopy(original))
restored = []
monkeypatch.setattr(quest, "restore_mixed_quest", lambda model, value: restored.append(value))
wrapper._setup_quest([])
assert restored == [original]
assert wrapper._quest_metadata == original
assert wrapper.model.config.soup_quest == original
def test_warm_start_rejects_a_declaration_without_a_local_artifact(tmp_path, monkeypatch):
original = metadata()
wrapper = wrapper_for(tmp_path / "org" / "hub-model", monkeypatch, copy.deepcopy(original))
monkeypatch.setattr(
quest, "restore_mixed_quest", lambda *_: pytest.fail("non-local artifact restored")
)
with pytest.raises(ValueError, match="requires a local artifact"):
wrapper._setup_quest([])
@pytest.mark.parametrize("mutation", ["missing-sidecar", "missing-config", "different-config"])
def test_warm_start_rejects_incomplete_or_different_declarations(tmp_path, monkeypatch, mutation):
original = metadata()
if mutation != "missing-sidecar":
quest.write_metadata(tmp_path, original)
declaration = None if mutation == "missing-config" else copy.deepcopy(original)
if mutation == "different-config":
declaration["activation_scales"][quest.EXPECTED_MODULES[0]] = 4.0
wrapper = wrapper_for(tmp_path, monkeypatch, declaration)
monkeypatch.setattr(
quest, "restore_mixed_quest", lambda *_: pytest.fail("bad artifact restored")
)
with pytest.raises((ValueError, FileNotFoundError), match="(?i)(sidecar|metadata|declaration)"):
wrapper._setup_quest([])
def test_warm_start_rejects_base_changed_during_loading(tmp_path, monkeypatch):
original = metadata()
quest.write_metadata(tmp_path, original)
wrapper = wrapper_for(tmp_path, monkeypatch, original)
wrapper._quest_base_identity_before = "local-sha256:" + "d" * 64
monkeypatch.setattr(
quest, "restore_mixed_quest", lambda *_: pytest.fail("changed base restored")
)
with pytest.raises(ValueError, match="base changed"):
wrapper._setup_quest([])