| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # ruff: noqa: E501 |
| 4 | """This example shows how to use vLLM for running offline inference with |
| 5 | the correct prompt format on vision language models for multimodal embedding. |
| 6 | |
| 7 | For most models, the prompt format should follow corresponding examples |
| 8 | on HuggingFace model repository. |
| 9 | """ |
| 10 | |
| 11 | import argparse |
| 12 | from pathlib import Path |
| 13 | |
| 14 | from PIL.Image import Image |
| 15 | |
| 16 | from vllm import LLM |
| 17 | from vllm.multimodal.utils import fetch_image |
| 18 | from vllm.utils.print_utils import print_embeddings |
| 19 | |
| 20 | ROOT_DIR = Path(__file__).parent.parent.parent |
| 21 | EMBED_TEMPLATE_DIR = ROOT_DIR / "pooling/embed/template/" |
| 22 | |
| 23 | image_url = "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/cat_snow.jpg" |
| 24 | text = "A cat standing in the snow." |
| 25 | multi_modal_data = {"image": fetch_image(image_url)} |
| 26 | |
| 27 | |
| 28 | def run_clip(seed: int): |
| 29 | llm = LLM( |
| 30 | model="openai/clip-vit-base-patch32", |
| 31 | runner="pooling", |
| 32 | limit_mm_per_prompt={"image": 1}, |
| 33 | seed=seed, |
| 34 | ) |
| 35 | |
| 36 | print("Text embedding output:") |
| 37 | outputs = llm.embed(text, use_tqdm=False) |
| 38 | print_embeddings(outputs[0].outputs.embedding) |
| 39 | |
| 40 | print("Image embedding output:") |
| 41 | prompt = "" # For image input, make sure that the prompt text is empty |
| 42 | outputs = llm.embed( |
| 43 | { |
| 44 | "prompt": prompt, |
| 45 | "multi_modal_data": multi_modal_data, |
| 46 | }, |
| 47 | use_tqdm=False, |
| 48 | ) |
| 49 | print_embeddings(outputs[0].outputs.embedding) |
| 50 | |
| 51 | |
| 52 | def run_e5_v(seed: int): |
| 53 | llm = LLM( |
| 54 | model="royokong/e5-v", |
| 55 | runner="pooling", |
| 56 | max_model_len=4096, |
| 57 | limit_mm_per_prompt={"image": 1}, |
| 58 | seed=seed, |
| 59 | ) |
| 60 | |
| 61 | llama3_template = "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n \n" # noqa: E501 |
| 62 | |
| 63 | print("Text embedding output:") |
| 64 | prompt_text = llama3_template.format( |
| 65 | f"{text}\nSummary above sentence in one word: " |
| 66 | ) |
| 67 | outputs = llm.embed(prompt_text, use_tqdm=False) |
| 68 | print_embeddings(outputs[0].outputs.embedding) |
| 69 | |
| 70 | print("Image embedding output:") |
| 71 | prompt_image = llama3_template.format("<image>\nSummary above image in one word: ") |
| 72 | outputs = llm.embed( |
| 73 | { |
| 74 | "prompt": prompt_image, |
| 75 | "multi_modal_data": multi_modal_data, |
| 76 | }, |
| 77 | use_tqdm=False, |
| 78 | ) |
| 79 | print_embeddings(outputs[0].outputs.embedding) |
| 80 | |
| 81 | |
| 82 | def run_qwen3_vl(seed: int): |
| 83 | try: |
| 84 | from qwen_vl_utils import smart_resize |
| 85 | except ModuleNotFoundError: |
| 86 | print( |
| 87 | "WARNING: `qwen-vl-utils` not installed, input images will not " |
| 88 | "be automatically resized. This can cause different results " |
| 89 | "comparing with HF repo's example. " |
| 90 | "You can enable this functionality by `pip install qwen-vl-utils`." |
| 91 | ) |
| 92 | smart_resize = None |
| 93 | |
| 94 | if smart_resize is not None: |
| 95 | |
| 96 | def post_process_image(image: Image) -> Image: |
| 97 | width, height = image.size |
| 98 | resized_height, resized_width = smart_resize( |
| 99 | height, |
| 100 | width, |
| 101 | factor=32, |
| 102 | ) |
| 103 | return image.resize((resized_width, resized_height)) |
| 104 | |
| 105 | multi_modal_data["image"] = post_process_image(multi_modal_data["image"]) |
| 106 | |
| 107 | default_instruction = "Represent the user's input." |
| 108 | image_placeholder = "<|vision_start|><|image_pad|><|vision_end|>" |
| 109 | prompt_text = f"<|im_start|>system\n{default_instruction}<|im_end|>\n<|im_start|>user\n{text}<|im_end|>\n<|im_start|>assistant\n" |
| 110 | prompt_image = f"<|im_start|>system\n{default_instruction}<|im_end|>\n<|im_start|>user\n{image_placeholder}<|im_end|>\n<|im_start|>assistant\n" |
| 111 | prompt_image_text = f"<|im_start|>system\n{default_instruction}<|im_end|>\n<|im_start|>user\n{image_placeholder}{text}<|im_end|>\n<|im_start|>assistant\n" |
| 112 | |
| 113 | llm = LLM( |
| 114 | model="Qwen/Qwen3-VL-Embedding-2B", |
| 115 | runner="pooling", |
| 116 | max_model_len=8192, |
| 117 | limit_mm_per_prompt={"image": 1}, |
| 118 | mm_processor_kwargs={"do_resize": False} if smart_resize is not None else None, |
| 119 | seed=seed, |
| 120 | ) |
| 121 | |
| 122 | print("Text embedding output:") |
| 123 | outputs = llm.embed(prompt_text, use_tqdm=False) |
| 124 | print_embeddings(outputs[0].outputs.embedding) |
| 125 | |
| 126 | print("Image embedding output:") |
| 127 | outputs = llm.embed( |
| 128 | { |
| 129 | "prompt": prompt_image, |
| 130 | "multi_modal_data": multi_modal_data, |
| 131 | }, |
| 132 | use_tqdm=False, |
| 133 | ) |
| 134 | print_embeddings(outputs[0].outputs.embedding) |
| 135 | |
| 136 | print("Image+Text embedding output:") |
| 137 | outputs = llm.embed( |
| 138 | { |
| 139 | "prompt": prompt_image_text, |
| 140 | "multi_modal_data": multi_modal_data, |
| 141 | }, |
| 142 | use_tqdm=False, |
| 143 | ) |
| 144 | print_embeddings(outputs[0].outputs.embedding) |
| 145 | |
| 146 | |
| 147 | def run_siglip(seed: int): |
| 148 | llm = LLM( |
| 149 | model="google/siglip-base-patch16-224", |
| 150 | runner="pooling", |
| 151 | limit_mm_per_prompt={"image": 1}, |
| 152 | seed=seed, |
| 153 | ) |
| 154 | |
| 155 | print("Text embedding output:") |
| 156 | # SigLIP is trained with padding to a fixed length and no attention mask, so |
| 157 | # text embeddings only line up with image embeddings when they are padded. |
| 158 | outputs = llm.embed( |
| 159 | text, |
| 160 | tokenization_kwargs={"padding": "max_length"}, |
| 161 | use_tqdm=False, |
| 162 | ) |
| 163 | print_embeddings(outputs[0].outputs.embedding) |
| 164 | |
| 165 | print("Image embedding output:") |
| 166 | prompt = "" # For image input, make sure that the prompt text is empty |
| 167 | outputs = llm.embed( |
| 168 | { |
| 169 | "prompt": prompt, |
| 170 | "multi_modal_data": multi_modal_data, |
| 171 | }, |
| 172 | use_tqdm=False, |
| 173 | ) |
| 174 | print_embeddings(outputs[0].outputs.embedding) |
| 175 | |
| 176 | |
| 177 | def run_vlm2vec_phi3v(seed: int): |
| 178 | llm = LLM( |
| 179 | model="TIGER-Lab/VLM2Vec-Full", |
| 180 | runner="pooling", |
| 181 | max_model_len=4096, |
| 182 | trust_remote_code=True, |
| 183 | mm_processor_kwargs={"num_crops": 4}, |
| 184 | limit_mm_per_prompt={"image": 1}, |
| 185 | seed=seed, |
| 186 | ) |
| 187 | image_token = "<|image_1|>" |
| 188 | |
| 189 | print("Text embedding output:") |
| 190 | prompt_text = f"Find me an everyday image that matches the given caption: {text}" |
| 191 | outputs = llm.embed(prompt_text, use_tqdm=False) |
| 192 | print_embeddings(outputs[0].outputs.embedding) |
| 193 | |
| 194 | print("Image embedding output:") |
| 195 | prompt_image = f"{image_token} Find a day-to-day image that looks similar to the provided image." # noqa: E501 |
| 196 | outputs = llm.embed( |
| 197 | { |
| 198 | "prompt": prompt_image, |
| 199 | "multi_modal_data": multi_modal_data, |
| 200 | }, |
| 201 | use_tqdm=False, |
| 202 | ) |
| 203 | print_embeddings(outputs[0].outputs.embedding) |
| 204 | |
| 205 | print("Image+Text embedding output:") |
| 206 | prompt_image_text = ( |
| 207 | f"{image_token} Represent the given image with the following question: {text}" # noqa: E501 |
| 208 | ) |
| 209 | outputs = llm.embed( |
| 210 | { |
| 211 | "prompt": prompt_image_text, |
| 212 | "multi_modal_data": multi_modal_data, |
| 213 | }, |
| 214 | use_tqdm=False, |
| 215 | ) |
| 216 | print_embeddings(outputs[0].outputs.embedding) |
| 217 | |
| 218 | |
| 219 | def run_vlm2vec_qwen2vl(seed: int): |
| 220 | # vLLM does not support LoRA adapters on multi-modal encoder, |
| 221 | # so we merge the weights first |
| 222 | from huggingface_hub.constants import HF_HUB_CACHE |
| 223 | from peft import PeftConfig, PeftModel |
| 224 | from transformers import AutoModelForImageTextToText, AutoProcessor |
| 225 | |
| 226 | from vllm.entrypoints.chat_utils import load_chat_template |
| 227 | |
| 228 | model_id = "TIGER-Lab/VLM2Vec-Qwen2VL-2B" |
| 229 | |
| 230 | base_model = AutoModelForImageTextToText.from_pretrained(model_id) |
| 231 | lora_model = PeftModel.from_pretrained( |
| 232 | base_model, |
| 233 | model_id, |
| 234 | config=PeftConfig.from_pretrained(model_id), |
| 235 | ) |
| 236 | model = lora_model.merge_and_unload().to(dtype=base_model.dtype) |
| 237 | model._hf_peft_config_loaded = False # Needed to save the merged model |
| 238 | |
| 239 | processor = AutoProcessor.from_pretrained( |
| 240 | model_id, |
| 241 | # `min_pixels` and `max_pixels` are deprecated for |
| 242 | # transformers `preprocessor_config.json` |
| 243 | size={"shortest_edge": 3136, "longest_edge": 12845056}, |
| 244 | ) |
| 245 | processor.chat_template = load_chat_template( |
| 246 | # The original chat template is not correct |
| 247 | EMBED_TEMPLATE_DIR / "vlm2vec_qwen2vl.jinja", |
| 248 | ) |
| 249 | |
| 250 | merged_path = str( |
| 251 | Path(HF_HUB_CACHE) / ("models--" + model_id.replace("/", "--") + "-vllm") |
| 252 | ) |
| 253 | print(f"Saving merged model to {merged_path}...") |
| 254 | print( |
| 255 | "NOTE: This directory is not tracked by `huggingface_hub` " |
| 256 | "so you have to delete this manually if you don't want it anymore." |
| 257 | ) |
| 258 | model.save_pretrained(merged_path) |
| 259 | processor.save_pretrained(merged_path) |
| 260 | print("Done!") |
| 261 | |
| 262 | llm = LLM( |
| 263 | model=merged_path, |
| 264 | runner="pooling", |
| 265 | max_model_len=4096, |
| 266 | mm_processor_kwargs={ |
| 267 | "min_pixels": 3136, |
| 268 | "max_pixels": 12845056, |
| 269 | }, |
| 270 | limit_mm_per_prompt={"image": 1}, |
| 271 | seed=seed, |
| 272 | ) |
| 273 | image_token = "<|image_pad|>" |
| 274 | |
| 275 | print("Text embedding output:") |
| 276 | prompt_text = f"Find me an everyday image that matches the given caption: {text}" |
| 277 | outputs = llm.embed(prompt_text, use_tqdm=False) |
| 278 | print_embeddings(outputs[0].outputs.embedding) |
| 279 | |
| 280 | print("Image embedding output:") |
| 281 | prompt_image = f"{image_token} Find a day-to-day image that looks similar to the provided image." # noqa: E501 |
| 282 | outputs = llm.embed( |
| 283 | { |
| 284 | "prompt": prompt_image, |
| 285 | "multi_modal_data": multi_modal_data, |
| 286 | }, |
| 287 | use_tqdm=False, |
| 288 | ) |
| 289 | print_embeddings(outputs[0].outputs.embedding) |
| 290 | |
| 291 | print("Image+Text embedding output:") |
| 292 | prompt_image_text = ( |
| 293 | f"{image_token} Represent the given image with the following question: {text}" # noqa: E501 |
| 294 | ) |
| 295 | outputs = llm.embed( |
| 296 | { |
| 297 | "prompt": prompt_image_text, |
| 298 | "multi_modal_data": multi_modal_data, |
| 299 | }, |
| 300 | use_tqdm=False, |
| 301 | ) |
| 302 | print_embeddings(outputs[0].outputs.embedding) |
| 303 | |
| 304 | |
| 305 | model_example_map = { |
| 306 | "clip": run_clip, |
| 307 | "e5_v": run_e5_v, |
| 308 | "qwen3_vl": run_qwen3_vl, |
| 309 | "siglip": run_siglip, |
| 310 | "vlm2vec_phi3v": run_vlm2vec_phi3v, |
| 311 | "vlm2vec_qwen2vl": run_vlm2vec_qwen2vl, |
| 312 | } |
| 313 | |
| 314 | |
| 315 | def parse_args(): |
| 316 | parser = argparse.ArgumentParser( |
| 317 | "Script to run a specified VLM through vLLM offline api." |
| 318 | ) |
| 319 | parser.add_argument( |
| 320 | "--model", |
| 321 | "-m", |
| 322 | type=str, |
| 323 | default="vlm2vec_phi3v", |
| 324 | choices=model_example_map.keys(), |
| 325 | help="The name of the embedding model.", |
| 326 | ) |
| 327 | parser.add_argument( |
| 328 | "--seed", |
| 329 | type=int, |
| 330 | default=0, |
| 331 | help="Set the seed when initializing `vllm.LLM`.", |
| 332 | ) |
| 333 | return parser.parse_args() |
| 334 | |
| 335 | |
| 336 | def main(args): |
| 337 | model_example_map[args.model](args.seed) |
| 338 | |
| 339 | |
| 340 | if __name__ == "__main__": |
| 341 | args = parse_args() |
| 342 | main(args) |