| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import asyncio |
| 5 | import os |
| 6 | from collections.abc import Callable |
| 7 | from concurrent.futures import Future |
| 8 | from typing import Any |
| 9 | |
| 10 | import pytest |
| 11 | |
| 12 | from vllm.distributed.ec_transfer.ec_connector.utils import ECOutputAggregator |
| 13 | from vllm.distributed.kv_transfer.kv_connector.utils import KVOutputAggregator |
| 14 | from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs |
| 15 | from vllm.sampling_params import SamplingParams |
| 16 | from vllm.v1.engine.async_llm import AsyncLLM |
| 17 | from vllm.v1.engine.llm_engine import LLMEngine |
| 18 | from vllm.v1.executor import multiproc_executor as multiproc_executor_module |
| 19 | from vllm.v1.executor.abstract import Executor |
| 20 | from vllm.v1.executor.multiproc_executor import MultiprocExecutor |
| 21 | from vllm.v1.executor.uniproc_executor import ( |
| 22 | ExecutorWithExternalLauncher, |
| 23 | UniProcExecutor, |
| 24 | ) |
| 25 | |
| 26 | |
| 27 | class Mock: ... |
| 28 | |
| 29 | |
| 30 | def test_supports_async_scheduling_base_executor(): |
| 31 | assert Executor.supports_async_scheduling() is False |
| 32 | |
| 33 | |
| 34 | def test_supports_async_scheduling_uniproc_executor(): |
| 35 | assert UniProcExecutor.supports_async_scheduling() is True |
| 36 | |
| 37 | |
| 38 | def test_supports_async_scheduling_executor_with_external_launcher(): |
| 39 | # ExecutorWithExternalLauncher inherits from UniProcExecutor and does not |
| 40 | # override supports_async_scheduling, so it should return True. |
| 41 | assert ExecutorWithExternalLauncher.supports_async_scheduling() is True |
| 42 | |
| 43 | |
| 44 | def test_supports_async_scheduling_multiproc_executor(): |
| 45 | assert MultiprocExecutor.supports_async_scheduling() is True |
| 46 | |
| 47 | |
| 48 | class _FakeClock: |
| 49 | def __init__(self) -> None: |
| 50 | self.now = 0.0 |
| 51 | |
| 52 | def time(self) -> float: |
| 53 | return self.now |
| 54 | |
| 55 | def sleep(self, seconds: float) -> None: |
| 56 | self.now += seconds |
| 57 | |
| 58 | |
| 59 | class _FakeProcess: |
| 60 | def __init__(self, clock: _FakeClock, exits_at: float) -> None: |
| 61 | self.clock = clock |
| 62 | self.exits_at = exits_at |
| 63 | self.terminate_called = False |
| 64 | |
| 65 | def is_alive(self) -> bool: |
| 66 | return self.clock.time() < self.exits_at |
| 67 | |
| 68 | def terminate(self) -> None: |
| 69 | self.terminate_called = True |
| 70 | |
| 71 | |
| 72 | @pytest.mark.parametrize( |
| 73 | ("timeout", "exits_at", "expected_terminate"), |
| 74 | [ |
| 75 | pytest.param(6, 5, False, id="worker-exits-before-timeout"), |
| 76 | pytest.param(6, 7, True, id="worker-exceeds-timeout"), |
| 77 | ], |
| 78 | ) |
| 79 | def test_multiproc_executor_worker_termination_timeout( |
| 80 | monkeypatch, timeout, exits_at, expected_terminate |
| 81 | ): |
| 82 | monkeypatch.setenv("VLLM_WORKER_SHUTDOWN_TIMEOUT_SECONDS", str(timeout)) |
| 83 | clock = _FakeClock() |
| 84 | monkeypatch.setattr(multiproc_executor_module.time, "time", clock.time) |
| 85 | monkeypatch.setattr(multiproc_executor_module.time, "sleep", clock.sleep) |
| 86 | executor = MultiprocExecutor.__new__(MultiprocExecutor) |
| 87 | proc = _FakeProcess(clock, exits_at=exits_at) |
| 88 | executor._ensure_worker_termination([proc]) # type: ignore[list-item] |
| 89 | assert proc.terminate_called is expected_terminate |
| 90 | |
| 91 | |
| 92 | class CustomMultiprocExecutor(MultiprocExecutor): |
| 93 | def collective_rpc( # type: ignore[override] |
| 94 | self, |
| 95 | method: str | Callable, |
| 96 | timeout: float | None = None, |
| 97 | args: tuple = (), |
| 98 | kwargs: dict | None = None, |
| 99 | non_block: bool = False, |
| 100 | unique_reply_rank: int | None = None, |
| 101 | kv_output_aggregator: KVOutputAggregator | None = None, |
| 102 | ec_output_aggregator: ECOutputAggregator | None = None, |
| 103 | ) -> Any | list[Any] | Future[Any | list[Any]]: |
| 104 | # Drop marker to show that this was run |
| 105 | with open(".marker", "w"): |
| 106 | ... |
| 107 | return super().collective_rpc( |
| 108 | method, |
| 109 | timeout, |
| 110 | args, |
| 111 | kwargs, |
| 112 | non_block, |
| 113 | unique_reply_rank, |
| 114 | kv_output_aggregator, |
| 115 | ec_output_aggregator, |
| 116 | ) |
| 117 | |
| 118 | |
| 119 | CustomMultiprocExecutorAsync = CustomMultiprocExecutor |
| 120 | MODEL = "Qwen/Qwen3-0.6B" |
| 121 | |
| 122 | |
| 123 | def test_custom_executor_type_checking(): |
| 124 | with pytest.raises(ValueError): |
| 125 | engine_args = EngineArgs( |
| 126 | model=MODEL, |
| 127 | gpu_memory_utilization=0.2, |
| 128 | max_model_len=8192, |
| 129 | distributed_executor_backend=Mock, |
| 130 | ) |
| 131 | LLMEngine.from_engine_args(engine_args) |
| 132 | with pytest.raises(ValueError): |
| 133 | engine_args = AsyncEngineArgs( |
| 134 | model=MODEL, |
| 135 | gpu_memory_utilization=0.2, |
| 136 | max_model_len=8192, |
| 137 | distributed_executor_backend=Mock, |
| 138 | ) |
| 139 | AsyncLLM.from_engine_args(engine_args) |
| 140 | |
| 141 | |
| 142 | @pytest.mark.parametrize( |
| 143 | "distributed_executor_backend", |
| 144 | [ |
| 145 | CustomMultiprocExecutor, |
| 146 | "tests.v1.executor.test_executor.CustomMultiprocExecutor", |
| 147 | ], |
| 148 | ) |
| 149 | def test_custom_executor(distributed_executor_backend, tmp_path): |
| 150 | cwd = os.path.abspath(".") |
| 151 | os.chdir(tmp_path) |
| 152 | try: |
| 153 | assert not os.path.exists(".marker") |
| 154 | |
| 155 | engine_args = EngineArgs( |
| 156 | model=MODEL, |
| 157 | gpu_memory_utilization=0.2, |
| 158 | max_model_len=8192, |
| 159 | distributed_executor_backend=distributed_executor_backend, |
| 160 | enforce_eager=True, # reduce test time |
| 161 | ) |
| 162 | engine = LLMEngine.from_engine_args(engine_args) |
| 163 | sampling_params = SamplingParams(max_tokens=1) |
| 164 | |
| 165 | engine.add_request("0", "foo", sampling_params) |
| 166 | engine.step() |
| 167 | |
| 168 | assert os.path.exists(".marker") |
| 169 | finally: |
| 170 | os.chdir(cwd) |
| 171 | |
| 172 | |
| 173 | @pytest.mark.parametrize( |
| 174 | "distributed_executor_backend", |
| 175 | [ |
| 176 | CustomMultiprocExecutorAsync, |
| 177 | "tests.v1.executor.test_executor.CustomMultiprocExecutorAsync", |
| 178 | ], |
| 179 | ) |
| 180 | def test_custom_executor_async(distributed_executor_backend, tmp_path): |
| 181 | cwd = os.path.abspath(".") |
| 182 | os.chdir(tmp_path) |
| 183 | try: |
| 184 | assert not os.path.exists(".marker") |
| 185 | |
| 186 | engine_args = AsyncEngineArgs( |
| 187 | model=MODEL, |
| 188 | gpu_memory_utilization=0.2, |
| 189 | max_model_len=8192, |
| 190 | distributed_executor_backend=distributed_executor_backend, |
| 191 | enforce_eager=True, # reduce test time |
| 192 | ) |
| 193 | engine = AsyncLLM.from_engine_args(engine_args) |
| 194 | sampling_params = SamplingParams(max_tokens=1) |
| 195 | |
| 196 | async def t(): |
| 197 | stream = engine.generate( |
| 198 | request_id="0", prompt="foo", sampling_params=sampling_params |
| 199 | ) |
| 200 | async for x in stream: |
| 201 | ... |
| 202 | |
| 203 | asyncio.run(t()) |
| 204 | |
| 205 | assert os.path.exists(".marker") |
| 206 | finally: |
| 207 | os.chdir(cwd) |