| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | |
| 6 | from tests.utils import single_gpu_only |
| 7 | from vllm import SamplingParams |
| 8 | |
| 9 | from .._correctness import check_mtp_correctness |
| 10 | |
| 11 | |
| 12 | @pytest.mark.parametrize( |
| 13 | ["model_setup", "mm_enabled", "expected_accuracy_threshold"], |
| 14 | [ |
| 15 | ( |
| 16 | ("mtp", "google/gemma-4-E4B-it", 1, "google/gemma-4-E4B-it-assistant"), |
| 17 | False, |
| 18 | 0.50, |
| 19 | ), # gemma4 MTP with assistant model, ref: ~62% |
| 20 | ], |
| 21 | ids=["gemma4-e4b"], |
| 22 | ) |
| 23 | @single_gpu_only |
| 24 | def test_mtp_correctness( |
| 25 | monkeypatch: pytest.MonkeyPatch, |
| 26 | sampling_config: SamplingParams, |
| 27 | model_setup: tuple[str, str, int, str | None], |
| 28 | mm_enabled: bool, |
| 29 | expected_accuracy_threshold: float, |
| 30 | vllm_runner, |
| 31 | ): |
| 32 | check_mtp_correctness( |
| 33 | monkeypatch, |
| 34 | sampling_config, |
| 35 | model_setup, |
| 36 | mm_enabled, |
| 37 | expected_accuracy_threshold, |
| 38 | vllm_runner, |
| 39 | ) |