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
"""v0.41.0 Part A — Optimizer Zoo tests."""
from __future__ import annotations
import pytest
from pydantic import ValidationError
from soup_cli.config.schema import TrainingConfig
from soup_cli.utils.optimizer_zoo import (
MLX_ONLY_OPTIMIZERS,
SUPPORTED_OPTIMIZERS,
is_new_v0_41_optimizer,
required_package,
validate_optimizer_name,
)
class TestSupportedOptimizers:
def test_default_in_allowlist(self):
assert "adamw_torch" in SUPPORTED_OPTIMIZERS
def test_new_v0_41_entries(self):
# The 4.x-era additions transformers 5.x rejects were retired (#1269);
# only the survivors remain in the allowlist.
for name in (
"apollo_adamw",
"lomo",
"adalomo",
"grokadamw",
"schedule_free_adamw",
"schedule_free_sgd",
):
assert name in SUPPORTED_OPTIMIZERS, name
assert is_new_v0_41_optimizer(name)
def test_torch_rejected_entries_retired(self):
# #1269: these eight used to be here but transformers 5.x refuses them
# in TrainingArguments, so they are refused at config load instead.
# (`muon` and `adamw_hf` are rejected by transformers too, but the MLX
# backend builds them itself, so they stay allowlisted:
# test_mlx_only_entries_stay_allowlisted.)
for name in (
"badam",
"adam_mini",
"dion",
"came_pytorch",
"ao_adamw_fp8",
"ao_adamw_4bit",
"ao_adamw_8bit",
"adamw_apex_fused",
):
assert name not in SUPPORTED_OPTIMIZERS, name
with pytest.raises(ValueError, match="no longer supported"):
validate_optimizer_name(name)
def test_mlx_only_entries_stay_allowlisted(self):
# #1283: refused per backend by SoupConfig, not by name.
assert MLX_ONLY_OPTIMIZERS == {"adamw_hf", "muon"}
for name in sorted(MLX_ONLY_OPTIMIZERS):
assert name in SUPPORTED_OPTIMIZERS, name
assert validate_optimizer_name(name) == name
def test_bnb_entries_present(self):
assert "adamw_bnb_8bit" in SUPPORTED_OPTIMIZERS
assert "paged_adamw_8bit" in SUPPORTED_OPTIMIZERS
def test_frozenset_immutable(self):
with pytest.raises(AttributeError):
SUPPORTED_OPTIMIZERS.add("evil") # type: ignore[attr-defined]
class TestValidateOptimizerName:
def test_default_passes(self):
assert validate_optimizer_name("adamw_torch") == "adamw_torch"
def test_uppercase_normalised(self):
assert validate_optimizer_name("GROKADAMW") == "grokadamw"
def test_unknown_rejected(self):
with pytest.raises(ValueError, match="not in the supported allowlist"):
validate_optimizer_name("magicoptimizer")
def test_non_string_rejected(self):
with pytest.raises(ValueError, match="must be a string"):
validate_optimizer_name(123) # type: ignore[arg-type]
def test_empty_rejected(self):
with pytest.raises(ValueError, match="non-empty"):
validate_optimizer_name("")
def test_null_byte_rejected(self):
with pytest.raises(ValueError, match="null bytes"):
validate_optimizer_name("adamw\x00")
def test_too_long_rejected(self):
with pytest.raises(ValueError, match="exceeds"):
validate_optimizer_name("a" * 65)
class TestRequiredPackage:
def test_native_returns_none(self):
assert required_package("adamw_torch") is None
def test_bnb_returns_none(self):
assert required_package("adamw_bnb_8bit") is None
def test_new_returns_pkg(self):
assert required_package("apollo_adamw") == "apollo-torch"
assert required_package("schedule_free_adamw") == "schedulefree"
class TestSchemaIntegration:
def test_default_optimizer_passes(self):
TrainingConfig()
def test_new_optimizer_accepted(self):
cfg = TrainingConfig(optimizer="grokadamw")
assert cfg.optimizer == "grokadamw"
def test_unknown_optimizer_rejected(self):
with pytest.raises(ValidationError) as exc:
TrainingConfig(optimizer="lookmaisnewname")
assert "not in the supported allowlist" in str(exc.value)
def test_uppercase_normalised(self):
cfg = TrainingConfig(optimizer="GrokAdamW")
assert cfg.optimizer == "grokadamw"
def test_null_byte_rejected(self):
with pytest.raises(ValidationError, match="null bytes"):
TrainingConfig(optimizer="adamw\x00")
class TestIsNewV041Optimizer:
def test_legacy_returns_false(self):
assert is_new_v0_41_optimizer("adamw_torch") is False
assert is_new_v0_41_optimizer("adafactor") is False
def test_bnb_returns_false(self):
assert is_new_v0_41_optimizer("adamw_bnb_8bit") is False
def test_unknown_returns_false(self):
assert is_new_v0_41_optimizer("not_an_optimizer") is False
def test_non_string_returns_false(self):
assert is_new_v0_41_optimizer(123) is False
assert is_new_v0_41_optimizer(None) is False
def test_case_insensitive(self):
assert is_new_v0_41_optimizer("GrokAdamW") is True
class TestRequiredPackageFull:
# #1269: the transformers-5-rejected additions were retired, so only the
# surviving v0.41.0 additions still map to a package.
@pytest.mark.parametrize("name,pkg", [
("lomo", "lomo-optim"),
("adalomo", "lomo-optim"),
("grokadamw", "grokadamw"),
("schedule_free_adamw", "schedulefree"),
("schedule_free_sgd", "schedulefree"),
])
def test_each_pkg(self, name, pkg):
assert required_package(name) == pkg
def test_unknown_returns_none(self):
assert required_package("not_an_optimizer") is None