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
import pytest
from soup_cli.config.loader import load_config_from_string
@pytest.mark.parametrize("task", ["sft", "pretrain", "dpo"])
def test_bitnet_training_is_rejected_at_config_load(task: str) -> None:
yaml = (
"base: tiiuae/Falcon-E-1B-Instruct\n"
f"task: {task}\n"
"data: {train: ./d.jsonl}\n"
"training: {quantization: bitnet_1.58}\n"
)
with pytest.raises(ValueError) as exc_info:
load_config_from_string(yaml)
message = str(exc_info.value)
assert "BitNet 1.58 training is not implemented yet" in message
assert "soup export --format tq1_0" in message