mlflow--mlflow
wehub-resource-sync
143 行
4.3 KiB
Python
143 行
4.3 KiB
Python
import importlib.metadata
|
|
|
|
import dspy
|
|
import pandas as pd
|
|
import pytest
|
|
from dspy.utils.dummies import DummyLM
|
|
from packaging.version import Version
|
|
|
|
import mlflow
|
|
from mlflow.tracing.constant import TraceMetadataKey
|
|
|
|
from tests.openai.test_openai_evaluate import purge_traces
|
|
from tests.tracing.helper import get_traces, reset_autolog_state # noqa: F401
|
|
|
|
if Version(importlib.metadata.version("dspy")) < Version("2.5.17"):
|
|
pytest.skip("Evaluation test requires dspy>=2.5.17", allow_module_level=True)
|
|
|
|
_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.",
|
|
],
|
|
})
|
|
|
|
|
|
def get_fake_model():
|
|
dspy.settings.configure(
|
|
lm=DummyLM({
|
|
"What is MLflow?": {
|
|
"answer": "MLflow is an open-source platform to manage the ML lifecycle.",
|
|
"reasoning": "No reasoning provided.",
|
|
},
|
|
"What is Spark?": {
|
|
"answer": "Spark is a unified analytics engine for big data processing.",
|
|
"reasoning": "No reasoning provided.",
|
|
},
|
|
})
|
|
)
|
|
|
|
class CoT(dspy.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.generate_answer = dspy.ChainOfThought("question -> answer")
|
|
|
|
def forward(self, question):
|
|
prediction = self.generate_answer(question=question)
|
|
return dspy.Prediction(answer=prediction.answer)
|
|
|
|
return CoT()
|
|
|
|
|
|
@pytest.mark.parametrize(
|
|
"config",
|
|
[
|
|
None,
|
|
{"log_traces": False},
|
|
{"log_traces": True},
|
|
],
|
|
)
|
|
@pytest.mark.usefixtures("reset_autolog_state")
|
|
def test_dspy_evaluate(config):
|
|
if config:
|
|
mlflow.dspy.autolog(**config)
|
|
|
|
is_trace_disabled = config and not config.get("log_traces", True)
|
|
is_trace_enabled = config and config.get("log_traces", True)
|
|
|
|
cot = get_fake_model()
|
|
|
|
def model(inputs):
|
|
return [cot(question).answer for question in inputs["inputs"]]
|
|
|
|
with mlflow.start_run() as run:
|
|
eval_result = mlflow.evaluate(
|
|
model,
|
|
data=_EVAL_DATA,
|
|
targets="ground_truth",
|
|
extra_metrics=[mlflow.metrics.exact_match()],
|
|
)
|
|
assert eval_result.metrics["exact_match/v1"] == 1.0
|
|
|
|
# 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
|
|
cot(question="What is MLflow?")
|
|
assert len(get_traces()) == (1 if is_trace_enabled else 0)
|
|
|
|
|
|
@pytest.mark.skip(
|
|
reason="DSPy pyfunc wrapper does not support batch inputs, which is required for eval. "
|
|
"Unskip when we add support for batch inputs."
|
|
)
|
|
@pytest.mark.usefixtures("reset_autolog_state")
|
|
def test_dspy_pyfunc_evaluate():
|
|
with mlflow.start_run() as run:
|
|
model_info = mlflow.dspy.log_model(get_fake_model(), name="model")
|
|
eval_result = mlflow.evaluate(
|
|
model_info.model_uri,
|
|
data=_EVAL_DATA,
|
|
targets="ground_truth",
|
|
extra_metrics=[mlflow.metrics.exact_match()],
|
|
)
|
|
assert eval_result.metrics["exact_match/v1"] == 1.0
|
|
|
|
# Traces should be automatically enabled during evaluation
|
|
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_dspy_evaluate_should_not_log_traces_when_disabled(globally_disabled):
|
|
if globally_disabled:
|
|
mlflow.autolog(disable=True)
|
|
else:
|
|
mlflow.dspy.autolog(disable=True)
|
|
|
|
cot = get_fake_model()
|
|
|
|
def model(inputs):
|
|
return [cot(question).answer for question in inputs["inputs"]]
|
|
|
|
with mlflow.start_run():
|
|
eval_result = mlflow.evaluate(
|
|
model,
|
|
data=_EVAL_DATA,
|
|
targets="ground_truth",
|
|
extra_metrics=[mlflow.metrics.exact_match()],
|
|
)
|
|
assert eval_result.metrics["exact_match/v1"] == 1.0
|
|
assert len(get_traces()) == 0
|