| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """This example shows how to use vLLM for running offline inference with |
| 4 | multi-image input on vision language models for text generation, |
| 5 | using the chat template defined by the model. |
| 6 | """ |
| 7 | |
| 8 | import os |
| 9 | from argparse import Namespace |
| 10 | from typing import NamedTuple |
| 11 | |
| 12 | from huggingface_hub import snapshot_download |
| 13 | from PIL.Image import Image |
| 14 | from transformers import AutoProcessor, AutoTokenizer |
| 15 | |
| 16 | from vllm import LLM, EngineArgs, SamplingParams |
| 17 | from vllm.lora.request import LoRARequest |
| 18 | from vllm.multimodal.utils import fetch_image |
| 19 | from vllm.platforms import current_platform |
| 20 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 21 | |
| 22 | QUESTION = "What is the content of each image?" |
| 23 | IMAGE_URLS = [ |
| 24 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/duck.jpg", |
| 25 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/lion.jpg", |
| 26 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/flycatcher.jpeg", |
| 27 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/somefish.jpg", |
| 28 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/starfish.jpg", |
| 29 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/snail.jpg", |
| 30 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/thistle.jpg", |
| 31 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/husky.jpg", |
| 32 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/orangetabbycat.jpg", |
| 33 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/guineapig.jpg", |
| 34 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/rabbit.jpg", |
| 35 | "https://vllm-public-assets.s3.us-west-2.amazonaws.com/multimodal_asset/horsepony.jpg", |
| 36 | ] |
| 37 | |
| 38 | |
| 39 | class ModelRequestData(NamedTuple): |
| 40 | engine_args: EngineArgs |
| 41 | prompt: str |
| 42 | image_data: list[Image] |
| 43 | stop_token_ids: list[int] | None = None |
| 44 | chat_template: str | None = None |
| 45 | lora_requests: list[LoRARequest] | None = None |
| 46 | sampling_params: SamplingParams | None = None |
| 47 | |
| 48 | |
| 49 | # NOTE: The default `max_num_seqs` and `max_model_len` may result in OOM on |
| 50 | # lower-end GPUs. |
| 51 | # Unless specified, these settings have been tested to work on a single L4. |
| 52 | |
| 53 | |
| 54 | def load_aria(question: str, image_urls: list[str]) -> ModelRequestData: |
| 55 | model_name = "rhymes-ai/Aria" |
| 56 | engine_args = EngineArgs( |
| 57 | model=model_name, |
| 58 | trust_remote_code=True, |
| 59 | dtype="bfloat16", |
| 60 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 61 | ) |
| 62 | placeholders = "<fim_prefix><|img|><fim_suffix>\n" * len(image_urls) |
| 63 | prompt = ( |
| 64 | f"<|im_start|>user\n{placeholders}{question}<|im_end|>\n<|im_start|>assistant\n" |
| 65 | ) |
| 66 | stop_token_ids = [93532, 93653, 944, 93421, 1019, 93653, 93519] |
| 67 | |
| 68 | return ModelRequestData( |
| 69 | engine_args=engine_args, |
| 70 | prompt=prompt, |
| 71 | stop_token_ids=stop_token_ids, |
| 72 | image_data=[fetch_image(url) for url in image_urls], |
| 73 | ) |
| 74 | |
| 75 | |
| 76 | def load_bee(question: str, image_urls: list[str]) -> ModelRequestData: |
| 77 | model_name = "Open-Bee/Bee-8B-RL" |
| 78 | |
| 79 | engine_args = EngineArgs( |
| 80 | model=model_name, |
| 81 | max_model_len=16384, |
| 82 | max_num_seqs=16, |
| 83 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 84 | trust_remote_code=True, |
| 85 | ) |
| 86 | |
| 87 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 88 | messages = [ |
| 89 | { |
| 90 | "role": "user", |
| 91 | "content": [ |
| 92 | *placeholders, |
| 93 | {"type": "text", "text": question}, |
| 94 | ], |
| 95 | } |
| 96 | ] |
| 97 | |
| 98 | processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True) |
| 99 | |
| 100 | prompt = processor.apply_chat_template( |
| 101 | messages, tokenize=False, add_generation_prompt=True |
| 102 | ) |
| 103 | |
| 104 | return ModelRequestData( |
| 105 | engine_args=engine_args, |
| 106 | prompt=prompt, |
| 107 | image_data=[fetch_image(url) for url in image_urls], |
| 108 | ) |
| 109 | |
| 110 | |
| 111 | def load_command_a_vision(question: str, image_urls: list[str]) -> ModelRequestData: |
| 112 | model_name = "CohereLabs/command-a-vision-07-2025" |
| 113 | |
| 114 | # NOTE: This model is 122B parameters and requires tensor parallelism |
| 115 | # Recommended to use tp=4 on H100 GPUs |
| 116 | engine_args = EngineArgs( |
| 117 | model=model_name, |
| 118 | max_model_len=32768, |
| 119 | tensor_parallel_size=4, |
| 120 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 121 | ) |
| 122 | |
| 123 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 124 | messages = [ |
| 125 | { |
| 126 | "role": "user", |
| 127 | "content": [ |
| 128 | *placeholders, |
| 129 | {"type": "text", "text": question}, |
| 130 | ], |
| 131 | } |
| 132 | ] |
| 133 | |
| 134 | processor = AutoProcessor.from_pretrained(model_name) |
| 135 | |
| 136 | prompt = processor.apply_chat_template( |
| 137 | messages, tokenize=False, add_generation_prompt=True |
| 138 | ) |
| 139 | |
| 140 | return ModelRequestData( |
| 141 | engine_args=engine_args, |
| 142 | prompt=prompt, |
| 143 | image_data=[fetch_image(url) for url in image_urls], |
| 144 | ) |
| 145 | |
| 146 | |
| 147 | def load_deepseek_vl2(question: str, image_urls: list[str]) -> ModelRequestData: |
| 148 | model_name = "deepseek-ai/deepseek-vl2-tiny" |
| 149 | |
| 150 | engine_args = EngineArgs( |
| 151 | model=model_name, |
| 152 | max_model_len=4096, |
| 153 | max_num_seqs=2, |
| 154 | hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]}, |
| 155 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 156 | ) |
| 157 | |
| 158 | placeholder = "".join( |
| 159 | f"image_{i}:<image>\n" for i, _ in enumerate(image_urls, start=1) |
| 160 | ) |
| 161 | prompt = f"<|User|>: {placeholder}{question}\n\n<|Assistant|>:" |
| 162 | |
| 163 | return ModelRequestData( |
| 164 | engine_args=engine_args, |
| 165 | prompt=prompt, |
| 166 | image_data=[fetch_image(url) for url in image_urls], |
| 167 | ) |
| 168 | |
| 169 | |
| 170 | def load_deepseek_ocr(question: str, image_urls: list[str]) -> ModelRequestData: |
| 171 | from vllm.model_executor.models.deepseek_ocr import NGramPerReqLogitsProcessor |
| 172 | |
| 173 | model_name = "deepseek-ai/DeepSeek-OCR" |
| 174 | |
| 175 | engine_args = EngineArgs( |
| 176 | model=model_name, |
| 177 | max_num_seqs=2, |
| 178 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 179 | logits_processors=[NGramPerReqLogitsProcessor], |
| 180 | ) |
| 181 | |
| 182 | placeholder = "<image>\n" * len(image_urls) |
| 183 | prompt = placeholder + question |
| 184 | |
| 185 | # The following sampling params config is taken from |
| 186 | # the official Deepseek-OCR inference example. |
| 187 | # (IMPORTANT) Use the custom logits processor and avoid skipping |
| 188 | # special tokens for this model for the optimal OCR performance. |
| 189 | sampling_params = SamplingParams( |
| 190 | temperature=0.0, |
| 191 | max_tokens=8192, |
| 192 | # ngram logit processor args |
| 193 | extra_args=dict( |
| 194 | ngram_size=30, |
| 195 | window_size=90, |
| 196 | # whitelist: <td>, </td> |
| 197 | whitelist_token_ids={128821, 128822}, |
| 198 | ), |
| 199 | skip_special_tokens=False, |
| 200 | ) |
| 201 | |
| 202 | return ModelRequestData( |
| 203 | engine_args=engine_args, |
| 204 | prompt=prompt, |
| 205 | image_data=[fetch_image(url) for url in image_urls], |
| 206 | sampling_params=sampling_params, |
| 207 | ) |
| 208 | |
| 209 | |
| 210 | # exaone4_5 |
| 211 | def load_exaone4_5(question: str, image_urls: list[str]) -> ModelRequestData: |
| 212 | model_name = "LGAI-EXAONE/EXAONE-4.5-33B" |
| 213 | |
| 214 | engine_args = EngineArgs( |
| 215 | model=model_name, |
| 216 | max_model_len=8192, |
| 217 | max_num_seqs=2, |
| 218 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 219 | ) |
| 220 | |
| 221 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 222 | messages = [ |
| 223 | { |
| 224 | "role": "user", |
| 225 | "content": [ |
| 226 | *placeholders, |
| 227 | {"type": "text", "text": question}, |
| 228 | ], |
| 229 | } |
| 230 | ] |
| 231 | |
| 232 | processor = AutoProcessor.from_pretrained(model_name) |
| 233 | |
| 234 | prompt = processor.apply_chat_template( |
| 235 | messages, tokenize=False, add_generation_prompt=True |
| 236 | ) |
| 237 | |
| 238 | return ModelRequestData( |
| 239 | engine_args=engine_args, |
| 240 | prompt=prompt, |
| 241 | image_data=[fetch_image(url) for url in image_urls], |
| 242 | ) |
| 243 | |
| 244 | |
| 245 | def load_gemma3(question: str, image_urls: list[str]) -> ModelRequestData: |
| 246 | model_name = "google/gemma-3-4b-it" |
| 247 | |
| 248 | engine_args = EngineArgs( |
| 249 | model=model_name, |
| 250 | max_model_len=8192, |
| 251 | max_num_seqs=2, |
| 252 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 253 | ) |
| 254 | |
| 255 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 256 | messages = [ |
| 257 | { |
| 258 | "role": "user", |
| 259 | "content": [ |
| 260 | *placeholders, |
| 261 | {"type": "text", "text": question}, |
| 262 | ], |
| 263 | } |
| 264 | ] |
| 265 | |
| 266 | processor = AutoProcessor.from_pretrained(model_name) |
| 267 | |
| 268 | prompt = processor.apply_chat_template( |
| 269 | messages, tokenize=False, add_generation_prompt=True |
| 270 | ) |
| 271 | |
| 272 | return ModelRequestData( |
| 273 | engine_args=engine_args, |
| 274 | prompt=prompt, |
| 275 | image_data=[fetch_image(url) for url in image_urls], |
| 276 | ) |
| 277 | |
| 278 | |
| 279 | def load_granite4_vision(question: str, image_urls: list[str]) -> ModelRequestData: |
| 280 | model_name = "ibm-granite/granite-vision-4.1-4b" |
| 281 | engine_args = EngineArgs( |
| 282 | model=model_name, |
| 283 | max_model_len=4096, |
| 284 | max_num_seqs=16, |
| 285 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 286 | ) |
| 287 | |
| 288 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 289 | messages = [ |
| 290 | { |
| 291 | "role": "user", |
| 292 | "content": [ |
| 293 | *placeholders, |
| 294 | {"type": "text", "text": question}, |
| 295 | ], |
| 296 | } |
| 297 | ] |
| 298 | |
| 299 | processor = AutoProcessor.from_pretrained(model_name) |
| 300 | prompt = processor.apply_chat_template( |
| 301 | messages, tokenize=False, add_generation_prompt=True |
| 302 | ) |
| 303 | |
| 304 | return ModelRequestData( |
| 305 | engine_args=engine_args, |
| 306 | prompt=prompt, |
| 307 | image_data=[fetch_image(url) for url in image_urls], |
| 308 | ) |
| 309 | |
| 310 | |
| 311 | def load_h2ovl(question: str, image_urls: list[str]) -> ModelRequestData: |
| 312 | model_name = "h2oai/h2ovl-mississippi-800m" |
| 313 | |
| 314 | engine_args = EngineArgs( |
| 315 | model=model_name, |
| 316 | trust_remote_code=True, |
| 317 | max_model_len=8192, |
| 318 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 319 | mm_processor_kwargs={"max_dynamic_patch": 4}, |
| 320 | ) |
| 321 | |
| 322 | placeholders = "\n".join( |
| 323 | f"Image-{i}: <image>\n" for i, _ in enumerate(image_urls, start=1) |
| 324 | ) |
| 325 | messages = [{"role": "user", "content": f"{placeholders}\n{question}"}] |
| 326 | |
| 327 | tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True) |
| 328 | prompt = tokenizer.apply_chat_template( |
| 329 | messages, tokenize=False, add_generation_prompt=True |
| 330 | ) |
| 331 | |
| 332 | # Stop tokens for H2OVL-Mississippi |
| 333 | # https://huggingface.co/h2oai/h2ovl-mississippi-800m |
| 334 | stop_token_ids = [tokenizer.eos_token_id] |
| 335 | |
| 336 | return ModelRequestData( |
| 337 | engine_args=engine_args, |
| 338 | prompt=prompt, |
| 339 | stop_token_ids=stop_token_ids, |
| 340 | image_data=[fetch_image(url) for url in image_urls], |
| 341 | ) |
| 342 | |
| 343 | |
| 344 | # HunyuanOCR |
| 345 | def load_hunyuan_vl(question: str, image_urls: list[str]) -> ModelRequestData: |
| 346 | model_name = "tencent/HunyuanOCR" |
| 347 | |
| 348 | engine_args = EngineArgs( |
| 349 | model=model_name, |
| 350 | max_model_len=8192, |
| 351 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 352 | ) |
| 353 | |
| 354 | placeholder = ( |
| 355 | "<｜hy_place▁holder▁no▁100｜><｜hy_place▁holder▁no▁102｜><｜hy_place▁holder▁no▁101｜>" # noqa: E501 |
| 356 | ) * len(image_urls) |
| 357 | prompt = f"<｜hy_begin▁of▁sentence｜>{placeholder}{question}<｜hy_User｜>" |
| 358 | |
| 359 | return ModelRequestData( |
| 360 | engine_args=engine_args, |
| 361 | prompt=prompt, |
| 362 | image_data=[fetch_image(url) for url in image_urls], |
| 363 | ) |
| 364 | |
| 365 | |
| 366 | def load_idefics3(question: str, image_urls: list[str]) -> ModelRequestData: |
| 367 | model_name = "HuggingFaceM4/Idefics3-8B-Llama3" |
| 368 | |
| 369 | # The configuration below has been confirmed to launch on a single L40 GPU. |
| 370 | engine_args = EngineArgs( |
| 371 | model=model_name, |
| 372 | max_model_len=8192, |
| 373 | max_num_seqs=16, |
| 374 | enforce_eager=True, |
| 375 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 376 | # if you are running out of memory, you can reduce the "longest_edge". |
| 377 | # see: https://huggingface.co/HuggingFaceM4/Idefics3-8B-Llama3#model-optimizations |
| 378 | mm_processor_kwargs={ |
| 379 | "size": {"longest_edge": 2 * 364}, |
| 380 | }, |
| 381 | ) |
| 382 | |
| 383 | placeholders = "\n".join( |
| 384 | f"Image-{i}: <image>\n" for i, _ in enumerate(image_urls, start=1) |
| 385 | ) |
| 386 | prompt = f"<|begin_of_text|>User:{placeholders}\n{question}<end_of_utterance>\nAssistant:" # noqa: E501 |
| 387 | return ModelRequestData( |
| 388 | engine_args=engine_args, |
| 389 | prompt=prompt, |
| 390 | image_data=[fetch_image(url) for url in image_urls], |
| 391 | ) |
| 392 | |
| 393 | |
| 394 | def load_interns1(question: str, image_urls: list[str]) -> ModelRequestData: |
| 395 | model_name = "internlm/Intern-S1-mini" |
| 396 | |
| 397 | engine_args = EngineArgs( |
| 398 | model=model_name, |
| 399 | trust_remote_code=True, |
| 400 | max_model_len=4096, |
| 401 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 402 | ) |
| 403 | |
| 404 | placeholders = "\n".join( |
| 405 | f"Image-{i}: <IMG_CONTEXT>\n" for i, _ in enumerate(image_urls, start=1) |
| 406 | ) |
| 407 | messages = [{"role": "user", "content": f"{placeholders}\n{question}"}] |
| 408 | |
| 409 | tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True) |
| 410 | prompt = tokenizer.apply_chat_template( |
| 411 | messages, tokenize=False, add_generation_prompt=True |
| 412 | ) |
| 413 | |
| 414 | return ModelRequestData( |
| 415 | engine_args=engine_args, |
| 416 | prompt=prompt, |
| 417 | image_data=[fetch_image(url) for url in image_urls], |
| 418 | ) |
| 419 | |
| 420 | |
| 421 | def load_internvl(question: str, image_urls: list[str]) -> ModelRequestData: |
| 422 | model_name = "OpenGVLab/InternVL2-2B" |
| 423 | |
| 424 | engine_args = EngineArgs( |
| 425 | model=model_name, |
| 426 | trust_remote_code=True, |
| 427 | max_model_len=4096, |
| 428 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 429 | mm_processor_kwargs={"max_dynamic_patch": 4}, |
| 430 | ) |
| 431 | |
| 432 | placeholders = "\n".join( |
| 433 | f"Image-{i}: <image>\n" for i, _ in enumerate(image_urls, start=1) |
| 434 | ) |
| 435 | messages = [{"role": "user", "content": f"{placeholders}\n{question}"}] |
| 436 | |
| 437 | tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True) |
| 438 | prompt = tokenizer.apply_chat_template( |
| 439 | messages, tokenize=False, add_generation_prompt=True |
| 440 | ) |
| 441 | |
| 442 | # Stop tokens for InternVL |
| 443 | # models variants may have different stop tokens |
| 444 | # please refer to the model card for the correct "stop words": |
| 445 | # https://huggingface.co/OpenGVLab/InternVL2-2B/blob/main/conversation.py |
| 446 | stop_tokens = ["<|endoftext|>", "<|im_start|>", "<|im_end|>", "<|end|>"] |
| 447 | stop_token_ids = [tokenizer.convert_tokens_to_ids(i) for i in stop_tokens] |
| 448 | |
| 449 | return ModelRequestData( |
| 450 | engine_args=engine_args, |
| 451 | prompt=prompt, |
| 452 | stop_token_ids=stop_token_ids, |
| 453 | image_data=[fetch_image(url) for url in image_urls], |
| 454 | ) |
| 455 | |
| 456 | |
| 457 | def load_keye_vl(question: str, image_urls: list[str]) -> ModelRequestData: |
| 458 | model_name = "Kwai-Keye/Keye-VL-8B-Preview" |
| 459 | |
| 460 | engine_args = EngineArgs( |
| 461 | model=model_name, |
| 462 | trust_remote_code=True, |
| 463 | max_model_len=8192, |
| 464 | max_num_seqs=5, |
| 465 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 466 | ) |
| 467 | |
| 468 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 469 | messages = [ |
| 470 | { |
| 471 | "role": "user", |
| 472 | "content": [ |
| 473 | *placeholders, |
| 474 | {"type": "text", "text": question}, |
| 475 | ], |
| 476 | }, |
| 477 | ] |
| 478 | |
| 479 | processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True) |
| 480 | |
| 481 | prompt = processor.apply_chat_template( |
| 482 | messages, tokenize=False, add_generation_prompt=True |
| 483 | ) |
| 484 | |
| 485 | image_data = [fetch_image(url) for url in image_urls] |
| 486 | |
| 487 | return ModelRequestData( |
| 488 | engine_args=engine_args, |
| 489 | prompt=prompt, |
| 490 | image_data=image_data, |
| 491 | ) |
| 492 | |
| 493 | |
| 494 | def load_keye_vl1_5(question: str, image_urls: list[str]) -> ModelRequestData: |
| 495 | model_name = "Kwai-Keye/Keye-VL-1_5-8B" |
| 496 | |
| 497 | engine_args = EngineArgs( |
| 498 | model=model_name, |
| 499 | trust_remote_code=True, |
| 500 | max_model_len=32768, |
| 501 | max_num_seqs=5, |
| 502 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 503 | ) |
| 504 | |
| 505 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 506 | messages = [ |
| 507 | { |
| 508 | "role": "user", |
| 509 | "content": [ |
| 510 | *placeholders, |
| 511 | {"type": "text", "text": question}, |
| 512 | ], |
| 513 | }, |
| 514 | ] |
| 515 | |
| 516 | processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True) |
| 517 | |
| 518 | prompt = processor.apply_chat_template( |
| 519 | messages, tokenize=False, add_generation_prompt=True |
| 520 | ) |
| 521 | |
| 522 | image_data = [fetch_image(url) for url in image_urls] |
| 523 | |
| 524 | return ModelRequestData( |
| 525 | engine_args=engine_args, |
| 526 | prompt=prompt, |
| 527 | image_data=image_data, |
| 528 | ) |
| 529 | |
| 530 | |
| 531 | def load_kimi_vl(question: str, image_urls: list[str]) -> ModelRequestData: |
| 532 | model_name = "moonshotai/Kimi-VL-A3B-Instruct" |
| 533 | |
| 534 | engine_args = EngineArgs( |
| 535 | model=model_name, |
| 536 | trust_remote_code=True, |
| 537 | max_model_len=4096, |
| 538 | max_num_seqs=4, |
| 539 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 540 | ) |
| 541 | |
| 542 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 543 | messages = [ |
| 544 | { |
| 545 | "role": "user", |
| 546 | "content": [ |
| 547 | *placeholders, |
| 548 | {"type": "text", "text": question}, |
| 549 | ], |
| 550 | } |
| 551 | ] |
| 552 | |
| 553 | processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True) |
| 554 | |
| 555 | prompt = processor.apply_chat_template( |
| 556 | messages, tokenize=False, add_generation_prompt=True |
| 557 | ) |
| 558 | |
| 559 | return ModelRequestData( |
| 560 | engine_args=engine_args, |
| 561 | prompt=prompt, |
| 562 | image_data=[fetch_image(url) for url in image_urls], |
| 563 | ) |
| 564 | |
| 565 | |
| 566 | def load_llama4(question: str, image_urls: list[str]) -> ModelRequestData: |
| 567 | model_name = "meta-llama/Llama-4-Scout-17B-16E-Instruct" |
| 568 | |
| 569 | engine_args = EngineArgs( |
| 570 | model=model_name, |
| 571 | max_model_len=131072, |
| 572 | tensor_parallel_size=8, |
| 573 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 574 | ) |
| 575 | |
| 576 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 577 | messages = [ |
| 578 | { |
| 579 | "role": "user", |
| 580 | "content": [ |
| 581 | *placeholders, |
| 582 | {"type": "text", "text": question}, |
| 583 | ], |
| 584 | } |
| 585 | ] |
| 586 | |
| 587 | processor = AutoProcessor.from_pretrained(model_name) |
| 588 | |
| 589 | prompt = processor.apply_chat_template( |
| 590 | messages, tokenize=False, add_generation_prompt=True |
| 591 | ) |
| 592 | |
| 593 | return ModelRequestData( |
| 594 | engine_args=engine_args, |
| 595 | prompt=prompt, |
| 596 | image_data=[fetch_image(url) for url in image_urls], |
| 597 | ) |
| 598 | |
| 599 | |
| 600 | def load_llava(question: str, image_urls: list[str]) -> ModelRequestData: |
| 601 | # NOTE: CAUTION! Original Llava models wasn't really trained on multi-image inputs, |
| 602 | # it will generate poor response for multi-image inputs! |
| 603 | model_name = "llava-hf/llava-1.5-7b-hf" |
| 604 | engine_args = EngineArgs( |
| 605 | model=model_name, |
| 606 | max_num_seqs=16, |
| 607 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 608 | ) |
| 609 | |
| 610 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 611 | messages = [ |
| 612 | { |
| 613 | "role": "user", |
| 614 | "content": [ |
| 615 | *placeholders, |
| 616 | {"type": "text", "text": question}, |
| 617 | ], |
| 618 | } |
| 619 | ] |
| 620 | |
| 621 | processor = AutoProcessor.from_pretrained(model_name) |
| 622 | |
| 623 | prompt = processor.apply_chat_template( |
| 624 | messages, tokenize=False, add_generation_prompt=True |
| 625 | ) |
| 626 | |
| 627 | return ModelRequestData( |
| 628 | engine_args=engine_args, |
| 629 | prompt=prompt, |
| 630 | image_data=[fetch_image(url) for url in image_urls], |
| 631 | ) |
| 632 | |
| 633 | |
| 634 | def load_llava_next(question: str, image_urls: list[str]) -> ModelRequestData: |
| 635 | model_name = "llava-hf/llava-v1.6-mistral-7b-hf" |
| 636 | engine_args = EngineArgs( |
| 637 | model=model_name, |
| 638 | max_model_len=8192, |
| 639 | max_num_seqs=16, |
| 640 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 641 | ) |
| 642 | |
| 643 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 644 | messages = [ |
| 645 | { |
| 646 | "role": "user", |
| 647 | "content": [ |
| 648 | *placeholders, |
| 649 | {"type": "text", "text": question}, |
| 650 | ], |
| 651 | } |
| 652 | ] |
| 653 | |
| 654 | processor = AutoProcessor.from_pretrained(model_name) |
| 655 | |
| 656 | prompt = processor.apply_chat_template( |
| 657 | messages, tokenize=False, add_generation_prompt=True |
| 658 | ) |
| 659 | |
| 660 | return ModelRequestData( |
| 661 | engine_args=engine_args, |
| 662 | prompt=prompt, |
| 663 | image_data=[fetch_image(url) for url in image_urls], |
| 664 | ) |
| 665 | |
| 666 | |
| 667 | def load_llava_onevision(question: str, image_urls: list[str]) -> ModelRequestData: |
| 668 | model_name = "llava-hf/llava-onevision-qwen2-7b-ov-hf" |
| 669 | engine_args = EngineArgs( |
| 670 | model=model_name, |
| 671 | max_model_len=16384, |
| 672 | max_num_seqs=16, |
| 673 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 674 | ) |
| 675 | |
| 676 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 677 | messages = [ |
| 678 | { |
| 679 | "role": "user", |
| 680 | "content": [ |
| 681 | *placeholders, |
| 682 | {"type": "text", "text": question}, |
| 683 | ], |
| 684 | } |
| 685 | ] |
| 686 | |
| 687 | processor = AutoProcessor.from_pretrained(model_name) |
| 688 | |
| 689 | prompt = processor.apply_chat_template( |
| 690 | messages, tokenize=False, add_generation_prompt=True |
| 691 | ) |
| 692 | |
| 693 | return ModelRequestData( |
| 694 | engine_args=engine_args, |
| 695 | prompt=prompt, |
| 696 | image_data=[fetch_image(url) for url in image_urls], |
| 697 | ) |
| 698 | |
| 699 | |
| 700 | def load_mistral3(question: str, image_urls: list[str]) -> ModelRequestData: |
| 701 | model_name = "mistralai/Mistral-Small-3.1-24B-Instruct-2503" |
| 702 | |
| 703 | # Adjust this as necessary to fit in GPU |
| 704 | engine_args = EngineArgs( |
| 705 | model=model_name, |
| 706 | max_model_len=8192, |
| 707 | max_num_seqs=2, |
| 708 | tensor_parallel_size=2, |
| 709 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 710 | ignore_patterns=["consolidated.safetensors"], |
| 711 | ) |
| 712 | |
| 713 | placeholders = "[IMG]" * len(image_urls) |
| 714 | prompt = f"<s>[INST]{question}\n{placeholders}[/INST]" |
| 715 | |
| 716 | return ModelRequestData( |
| 717 | engine_args=engine_args, |
| 718 | prompt=prompt, |
| 719 | image_data=[fetch_image(url) for url in image_urls], |
| 720 | ) |
| 721 | |
| 722 | |
| 723 | def load_nvlm_d(question: str, image_urls: list[str]) -> ModelRequestData: |
| 724 | model_name = "nvidia/NVLM-D-72B" |
| 725 | |
| 726 | # Adjust this as necessary to fit in GPU |
| 727 | engine_args = EngineArgs( |
| 728 | model=model_name, |
| 729 | trust_remote_code=True, |
| 730 | max_model_len=8192, |
| 731 | tensor_parallel_size=4, |
| 732 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 733 | mm_processor_kwargs={"max_dynamic_patch": 4}, |
| 734 | ) |
| 735 | |
| 736 | placeholders = "\n".join( |
| 737 | f"Image-{i}: <image>\n" for i, _ in enumerate(image_urls, start=1) |
| 738 | ) |
| 739 | messages = [{"role": "user", "content": f"{placeholders}\n{question}"}] |
| 740 | |
| 741 | tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True) |
| 742 | prompt = tokenizer.apply_chat_template( |
| 743 | messages, tokenize=False, add_generation_prompt=True |
| 744 | ) |
| 745 | |
| 746 | return ModelRequestData( |
| 747 | engine_args=engine_args, |
| 748 | prompt=prompt, |
| 749 | image_data=[fetch_image(url) for url in image_urls], |
| 750 | ) |
| 751 | |
| 752 | |
| 753 | # OpenPangu |
| 754 | def load_openpangu_vl(question: str, image_urls: list[str]) -> ModelRequestData: |
| 755 | model_name = "FreedomIntelligence/openPangu-VL-7B" |
| 756 | |
| 757 | engine_args = EngineArgs( |
| 758 | model=model_name, |
| 759 | trust_remote_code=True, |
| 760 | max_model_len=8192, |
| 761 | max_num_seqs=2, |
| 762 | enforce_eager=True, |
| 763 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 764 | ) |
| 765 | |
| 766 | placeholders = "[unused18][unused19][unused20]" * len(image_urls) |
| 767 | prompt = ( |
| 768 | f"<s>[unused9]系统：[unused10][unused9]用户：{question}{placeholders}" |
| 769 | "[unused10][unused9]助手：" |
| 770 | ) |
| 771 | |
| 772 | return ModelRequestData( |
| 773 | engine_args=engine_args, |
| 774 | prompt=prompt, |
| 775 | image_data=[fetch_image(url) for url in image_urls], |
| 776 | ) |
| 777 | |
| 778 | |
| 779 | # Ovis |
| 780 | def load_ovis(question: str, image_urls: list[str]) -> ModelRequestData: |
| 781 | model_name = "AIDC-AI/Ovis2-1B" |
| 782 | |
| 783 | engine_args = EngineArgs( |
| 784 | model=model_name, |
| 785 | max_model_len=8192, |
| 786 | max_num_seqs=2, |
| 787 | trust_remote_code=True, |
| 788 | dtype="half", |
| 789 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 790 | ) |
| 791 | |
| 792 | placeholders = "\n".join( |
| 793 | f"Image-{i}: <image>\n" for i, _ in enumerate(image_urls, start=1) |
| 794 | ) |
| 795 | messages = [{"role": "user", "content": f"{placeholders}\n{question}"}] |
| 796 | |
| 797 | tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True) |
| 798 | prompt = tokenizer.apply_chat_template( |
| 799 | messages, tokenize=False, add_generation_prompt=True |
| 800 | ) |
| 801 | |
| 802 | return ModelRequestData( |
| 803 | engine_args=engine_args, |
| 804 | prompt=prompt, |
| 805 | image_data=[fetch_image(url) for url in image_urls], |
| 806 | ) |
| 807 | |
| 808 | |
| 809 | # ovis2_5 |
| 810 | def load_ovis2_5(question: str, image_urls: list[str]) -> ModelRequestData: |
| 811 | model_name = "AIDC-AI/Ovis2.5-2B" |
| 812 | |
| 813 | engine_args = EngineArgs( |
| 814 | model=model_name, |
| 815 | max_model_len=8192, |
| 816 | max_num_seqs=2, |
| 817 | trust_remote_code=True, |
| 818 | dtype="half", |
| 819 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 820 | ) |
| 821 | |
| 822 | placeholders = "\n".join( |
| 823 | f"Image-{i}: <image>\n" for i, _ in enumerate(image_urls, start=1) |
| 824 | ) |
| 825 | prompt = ( |
| 826 | f"<|im_start|>user\n\n{placeholders}\n{question}<|im_end|>\n" |
| 827 | "<|im_start|>assistant\n" |
| 828 | ) |
| 829 | |
| 830 | return ModelRequestData( |
| 831 | engine_args=engine_args, |
| 832 | prompt=prompt, |
| 833 | image_data=[fetch_image(url) for url in image_urls], |
| 834 | ) |
| 835 | |
| 836 | |
| 837 | def load_paddleocr_vl(question: str, image_urls: list[str]) -> ModelRequestData: |
| 838 | model_name = "PaddlePaddle/PaddleOCR-VL" |
| 839 | |
| 840 | engine_args = EngineArgs( |
| 841 | model=model_name, |
| 842 | trust_remote_code=True, |
| 843 | max_model_len=8192, |
| 844 | max_num_seqs=2, |
| 845 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 846 | ) |
| 847 | |
| 848 | placeholders = "<|IMAGE_START|><|IMAGE_PLACEHOLDER|><|IMAGE_END|>" * len(image_urls) |
| 849 | prompt = f"<|begin_of_sentence|>User: {question}{placeholders}\nAssistant: " |
| 850 | |
| 851 | return ModelRequestData( |
| 852 | engine_args=engine_args, |
| 853 | prompt=prompt, |
| 854 | image_data=[fetch_image(url) for url in image_urls], |
| 855 | ) |
| 856 | |
| 857 | |
| 858 | def load_pixtral_hf(question: str, image_urls: list[str]) -> ModelRequestData: |
| 859 | model_name = "mistral-community/pixtral-12b" |
| 860 | |
| 861 | # Adjust this as necessary to fit in GPU |
| 862 | engine_args = EngineArgs( |
| 863 | model=model_name, |
| 864 | max_model_len=8192, |
| 865 | max_num_seqs=2, |
| 866 | tensor_parallel_size=2, |
| 867 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 868 | ) |
| 869 | |
| 870 | placeholders = "[IMG]" * len(image_urls) |
| 871 | prompt = f"<s>[INST]{question}\n{placeholders}[/INST]" |
| 872 | |
| 873 | return ModelRequestData( |
| 874 | engine_args=engine_args, |
| 875 | prompt=prompt, |
| 876 | image_data=[fetch_image(url) for url in image_urls], |
| 877 | ) |
| 878 | |
| 879 | |
| 880 | def load_phi3v(question: str, image_urls: list[str]) -> ModelRequestData: |
| 881 | # num_crops is an override kwarg to the multimodal image processor; |
| 882 | # For some models, e.g., Phi-3.5-vision-instruct, it is recommended |
| 883 | # to use 16 for single frame scenarios, and 4 for multi-frame. |
| 884 | # |
| 885 | # Generally speaking, a larger value for num_crops results in more |
| 886 | # tokens per image instance, because it may scale the image more in |
| 887 | # the image preprocessing. Some references in the model docs and the |
| 888 | # formula for image tokens after the preprocessing |
| 889 | # transform can be found below. |
| 890 | # |
| 891 | # https://huggingface.co/microsoft/Phi-3.5-vision-instruct#loading-the-model-locally |
| 892 | # https://huggingface.co/microsoft/Phi-3.5-vision-instruct/blob/main/processing_phi3_v.py#L194 |
| 893 | engine_args = EngineArgs( |
| 894 | model="microsoft/Phi-3.5-vision-instruct", |
| 895 | trust_remote_code=True, |
| 896 | max_model_len=4096, |
| 897 | max_num_seqs=2, |
| 898 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 899 | mm_processor_kwargs={"num_crops": 4}, |
| 900 | ) |
| 901 | placeholders = "\n".join( |
| 902 | f"<|image_{i}|>" for i, _ in enumerate(image_urls, start=1) |
| 903 | ) |
| 904 | prompt = f"<|user|>\n{placeholders}\n{question}<|end|>\n<|assistant|>\n" |
| 905 | |
| 906 | return ModelRequestData( |
| 907 | engine_args=engine_args, |
| 908 | prompt=prompt, |
| 909 | image_data=[fetch_image(url) for url in image_urls], |
| 910 | ) |
| 911 | |
| 912 | |
| 913 | def load_phi4mm(question: str, image_urls: list[str]) -> ModelRequestData: |
| 914 | """Phi-4-multimodal-instruct supports both image and audio inputs. Here, we |
| 915 | show how to process multi images inputs. |
| 916 | """ |
| 917 | model_path = snapshot_download("microsoft/Phi-4-multimodal-instruct") |
| 918 | # Since the vision-lora and speech-lora co-exist with the base model, |
| 919 | # we have to manually specify the path of the lora weights. |
| 920 | vision_lora_path = os.path.join(model_path, "vision-lora") |
| 921 | engine_args = EngineArgs( |
| 922 | model=model_path, |
| 923 | trust_remote_code=True, |
| 924 | max_model_len=4096, |
| 925 | max_num_seqs=2, |
| 926 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 927 | enable_lora=True, |
| 928 | max_lora_rank=320, |
| 929 | # Note - mm_processor_kwargs can also be passed to generate/chat calls |
| 930 | mm_processor_kwargs={"dynamic_hd": 4}, |
| 931 | ) |
| 932 | |
| 933 | placeholders = "".join(f"<|image_{i}|>" for i, _ in enumerate(image_urls, start=1)) |
| 934 | prompt = f"<|user|>{placeholders}{question}<|end|><|assistant|>" |
| 935 | |
| 936 | return ModelRequestData( |
| 937 | engine_args=engine_args, |
| 938 | prompt=prompt, |
| 939 | image_data=[fetch_image(url) for url in image_urls], |
| 940 | lora_requests=[LoRARequest("vision", 1, vision_lora_path)], |
| 941 | ) |
| 942 | |
| 943 | |
| 944 | def load_phi4siglip(question: str, image_urls: list[str]) -> ModelRequestData: |
| 945 | model_name = "microsoft/Phi-4-reasoning-vision-15B" |
| 946 | placeholders = "\n".join("<image>" for _ in image_urls) |
| 947 | prompt = f"<|user|>\n{placeholders}\n{question}<|end|>\n<|assistant|>\n" |
| 948 | engine_args = EngineArgs( |
| 949 | model=model_name, |
| 950 | trust_remote_code=True, |
| 951 | max_model_len=8192, |
| 952 | max_num_seqs=2, |
| 953 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 954 | ) |
| 955 | return ModelRequestData( |
| 956 | engine_args=engine_args, |
| 957 | prompt=prompt, |
| 958 | image_data=[fetch_image(url) for url in image_urls], |
| 959 | ) |
| 960 | |
| 961 | |
| 962 | def load_qwen2_vl(question: str, image_urls: list[str]) -> ModelRequestData: |
| 963 | try: |
| 964 | from qwen_vl_utils import smart_resize |
| 965 | except ModuleNotFoundError: |
| 966 | print( |
| 967 | "WARNING: `qwen-vl-utils` not installed, input images will not " |
| 968 | "be automatically resized. You can enable this functionality by " |
| 969 | "`pip install qwen-vl-utils`." |
| 970 | ) |
| 971 | smart_resize = None |
| 972 | |
| 973 | model_name = "Qwen/Qwen2-VL-7B-Instruct" |
| 974 | |
| 975 | # Tested on L40 |
| 976 | engine_args = EngineArgs( |
| 977 | model=model_name, |
| 978 | max_model_len=32768 if smart_resize is None else 4096, |
| 979 | max_num_seqs=5, |
| 980 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 981 | ) |
| 982 | |
| 983 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 984 | messages = [ |
| 985 | {"role": "system", "content": "You are a helpful assistant."}, |
| 986 | { |
| 987 | "role": "user", |
| 988 | "content": [ |
| 989 | *placeholders, |
| 990 | {"type": "text", "text": question}, |
| 991 | ], |
| 992 | }, |
| 993 | ] |
| 994 | |
| 995 | processor = AutoProcessor.from_pretrained(model_name) |
| 996 | |
| 997 | prompt = processor.apply_chat_template( |
| 998 | messages, tokenize=False, add_generation_prompt=True |
| 999 | ) |
| 1000 | |
| 1001 | if smart_resize is None: |
| 1002 | image_data = [fetch_image(url) for url in image_urls] |
| 1003 | else: |
| 1004 | |
| 1005 | def post_process_image(image: Image) -> Image: |
| 1006 | width, height = image.size |
| 1007 | resized_height, resized_width = smart_resize( |
| 1008 | height, width, max_pixels=1024 * 28 * 28 |
| 1009 | ) |
| 1010 | return image.resize((resized_width, resized_height)) |
| 1011 | |
| 1012 | image_data = [post_process_image(fetch_image(url)) for url in image_urls] |
| 1013 | |
| 1014 | return ModelRequestData( |
| 1015 | engine_args=engine_args, |
| 1016 | prompt=prompt, |
| 1017 | image_data=image_data, |
| 1018 | ) |
| 1019 | |
| 1020 | |
| 1021 | def load_qwen2_5_vl(question: str, image_urls: list[str]) -> ModelRequestData: |
| 1022 | try: |
| 1023 | from qwen_vl_utils import smart_resize |
| 1024 | except ModuleNotFoundError: |
| 1025 | print( |
| 1026 | "WARNING: `qwen-vl-utils` not installed, input images will not " |
| 1027 | "be automatically resized. You can enable this functionality by " |
| 1028 | "`pip install qwen-vl-utils`." |
| 1029 | ) |
| 1030 | smart_resize = None |
| 1031 | |
| 1032 | model_name = "Qwen/Qwen2.5-VL-3B-Instruct" |
| 1033 | |
| 1034 | engine_args = EngineArgs( |
| 1035 | model=model_name, |
| 1036 | max_model_len=32768 if smart_resize is None else 4096, |
| 1037 | max_num_seqs=5, |
| 1038 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 1039 | ) |
| 1040 | |
| 1041 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 1042 | messages = [ |
| 1043 | {"role": "system", "content": "You are a helpful assistant."}, |
| 1044 | { |
| 1045 | "role": "user", |
| 1046 | "content": [ |
| 1047 | *placeholders, |
| 1048 | {"type": "text", "text": question}, |
| 1049 | ], |
| 1050 | }, |
| 1051 | ] |
| 1052 | |
| 1053 | processor = AutoProcessor.from_pretrained(model_name) |
| 1054 | |
| 1055 | prompt = processor.apply_chat_template( |
| 1056 | messages, tokenize=False, add_generation_prompt=True |
| 1057 | ) |
| 1058 | |
| 1059 | if smart_resize is None: |
| 1060 | image_data = [fetch_image(url) for url in image_urls] |
| 1061 | else: |
| 1062 | |
| 1063 | def post_process_image(image: Image) -> Image: |
| 1064 | width, height = image.size |
| 1065 | resized_height, resized_width = smart_resize( |
| 1066 | height, width, max_pixels=1024 * 28 * 28 |
| 1067 | ) |
| 1068 | return image.resize((resized_width, resized_height)) |
| 1069 | |
| 1070 | image_data = [post_process_image(fetch_image(url)) for url in image_urls] |
| 1071 | |
| 1072 | return ModelRequestData( |
| 1073 | engine_args=engine_args, |
| 1074 | prompt=prompt, |
| 1075 | image_data=image_data, |
| 1076 | ) |
| 1077 | |
| 1078 | |
| 1079 | def load_r_vl(question: str, image_urls: list[str]) -> ModelRequestData: |
| 1080 | model_name = "YannQi/R-4B" |
| 1081 | engine_args = EngineArgs( |
| 1082 | model=model_name, |
| 1083 | max_model_len=16384, |
| 1084 | max_num_seqs=16, |
| 1085 | trust_remote_code=True, |
| 1086 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 1087 | ) |
| 1088 | |
| 1089 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 1090 | messages = [ |
| 1091 | { |
| 1092 | "role": "user", |
| 1093 | "content": [ |
| 1094 | *placeholders, |
| 1095 | {"type": "text", "text": question}, |
| 1096 | ], |
| 1097 | } |
| 1098 | ] |
| 1099 | |
| 1100 | processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True) |
| 1101 | |
| 1102 | prompt = processor.apply_chat_template( |
| 1103 | messages, tokenize=False, add_generation_prompt=True |
| 1104 | ) |
| 1105 | |
| 1106 | return ModelRequestData( |
| 1107 | engine_args=engine_args, |
| 1108 | prompt=prompt, |
| 1109 | image_data=[fetch_image(url) for url in image_urls], |
| 1110 | ) |
| 1111 | |
| 1112 | |
| 1113 | def load_smolvlm(question: str, image_urls: list[str]) -> ModelRequestData: |
| 1114 | model_name = "HuggingFaceTB/SmolVLM2-2.2B-Instruct" |
| 1115 | |
| 1116 | # The configuration below has been confirmed to launch on a single L40 GPU. |
| 1117 | engine_args = EngineArgs( |
| 1118 | model=model_name, |
| 1119 | max_model_len=8192, |
| 1120 | max_num_seqs=16, |
| 1121 | enforce_eager=True, |
| 1122 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 1123 | mm_processor_kwargs={ |
| 1124 | "max_image_size": {"longest_edge": 384}, |
| 1125 | }, |
| 1126 | ) |
| 1127 | |
| 1128 | placeholders = "\n".join( |
| 1129 | f"Image-{i}: <image>\n" for i, _ in enumerate(image_urls, start=1) |
| 1130 | ) |
| 1131 | prompt = ( |
| 1132 | f"<|im_start|>User:{placeholders}\n{question}<end_of_utterance>\nAssistant:" # noqa: E501 |
| 1133 | ) |
| 1134 | return ModelRequestData( |
| 1135 | engine_args=engine_args, |
| 1136 | prompt=prompt, |
| 1137 | image_data=[fetch_image(url) for url in image_urls], |
| 1138 | ) |
| 1139 | |
| 1140 | |
| 1141 | def load_step3(question: str, image_urls: list[str]) -> ModelRequestData: |
| 1142 | model_name = "stepfun-ai/step3-fp8" |
| 1143 | |
| 1144 | # NOTE: Below are verified configurations for step3-fp8 |
| 1145 | # on 8xH100 GPUs. |
| 1146 | engine_args = EngineArgs( |
| 1147 | model=model_name, |
| 1148 | max_num_batched_tokens=4096, |
| 1149 | gpu_memory_utilization=0.85, |
| 1150 | tensor_parallel_size=8, |
| 1151 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 1152 | reasoning_parser="step3", |
| 1153 | ) |
| 1154 | |
| 1155 | prompt = ( |
| 1156 | "<｜begin▁of▁sentence｜> You are a helpful assistant. <|BOT|>user\n " |
| 1157 | f"{'<im_patch>' * len(image_urls)}{question} <|EOT|><|BOT|" |
| 1158 | ">assistant\n<think>\n" |
| 1159 | ) |
| 1160 | image_data = [fetch_image(url) for url in image_urls] |
| 1161 | |
| 1162 | return ModelRequestData( |
| 1163 | engine_args=engine_args, |
| 1164 | prompt=prompt, |
| 1165 | image_data=image_data, |
| 1166 | ) |
| 1167 | |
| 1168 | |
| 1169 | def load_step_vl(question: str, image_urls: list[str]) -> ModelRequestData: |
| 1170 | model_name = "stepfun-ai/Step3-VL-10B" |
| 1171 | |
| 1172 | engine_args = EngineArgs( |
| 1173 | model=model_name, |
| 1174 | max_num_batched_tokens=4096, |
| 1175 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 1176 | hf_overrides={"vision_config": {"enable_patch": False}}, |
| 1177 | trust_remote_code=True, |
| 1178 | reasoning_parser="deepseek_r1", |
| 1179 | ) |
| 1180 | |
| 1181 | prompt = ( |
| 1182 | "<｜begin▁of▁sentence｜> You are a helpful assistant.<|BOT|>user\n " |
| 1183 | f"{'<im_patch>' * len(image_urls)}{question}<|EOT|><|BOT|>" |
| 1184 | "assistant\n<think>\n" |
| 1185 | ) |
| 1186 | image_data = [fetch_image(url) for url in image_urls] |
| 1187 | |
| 1188 | return ModelRequestData( |
| 1189 | engine_args=engine_args, |
| 1190 | prompt=prompt, |
| 1191 | image_data=image_data, |
| 1192 | ) |
| 1193 | |
| 1194 | |
| 1195 | # GLM-4.1V |
| 1196 | def load_glm4_1v(question: str, image_urls: list[str]) -> ModelRequestData: |
| 1197 | model_name = "zai-org/GLM-4.1V-9B-Thinking" |
| 1198 | |
| 1199 | engine_args = EngineArgs( |
| 1200 | model=model_name, |
| 1201 | max_model_len=45082, |
| 1202 | max_num_seqs=2, |
| 1203 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 1204 | enforce_eager=True, |
| 1205 | ) |
| 1206 | |
| 1207 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 1208 | messages = [ |
| 1209 | { |
| 1210 | "role": "user", |
| 1211 | "content": [ |
| 1212 | *placeholders, |
| 1213 | {"type": "text", "text": question}, |
| 1214 | ], |
| 1215 | } |
| 1216 | ] |
| 1217 | |
| 1218 | processor = AutoProcessor.from_pretrained(model_name) |
| 1219 | prompt = processor.apply_chat_template( |
| 1220 | messages, tokenize=False, add_generation_prompt=True |
| 1221 | ) |
| 1222 | image_data = [fetch_image(url) for url in image_urls] |
| 1223 | |
| 1224 | return ModelRequestData( |
| 1225 | engine_args=engine_args, |
| 1226 | prompt=prompt, |
| 1227 | image_data=image_data, |
| 1228 | ) |
| 1229 | |
| 1230 | |
| 1231 | # GLM-4.5V |
| 1232 | def load_glm4_5v(question: str, image_urls: list[str]) -> ModelRequestData: |
| 1233 | model_name = "zai-org/GLM-4.5V" |
| 1234 | |
| 1235 | engine_args = EngineArgs( |
| 1236 | model=model_name, |
| 1237 | max_model_len=32768, |
| 1238 | max_num_seqs=2, |
| 1239 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 1240 | enforce_eager=True, |
| 1241 | tensor_parallel_size=4, |
| 1242 | ) |
| 1243 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 1244 | messages = [ |
| 1245 | { |
| 1246 | "role": "user", |
| 1247 | "content": [ |
| 1248 | *placeholders, |
| 1249 | {"type": "text", "text": question}, |
| 1250 | ], |
| 1251 | } |
| 1252 | ] |
| 1253 | processor = AutoProcessor.from_pretrained(model_name) |
| 1254 | prompt = processor.apply_chat_template( |
| 1255 | messages, tokenize=False, add_generation_prompt=True |
| 1256 | ) |
| 1257 | image_data = [fetch_image(url) for url in image_urls] |
| 1258 | |
| 1259 | return ModelRequestData( |
| 1260 | engine_args=engine_args, |
| 1261 | prompt=prompt, |
| 1262 | image_data=image_data, |
| 1263 | ) |
| 1264 | |
| 1265 | |
| 1266 | # GLM-4.5V-FP8 |
| 1267 | def load_glm4_5v_fp8(question: str, image_urls: list[str]) -> ModelRequestData: |
| 1268 | model_name = "zai-org/GLM-4.5V-FP8" |
| 1269 | |
| 1270 | engine_args = EngineArgs( |
| 1271 | model=model_name, |
| 1272 | max_model_len=32768, |
| 1273 | max_num_seqs=2, |
| 1274 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 1275 | enforce_eager=True, |
| 1276 | tensor_parallel_size=4, |
| 1277 | ) |
| 1278 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 1279 | messages = [ |
| 1280 | { |
| 1281 | "role": "user", |
| 1282 | "content": [ |
| 1283 | *placeholders, |
| 1284 | {"type": "text", "text": question}, |
| 1285 | ], |
| 1286 | } |
| 1287 | ] |
| 1288 | processor = AutoProcessor.from_pretrained(model_name) |
| 1289 | prompt = processor.apply_chat_template( |
| 1290 | messages, tokenize=False, add_generation_prompt=True |
| 1291 | ) |
| 1292 | image_data = [fetch_image(url) for url in image_urls] |
| 1293 | |
| 1294 | return ModelRequestData( |
| 1295 | engine_args=engine_args, |
| 1296 | prompt=prompt, |
| 1297 | image_data=image_data, |
| 1298 | ) |
| 1299 | |
| 1300 | |
| 1301 | def load_molmo2(question: str, image_urls: list[str]) -> ModelRequestData: |
| 1302 | model_name = "allenai/Molmo2-8B" |
| 1303 | |
| 1304 | engine_args = EngineArgs( |
| 1305 | model=model_name, |
| 1306 | trust_remote_code=True, |
| 1307 | dtype="bfloat16", |
| 1308 | limit_mm_per_prompt={"image": len(image_urls)}, |
| 1309 | max_num_batched_tokens=36864, |
| 1310 | ) |
| 1311 | |
| 1312 | placeholders = [{"type": "image", "image": url} for url in image_urls] |
| 1313 | messages = [ |
| 1314 | { |
| 1315 | "role": "user", |
| 1316 | "content": [ |
| 1317 | *placeholders, |
| 1318 | {"type": "text", "text": question}, |
| 1319 | ], |
| 1320 | }, |
| 1321 | ] |
| 1322 | |
| 1323 | processor = AutoProcessor.from_pretrained(model_name) |
| 1324 | |
| 1325 | prompt = processor.apply_chat_template( |
| 1326 | messages, tokenize=False, add_generation_prompt=True |
| 1327 | ) |
| 1328 | |
| 1329 | image_data = [fetch_image(url) for url in image_urls] |
| 1330 | |
| 1331 | return ModelRequestData( |
| 1332 | engine_args=engine_args, |
| 1333 | prompt=prompt, |
| 1334 | image_data=image_data, |
| 1335 | ) |
| 1336 | |
| 1337 | |
| 1338 | model_example_map = { |
| 1339 | "aria": load_aria, |
| 1340 | "bee": load_bee, |
| 1341 | "command_a_vision": load_command_a_vision, |
| 1342 | "deepseek_vl_v2": load_deepseek_vl2, |
| 1343 | "deepseek_ocr": load_deepseek_ocr, |
| 1344 | "exaone4_5": load_exaone4_5, |
| 1345 | "gemma3": load_gemma3, |
| 1346 | "granite4_vision": load_granite4_vision, |
| 1347 | "h2ovl_chat": load_h2ovl, |
| 1348 | "hunyuan_vl": load_hunyuan_vl, |
| 1349 | "idefics3": load_idefics3, |
| 1350 | "interns1": load_interns1, |
| 1351 | "internvl_chat": load_internvl, |
| 1352 | "keye_vl": load_keye_vl, |
| 1353 | "keye_vl1_5": load_keye_vl1_5, |
| 1354 | "kimi_vl": load_kimi_vl, |
| 1355 | "llama4": load_llama4, |
| 1356 | "llava": load_llava, |
| 1357 | "llava-next": load_llava_next, |
| 1358 | "llava-onevision": load_llava_onevision, |
| 1359 | "mistral3": load_mistral3, |
| 1360 | "molmo2": load_molmo2, |
| 1361 | "NVLM_D": load_nvlm_d, |
| 1362 | "openpangu_vl": load_openpangu_vl, |
| 1363 | "ovis": load_ovis, |
| 1364 | "ovis2_5": load_ovis2_5, |
| 1365 | "paddleocr_vl": load_paddleocr_vl, |
| 1366 | "phi3_v": load_phi3v, |
| 1367 | "phi4_mm": load_phi4mm, |
| 1368 | "phi4_siglip": load_phi4siglip, |
| 1369 | "pixtral_hf": load_pixtral_hf, |
| 1370 | "qwen2_vl": load_qwen2_vl, |
| 1371 | "qwen2_5_vl": load_qwen2_5_vl, |
| 1372 | "rvl": load_r_vl, |
| 1373 | "smolvlm": load_smolvlm, |
| 1374 | "step3": load_step3, |
| 1375 | "stepvl": load_step_vl, |
| 1376 | "glm4_1v": load_glm4_1v, |
| 1377 | "glm4_5v": load_glm4_5v, |
| 1378 | "glm4_5v_fp8": load_glm4_5v_fp8, |
| 1379 | } |
| 1380 | |
| 1381 | |
| 1382 | def run_generate( |
| 1383 | model, |
| 1384 | question: str, |
| 1385 | image_urls: list[str], |
| 1386 | seed: int, |
| 1387 | tensor_parallel_size: int | None, |
| 1388 | ): |
| 1389 | req_data = model_example_map[model](question, image_urls) |
| 1390 | |
| 1391 | engine_args = req_data.engine_args |
| 1392 | engine_args.seed = seed |
| 1393 | if tensor_parallel_size is not None: |
| 1394 | engine_args.tensor_parallel_size = tensor_parallel_size |
| 1395 | if current_platform.is_rocm(): |
| 1396 | os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" |
| 1397 | llm = LLM.from_engine_args(engine_args) |
| 1398 | |
| 1399 | sampling_params = SamplingParams( |
| 1400 | temperature=0.0, max_tokens=256, stop_token_ids=req_data.stop_token_ids |
| 1401 | ) |
| 1402 | |
| 1403 | outputs = llm.generate( |
| 1404 | { |
| 1405 | "prompt": req_data.prompt, |
| 1406 | "multi_modal_data": {"image": req_data.image_data}, |
| 1407 | }, |
| 1408 | sampling_params=sampling_params, |
| 1409 | lora_request=req_data.lora_requests, |
| 1410 | ) |
| 1411 | |
| 1412 | print("-" * 50) |
| 1413 | for o in outputs: |
| 1414 | generated_text = o.outputs[0].text |
| 1415 | print(generated_text) |
| 1416 | print("-" * 50) |
| 1417 | |
| 1418 | |
| 1419 | def run_chat( |
| 1420 | model: str, |
| 1421 | question: str, |
| 1422 | image_urls: list[str], |
| 1423 | seed: int, |
| 1424 | tensor_parallel_size: int | None, |
| 1425 | ): |
| 1426 | req_data = model_example_map[model](question, image_urls) |
| 1427 | |
| 1428 | # Disable other modalities to save memory |
| 1429 | default_limits = {"image": 0, "video": 0, "audio": 0} |
| 1430 | req_data.engine_args.limit_mm_per_prompt = default_limits | dict( |
| 1431 | req_data.engine_args.limit_mm_per_prompt or {} |
| 1432 | ) |
| 1433 | |
| 1434 | engine_args = req_data.engine_args |
| 1435 | engine_args.seed = seed |
| 1436 | if tensor_parallel_size is not None: |
| 1437 | engine_args.tensor_parallel_size = tensor_parallel_size |
| 1438 | if current_platform.is_rocm(): |
| 1439 | os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" |
| 1440 | llm = LLM.from_engine_args(engine_args) |
| 1441 | |
| 1442 | sampling_params = ( |
| 1443 | SamplingParams( |
| 1444 | temperature=0.0, max_tokens=256, stop_token_ids=req_data.stop_token_ids |
| 1445 | ) |
| 1446 | if req_data.sampling_params is None |
| 1447 | else req_data.sampling_params |
| 1448 | ) |
| 1449 | outputs = llm.chat( |
| 1450 | [ |
| 1451 | { |
| 1452 | "role": "user", |
| 1453 | "content": [ |
| 1454 | { |
| 1455 | "type": "text", |
| 1456 | "text": question, |
| 1457 | }, |
| 1458 | *( |
| 1459 | { |
| 1460 | "type": "image_url", |
| 1461 | "image_url": {"url": image_url}, |
| 1462 | } |
| 1463 | for image_url in image_urls |
| 1464 | ), |
| 1465 | ], |
| 1466 | } |
| 1467 | ], |
| 1468 | sampling_params=sampling_params, |
| 1469 | chat_template=req_data.chat_template, |
| 1470 | lora_request=req_data.lora_requests, |
| 1471 | ) |
| 1472 | |
| 1473 | print("-" * 50) |
| 1474 | for o in outputs: |
| 1475 | generated_text = o.outputs[0].text |
| 1476 | print(generated_text) |
| 1477 | print("-" * 50) |
| 1478 | |
| 1479 | |
| 1480 | def parse_args(): |
| 1481 | parser = FlexibleArgumentParser( |
| 1482 | description="Demo on using vLLM for offline inference with " |
| 1483 | "vision language models that support multi-image input for text " |
| 1484 | "generation" |
| 1485 | ) |
| 1486 | parser.add_argument( |
| 1487 | "--model-type", |
| 1488 | "-m", |
| 1489 | type=str, |
| 1490 | default="phi3_v", |
| 1491 | choices=model_example_map.keys(), |
| 1492 | help='Huggingface "model_type".', |
| 1493 | ) |
| 1494 | parser.add_argument( |
| 1495 | "--method", |
| 1496 | type=str, |
| 1497 | default="generate", |
| 1498 | choices=["generate", "chat"], |
| 1499 | help="The method to run in `vllm.LLM`.", |
| 1500 | ) |
| 1501 | parser.add_argument( |
| 1502 | "--seed", |
| 1503 | type=int, |
| 1504 | default=0, |
| 1505 | help="Set the seed when initializing `vllm.LLM`.", |
| 1506 | ) |
| 1507 | parser.add_argument( |
| 1508 | "--num-images", |
| 1509 | "-n", |
| 1510 | type=int, |
| 1511 | choices=list(range(1, len(IMAGE_URLS) + 1)), # the max number of images |
| 1512 | default=2, |
| 1513 | help="Number of images to use for the demo.", |
| 1514 | ) |
| 1515 | parser.add_argument( |
| 1516 | "--tensor-parallel-size", |
| 1517 | "-tp", |
| 1518 | type=int, |
| 1519 | default=None, |
| 1520 | help="Tensor parallel size to override the model's default setting. ", |
| 1521 | ) |
| 1522 | return parser.parse_args() |
| 1523 | |
| 1524 | |
| 1525 | def main(args: Namespace): |
| 1526 | model = args.model_type |
| 1527 | method = args.method |
| 1528 | seed = args.seed |
| 1529 | tensor_parallel_size = args.tensor_parallel_size |
| 1530 | |
| 1531 | if tensor_parallel_size is not None and tensor_parallel_size < 1: |
| 1532 | raise ValueError( |
| 1533 | f"tensor_parallel_size must be a positive integer, " |
| 1534 | f"got {tensor_parallel_size}" |
| 1535 | ) |
| 1536 | |
| 1537 | image_urls = IMAGE_URLS[: args.num_images] |
| 1538 | |
| 1539 | if method == "generate": |
| 1540 | run_generate(model, QUESTION, image_urls, seed, tensor_parallel_size) |
| 1541 | elif method == "chat": |
| 1542 | run_chat(model, QUESTION, image_urls, seed, tensor_parallel_size) |
| 1543 | else: |
| 1544 | raise ValueError(f"Invalid method: {method}") |
| 1545 | |
| 1546 | |
| 1547 | if __name__ == "__main__": |
| 1548 | args = parse_args() |
| 1549 | main(args) |