mlflow--mlflow
wehub-resource-sync
138 行
3.9 KiB
Python
138 行
3.9 KiB
Python
from unittest import mock
|
|
|
|
import openai
|
|
import pandas as pd
|
|
import pytest
|
|
|
|
import mlflow
|
|
from mlflow.models.evaluation import evaluate
|
|
from mlflow.tracing.constant import TraceMetadataKey
|
|
|
|
from tests.tracing.helper import get_traces, purge_traces, reset_autolog_state # noqa: F401
|
|
|
|
_EVAL_DATA = pd.DataFrame({
|
|
"inputs": [
|
|
"What is MLflow?",
|
|
"What is Spark?",
|
|
],
|
|
"ground_truth": [
|
|
"MLflow is an open-source platform to manage the ML lifecycle.",
|
|
"Spark is a unified analytics engine for big data processing.",
|
|
],
|
|
})
|
|
|
|
|
|
@pytest.fixture
|
|
def client(monkeypatch, mock_openai):
|
|
monkeypatch.setenv("OPENAI_API_KEY", "test")
|
|
monkeypatch.setenv("OPENAI_API_BASE", mock_openai)
|
|
return openai.OpenAI(api_key="test", base_url=mock_openai)
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
"config",
|
|
[
|
|
None,
|
|
{"log_traces": False},
|
|
{"log_traces": True},
|
|
],
|
|
)
|
|
@pytest.mark.usefixtures("reset_autolog_state")
|
|
def test_openai_evaluate(client, config):
|
|
if config:
|
|
mlflow.openai.autolog(**config)
|
|
|
|
is_trace_disabled = config and not config.get("log_traces", True)
|
|
is_trace_enabled = config and config.get("log_traces", True)
|
|
|
|
def model(inputs):
|
|
return [
|
|
client.chat.completions
|
|
.create(
|
|
messages=[{"role": "user", "content": question}],
|
|
model="gpt-4o-mini",
|
|
temperature=0.0,
|
|
)
|
|
.choices[0]
|
|
.message.content
|
|
for question in inputs["inputs"]
|
|
]
|
|
|
|
with mock.patch("mlflow.openai.log_model") as log_model_mock:
|
|
with mlflow.start_run() as run:
|
|
evaluate(
|
|
model,
|
|
data=_EVAL_DATA,
|
|
targets="ground_truth",
|
|
extra_metrics=[mlflow.metrics.exact_match()],
|
|
)
|
|
log_model_mock.assert_not_called()
|
|
|
|
# Traces should not be logged when disabled explicitly
|
|
if is_trace_disabled:
|
|
assert len(get_traces()) == 0
|
|
else:
|
|
assert len(get_traces()) == 2
|
|
assert run.info.run_id == get_traces()[0].info.request_metadata[TraceMetadataKey.SOURCE_RUN]
|
|
|
|
purge_traces()
|
|
|
|
# Test original autolog configs is restored
|
|
client.chat.completions.create(
|
|
messages=[{"role": "user", "content": "hi"}], model="gpt-4o-mini"
|
|
)
|
|
|
|
assert len(get_traces()) == (1 if is_trace_enabled else 0)
|
|
|
|
|
|
@pytest.mark.usefixtures("reset_autolog_state")
|
|
def test_openai_pyfunc_evaluate(client):
|
|
with mlflow.start_run() as run:
|
|
model_info = mlflow.openai.log_model(
|
|
"gpt-4o-mini",
|
|
"chat.completions",
|
|
name="model",
|
|
messages=[{"role": "system", "content": "You are an MLflow expert."}],
|
|
)
|
|
|
|
evaluate(
|
|
model_info.model_uri,
|
|
data=_EVAL_DATA,
|
|
targets="ground_truth",
|
|
extra_metrics=[mlflow.metrics.exact_match()],
|
|
)
|
|
assert len(get_traces()) == 2
|
|
assert run.info.run_id == get_traces()[0].info.request_metadata[TraceMetadataKey.SOURCE_RUN]
|
|
|
|
|
|
@pytest.mark.parametrize("globally_disabled", [True, False])
|
|
@pytest.mark.usefixtures("reset_autolog_state")
|
|
def test_openai_evaluate_should_not_log_traces_when_disabled(client, globally_disabled):
|
|
if globally_disabled:
|
|
mlflow.autolog(disable=True)
|
|
else:
|
|
mlflow.openai.autolog(disable=True)
|
|
|
|
def model(inputs):
|
|
return [
|
|
client.chat.completions
|
|
.create(
|
|
messages=[{"role": "user", "content": question}],
|
|
model="gpt-4o-mini",
|
|
temperature=0.0,
|
|
)
|
|
.choices[0]
|
|
.message.content
|
|
for question in inputs["inputs"]
|
|
]
|
|
|
|
with mlflow.start_run():
|
|
evaluate(
|
|
model,
|
|
data=_EVAL_DATA,
|
|
targets="ground_truth",
|
|
extra_metrics=[mlflow.metrics.exact_match()],
|
|
)
|
|
|
|
assert len(get_traces()) == 0
|