| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import tempfile |
| 5 | from collections import OrderedDict |
| 6 | from importlib import reload |
| 7 | from unittest.mock import MagicMock |
| 8 | |
| 9 | import pytest |
| 10 | import torch |
| 11 | import torch.nn as nn |
| 12 | |
| 13 | from vllm.distributed import ( |
| 14 | cleanup_dist_env_and_memory, |
| 15 | init_distributed_environment, |
| 16 | initialize_model_parallel, |
| 17 | ) |
| 18 | from vllm.model_executor.layers.linear import ( |
| 19 | ColumnParallelLinear, |
| 20 | MergedColumnParallelLinear, |
| 21 | RowParallelLinear, |
| 22 | ) |
| 23 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 24 | from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead |
| 25 | from vllm.model_executor.models.interfaces import SupportsLoRA |
| 26 | from vllm.platforms import current_platform |
| 27 | from vllm.transformers_utils.repo_utils import hf_api |
| 28 | |
| 29 | |
| 30 | @pytest.fixture() |
| 31 | def should_do_global_cleanup_after_test(request) -> bool: |
| 32 | """Allow subdirectories to skip global cleanup by overriding this fixture. |
| 33 | This can provide a ~10x speedup for non-GPU unit tests since they don't need |
| 34 | to initialize torch. |
| 35 | """ |
| 36 | return not request.node.get_closest_marker("skip_global_cleanup") |
| 37 | |
| 38 | |
| 39 | @pytest.fixture(autouse=True) |
| 40 | def cleanup_fixture(should_do_global_cleanup_after_test: bool): |
| 41 | yield |
| 42 | if should_do_global_cleanup_after_test: |
| 43 | cleanup_dist_env_and_memory(shutdown_ray=True) |
| 44 | |
| 45 | |
| 46 | @pytest.fixture |
| 47 | def maybe_enable_lora_dual_stream(monkeypatch: pytest.MonkeyPatch): |
| 48 | if current_platform.is_cuda(): |
| 49 | monkeypatch.setenv("VLLM_LORA_ENABLE_DUAL_STREAM", "1") |
| 50 | import vllm.lora.layers.base_linear |
| 51 | |
| 52 | if not hasattr(vllm.lora.layers.base_linear, "lora_linear_async"): |
| 53 | # Reload the module to ensure the environment variable takes effect. |
| 54 | reload(vllm.lora.layers.base_linear) |
| 55 | yield |
| 56 | |
| 57 | |
| 58 | @pytest.fixture |
| 59 | def dist_init(): |
| 60 | from tests.utils import ensure_current_vllm_config |
| 61 | |
| 62 | temp_file = tempfile.mkstemp()[1] |
| 63 | |
| 64 | backend = "gloo" if current_platform.is_tpu() else current_platform.dist_backend |
| 65 | |
| 66 | with ensure_current_vllm_config(): |
| 67 | init_distributed_environment( |
| 68 | world_size=1, |
| 69 | rank=0, |
| 70 | distributed_init_method=f"file://{temp_file}", |
| 71 | local_rank=0, |
| 72 | backend=backend, |
| 73 | ) |
| 74 | initialize_model_parallel(1, 1) |
| 75 | yield |
| 76 | cleanup_dist_env_and_memory(shutdown_ray=True) |
| 77 | |
| 78 | |
| 79 | class DummyLoRAModel(nn.Sequential, SupportsLoRA): |
| 80 | pass |
| 81 | |
| 82 | |
| 83 | @pytest.fixture |
| 84 | def dummy_model(default_vllm_config) -> nn.Module: |
| 85 | model = DummyLoRAModel( |
| 86 | OrderedDict( |
| 87 | [ |
| 88 | ("dense1", ColumnParallelLinear(764, 100)), |
| 89 | ("dense2", RowParallelLinear(100, 50)), |
| 90 | ( |
| 91 | "layer1", |
| 92 | nn.Sequential( |
| 93 | OrderedDict( |
| 94 | [ |
| 95 | ("dense1", ColumnParallelLinear(100, 10)), |
| 96 | ("dense2", RowParallelLinear(10, 50)), |
| 97 | ] |
| 98 | ) |
| 99 | ), |
| 100 | ), |
| 101 | ("act2", nn.ReLU()), |
| 102 | ("output", ColumnParallelLinear(50, 10)), |
| 103 | ("outact", nn.Sigmoid()), |
| 104 | # Special handling for lm_head & sampler |
| 105 | ("lm_head", ParallelLMHead(32064, 10)), |
| 106 | ("logits_processor", LogitsProcessor(32064)), |
| 107 | ] |
| 108 | ) |
| 109 | ) |
| 110 | model.config = MagicMock() |
| 111 | model.embedding_modules = {"lm_head": "lm_head"} |
| 112 | model.unpadded_vocab_size = 32064 |
| 113 | return model |
| 114 | |
| 115 | |
| 116 | @pytest.fixture |
| 117 | def dummy_model_gate_up(default_vllm_config) -> nn.Module: |
| 118 | model = DummyLoRAModel( |
| 119 | OrderedDict( |
| 120 | [ |
| 121 | ("dense1", ColumnParallelLinear(764, 100)), |
| 122 | ("dense2", RowParallelLinear(100, 50)), |
| 123 | ( |
| 124 | "layer1", |
| 125 | nn.Sequential( |
| 126 | OrderedDict( |
| 127 | [ |
| 128 | ("dense1", ColumnParallelLinear(100, 10)), |
| 129 | ("dense2", RowParallelLinear(10, 50)), |
| 130 | ] |
| 131 | ) |
| 132 | ), |
| 133 | ), |
| 134 | ("act2", nn.ReLU()), |
| 135 | ("gate_up_proj", MergedColumnParallelLinear(50, [5, 5])), |
| 136 | ("outact", nn.Sigmoid()), |
| 137 | # Special handling for lm_head & sampler |
| 138 | ("lm_head", ParallelLMHead(32064, 10)), |
| 139 | ("logits_processor", LogitsProcessor(32064)), |
| 140 | ] |
| 141 | ) |
| 142 | ) |
| 143 | model.config = MagicMock() |
| 144 | model.packed_modules_mapping = { |
| 145 | "gate_up_proj": [ |
| 146 | "gate_proj", |
| 147 | "up_proj", |
| 148 | ], |
| 149 | } |
| 150 | model.embedding_modules = {"lm_head": "lm_head"} |
| 151 | model.unpadded_vocab_size = 32064 |
| 152 | |
| 153 | return model |
| 154 | |
| 155 | |
| 156 | @pytest.fixture(scope="session") |
| 157 | def mixtral_lora_files(): |
| 158 | # Note: this module has incorrect adapter_config.json to test |
| 159 | # https://github.com/vllm-project/vllm/pull/5909/files. |
| 160 | return hf_api().snapshot_download(repo_id="SangBinCho/mixtral-lora") |
| 161 | |
| 162 | |
| 163 | @pytest.fixture(scope="session") |
| 164 | def chatglm3_lora_files(): |
| 165 | return hf_api().snapshot_download(repo_id="jeeejeee/chatglm3-text2sql-spider") |
| 166 | |
| 167 | |
| 168 | @pytest.fixture(scope="session") |
| 169 | def baichuan_lora_files(): |
| 170 | return hf_api().snapshot_download(repo_id="jeeejeee/baichuan7b-text2sql-spider") |
| 171 | |
| 172 | |
| 173 | @pytest.fixture(scope="session") |
| 174 | def baichuan_zero_lora_files(): |
| 175 | # all the lora_B weights are initialized to zero. |
| 176 | return hf_api().snapshot_download(repo_id="jeeejeee/baichuan7b-zero-init") |
| 177 | |
| 178 | |
| 179 | @pytest.fixture(scope="session") |
| 180 | def baichuan_regex_lora_files(): |
| 181 | return hf_api().snapshot_download(repo_id="jeeejeee/baichuan-7b-lora-zero-regex") |
| 182 | |
| 183 | |
| 184 | @pytest.fixture(scope="session") |
| 185 | def ilama_lora_files(): |
| 186 | return hf_api().snapshot_download(repo_id="jeeejeee/ilama-text2sql-spider") |
| 187 | |
| 188 | |
| 189 | @pytest.fixture(scope="session") |
| 190 | def minicpmv_lora_files(): |
| 191 | return hf_api().snapshot_download(repo_id="jeeejeee/minicpmv25-lora-pokemon") |
| 192 | |
| 193 | |
| 194 | @pytest.fixture(scope="session") |
| 195 | def qwen2vl_lora_files(): |
| 196 | return hf_api().snapshot_download(repo_id="jeeejeee/qwen2-vl-lora-pokemon") |
| 197 | |
| 198 | |
| 199 | @pytest.fixture(scope="session") |
| 200 | def qwen25vl_lora_files(): |
| 201 | return hf_api().snapshot_download(repo_id="jeeejeee/qwen25-vl-lora-pokemon") |
| 202 | |
| 203 | |
| 204 | @pytest.fixture(scope="session") |
| 205 | def qwen2vl_language_lora_files(): |
| 206 | return hf_api().snapshot_download( |
| 207 | repo_id="prashanth058/qwen2vl-flickr-lora-language" |
| 208 | ) |
| 209 | |
| 210 | |
| 211 | @pytest.fixture(scope="session") |
| 212 | def qwen2vl_vision_tower_connector_lora_files(): |
| 213 | return hf_api().snapshot_download( |
| 214 | repo_id="prashanth058/qwen2vl-flickr-lora-tower-connector" |
| 215 | ) |
| 216 | |
| 217 | |
| 218 | @pytest.fixture(scope="session") |
| 219 | def qwen2vl_vision_tower_lora_files(): |
| 220 | return hf_api().snapshot_download(repo_id="prashanth058/qwen2vl-flickr-lora-tower") |
| 221 | |
| 222 | |
| 223 | @pytest.fixture(scope="session") |
| 224 | def qwen25vl_vision_lora_files(): |
| 225 | return hf_api().snapshot_download( |
| 226 | repo_id="EpochEcho/qwen2.5-3b-vl-lora-vision-connector" |
| 227 | ) |
| 228 | |
| 229 | |
| 230 | @pytest.fixture(scope="session") |
| 231 | def qwen3vl_vision_lora_files(): |
| 232 | return hf_api().snapshot_download( |
| 233 | repo_id="EpochEcho/qwen3-4b-vl-lora-vision-connector" |
| 234 | ) |
| 235 | |
| 236 | |
| 237 | @pytest.fixture(scope="session") |
| 238 | def gemma4_vision_lora_files(): |
| 239 | return hf_api().snapshot_download(repo_id="EpochEcho/gemma4-e2b-it-lora-pokemon") |
| 240 | |
| 241 | |
| 242 | @pytest.fixture(scope="session") |
| 243 | def qwen3_meowing_lora_files(): |
| 244 | """Download Qwen3 Meow LoRA files once per test session.""" |
| 245 | return hf_api().snapshot_download(repo_id="Jackmin108/Qwen3-0.6B-Meow-LoRA") |
| 246 | |
| 247 | |
| 248 | @pytest.fixture(scope="session") |
| 249 | def qwen3_woofing_lora_files(): |
| 250 | """Download Qwen3 Woof LoRA files once per test session.""" |
| 251 | return hf_api().snapshot_download(repo_id="Jackmin108/Qwen3-0.6B-Woof-LoRA") |
| 252 | |
| 253 | |
| 254 | @pytest.fixture(scope="session") |
| 255 | def qwen3_guard_star_trek_lora_files(): |
| 256 | return hf_api().snapshot_download( |
| 257 | repo_id="geoffmunn/Qwen3Guard-StarTrek-Classification-0.6B", |
| 258 | allow_patterns=["adapter_config.json", "adapter_model.safetensors"], |
| 259 | ) |
| 260 | |
| 261 | |
| 262 | @pytest.fixture(scope="session") |
| 263 | def qwen3_guard_new_zealand_lora_files(): |
| 264 | return hf_api().snapshot_download( |
| 265 | repo_id="geoffmunn/Qwen3Guard-NewZealand-Classification-0.6B", |
| 266 | allow_patterns=["adapter_config.json", "adapter_model.safetensors"], |
| 267 | ) |
| 268 | |
| 269 | |
| 270 | @pytest.fixture(scope="session") |
| 271 | def skywork_qwen3_reward_lora_files(): |
| 272 | return hf_api().snapshot_download( |
| 273 | repo_id="AmirMohseni/skywork-qwen3-0.6b-reward-lora", |
| 274 | allow_patterns=["adapter_config.json", "adapter_model.safetensors"], |
| 275 | ) |
| 276 | |
| 277 | |
| 278 | @pytest.fixture(scope="session") |
| 279 | def deepseekv2_lora_files(): |
| 280 | return hf_api().snapshot_download(repo_id="wuchen01/DeepSeek-V2-Lite-Chat-All-LoRA") |
| 281 | |
| 282 | |
| 283 | @pytest.fixture(scope="session") |
| 284 | def gptoss20b_lora_files(): |
| 285 | return hf_api().snapshot_download( |
| 286 | repo_id="jeeejeee/gpt-oss-20b-lora-adapter-text2sql" |
| 287 | ) |
| 288 | |
| 289 | |
| 290 | @pytest.fixture(scope="session") |
| 291 | def qwen3moe_lora_files(): |
| 292 | return hf_api().snapshot_download(repo_id="jeeejeee/qwen3-moe-text2sql-spider") |
| 293 | |
| 294 | |
| 295 | @pytest.fixture(scope="session") |
| 296 | def olmoe_lora_files(): |
| 297 | return hf_api().snapshot_download(repo_id="jeeejeee/olmoe-instruct-text2sql-spider") |
| 298 | |
| 299 | |
| 300 | @pytest.fixture(scope="session") |
| 301 | def qwen3_lora_files(): |
| 302 | return hf_api().snapshot_download(repo_id="charent/self_cognition_Alice") |
| 303 | |
| 304 | |
| 305 | @pytest.fixture(scope="session") |
| 306 | def llama32_lora_huggingface_id(): |
| 307 | # huggingface repo id is used to test lora runtime downloading. |
| 308 | return "jeeejeee/llama32-3b-text2sql-spider" |
| 309 | |
| 310 | |
| 311 | @pytest.fixture(scope="session") |
| 312 | def llama32_lora_files(llama32_lora_huggingface_id): |
| 313 | return hf_api().snapshot_download(repo_id=llama32_lora_huggingface_id) |
| 314 | |
| 315 | |
| 316 | @pytest.fixture(scope="session") |
| 317 | def whisper_lora_files(): |
| 318 | return hf_api().snapshot_download( |
| 319 | repo_id="chengyili2005/whisper-small-mandarin-lora" |
| 320 | ) |
| 321 | |
| 322 | |
| 323 | @pytest.fixture(scope="session") |
| 324 | def qwen35_text_lora_files(): |
| 325 | return hf_api().snapshot_download(repo_id="jeeejeee/qwen35-4b-text-only-sql-lora") |
| 326 | |
| 327 | |
| 328 | @pytest.fixture(scope="session") |
| 329 | def qwen35_vl_lora_files(): |
| 330 | return hf_api().snapshot_download( |
| 331 | repo_id="jeeejeee/qwen35-4b-all-linear-pokemon-lora" |
| 332 | ) |
| 333 | |
| 334 | |
| 335 | @pytest.fixture(scope="session") |
| 336 | def qwen36_moe_2d_lora_files(): |
| 337 | return hf_api().snapshot_download( |
| 338 | repo_id="jeeejeee/qwen36-35ba3b-2d-weights-poken-lora" |
| 339 | ) |
| 340 | |
| 341 | |
| 342 | @pytest.fixture(scope="session") |
| 343 | def qwen36_moe_3d_lora_files(): |
| 344 | return hf_api().snapshot_download( |
| 345 | repo_id="jeeejeee/qwen36-35ba3b-moe-all-linear-poken-lora" |
| 346 | ) |
| 347 | |
| 348 | |
| 349 | @pytest.fixture |
| 350 | def reset_default_device(): |
| 351 | """Some tests, such as `test_punica_ops.py`, explicitly set the |
| 352 | default device, which can affect subsequent tests. Adding this fixture |
| 353 | helps avoid this problem. |
| 354 | """ |
| 355 | original_device = torch.get_default_device() |
| 356 | yield |
| 357 | torch.set_default_device(original_device) |