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
"""Regression coverage for streamed adapter base-model provenance (#531)."""
import json
from typer.testing import CliRunner
from soup_cli.cli import app
from soup_cli.commands.chat import _detect_base_model
from soup_cli.utils.layer_stream_runtime import build_meta_skeleton
MODEL_ID = "Qwen/Qwen3.8-27B"
def _blank_tiny_config():
from transformers import LlamaConfig
config = LlamaConfig(
vocab_size=64,
hidden_size=32,
intermediate_size=64,
num_hidden_layers=1,
num_attention_heads=4,
num_key_value_heads=2,
)
config._name_or_path = ""
return config
def test_streamed_peft_config_saves_exact_configured_base(monkeypatch, tmp_path):
"""Pin the saved PEFT contract, not merely the in-memory symptom."""
import transformers
from peft import LoraConfig, TaskType, get_peft_model
monkeypatch.setattr(
transformers.AutoConfig,
"from_pretrained",
lambda *_args, **_kwargs: _blank_tiny_config(),
)
model = build_meta_skeleton(MODEL_ID, dtype="float32")
assert model.__dict__["name_or_path"] == MODEL_ID
assert model.config._name_or_path == MODEL_ID
adapter = get_peft_model(
model,
LoraConfig(
r=4,
lora_alpha=8,
target_modules=["q_proj", "v_proj"],
task_type=TaskType.CAUSAL_LM,
),
)
output = tmp_path / "adapter"
adapter.peft_config["default"].save_pretrained(output)
saved = json.loads((output / "adapter_config.json").read_text(encoding="utf-8"))
assert saved["base_model_name_or_path"] == MODEL_ID
assert _detect_base_model(output / "adapter_config.json") == MODEL_ID
def test_blank_provenance_control_breaks_auto_detection(tmp_path):
adapter = tmp_path / "blank-adapter"
adapter.mkdir()
config_path = adapter / "adapter_config.json"
config_path.write_text(
json.dumps({"base_model_name_or_path": "", "peft_type": "LORA"}),
encoding="utf-8",
)
assert _detect_base_model(config_path) == ""
assert not _detect_base_model(config_path)
def test_explicit_base_still_bypasses_auto_detection(monkeypatch, tmp_path):
adapter = tmp_path / "blank-adapter"
adapter.mkdir()
(adapter / "adapter_config.json").write_text(
json.dumps({"base_model_name_or_path": "", "peft_type": "LORA"}),
encoding="utf-8",
)
monkeypatch.setattr(
"soup_cli.utils.trust_remote.model_requires_trust_remote_code",
lambda *_args, **_kwargs: False,
)
monkeypatch.setattr(
"soup_cli.utils.trust_remote.resolve_trust_remote_code",
lambda *_args, **_kwargs: False,
)
monkeypatch.setattr(
"soup_cli.commands.chat._load_model",
lambda **_kwargs: (object(), object()),
)
result = CliRunner().invoke(
app,
[
"chat",
"--model",
str(adapter),
"--base",
MODEL_ID,
"--device",
"cpu",
],
input="/quit\n",
)
assert result.exit_code == 0
assert "Cannot detect base model" not in result.output
assert MODEL_ID in result.output