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 tests for #816: --auto-quant must not guess from timer noise."""
from __future__ import annotations
import inspect
import re
import pytest
from typer.testing import CliRunner
def _plain(text: str) -> str:
return " ".join(re.sub(r"\x1b\[[0-9;]*m", "", text).split())
@pytest.mark.parametrize("backend", ["transformers", "vllm", "sglang", "mii"])
def test_auto_quant_refuses_before_any_backend_can_apply_a_guess(backend: str) -> None:
from soup_cli.cli import app
result = CliRunner().invoke(
app,
["serve", "--model", "unused", "--backend", backend, "--auto-quant"],
)
output = _plain(result.output)
assert result.exit_code == 2
assert "--auto-quant is unavailable" in output
assert "Refusing instead of guessing" in output
assert "picked:" not in output
assert "binding vLLM" not in output
def test_auto_quant_result_is_independent_of_perf_counter(monkeypatch: pytest.MonkeyPatch) -> None:
from soup_cli.cli import app
ticks = iter([1000.0, 0.0, 50.0, -50.0])
monkeypatch.setattr("time.perf_counter", lambda: next(ticks))
runner = CliRunner()
first = runner.invoke(app, ["serve", "--model", "unused", "--auto-quant"])
second = runner.invoke(app, ["serve", "--model", "unused", "--auto-quant"])
assert first.exit_code == second.exit_code == 2
assert _plain(first.output) == _plain(second.output)
def test_serve_has_no_stub_picker_or_quantization_forwarding() -> None:
from soup_cli.commands.serve import serve
source = inspect.getsource(serve)
assert "run_auto_quant_picker" not in source
assert "quant_name_to_vllm_kwargs" not in source
assert "return (\"\", True)" not in source
assert "quantization=None" in source
def test_help_describes_refusal_instead_of_a_live_eval() -> None:
from soup_cli.cli import app
from soup_cli.commands.serve import serve
result = CliRunner().invoke(app, ["serve", "--help"], terminal_width=220)
output = _plain(result.output)
help_text = inspect.signature(serve).parameters["auto_quant"].default.help
assert result.exit_code == 0
assert "--auto-quant" in output
assert "refuses instead of guessing" in help_text
assert "pick fastest-at-acceptable-quality" not in help_text