| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Entrypoints for wrapping the core run_test implementation for specific test |
| 4 | types / modalities. |
| 5 | """ |
| 6 | |
| 7 | import itertools |
| 8 | from pathlib import PosixPath |
| 9 | |
| 10 | from .....conftest import ( |
| 11 | AudioTestAssets, |
| 12 | HfRunner, |
| 13 | ImageTestAssets, |
| 14 | VideoTestAssets, |
| 15 | VllmRunner, |
| 16 | ) |
| 17 | from . import builders, core |
| 18 | from .types import ExpandableVLMTestArgs, VLMTestInfo |
| 19 | |
| 20 | |
| 21 | ####### Entrypoints for running different test types |
| 22 | def run_single_image_test( |
| 23 | *, |
| 24 | tmp_path: PosixPath, |
| 25 | model_test_info: VLMTestInfo, |
| 26 | test_case: ExpandableVLMTestArgs, |
| 27 | hf_runner: type[HfRunner], |
| 28 | vllm_runner: type[VllmRunner], |
| 29 | image_assets: ImageTestAssets, |
| 30 | ): |
| 31 | assert test_case.size_wrappers |
| 32 | inputs = list( |
| 33 | itertools.chain.from_iterable( |
| 34 | builders.build_single_image_inputs_from_test_info( |
| 35 | model_test_info, image_assets, size_wrapper, tmp_path |
| 36 | ) |
| 37 | for size_wrapper in test_case.size_wrappers |
| 38 | ) |
| 39 | ) |
| 40 | |
| 41 | core.run_test( |
| 42 | hf_runner=hf_runner, |
| 43 | vllm_runner=vllm_runner, |
| 44 | inputs=inputs, |
| 45 | model=test_case.model, |
| 46 | dtype=test_case.dtype, |
| 47 | max_tokens=test_case.max_tokens, |
| 48 | num_logprobs=test_case.num_logprobs, |
| 49 | limit_mm_per_prompt={"image": 1}, |
| 50 | distributed_executor_backend=test_case.distributed_executor_backend, |
| 51 | **model_test_info.get_non_parametrized_runner_kwargs(), |
| 52 | ) |
| 53 | |
| 54 | |
| 55 | def run_multi_image_test( |
| 56 | *, |
| 57 | tmp_path: PosixPath, |
| 58 | model_test_info: VLMTestInfo, |
| 59 | test_case: ExpandableVLMTestArgs, |
| 60 | hf_runner: type[HfRunner], |
| 61 | vllm_runner: type[VllmRunner], |
| 62 | image_assets: ImageTestAssets, |
| 63 | ): |
| 64 | assert test_case.size_wrappers |
| 65 | inputs = list( |
| 66 | itertools.chain.from_iterable( |
| 67 | builders.build_multi_image_inputs_from_test_info( |
| 68 | model_test_info, image_assets, size_wrapper, tmp_path |
| 69 | ) |
| 70 | for size_wrapper in test_case.size_wrappers |
| 71 | ) |
| 72 | ) |
| 73 | |
| 74 | core.run_test( |
| 75 | hf_runner=hf_runner, |
| 76 | vllm_runner=vllm_runner, |
| 77 | inputs=inputs, |
| 78 | model=test_case.model, |
| 79 | dtype=test_case.dtype, |
| 80 | max_tokens=test_case.max_tokens, |
| 81 | num_logprobs=test_case.num_logprobs, |
| 82 | limit_mm_per_prompt={"image": len(image_assets)}, |
| 83 | distributed_executor_backend=test_case.distributed_executor_backend, |
| 84 | **model_test_info.get_non_parametrized_runner_kwargs(), |
| 85 | ) |
| 86 | |
| 87 | |
| 88 | def run_embedding_test( |
| 89 | *, |
| 90 | model_test_info: VLMTestInfo, |
| 91 | test_case: ExpandableVLMTestArgs, |
| 92 | hf_runner: type[HfRunner], |
| 93 | vllm_runner: type[VllmRunner], |
| 94 | image_assets: ImageTestAssets, |
| 95 | ): |
| 96 | assert test_case.size_wrappers |
| 97 | inputs_and_embeddings = [ |
| 98 | builders.build_embedding_inputs_from_test_info( |
| 99 | model_test_info, image_assets, size_wrapper |
| 100 | ) |
| 101 | for size_wrapper in test_case.size_wrappers |
| 102 | ] |
| 103 | inputs = list( |
| 104 | itertools.chain.from_iterable(inputs for inputs, _ in inputs_and_embeddings) |
| 105 | ) |
| 106 | vllm_embeddings = list( |
| 107 | itertools.chain.from_iterable( |
| 108 | embeddings for _, embeddings in inputs_and_embeddings |
| 109 | ) |
| 110 | ) |
| 111 | |
| 112 | core.run_test( |
| 113 | hf_runner=hf_runner, |
| 114 | vllm_runner=vllm_runner, |
| 115 | inputs=inputs, |
| 116 | model=test_case.model, |
| 117 | dtype=test_case.dtype, |
| 118 | max_tokens=test_case.max_tokens, |
| 119 | num_logprobs=test_case.num_logprobs, |
| 120 | limit_mm_per_prompt={"image": 1}, |
| 121 | vllm_embeddings=vllm_embeddings, |
| 122 | distributed_executor_backend=test_case.distributed_executor_backend, |
| 123 | **model_test_info.get_non_parametrized_runner_kwargs(), |
| 124 | ) |
| 125 | |
| 126 | |
| 127 | def run_video_test( |
| 128 | *, |
| 129 | model_test_info: VLMTestInfo, |
| 130 | test_case: ExpandableVLMTestArgs, |
| 131 | hf_runner: type[HfRunner], |
| 132 | vllm_runner: type[VllmRunner], |
| 133 | video_assets: VideoTestAssets, |
| 134 | ): |
| 135 | assert test_case.size_wrappers |
| 136 | assert test_case.num_video_frames is not None |
| 137 | inputs = list( |
| 138 | itertools.chain.from_iterable( |
| 139 | builders.build_video_inputs_from_test_info( |
| 140 | model_test_info, |
| 141 | video_assets, |
| 142 | size_wrapper, |
| 143 | test_case.num_video_frames, |
| 144 | test_case.needs_video_metadata, |
| 145 | ) |
| 146 | for size_wrapper in test_case.size_wrappers |
| 147 | ) |
| 148 | ) |
| 149 | |
| 150 | core.run_test( |
| 151 | hf_runner=hf_runner, |
| 152 | vllm_runner=vllm_runner, |
| 153 | inputs=inputs, |
| 154 | model=test_case.model, |
| 155 | dtype=test_case.dtype, |
| 156 | max_tokens=test_case.max_tokens, |
| 157 | num_logprobs=test_case.num_logprobs, |
| 158 | limit_mm_per_prompt={"video": len(video_assets)}, |
| 159 | distributed_executor_backend=test_case.distributed_executor_backend, |
| 160 | **model_test_info.get_non_parametrized_runner_kwargs(), |
| 161 | ) |
| 162 | |
| 163 | |
| 164 | def run_audio_test( |
| 165 | *, |
| 166 | model_test_info: VLMTestInfo, |
| 167 | test_case: ExpandableVLMTestArgs, |
| 168 | hf_runner: type[HfRunner], |
| 169 | vllm_runner: type[VllmRunner], |
| 170 | audio_assets: AudioTestAssets, |
| 171 | ): |
| 172 | inputs = builders.build_audio_inputs_from_test_info(model_test_info, audio_assets) |
| 173 | |
| 174 | core.run_test( |
| 175 | hf_runner=hf_runner, |
| 176 | vllm_runner=vllm_runner, |
| 177 | inputs=inputs, |
| 178 | model=test_case.model, |
| 179 | dtype=test_case.dtype, |
| 180 | max_tokens=test_case.max_tokens, |
| 181 | num_logprobs=test_case.num_logprobs, |
| 182 | limit_mm_per_prompt={"audio": 1}, |
| 183 | distributed_executor_backend=test_case.distributed_executor_backend, |
| 184 | **model_test_info.get_non_parametrized_runner_kwargs(), |
| 185 | ) |
| 186 | |
| 187 | |
| 188 | def run_custom_inputs_test( |
| 189 | *, |
| 190 | model_test_info: VLMTestInfo, |
| 191 | test_case: ExpandableVLMTestArgs, |
| 192 | hf_runner: type[HfRunner], |
| 193 | vllm_runner: type[VllmRunner], |
| 194 | ): |
| 195 | # Custom test cases can provide inputs directly, but they need to |
| 196 | # explicitly provided a CustomTestConfig, which wraps the inputs and |
| 197 | # the limit_mm_per_prompt |
| 198 | assert test_case.custom_test_opts is not None |
| 199 | |
| 200 | inputs = test_case.custom_test_opts.inputs |
| 201 | limit_mm_per_prompt = test_case.custom_test_opts.limit_mm_per_prompt |
| 202 | # Inputs and limit_mm_per_prompt should all be set |
| 203 | assert inputs is not None |
| 204 | assert limit_mm_per_prompt is not None |
| 205 | |
| 206 | core.run_test( |
| 207 | hf_runner=hf_runner, |
| 208 | vllm_runner=vllm_runner, |
| 209 | inputs=inputs, |
| 210 | model=test_case.model, |
| 211 | dtype=test_case.dtype, |
| 212 | max_tokens=test_case.max_tokens, |
| 213 | num_logprobs=test_case.num_logprobs, |
| 214 | limit_mm_per_prompt=limit_mm_per_prompt, |
| 215 | distributed_executor_backend=test_case.distributed_executor_backend, |
| 216 | **model_test_info.get_non_parametrized_runner_kwargs(), |
| 217 | ) |