"""
This file test accuracy via LMEval.
It uses local-completions, which interacts with vLLM
through the OAI API with N concurrent connections.
This simulates real work usage of the API and makes
sure that the zmq frontend mp RPC message passing and
AsyncLLMEngine are working correctly.
"""
import lm_eval
import pytest
MODEL_NAMES = ["Qwen/Qwen3-0.6B", "vllm-ascend/DeepSeek-V2-Lite-W8A8"]
NUM_CONCURRENT = 500
TASK = "gsm8k"
FILTER = "exact_match,strict-match"
RTOL = 0.03
EXPECTED_VALUES = {"Qwen/Qwen3-0.6B": 0.414, "vllm-ascend/DeepSeek-V2-Lite-W8A8": 0.34}
def run_test(model_name, more_args=None):
"""Run the end to end accuracy test."""
model_args = f"pretrained={model_name},max_model_len=4096,enforce_eager=True"
if more_args is not None:
model_args = "{},{}".format(model_args, more_args)
results = lm_eval.simple_evaluate(
model="vllm",
model_args=model_args,
tasks="gsm8k",
batch_size="auto",
)
measured_value = results["results"][TASK][FILTER]
assert model_name in EXPECTED_VALUES, f"Cannot find the expected value for the model {model_name=}"
expected_value = EXPECTED_VALUES[model_name]
assert measured_value - RTOL < expected_value and measured_value + RTOL > expected_value, (
f"Expected: {expected_value} | Measured: {measured_value}"
)
@pytest.mark.parametrize("model", MODEL_NAMES)
def test_lm_eval_accuracy(model):
"""Run with the V1 Engine."""
more_args = None
run_test(model, more_args)