| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Core test implementation to be shared across modalities.""" |
| 4 | |
| 5 | from collections.abc import Callable |
| 6 | from typing import Any |
| 7 | |
| 8 | import torch |
| 9 | from transformers.models.auto.auto_factory import _BaseAutoModelClass |
| 10 | |
| 11 | from vllm.config.model import RunnerOption |
| 12 | from vllm.tokenizers import TokenizerLike |
| 13 | |
| 14 | from .....conftest import HfRunner, VllmRunner |
| 15 | from ....registry import HF_EXAMPLE_MODELS |
| 16 | from .types import PromptWithMultiModalInput, RunnerOutput |
| 17 | |
| 18 | |
| 19 | def run_test( |
| 20 | *, |
| 21 | hf_runner: type[HfRunner], |
| 22 | vllm_runner: type[VllmRunner], |
| 23 | inputs: list[PromptWithMultiModalInput], |
| 24 | model: str, |
| 25 | dtype: str, |
| 26 | max_tokens: int, |
| 27 | num_logprobs: int, |
| 28 | enforce_eager: bool, |
| 29 | max_model_len: int, |
| 30 | max_num_seqs: int, |
| 31 | hf_output_post_proc: Callable[[RunnerOutput, str], Any] | None, |
| 32 | vllm_output_post_proc: Callable[[RunnerOutput, str], Any] | None, |
| 33 | auto_cls: type[_BaseAutoModelClass], |
| 34 | use_tokenizer_eos: bool, |
| 35 | comparator: Callable[..., None], |
| 36 | get_stop_token_ids: Callable[[TokenizerLike], list[int]] | None, |
| 37 | stop_str: list[str] | None, |
| 38 | limit_mm_per_prompt: dict[str, int], |
| 39 | vllm_runner_kwargs: dict[str, Any] | None, |
| 40 | hf_model_kwargs: dict[str, Any] | None, |
| 41 | hf_processor: Callable[[str], Any] | None, |
| 42 | patch_hf_runner: Callable[[HfRunner], HfRunner] | None, |
| 43 | runner: RunnerOption = "auto", |
| 44 | distributed_executor_backend: str | None = None, |
| 45 | tensor_parallel_size: int = 1, |
| 46 | vllm_embeddings: torch.Tensor | None = None, |
| 47 | ): |
| 48 | """Modality agnostic test executor for comparing HF/vLLM outputs.""" |
| 49 | # In the case of embeddings, vLLM takes separate input tensors |
| 50 | vllm_inputs = vllm_embeddings if vllm_embeddings is not None else inputs |
| 51 | |
| 52 | model_info = HF_EXAMPLE_MODELS.find_hf_info(model) |
| 53 | model_info.check_available_online(on_fail="skip") |
| 54 | model_info.check_transformers_version(on_fail="skip") |
| 55 | |
| 56 | # Disable other modalities to save memory |
| 57 | default_limits = {"image": 0, "video": 0, "audio": 0} |
| 58 | limit_mm_per_prompt = default_limits | limit_mm_per_prompt |
| 59 | |
| 60 | vllm_outputs_per_mm = [] |
| 61 | hf_outputs_per_mm = [] |
| 62 | |
| 63 | # NOTE: take care of the order. run vLLM first, and then run HF. |
| 64 | # vLLM needs a fresh new process without cuda initialization. |
| 65 | # if we run HF first, the cuda initialization will be done and it |
| 66 | # will hurt multiprocessing backend with fork method (the default method). |
| 67 | |
| 68 | vllm_runner_kwargs_: dict[str, Any] = {"mm_processor_cache_gb": 0} |
| 69 | if model_info.tokenizer: |
| 70 | vllm_runner_kwargs_["tokenizer_name"] = model_info.tokenizer |
| 71 | if model_info.tokenizer_mode: |
| 72 | vllm_runner_kwargs_["tokenizer_mode"] = model_info.tokenizer_mode |
| 73 | if model_info.hf_overrides: |
| 74 | vllm_runner_kwargs_["hf_overrides"] = model_info.hf_overrides |
| 75 | if model_info.require_embed_inputs: |
| 76 | for k in ("skip_tokenizer_init", "enable_prompt_embeds", "enable_mm_embeds"): |
| 77 | vllm_runner_kwargs_[k] = model_info.require_embed_inputs |
| 78 | if not model_info.enable_prefix_caching: |
| 79 | vllm_runner_kwargs_["enable_prefix_caching"] = False |
| 80 | |
| 81 | if vllm_runner_kwargs: |
| 82 | vllm_runner_kwargs_.update(vllm_runner_kwargs) |
| 83 | |
| 84 | # Avoid passing limit_mm_per_prompt twice when vllm_runner_kwargs |
| 85 | # already contains it (e.g. gemma4 sets it via vllm_runner_kwargs). |
| 86 | if "limit_mm_per_prompt" in vllm_runner_kwargs_: |
| 87 | limit_mm_per_prompt = vllm_runner_kwargs_.pop("limit_mm_per_prompt") |
| 88 | |
| 89 | with vllm_runner( |
| 90 | model, |
| 91 | max_model_len=max_model_len, |
| 92 | max_num_seqs=max_num_seqs, |
| 93 | dtype=dtype, |
| 94 | limit_mm_per_prompt=limit_mm_per_prompt, |
| 95 | tensor_parallel_size=tensor_parallel_size, |
| 96 | distributed_executor_backend=distributed_executor_backend, |
| 97 | enforce_eager=enforce_eager, |
| 98 | runner=runner, |
| 99 | **vllm_runner_kwargs_, |
| 100 | ) as vllm_model: |
| 101 | tokenizer = vllm_model.llm.get_tokenizer() |
| 102 | |
| 103 | vllm_kwargs: dict[str, Any] = {} |
| 104 | if get_stop_token_ids is not None: |
| 105 | vllm_kwargs["stop_token_ids"] = get_stop_token_ids(tokenizer) |
| 106 | if stop_str: |
| 107 | vllm_kwargs["stop"] = stop_str |
| 108 | |
| 109 | for prompts, image_data, video_data, audio_data in vllm_inputs: |
| 110 | mm_data = dict(images=image_data, videos=video_data, audios=audio_data) |
| 111 | vllm_kwargs_with_mm_data = vllm_kwargs | mm_data |
| 112 | vllm_output = vllm_model.generate_greedy_logprobs( |
| 113 | prompts, |
| 114 | max_tokens, |
| 115 | num_logprobs=num_logprobs, |
| 116 | **vllm_kwargs_with_mm_data, |
| 117 | ) |
| 118 | vllm_outputs_per_mm.append(vllm_output) |
| 119 | |
| 120 | hf_runner_kwargs: dict[str, Any] = {} |
| 121 | if model_info.tokenizer: |
| 122 | hf_runner_kwargs["tokenizer_name"] = model_info.tokenizer |
| 123 | if hf_processor is not None: |
| 124 | hf_runner_kwargs["processor"] = hf_processor(model) |
| 125 | |
| 126 | hf_model = hf_runner( |
| 127 | model, |
| 128 | dtype=dtype, |
| 129 | auto_cls=auto_cls, |
| 130 | model_kwargs=hf_model_kwargs, |
| 131 | **hf_runner_kwargs, |
| 132 | ) |
| 133 | |
| 134 | # Some models need to patch things like the model processor, e.g., internvl |
| 135 | if patch_hf_runner is not None: |
| 136 | hf_model = patch_hf_runner(hf_model) |
| 137 | |
| 138 | with hf_model, torch.no_grad(): |
| 139 | tokenizer = hf_model.tokenizer |
| 140 | |
| 141 | # Some models need to explicitly pass the eos_token_id off the tokenizer |
| 142 | # or processor for a good comparison; |
| 143 | # currently assume processor/tokenizer agree on the EOS, and pull it off |
| 144 | # the tokenizer if requested. |
| 145 | hf_kwargs = {} |
| 146 | if use_tokenizer_eos: |
| 147 | hf_kwargs["eos_token_id"] = tokenizer.eos_token_id |
| 148 | if stop_str: |
| 149 | hf_kwargs["stop_strings"] = stop_str |
| 150 | |
| 151 | for prompts, image_data, video_data, audio_data in inputs: |
| 152 | mm_data = dict(images=image_data, videos=video_data, audios=audio_data) |
| 153 | hf_kwargs_with_mm_data = hf_kwargs | mm_data |
| 154 | hf_output = hf_model.generate_greedy_logprobs_limit( |
| 155 | prompts, |
| 156 | max_tokens, |
| 157 | num_logprobs=num_logprobs, |
| 158 | tokenizer=tokenizer, |
| 159 | **hf_kwargs_with_mm_data, |
| 160 | ) |
| 161 | hf_outputs_per_mm.append(hf_output) |
| 162 | |
| 163 | # Apply output processing / sanitation to the vLLM and HF runner results |
| 164 | hf_outputs_per_mm, vllm_outputs_per_mm = process_runner_outputs( |
| 165 | model, |
| 166 | first_runner_outputs=hf_outputs_per_mm, |
| 167 | second_runner_outputs=vllm_outputs_per_mm, |
| 168 | first_runner_processor=hf_output_post_proc, |
| 169 | second_runner_processor=vllm_output_post_proc, |
| 170 | ) |
| 171 | |
| 172 | for hf_outputs, vllm_outputs in zip(hf_outputs_per_mm, vllm_outputs_per_mm): |
| 173 | # This is usually check_logprobs_close, but it's passed through to |
| 174 | # allow things like check_outputs_equal where needed |
| 175 | comparator( |
| 176 | outputs_0_lst=hf_outputs, |
| 177 | outputs_1_lst=vllm_outputs, |
| 178 | name_0="hf", |
| 179 | name_1="vllm", |
| 180 | ) |
| 181 | |
| 182 | |
| 183 | def process_runner_outputs( |
| 184 | model, |
| 185 | first_runner_outputs, |
| 186 | second_runner_outputs, |
| 187 | first_runner_processor=None, |
| 188 | second_runner_processor=None, |
| 189 | ): |
| 190 | """Applies the runner processor(s) to the runner outputs, if any.""" |
| 191 | if first_runner_processor is not None: |
| 192 | first_runner_outputs = process_outputs( |
| 193 | first_runner_processor, model, first_runner_outputs |
| 194 | ) |
| 195 | if second_runner_processor is not None: |
| 196 | second_runner_outputs = process_outputs( |
| 197 | second_runner_processor, model, second_runner_outputs |
| 198 | ) |
| 199 | return first_runner_outputs, second_runner_outputs |
| 200 | |
| 201 | |
| 202 | def process_outputs(output_processor, model, outputs_per_image): |
| 203 | """Applies a model specific post-processor function to a runner's output.""" |
| 204 | return [ |
| 205 | [output_processor(res, model) for res in outputs] |
| 206 | for outputs in outputs_per_image |
| 207 | ] |