test_models.py

"""Compare the outputs of HF and vLLM when using greedy sampling.

Run `pytest tests/models/test_models.py`.
"""
import pytest
import os

from ...utils import check_logprobs_close
from ....utils import models_path_prefix


@pytest.mark.parametrize(
    "model",
    [
        pytest.param(
            os.path.join(models_path_prefix, "bigscience/bloom-560m"),  # bloom - testing alibi slopes
            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
        ),
        pytest.param(
            os.path.join(models_path_prefix, "openai-community/gpt2"),  # gpt2
            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
        ),
        pytest.param(os.path.join(models_path_prefix, "Milos/slovak-gpt-j-405M")),  # gptj
        pytest.param(os.path.join(models_path_prefix, "bigcode/tiny_starcoder_py")),  # gpt_bigcode
        pytest.param(os.path.join(models_path_prefix, "EleutherAI/pythia-70m")),  # gpt_neox
        pytest.param(
            os.path.join(models_path_prefix, "google/gemma-1.1-2b-it"),  # gemma
            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
        ),
        pytest.param(
            os.path.join(models_path_prefix, "meta-llama/Llama-3.2-1B-Instruct"),  # llama
            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
        ),
        pytest.param(
            os.path.join(models_path_prefix, "openbmb/MiniCPM3-4B"),
            # fused_moe not supported on CPU
            marks=[pytest.mark.core_model],
        ),
        pytest.param(
            os.path.join(models_path_prefix, "facebook/opt-125m"),  # opt
            marks=[pytest.mark.core_model, pytest.mark.cpu_model],
        ),
        pytest.param(
            os.path.join(models_path_prefix, "microsoft/phi-2"),  # phi
            marks=[pytest.mark.core_model],
        ),
        pytest.param(
            os.path.join(models_path_prefix, "Qwen/Qwen2.5-0.5B-Instruct"),  # qwen2
            marks=[pytest.mark.core_model],
        ),
        pytest.param(os.path.join(models_path_prefix, "stabilityai/stablelm-3b-4e1t")),  # stablelm
        pytest.param(os.path.join(models_path_prefix, "bigcode/starcoder2-3b")),  # starcoder2
    ])
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("max_tokens", [32])
@pytest.mark.parametrize("num_logprobs", [5])
def test_models(
    hf_runner,
    vllm_runner,
    example_prompts,
    model: str,
    dtype: str,
    max_tokens: int,
    num_logprobs: int,
) -> None:

    with hf_runner(model, dtype=dtype) as hf_model:
        hf_outputs = hf_model.generate_greedy_logprobs_limit(
            example_prompts, max_tokens, num_logprobs)

    with vllm_runner(model, dtype=dtype) as vllm_model:
        vllm_outputs = vllm_model.generate_greedy_logprobs(
            example_prompts, max_tokens, num_logprobs)
        # This test is for verifying whether the model's extra_repr
        # can be printed correctly.
        print(vllm_model.model.llm_engine.model_executor.driver_worker.
              model_runner.model)

    check_logprobs_close(
        outputs_0_lst=hf_outputs,
        outputs_1_lst=vllm_outputs,
        name_0="hf",
        name_1="vllm",
    )