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
#!/usr/bin/env python3
"""Regenerate tests/fixtures/recipe_config_snapshots.json (#621, #638).
Deliberately a separate, explicit step rather than auto-updating on a failed
test: run this and review the diff, don't silently regenerate.
#638: the fixture no longer stores each recipe's full resolved config (162
copies of the same ~296 schema defaults). It stores the shared defaults
ONCE as "baseline", and per recipe only the fields that differ from that
baseline. A schema-default change then moves the one committed baseline
value, not 149-odd identical copies of it.
"""
from __future__ import annotations
import json
import sys
from pathlib import Path
_REPO_ROOT = Path(__file__).resolve().parent.parent
_FIXTURE_PATH = _REPO_ROOT / "tests" / "fixtures" / "recipe_config_snapshots.json"
# Minimal config: only the fields every SoupConfig requires (base, task,
# data.train, data.format) are set, to placeholder values no real recipe
# shares. Everything else resolves to its schema default. Those four
# placeholder-driven fields end up in EVERY recipe's delta automatically —
# no special-casing needed to keep genuinely-per-recipe fields out of the
# shared baseline, since no real recipe's value ever matches a placeholder.
_BASELINE_YAML = (
"base: __BASELINE_PLACEHOLDER_MODEL__\n"
"task: sft\n"
"data: {train: __baseline_placeholder__.jsonl, format: chatml}\n"
)
_MISSING = object()
def _ensure_src_on_path() -> None:
"""Make ``soup_cli`` importable, without piling up duplicate entries.
Called from every function that needs the import — each is usable on
its own — so membership is checked rather than inserting unconditionally.
"""
src = str(_REPO_ROOT / "src")
if src not in sys.path:
sys.path.insert(0, src)
def flatten(d: dict, prefix: str = "") -> dict[str, object]:
"""Flatten a nested dict to ``{"training.lora.r": 16, ...}``."""
out: dict[str, object] = {}
for key, value in d.items():
path = f"{prefix}.{key}" if prefix else key
if isinstance(value, dict):
out.update(flatten(value, path))
else:
out[path] = value
return out
def delta(resolved: dict[str, object], baseline: dict[str, object]) -> dict[str, object]:
"""Fields in ``resolved`` whose value differs from ``baseline``.
Public: this is part of the contract between the generator and the test
that verifies the fixture, not an internal of either.
A field present in ``resolved`` but absent from ``baseline`` (a nested
object a recipe fills in that the baseline leaves ``None``) counts as
differing — compared against a sentinel, never a value a real config
could hold.
"""
return {
path: value
for path, value in resolved.items()
if value != baseline.get(path, _MISSING)
}
def build_baseline() -> dict[str, object]:
"""The shared schema-default baseline every recipe's delta is relative to."""
_ensure_src_on_path()
from soup_cli.config.loader import load_config_from_string
return flatten(load_config_from_string(_BASELINE_YAML).model_dump(mode="json"))
def build_snapshot() -> dict[str, dict]:
_ensure_src_on_path()
from soup_cli.config.loader import load_config_from_string
from soup_cli.recipes.catalog import RECIPES
baseline = build_baseline()
recipes = {}
for name, recipe in sorted(RECIPES.items()):
try:
config = load_config_from_string(recipe.yaml_str)
except ValueError as exc:
if not recipe.unavailable_reason or recipe.unavailable_reason not in str(exc):
raise
recipes[name] = {"$unavailable": recipe.unavailable_reason}
else:
if recipe.unavailable_reason:
raise ValueError(f"{name} is marked unavailable but its config loads")
recipes[name] = delta(flatten(config.model_dump(mode="json")), baseline)
return {"baseline": baseline, "recipes": recipes}
def main() -> int:
snapshot = build_snapshot()
_FIXTURE_PATH.write_text(
json.dumps(snapshot, indent=2, sort_keys=True) + "\n",
encoding="utf-8",
newline="\n",
)
print(
f"Wrote baseline ({len(snapshot['baseline'])} fields) and "
f"{len(snapshot['recipes'])} recipe deltas to {_FIXTURE_PATH}"
)
return 0
if __name__ == "__main__":
raise SystemExit(main())