| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from typing import cast |
| 4 | |
| 5 | import torch |
| 6 | from datasets import load_dataset |
| 7 | from transformers import AutoModelForImageTextToText, AutoProcessor |
| 8 | |
| 9 | import tests.ci_envs as ci_envs |
| 10 | from tests.models.utils import ( |
| 11 | GenerateModelInfo, |
| 12 | TokensTextLogprobsPromptLogprobs, |
| 13 | get_vllm_extra_kwargs, |
| 14 | ) |
| 15 | from vllm.logprobs import Logprob |
| 16 | |
| 17 | PPL_TOL = 0.01 |
| 18 | |
| 19 | |
| 20 | @torch.inference_mode |
| 21 | def vqa_ppl_test( |
| 22 | hf_runner, |
| 23 | vllm_runner, |
| 24 | model_info: GenerateModelInfo, |
| 25 | mm_processor_kwargs=None, |
| 26 | vllm_extra_kwargs=None, |
| 27 | tol=PPL_TOL, |
| 28 | ): |
| 29 | dataset = load_dataset("lmms-lab-encoder/llava-bench-in-the-wild", split="train") |
| 30 | |
| 31 | vllm_extra_kwargs = get_vllm_extra_kwargs(model_info, vllm_extra_kwargs) |
| 32 | |
| 33 | if mm_processor_kwargs is None: |
| 34 | mm_processor_kwargs = {} |
| 35 | |
| 36 | processor = AutoProcessor.from_pretrained(model_info.name) |
| 37 | image_token_id = processor.image_token_id |
| 38 | |
| 39 | images = [] |
| 40 | prompts = [] |
| 41 | for row in dataset: |
| 42 | image = row["image"] |
| 43 | images.append(image) |
| 44 | |
| 45 | question = row["question"] |
| 46 | answer = row["gpt_answer"] |
| 47 | |
| 48 | conversation = [ |
| 49 | { |
| 50 | "role": "user", |
| 51 | "content": [ |
| 52 | { |
| 53 | "type": "image", |
| 54 | }, |
| 55 | {"type": "text", "text": question}, |
| 56 | ], |
| 57 | }, |
| 58 | {"role": "assistant", "content": answer}, |
| 59 | ] |
| 60 | |
| 61 | prompt = processor.apply_chat_template(conversation, add_generation_prompt=True) |
| 62 | prompts.append(prompt) |
| 63 | |
| 64 | with vllm_runner( |
| 65 | model_info.name, |
| 66 | gpu_memory_utilization=0.7, |
| 67 | max_num_seqs=1, |
| 68 | max_model_len=4096, |
| 69 | mm_processor_kwargs=mm_processor_kwargs, |
| 70 | **vllm_extra_kwargs, |
| 71 | ) as vllm_model: |
| 72 | # Use max_num_seqs=1 to avoid OOM, |
| 73 | # and avoid batch different requests together. |
| 74 | |
| 75 | model_config = vllm_model.llm.llm_engine.model_config |
| 76 | mm_device_do_normalize = model_config.multimodal_config.mm_device_do_normalize |
| 77 | |
| 78 | # Confirm whether vllm is using the correct architecture |
| 79 | if model_info.architecture: |
| 80 | assert model_info.architecture in model_config.architectures |
| 81 | |
| 82 | nll_sum = torch.tensor(0.0, dtype=torch.float32, device="cpu") |
| 83 | n_tokens = 0 |
| 84 | |
| 85 | outputs = vllm_model.generate_greedy_logprobs( |
| 86 | prompts=prompts, |
| 87 | images=images, |
| 88 | max_tokens=1, |
| 89 | num_logprobs=None, |
| 90 | num_prompt_logprobs=0, |
| 91 | use_tqdm=False, |
| 92 | ) |
| 93 | |
| 94 | for output in outputs: |
| 95 | output = cast(TokensTextLogprobsPromptLogprobs, output) |
| 96 | token_datas = cast(list[dict[int, Logprob] | None], output[3]) |
| 97 | |
| 98 | assert token_datas[0] is None |
| 99 | token_log_probs = [] |
| 100 | for token_data in token_datas[1:]: |
| 101 | assert token_data is not None |
| 102 | assert len(token_data) == 1 |
| 103 | token_id, token_log_prob = next(iter(token_data.items())) |
| 104 | if token_id == image_token_id: |
| 105 | continue |
| 106 | token_log_probs.append(token_log_prob.logprob) |
| 107 | |
| 108 | neg_log_likelihood = -torch.tensor( |
| 109 | token_log_probs, dtype=torch.float32, device="cpu" |
| 110 | ).sum() |
| 111 | nll_sum += neg_log_likelihood |
| 112 | n_tokens += len(token_log_probs) |
| 113 | vllm_ppl = float(torch.exp(nll_sum / n_tokens)) |
| 114 | vllm_dtype = model_config.dtype |
| 115 | head_dtype = model_config.head_dtype |
| 116 | |
| 117 | # Accelerate ppl test by setting Transformers ppl score to a constant |
| 118 | if model_info.hf_ppl is None: |
| 119 | with hf_runner( |
| 120 | model_info.name, |
| 121 | dtype=ci_envs.VLLM_CI_HF_DTYPE or model_info.hf_dtype, |
| 122 | auto_cls=AutoModelForImageTextToText, |
| 123 | ) as hf_model: |
| 124 | nll_sum = torch.tensor(0.0, dtype=torch.float32, device="cpu") |
| 125 | n_tokens = 0 |
| 126 | |
| 127 | for prompt, image in zip(prompts, images): |
| 128 | inputs = processor( |
| 129 | text=[prompt], |
| 130 | images=[image], |
| 131 | padding=True, |
| 132 | return_tensors="pt", |
| 133 | **mm_processor_kwargs, |
| 134 | ) |
| 135 | inputs = inputs.to("cuda") |
| 136 | input_ids = inputs["input_ids"] |
| 137 | |
| 138 | labels = input_ids.masked_fill(input_ids == image_token_id, -100) |
| 139 | outputs = hf_model.model(**inputs, labels=labels) |
| 140 | neg_log_likelihood = outputs.loss |
| 141 | neg_log_likelihood = neg_log_likelihood.to(torch.float32).cpu() |
| 142 | |
| 143 | num_loss_tokens = int((labels[:, 1:] != -100).sum()) |
| 144 | nll_sum += neg_log_likelihood * num_loss_tokens |
| 145 | n_tokens += num_loss_tokens |
| 146 | |
| 147 | hf_ppl = float(torch.exp(nll_sum / n_tokens)) |
| 148 | hf_dtype = next(hf_model.model.parameters()).dtype |
| 149 | else: |
| 150 | hf_ppl = model_info.hf_ppl |
| 151 | hf_dtype = "Constant" |
| 152 | |
| 153 | differ = (vllm_ppl - hf_ppl) / hf_ppl |
| 154 | print("Model:", model_info.name) |
| 155 | print("mm_device_do_normalize:", mm_device_do_normalize) |
| 156 | print("VLLM:", f"dtype:{vllm_dtype}", f"head_dtype:{head_dtype}", vllm_ppl) |
| 157 | print("Transformers:", hf_dtype, hf_ppl) |
| 158 | print("Difference (%):", differ * 100) |
| 159 | |
| 160 | assert abs(differ) < tol |
| 161 | return vllm_ppl |