| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Mooncake P/D accuracy with same-host HIP transport on ROCm.""" |
| 4 | |
| 5 | import pytest |
| 6 | |
| 7 | from tests.utils import spawn_new_process_for_each_test |
| 8 | from tests.v1.kv_connector.rocm_pd_accuracy_utils import ( |
| 9 | QWEN_MODEL, |
| 10 | run_rocm_pd_accuracy, |
| 11 | ) |
| 12 | from vllm.platforms import current_platform |
| 13 | |
| 14 | pytestmark = [ |
| 15 | pytest.mark.skipif(not current_platform.is_rocm(), reason="ROCm P/D equivalents"), |
| 16 | pytest.mark.distributed(num_gpus=4), |
| 17 | ] |
| 18 | |
| 19 | |
| 20 | @pytest.mark.parametrize( |
| 21 | "model,prefill_tp,decode_tp", |
| 22 | [ |
| 23 | pytest.param(QWEN_MODEL, 1, 1, id="mooncake-hip-qwen-p1-d1"), |
| 24 | pytest.param(QWEN_MODEL, 2, 2, id="mooncake-hip-qwen-p2-d2"), |
| 25 | ], |
| 26 | ) |
| 27 | @spawn_new_process_for_each_test |
| 28 | def test_rocm_pd_accuracy(model, prefill_tp, decode_tp, monkeypatch): |
| 29 | run_rocm_pd_accuracy("mooncake", model, prefill_tp, decode_tp, monkeypatch) |