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 #427's deliberate Qwen3.5 text-only recipes."""
from __future__ import annotations
import pytest
import yaml
from soup_cli.config.loader import load_config_from_string
from soup_cli.recipes.catalog import RECIPES
EXPECTED_QWEN35_TEXT_RECIPES = {
"qwen3.5-0.8b-sft",
"qwen3.5-0.8b-grpo",
"qwen3.5-2b-sft",
"qwen3.5-2b-dpo",
"qwen3.5-2b-grpo",
"qwen3.5-4b-sft",
"qwen3.5-4b-dpo",
"qwen3.5-4b-pretrain",
"qwen3.5-9b-sft",
"qwen3.5-9b-grpo",
"qwen3.5-9b-dpo",
"qwen3.5-27b-sft",
"qwen3.5-35b-a3b-sft",
"qwen3.5-35b-a3b-dpo",
"qwen3.5-35b-a3b-grpo",
"qwen3.5-122b-a10b-sft",
"qwen3.5-397b-a17b-sft",
# Qwen3.6 exposes the same qwen3_5 / qwen3_5_moe architecture on the Hub.
"qwen3.6-27b-sft",
"qwen3.6-27b-dpo",
"qwen3.6-27b-grpo",
"qwen3.6-35b-a3b-sft",
"qwen3.6-35b-a3b-dpo",
"qwen3.6-35b-a3b-grpo",
# Qwen3.8-27B exposes the same qwen3_5 text tower.
"qwen3.8-27b-sft",
}
def _qwen35_family_recipe_names() -> tuple[str, ...]:
"""Select every catalog recipe covered by the measured architecture decision."""
prefixes = ("Qwen/Qwen3.5-", "Qwen/Qwen3.6-", "Qwen/Qwen3.8-")
return tuple(name for name, recipe in RECIPES.items() if recipe.model.startswith(prefixes))
QWEN35_TEXT_RECIPES = _qwen35_family_recipe_names()
def test_qwen35_family_audit_covers_every_current_recipe() -> None:
"""A newly catalogued sibling must join the explicit-modality contract."""
assert set(QWEN35_TEXT_RECIPES) == EXPECTED_QWEN35_TEXT_RECIPES
@pytest.mark.parametrize("name", QWEN35_TEXT_RECIPES)
def test_qwen35_family_recipes_explicitly_select_language_tower(name: str) -> None:
"""Do not let the decoder-only decision fall back to the schema default."""
recipe = RECIPES[name]
raw = yaml.safe_load(recipe.yaml_str)
config = load_config_from_string(recipe.yaml_str)
assert raw["modality"] == "text"
assert config.modality == "text"