| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from contextlib import suppress |
| 5 | from dataclasses import dataclass, field |
| 6 | from http import HTTPStatus |
| 7 | from unittest.mock import AsyncMock, MagicMock |
| 8 | |
| 9 | import pytest |
| 10 | |
| 11 | from vllm.config.multimodal import MultiModalConfig |
| 12 | from vllm.entrypoints.openai.completion.protocol import CompletionRequest |
| 13 | from vllm.entrypoints.openai.completion.serving import OpenAIServingCompletion |
| 14 | from vllm.entrypoints.openai.models.protocol import BaseModelPath |
| 15 | from vllm.entrypoints.openai.models.serving import OpenAIServingModels |
| 16 | from vllm.entrypoints.serve.engine.protocol import ErrorResponse |
| 17 | from vllm.lora.request import LoRARequest |
| 18 | from vllm.lora.resolver import LoRAResolver, LoRAResolverRegistry |
| 19 | from vllm.renderers.hf import HfRenderer |
| 20 | from vllm.renderers.online_renderer import OnlineRenderer |
| 21 | from vllm.tokenizers.registry import cached_tokenizer_from_config |
| 22 | from vllm.v1.engine.async_llm import AsyncLLM |
| 23 | |
| 24 | MODEL_NAME = "openai-community/gpt2" |
| 25 | BASE_MODEL_PATHS = [BaseModelPath(name=MODEL_NAME, model_path=MODEL_NAME)] |
| 26 | |
| 27 | MOCK_RESOLVER_NAME = "mock_test_resolver" |
| 28 | |
| 29 | |
| 30 | @dataclass |
| 31 | class MockHFConfig: |
| 32 | model_type: str = "any" |
| 33 | |
| 34 | |
| 35 | @dataclass |
| 36 | class MockModelConfig: |
| 37 | """Minimal mock ModelConfig for testing.""" |
| 38 | |
| 39 | model: str = MODEL_NAME |
| 40 | runner_type = "generate" |
| 41 | tokenizer: str = MODEL_NAME |
| 42 | trust_remote_code: bool = False |
| 43 | tokenizer_mode: str = "auto" |
| 44 | max_model_len: int = 100 |
| 45 | tokenizer_revision: str | None = None |
| 46 | multimodal_config: MultiModalConfig = field(default_factory=MultiModalConfig) |
| 47 | hf_config: MockHFConfig = field(default_factory=MockHFConfig) |
| 48 | logits_processors: list[str] | None = None |
| 49 | diff_sampling_param: dict | None = None |
| 50 | allowed_local_media_path: str = "" |
| 51 | allowed_media_domains: list[str] | None = None |
| 52 | encoder_config = None |
| 53 | generation_config: str = "auto" |
| 54 | skip_tokenizer_init: bool = False |
| 55 | is_encoder_decoder: bool = False |
| 56 | is_multimodal_model: bool = False |
| 57 | supports_multimodal_inputs: bool = False |
| 58 | renderer_num_workers: int = 1 |
| 59 | |
| 60 | def get_diff_sampling_param(self): |
| 61 | return self.diff_sampling_param or {} |
| 62 | |
| 63 | |
| 64 | @dataclass |
| 65 | class MockParallelConfig: |
| 66 | _api_process_rank: int = 0 |
| 67 | |
| 68 | |
| 69 | @dataclass |
| 70 | class MockVllmConfig: |
| 71 | model_config: MockModelConfig |
| 72 | parallel_config: MockParallelConfig |
| 73 | |
| 74 | |
| 75 | class MockLoRAResolver(LoRAResolver): |
| 76 | async def resolve_lora( |
| 77 | self, base_model_name: str, lora_name: str |
| 78 | ) -> LoRARequest | None: |
| 79 | if lora_name == "test-lora": |
| 80 | return LoRARequest( |
| 81 | lora_name="test-lora", |
| 82 | lora_int_id=1, |
| 83 | lora_path="/fake/path/test-lora", |
| 84 | ) |
| 85 | elif lora_name == "invalid-lora": |
| 86 | return LoRARequest( |
| 87 | lora_name="invalid-lora", |
| 88 | lora_int_id=2, |
| 89 | lora_path="/fake/path/invalid-lora", |
| 90 | ) |
| 91 | return None |
| 92 | |
| 93 | |
| 94 | @pytest.fixture(autouse=True) |
| 95 | def register_mock_resolver(): |
| 96 | """Fixture to register and unregister the mock LoRA resolver.""" |
| 97 | resolver = MockLoRAResolver() |
| 98 | LoRAResolverRegistry.register_resolver(MOCK_RESOLVER_NAME, resolver) |
| 99 | yield |
| 100 | # Cleanup: remove the resolver after the test runs |
| 101 | if MOCK_RESOLVER_NAME in LoRAResolverRegistry.resolvers: |
| 102 | del LoRAResolverRegistry.resolvers[MOCK_RESOLVER_NAME] |
| 103 | |
| 104 | |
| 105 | def _build_renderer(model_config: MockModelConfig): |
| 106 | return HfRenderer( |
| 107 | MockVllmConfig(model_config, parallel_config=MockParallelConfig()), |
| 108 | cached_tokenizer_from_config(model_config), |
| 109 | ) |
| 110 | |
| 111 | |
| 112 | @pytest.fixture |
| 113 | def mock_serving_setup(): |
| 114 | """Provides a mocked engine and serving completion instance.""" |
| 115 | mock_engine = MagicMock(spec=AsyncLLM) |
| 116 | mock_engine.errored = False |
| 117 | |
| 118 | async def mock_add_lora_side_effect(lora_request: LoRARequest): |
| 119 | """Simulate engine behavior when adding LoRAs.""" |
| 120 | if lora_request.lora_name == "test-lora": |
| 121 | # Simulate successful addition |
| 122 | return True |
| 123 | if lora_request.lora_name == "invalid-lora": |
| 124 | # Simulate failure during addition (e.g. invalid format) |
| 125 | raise ValueError(f"Simulated failure adding LoRA: {lora_request.lora_name}") |
| 126 | return True |
| 127 | |
| 128 | mock_engine.add_lora = AsyncMock(side_effect=mock_add_lora_side_effect) |
| 129 | |
| 130 | async def mock_generate(*args, **kwargs): |
| 131 | for _ in []: |
| 132 | yield _ |
| 133 | |
| 134 | mock_engine.generate = MagicMock(spec=AsyncLLM.generate, side_effect=mock_generate) |
| 135 | |
| 136 | mock_engine.generate.reset_mock() |
| 137 | mock_engine.add_lora.reset_mock() |
| 138 | |
| 139 | mock_engine.model_config = MockModelConfig() |
| 140 | mock_engine.input_processor = MagicMock() |
| 141 | mock_engine.renderer = _build_renderer(mock_engine.model_config) |
| 142 | |
| 143 | models = OpenAIServingModels( |
| 144 | engine_client=mock_engine, |
| 145 | base_model_paths=BASE_MODEL_PATHS, |
| 146 | ) |
| 147 | |
| 148 | online_renderer = OnlineRenderer( |
| 149 | model_config=mock_engine.model_config, |
| 150 | renderer=mock_engine.renderer, |
| 151 | request_logger=None, |
| 152 | chat_template=None, |
| 153 | chat_template_content_format="auto", |
| 154 | ) |
| 155 | serving_completion = OpenAIServingCompletion( |
| 156 | mock_engine, models, online_renderer=online_renderer, request_logger=None |
| 157 | ) |
| 158 | |
| 159 | return mock_engine, serving_completion |
| 160 | |
| 161 | |
| 162 | @pytest.mark.asyncio |
| 163 | async def test_serving_completion_with_lora_resolver(mock_serving_setup, monkeypatch): |
| 164 | monkeypatch.setenv("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "true") |
| 165 | |
| 166 | mock_engine, serving_completion = mock_serving_setup |
| 167 | |
| 168 | lora_model_name = "test-lora" |
| 169 | req_found = CompletionRequest( |
| 170 | model=lora_model_name, |
| 171 | prompt="Generate with LoRA", |
| 172 | ) |
| 173 | |
| 174 | # Suppress potential errors during the mocked generate call, |
| 175 | # as we are primarily checking for add_lora and generate calls |
| 176 | with suppress(Exception): |
| 177 | await serving_completion.create_completion(req_found) |
| 178 | |
| 179 | mock_engine.add_lora.assert_awaited_once() |
| 180 | called_lora_request = mock_engine.add_lora.call_args[0][0] |
| 181 | assert isinstance(called_lora_request, LoRARequest) |
| 182 | assert called_lora_request.lora_name == lora_model_name |
| 183 | |
| 184 | mock_engine.generate.assert_called_once() |
| 185 | called_lora_request = mock_engine.generate.call_args[1]["lora_request"] |
| 186 | assert isinstance(called_lora_request, LoRARequest) |
| 187 | assert called_lora_request.lora_name == lora_model_name |
| 188 | |
| 189 | |
| 190 | @pytest.mark.asyncio |
| 191 | async def test_serving_completion_resolver_not_found(mock_serving_setup, monkeypatch): |
| 192 | monkeypatch.setenv("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "true") |
| 193 | |
| 194 | mock_engine, serving_completion = mock_serving_setup |
| 195 | |
| 196 | non_existent_model = "non-existent-lora-adapter" |
| 197 | req = CompletionRequest( |
| 198 | model=non_existent_model, |
| 199 | prompt="what is 1+1?", |
| 200 | ) |
| 201 | |
| 202 | response = await serving_completion.create_completion(req) |
| 203 | |
| 204 | mock_engine.add_lora.assert_not_awaited() |
| 205 | mock_engine.generate.assert_not_called() |
| 206 | |
| 207 | assert isinstance(response, ErrorResponse) |
| 208 | assert response.error.code == HTTPStatus.NOT_FOUND.value |
| 209 | assert non_existent_model in response.error.message |
| 210 | |
| 211 | |
| 212 | @pytest.mark.asyncio |
| 213 | async def test_serving_completion_resolver_add_lora_fails( |
| 214 | mock_serving_setup, monkeypatch |
| 215 | ): |
| 216 | monkeypatch.setenv("VLLM_ALLOW_RUNTIME_LORA_UPDATING", "true") |
| 217 | |
| 218 | mock_engine, serving_completion = mock_serving_setup |
| 219 | |
| 220 | invalid_model = "invalid-lora" |
| 221 | req = CompletionRequest( |
| 222 | model=invalid_model, |
| 223 | prompt="what is 1+1?", |
| 224 | ) |
| 225 | |
| 226 | response = await serving_completion.create_completion(req) |
| 227 | |
| 228 | # Assert add_lora was called before the failure |
| 229 | mock_engine.add_lora.assert_awaited_once() |
| 230 | called_lora_request = mock_engine.add_lora.call_args[0][0] |
| 231 | assert isinstance(called_lora_request, LoRARequest) |
| 232 | assert called_lora_request.lora_name == invalid_model |
| 233 | |
| 234 | # Assert generate was *not* called due to the failure |
| 235 | mock_engine.generate.assert_not_called() |
| 236 | |
| 237 | # Assert the correct error response |
| 238 | assert isinstance(response, ErrorResponse) |
| 239 | assert response.error.code == HTTPStatus.BAD_REQUEST.value |
| 240 | assert invalid_model in response.error.message |
| 241 | |
| 242 | |
| 243 | @pytest.mark.asyncio |
| 244 | async def test_serving_completion_flag_not_set(mock_serving_setup): |
| 245 | mock_engine, serving_completion = mock_serving_setup |
| 246 | |
| 247 | lora_model_name = "test-lora" |
| 248 | req_found = CompletionRequest( |
| 249 | model=lora_model_name, |
| 250 | prompt="Generate with LoRA", |
| 251 | ) |
| 252 | |
| 253 | await serving_completion.create_completion(req_found) |
| 254 | |
| 255 | mock_engine.add_lora.assert_not_called() |
| 256 | mock_engine.generate.assert_not_called() |