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
"""Issue #889 — MCP data_doctor format validation matches data_validate."""
from __future__ import annotations
import json
import sys
from types import SimpleNamespace
import pytest
pytestmark = pytest.mark.unit
@pytest.fixture
def dataset(tmp_path, monkeypatch):
monkeypatch.chdir(tmp_path)
path = tmp_path / "d.jsonl"
path.write_text(
json.dumps({"instruction": "q", "output": "a"}) + "\n",
encoding="utf-8",
)
return path
def _stub_doctor_runtime(monkeypatch):
"""Keep these tests on format routing, not the optional tokenizer stack."""
from soup_cli.utils import data_doctor as doctor_module
monkeypatch.setitem(sys.modules, "transformers", SimpleNamespace())
monkeypatch.setattr(doctor_module, "resolve_tokenizer", lambda model, **kwargs: object())
class Report:
def __init__(self, fmt):
self.fmt = fmt
def to_dict(self):
return {"format": self.fmt}
monkeypatch.setattr(
doctor_module,
"run_doctor",
lambda rows, tok, *, fmt, max_length, sample_size: Report(fmt),
)
class TestInvalidFormatIsRefused:
@pytest.mark.parametrize("bad_format", ["", " ", "bogus", "ALPACA", "alpaca "])
def test_blank_and_unknown_values_are_rejected(self, dataset, bad_format):
from soup_cli.mcp_server.registry import McpToolError, tool_data_doctor
with pytest.raises(McpToolError) as exc_info:
tool_data_doctor(
{"data": str(dataset), "model": "fake/model", "format": bad_format}
)
message = str(exc_info.value)
assert f"unknown format {bad_format!r}" in message
def test_error_matches_data_validate_and_names_the_real_allowlist(self, dataset):
from soup_cli.data.formats import VALID_FORMATS
from soup_cli.mcp_server.registry import (
McpToolError,
tool_data_doctor,
tool_data_validate,
)
with pytest.raises(McpToolError) as validate_exc:
tool_data_validate({"data": str(dataset), "format": "bogus"})
with pytest.raises(McpToolError) as doctor_exc:
tool_data_doctor(
{"data": str(dataset), "model": "fake/model", "format": "bogus"}
)
message = str(doctor_exc.value)
assert message == str(validate_exc.value)
# The CLI's `soup data validate` prints auto first, then the declared
# VALID_FORMATS order. Both MCP tools should use that same live order.
assert message.endswith("accepted: auto, " + ", ".join(VALID_FORMATS))
assert "auto" in message
assert all(fmt in message for fmt in VALID_FORMATS)
assert str(dataset) not in message
assert len(message) < 2048
class TestSharedAllowlistAndAutoDetection:
def test_every_valid_format_reaches_the_doctor(self, dataset, monkeypatch):
from soup_cli.data.formats import VALID_FORMATS
from soup_cli.mcp_server.registry import tool_data_doctor
_stub_doctor_runtime(monkeypatch)
for fmt in sorted(VALID_FORMATS):
result = tool_data_doctor(
{"data": str(dataset), "model": "fake/model", "format": fmt}
)
assert result["format"] == fmt
def test_new_allowlist_member_needs_no_handler_change(self, dataset, monkeypatch):
from soup_cli.data import formats as formats_module
from soup_cli.mcp_server.registry import tool_data_doctor
_stub_doctor_runtime(monkeypatch)
monkeypatch.setattr(
formats_module,
"VALID_FORMATS",
set(formats_module.VALID_FORMATS) | {"zzz_new"},
)
result = tool_data_doctor(
{"data": str(dataset), "model": "fake/model", "format": "zzz_new"}
)
assert result["format"] == "zzz_new"
def test_omitted_format_still_auto_detects(self, dataset, monkeypatch):
from soup_cli.mcp_server.registry import tool_data_doctor
_stub_doctor_runtime(monkeypatch)
result = tool_data_doctor({"data": str(dataset), "model": "fake/model"})
assert result["format"] == "alpaca"