| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import sys |
| 4 | from unittest.mock import patch |
| 5 | |
| 6 | from vllm.config import VllmConfig |
| 7 | from vllm.engine.arg_utils import AsyncEngineArgs |
| 8 | from vllm.v1.engine.async_llm import AsyncLLM |
| 9 | |
| 10 | |
| 11 | def test_mp_reducer(): |
| 12 | """Test that _reduce_config reducer is registered when AsyncLLM is instantiated |
| 13 | without transformers_modules. This is a regression test for |
| 14 | https://github.com/vllm-project/vllm/pull/18640. |
| 15 | """ |
| 16 | # Ensure transformers_modules is not in sys.modules |
| 17 | if "transformers_modules" in sys.modules: |
| 18 | del sys.modules["transformers_modules"] |
| 19 | |
| 20 | with patch("multiprocessing.reducer.register") as mock_register: |
| 21 | engine_args = AsyncEngineArgs( |
| 22 | model="facebook/opt-125m", |
| 23 | max_model_len=32, |
| 24 | gpu_memory_utilization=0.1, |
| 25 | disable_log_stats=True, |
| 26 | ) |
| 27 | |
| 28 | async_llm = AsyncLLM.from_engine_args( |
| 29 | engine_args, |
| 30 | start_engine_loop=False, |
| 31 | ) |
| 32 | |
| 33 | assert mock_register.called, ( |
| 34 | "multiprocessing.reducer.register should have been called" |
| 35 | ) |
| 36 | |
| 37 | vllm_config_registered = False |
| 38 | for call_args in mock_register.call_args_list: |
| 39 | # Verify that a reducer for VllmConfig was registered |
| 40 | if len(call_args[0]) >= 2 and call_args[0][0] == VllmConfig: |
| 41 | vllm_config_registered = True |
| 42 | |
| 43 | reducer_func = call_args[0][1] |
| 44 | assert callable(reducer_func), "Reducer function should be callable" |
| 45 | break |
| 46 | |
| 47 | assert vllm_config_registered, ( |
| 48 | "VllmConfig should have been registered to multiprocessing.reducer" |
| 49 | ) |
| 50 | |
| 51 | async_llm.shutdown() |