| 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 | from vllm.platforms import current_platform |
| 9 | |
| 10 | from .._correctness import check_mtp_correctness |
| 11 | |
| 12 | |
| 13 | @pytest.mark.parametrize( |
| 14 | ["model_setup", "mm_enabled", "expected_accuracy_threshold"], |
| 15 | [ |
| 16 | pytest.param( |
| 17 | ("mtp", "ZixiQi/DeepSeek-V3-4layers-MTP-FP8", 1, None), |
| 18 | False, |
| 19 | 0.0, |
| 20 | marks=pytest.mark.skipif( |
| 21 | current_platform.is_device_capability_family(100), |
| 22 | reason="DeepSeek MTP: TRTLLM MoE top_k check fails on Blackwell", |
| 23 | ), |
| 24 | ), # dummy model |
| 25 | ], |
| 26 | ids=["deepseek"], |
| 27 | ) |
| 28 | @single_gpu_only |
| 29 | def test_mtp_correctness( |
| 30 | monkeypatch: pytest.MonkeyPatch, |
| 31 | sampling_config: SamplingParams, |
| 32 | model_setup: tuple[str, str, int, str | None], |
| 33 | mm_enabled: bool, |
| 34 | expected_accuracy_threshold: float, |
| 35 | vllm_runner, |
| 36 | ): |
| 37 | check_mtp_correctness( |
| 38 | monkeypatch, |
| 39 | sampling_config, |
| 40 | model_setup, |
| 41 | mm_enabled, |
| 42 | expected_accuracy_threshold, |
| 43 | vllm_runner, |
| 44 | ) |