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
"""soup push: each hub authenticates with its own credential."""
from __future__ import annotations
import re
import pytest
from typer.testing import CliRunner
runner = CliRunner()
_ANSI = re.compile(r"\x1b\[[0-9;]*[A-Za-z]")
def _clean(text: str) -> str:
return " ".join(_ANSI.sub("", text).split())
@pytest.fixture
def model_dir(tmp_path, monkeypatch):
monkeypatch.chdir(tmp_path)
out = tmp_path / "out"
out.mkdir()
(out / "config.json").write_text("{}", encoding="utf-8")
# No cached HF login and no hub tokens unless a test sets one.
empty = tmp_path / "home"
empty.mkdir()
for var in ("HOME", "USERPROFILE", "HF_HOME"):
monkeypatch.setenv(var, str(empty))
for var in ("HF_TOKEN", "HUGGINGFACE_HUB_TOKEN", "MODELSCOPE_API_TOKEN", "MODELERS_TOKEN"):
monkeypatch.delenv(var, raising=False)
return out
@pytest.fixture
def recorded(monkeypatch):
calls: list[dict] = []
def _fake_upload_repo(hub, repo, **kwargs):
calls.append({"hub": hub, "repo": repo, **kwargs})
monkeypatch.setattr("soup_cli.utils.hubs.upload_repo", _fake_upload_repo)
return calls
def _push(*extra):
from soup_cli.cli import app
return runner.invoke(app, ["push", "--model", "out", "--repo", "user/m", *extra])
class TestHubTokenEnvVar:
@pytest.mark.parametrize(
("hub", "var"),
[("hf", "HF_TOKEN"), ("modelscope", "MODELSCOPE_API_TOKEN"),
("modelers", "MODELERS_TOKEN"), ("ModelScope", "MODELSCOPE_API_TOKEN")],
)
def test_maps_each_hub(self, hub, var):
from soup_cli.utils.hubs import hub_token_env_var
assert hub_token_env_var(hub) == var
def test_unknown_hub_rejected(self):
from soup_cli.utils.hubs import hub_token_env_var
with pytest.raises(ValueError):
hub_token_env_var("evilhub")
def test_resolve_refuses_hf(self):
from soup_cli.utils.hubs import resolve_hub_token
with pytest.raises(ValueError):
resolve_hub_token("hf")
def test_resolve_ignores_hf_env(self, monkeypatch):
from soup_cli.utils.hubs import resolve_hub_token
monkeypatch.setenv("HF_TOKEN", "hf_SENTINEL")
monkeypatch.delenv("MODELSCOPE_API_TOKEN", raising=False)
assert resolve_hub_token("modelscope") is None
class TestNonHfPush:
@pytest.mark.parametrize(("hub", "var"), [("modelscope", "MODELSCOPE_API_TOKEN"),
("modelers", "MODELERS_TOKEN")])
def test_hf_env_token_never_reaches_other_hub(self, model_dir, recorded, monkeypatch,
hub, var):
monkeypatch.setenv("HF_TOKEN", "hf_SENTINEL")
monkeypatch.setenv(var, "hub_OWN")
result = _push("--hub", hub)
assert result.exit_code == 0, (result.output, repr(result.exception))
assert recorded and recorded[0]["token"] == "hub_OWN"
assert "hf_SENTINEL" not in repr(recorded)
@pytest.mark.parametrize(("hub", "var"), [("modelscope", "MODELSCOPE_API_TOKEN"),
("modelers", "MODELERS_TOKEN")])
def test_only_hf_credential_is_refused_naming_hub_variable(self, model_dir, recorded,
monkeypatch, hub, var):
monkeypatch.setenv("HF_TOKEN", "hf_SENTINEL")
result = _push("--hub", hub)
assert result.exit_code == 1, (result.output, repr(result.exception))
assert var in _clean(result.output)
assert "hf_SENTINEL" not in result.output
assert recorded == []
def test_hf_cached_login_is_not_read(self, model_dir, recorded, monkeypatch):
import soup_cli.utils.hf as hf_mod
def _boom(*args, **kwargs):
raise AssertionError("HF token resolution must not run for a non-HF hub")
monkeypatch.setattr(hf_mod, "resolve_token", _boom)
monkeypatch.setenv("MODELSCOPE_API_TOKEN", "hub_OWN")
result = _push("--hub", "modelscope")
assert result.exit_code == 0, (result.output, repr(result.exception))
assert recorded[0]["token"] == "hub_OWN"
def test_command_line_token_is_used(self, model_dir, recorded, monkeypatch):
monkeypatch.setenv("HF_TOKEN", "hf_SENTINEL")
result = _push("--hub", "modelscope", "--token", "ms_CLI")
assert result.exit_code == 0, (result.output, repr(result.exception))
assert recorded[0]["token"] == "ms_CLI"
def test_no_credentials_at_all_names_variable(self, model_dir, recorded):
result = _push("--hub", "modelers")
assert result.exit_code == 1, (result.output, repr(result.exception))
assert "MODELERS_TOKEN" in _clean(result.output)
assert "HuggingFace" not in _clean(result.output)
assert recorded == []
class TestHfPushUnchanged:
def test_hf_hub_still_requires_hf_token(self, model_dir, recorded):
result = _push()
assert result.exit_code == 1, (result.output, repr(result.exception))
assert "No HuggingFace token found" in _clean(result.output)