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
"""Judge URLs: the OpenAI key is attached only for the OpenAI API host."""
from __future__ import annotations
from unittest import mock
import pytest
BOUNDED_POST = "soup_cli.utils.http_response.post_capped"
@pytest.fixture(autouse=True)
def _key(monkeypatch):
monkeypatch.setenv("OPENAI_API_KEY", "sk-canary")
def _auth_header_for(url: str):
from soup_cli.eval.gate import _parse_judge_url
from soup_cli.eval.judge import JudgeEvaluator
provider, model, base = _parse_judge_url(url)
evaluator = JudgeEvaluator(provider=provider, model=model, api_base=base)
with mock.patch(BOUNDED_POST) as post:
post.return_value.status_code = 200
post.return_value.json.return_value = {
"choices": [{"message": {"content": '{"winner": "A"}'}}]
}
evaluator.compare_pair("p", "a", "b")
assert post.call_args is not None, "the judge was never called"
return post.call_args.args[0], post.call_args.kwargs["headers"].get("Authorization")
@pytest.mark.parametrize(
"url",
[
"https://judge.example.com/gpt-4o-mini",
"https://api.openai.com.example.net/gpt-4o-mini",
"https://api.groq.com/openai/llama3-70b-8192",
"https://example.com/api.openai.com/gpt-4o-mini",
],
)
def test_other_https_hosts_get_no_openai_key(url):
_target, auth = _auth_header_for(url)
assert "sk-canary" not in str(auth)
assert auth is None
@pytest.mark.parametrize(
"url", ["https://api.openai.com/gpt-4o-mini", "https://API.OPENAI.COM/gpt-4o-mini"]
)
def test_openai_host_keeps_key(url):
target, auth = _auth_header_for(url)
assert target == "https://" + url.split("/")[2] + "/v1/chat/completions"
assert auth == "Bearer sk-canary"
def test_other_https_host_is_server_provider():
from soup_cli.eval.gate import _parse_judge_url
assert _parse_judge_url("https://judge.example.com/m") == (
"server", "m", "https://judge.example.com"
)