File size: 36,201 Bytes
7d24555 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 | """Processor for Yasa2 that unifies text + media preprocessing."""
from __future__ import annotations
import urllib.request
from enum import Enum
from typing import Any, Dict, List, Literal, Optional, Tuple, Union
import numpy as np
import torch
from PIL import Image
from transformers import AutoTokenizer, ProcessorMixin
from transformers.processing_utils import MultiModalData
from .image_processing_yasa2 import (
Yasa2ImageProcessor,
estimate_num_tiles_llava_next,
estimate_num_tiles_llava_uhd,
image_rgb_decoder_pil,
image_rgb_decoder_pil_tiling,
process_anyres_image,
process_anyres_image_uhd,
)
from .video_processing_yasa2 import (
Yasa2VideoProcessor,
video_rgb_decoder_factory,
)
class MediaType(str, Enum):
IMAGE = "image"
VIDEO = "video"
REKA_IMG_TOKEN = "<REKA_IMG_TOKEN>"
IMAGE_START = "<image>"
IMAGE_END = "</image>"
VIDEO_START = "<video>"
VIDEO_END = "</video>"
SEP_TOKEN = "<sep>"
PAD_ID = 100257 # <|endoftext|>
def _read_bytes_from_uri(uri: str) -> bytes:
"""Read bytes from a local path or HTTP(S) URL.
Args:
uri: Local file path or HTTP(S) URL.
Returns:
Raw bytes content.
"""
if uri.startswith("http://") or uri.startswith("https://"):
with urllib.request.urlopen(uri) as response:
return response.read()
with open(uri, "rb") as f:
return f.read()
def _decode_image_payload(
payload: Union[str, bytes],
img_tiling: bool,
tiling_method: str,
tiling_size: int,
grid_pinpoints: List[Tuple[int, int]],
max_tiles_num: int,
patch_size: int,
) -> Dict[str, Any]:
"""Decode image payload bytes or path into a normalized pixel dict.
Args:
payload: Image path/URL or raw bytes.
img_tiling: Whether to enable tiling.
tiling_method: Tiling method identifier.
tiling_size: Base tile size.
grid_pinpoints: Candidate grid pinpoints.
max_tiles_num: Maximum tile count for UHD tiling.
patch_size: Patch size for UHD tiling.
Returns:
Dict with decoded image data and tiling metadata.
"""
if isinstance(payload, str):
payload = _read_bytes_from_uri(payload)
if img_tiling:
return image_rgb_decoder_pil_tiling(
payload,
size=tiling_size,
grid_pinpoints=grid_pinpoints,
max_tiles_num=max_tiles_num,
patch_size=patch_size,
tiling_method=tiling_method,
)
return image_rgb_decoder_pil(payload)
def _decode_video_payload(
payload: Union[str, bytes],
num_frames: int,
sampling: str,
) -> Dict[str, Any]:
"""Decode video payload bytes or path into sampled frames.
Args:
payload: Video path/URL or raw bytes.
num_frames: Number of frames to sample.
sampling: Sampling strategy.
Returns:
Dict with sampled frames and metadata.
"""
if isinstance(payload, str):
payload = _read_bytes_from_uri(payload)
decoder = video_rgb_decoder_factory(
num_frames=num_frames, sampling=sampling
)
return decoder(payload)
class Yasa2Processor(ProcessorMixin):
"""Processor that applies the Yasa2 dialog formatting and media decoding."""
attributes = ["tokenizer", "image_processor", "video_processor"]
tokenizer_class = "AutoTokenizer"
image_processor_class = "AutoImageProcessor"
video_processor_class = "AutoVideoProcessor"
def __init__(
self,
tokenizer: AutoTokenizer | None = None,
image_processor: Yasa2ImageProcessor | None = None,
video_processor: Yasa2VideoProcessor | None = None,
num_img_tokens: int = 64,
image_token_id: int = 100278,
num_video_frames: int = 6,
video_sampling: str = "chunk",
max_tokens: int = 8192,
**kwargs,
) -> None:
"""Initialize the processor with tokenizer and media processors.
Args:
tokenizer: Tokenizer for text encoding.
image_processor: Image processor for ConvNeXt inputs.
video_processor: Video processor for sampled frames.
num_img_tokens: Number of image content tokens per image.
image_token_id: Token ID for image content tokens.
num_video_frames: Number of frames to sample per video.
video_sampling: Video sampling strategy.
max_tokens: Maximum text token budget.
**kwargs: Passed to ProcessorMixin.
"""
if image_processor is None:
image_processor = Yasa2ImageProcessor()
if video_processor is None:
video_processor = Yasa2VideoProcessor(
num_frames=num_video_frames,
frame_sample_mode=video_sampling,
max_num_frames=num_video_frames,
)
super().__init__(
tokenizer=tokenizer,
image_processor=image_processor,
video_processor=video_processor,
**kwargs,
)
self.num_img_tokens = num_img_tokens
self.num_video_frames = num_video_frames
self.video_sampling = video_sampling
self.max_tokens = max_tokens
self.image_token_id = image_token_id
def _build_prompt_and_media(
self,
messages: List[Dict[str, Any]],
num_img_tokens: int,
num_video_frames: int,
video_sampling: str,
img_tiling: bool,
tiling_method: str,
tiling_size: int,
grid_pinpoints: List[Tuple[int, int]],
max_tiles_num: int,
patch_size: int,
add_generation_prompt: bool,
tools: Optional[List[Dict[str, Any]]] = None,
enable_thinking: Optional[bool] = None,
) -> Tuple[str, List[Tuple[MediaType, Dict[str, Any]]]]:
"""Build Yasa2 prompt text and decode media payloads in prompt order.
Prompt formatting is delegated to the tokenizer's shared chat template.
Args:
messages: Conversation messages in HF format.
num_img_tokens: Content tokens per image.
num_video_frames: Frames to sample per video.
video_sampling: Sampling strategy for videos.
img_tiling: Whether to enable tiling.
tiling_method: Tiling method identifier.
tiling_size: Base tile size.
grid_pinpoints: Candidate grid pinpoints.
max_tiles_num: Maximum tile count for UHD tiling.
patch_size: Patch size for UHD tiling.
add_generation_prompt: Whether to append an assistant prefix.
tools: Optional tool schema list for system prompt injection.
enable_thinking: Unused compatibility flag.
Returns:
Tuple of prompt string and list of decoded media items.
"""
media_items: List[Tuple[MediaType, Dict[str, Any]]] = []
def image_builder(item: Dict[str, Any]) -> List[str]:
"""Serialize an image placeholder sequence for the chat prompt.
Args:
item: Raw message dict with image metadata.
Returns:
List[str]: Tokens that represent the image placeholder.
"""
payload = item.get("image") or item.get("image_url")
if payload is None:
raise ValueError("Image content requires an 'image' field.")
image_datum = _decode_image_payload(
payload,
img_tiling=img_tiling,
tiling_method=tiling_method,
tiling_size=tiling_size,
grid_pinpoints=grid_pinpoints,
max_tiles_num=max_tiles_num,
patch_size=patch_size,
)
num_tiles = image_datum.get("num_tiles", 1)
repeat_tokens = num_img_tokens * num_tiles
media_items.append((MediaType.IMAGE, image_datum))
return (
[IMAGE_START] + [REKA_IMG_TOKEN] * repeat_tokens + [IMAGE_END]
)
def video_builder(item: Dict[str, Any]) -> List[str]:
"""Serialize a video placeholder sequence for the chat prompt.
Args:
item: Raw message dict with video metadata.
Returns:
List[str]: Tokens that represent the video placeholder.
"""
payload = item.get("video") or item.get("video_url")
if payload is None:
raise ValueError("Video content requires a 'video' field.")
video_datum = _decode_video_payload(
payload,
num_frames=num_video_frames,
sampling=video_sampling,
)
repeat_tokens = num_img_tokens * video_datum.get(
"num_frames", num_video_frames
)
media_items.append((MediaType.VIDEO, video_datum))
return (
[VIDEO_START] + [REKA_IMG_TOKEN] * repeat_tokens + [VIDEO_END]
)
if self.tokenizer is None:
raise ValueError(
"Yasa2Processor requires a tokenizer to build prompts."
)
prompt = self.tokenizer.build_chat_prompt(
messages,
add_generation_prompt=add_generation_prompt,
continue_final_message=False,
tools=tools,
image_token_builder=image_builder,
video_token_builder=video_builder,
enable_thinking=enable_thinking,
)
return prompt, media_items
def apply_chat_template(
self,
messages: List[Dict[str, Any]],
tokenize: bool = False,
add_generation_prompt: bool = True,
tools: Optional[List[Dict[str, Any]]] = None,
return_tensors: Optional[str] = None,
return_dict: bool = False,
max_length: Optional[int] = None,
padding: Union[bool, Literal["longest", "max_length"]] = False,
num_img_tokens: Optional[int] = None,
num_video_frames: Optional[int] = None,
video_sampling: Optional[str] = None,
enable_thinking: Optional[bool] = None,
img_tiling: bool = True,
tiling_method: str = "llava-uhd",
tiling_size: int = 512,
grid_pinpoints: Optional[List[Tuple[int, int]]] = None,
max_tiles_num: int = 4,
patch_size: int = 14,
return_prompt: bool = False,
**kwargs,
) -> Union[str, Dict[str, Any]]:
"""Apply the Yasa2 dialog template and optionally tokenize + decode media.
The chat template is produced via the tokenizer for consistency with
text-only formatting.
Args:
messages: Conversation messages in HF format.
tokenize: Whether to tokenize and return tensors.
add_generation_prompt: Whether to append an assistant prefix.
tools: Optional tool schema list for system prompt injection.
return_tensors: Tensor type for outputs (e.g., "pt").
return_dict: Whether to return a dict payload.
max_length: Optional max token length.
padding: Padding strategy (False/True/"longest"/"max_length").
num_img_tokens: Override for image content tokens.
num_video_frames: Override for video frame count.
video_sampling: Override for video sampling strategy.
enable_thinking: Unused compatibility flag.
img_tiling: Whether to enable tiling for images.
tiling_method: Tiling method identifier.
tiling_size: Base tile size.
grid_pinpoints: Candidate grid pinpoints.
max_tiles_num: Maximum tile count for UHD tiling.
patch_size: Patch size for UHD tiling.
return_prompt: Whether to include the prompt string in output.
**kwargs: Unused extra arguments for compatibility.
Returns:
Prompt string if tokenize is False, otherwise a dict of tensors.
"""
if grid_pinpoints is None:
grid_pinpoints = [
(2, 2),
(1, 2),
(2, 1),
(1, 3),
(3, 1),
(1, 4),
(4, 1),
]
num_img_tokens = num_img_tokens or self.num_img_tokens
num_video_frames = num_video_frames or self.num_video_frames
video_sampling = video_sampling or self.video_sampling
user_max_length = max_length
max_tokens = user_max_length or self.max_tokens
prompt, media_items = self._build_prompt_and_media(
messages=messages,
num_img_tokens=num_img_tokens,
num_video_frames=num_video_frames,
video_sampling=video_sampling,
img_tiling=img_tiling,
tiling_method=tiling_method,
tiling_size=tiling_size,
grid_pinpoints=grid_pinpoints,
max_tiles_num=max_tiles_num,
patch_size=patch_size,
add_generation_prompt=add_generation_prompt,
tools=tools,
enable_thinking=enable_thinking,
)
if not tokenize:
return prompt
expected_img_tokens = 0
for media_type, media_datum in media_items:
if media_type == MediaType.IMAGE:
expected_img_tokens += num_img_tokens * media_datum.get(
"num_tiles", 1
)
elif media_type == MediaType.VIDEO:
expected_img_tokens += num_img_tokens * media_datum.get(
"num_frames", num_video_frames
)
input_ids = self.tokenizer.tiktoken.encode(
prompt, allowed_special="all"
)
input_ids = input_ids[:max_tokens]
if expected_img_tokens:
actual_img_tokens = sum(
1 for token_id in input_ids if token_id == self.image_token_id
)
# Ensure truncation did not drop any media placeholder tokens.
if actual_img_tokens != expected_img_tokens:
raise ValueError(
"Prompt truncation dropped image placeholder tokens. "
"Increase max_length/max_tokens or reduce media inputs."
)
attention_mask = [1] * len(input_ids)
token_type_ids, mm_token_type_ids = self._build_mm_token_type_ids(
input_ids
)
if padding not in (False, True, "longest", "max_length"):
raise ValueError(f"Unsupported padding value: {padding}")
if padding in (True, "longest", "max_length"):
pad_to_length = (
max_tokens
if (padding == "max_length" or user_max_length)
else len(input_ids)
)
pad_len = pad_to_length - len(input_ids)
if pad_len > 0:
# GPT-style decoder-only LMs use absolute positions, so left-pad to
# keep real tokens aligned at the end and avoid position offsets.
input_ids = [PAD_ID] * pad_len + input_ids
attention_mask = [0] * pad_len + attention_mask
token_type_ids = [0] * pad_len + token_type_ids
mm_token_type_ids = [0] * pad_len + mm_token_type_ids
pixel_values_list = []
patch_attention_list = []
for media_type, media_datum in media_items:
if media_type == MediaType.IMAGE:
image_outputs = self.image_processor(
images=media_datum["pixel_values"], return_tensors="pt"
)
pixel_values_list.append(image_outputs["pixel_values"])
if "patch_attention_mask" in image_outputs:
patch_attention_list.append(
image_outputs["patch_attention_mask"]
)
elif media_type == MediaType.VIDEO:
video_outputs = self.video_processor.preprocess(
videos=media_datum["pixel_values"], return_tensors="pt"
)
pixel_values_list.append(video_outputs["pixel_values"])
patch_attention_list.append(
video_outputs["patch_attention_mask"]
)
else:
raise ValueError(f"Unsupported media type: {media_type}")
if pixel_values_list:
pixel_values = torch.cat(pixel_values_list, dim=0)
else:
pixel_values = torch.tensor([])
if patch_attention_list:
patch_attention_mask = torch.cat(patch_attention_list, dim=0)
else:
patch_attention_mask = torch.tensor([])
if return_tensors == "pt":
input_ids = torch.tensor(input_ids, dtype=torch.long)
attention_mask = torch.tensor(attention_mask, dtype=torch.long)
token_type_ids = torch.tensor(token_type_ids, dtype=torch.long)
mm_token_type_ids = torch.tensor(
mm_token_type_ids, dtype=torch.long
)
if input_ids.dim() == 1:
input_ids = input_ids.unsqueeze(0)
if attention_mask.dim() == 1:
attention_mask = attention_mask.unsqueeze(0)
if token_type_ids.dim() == 1:
token_type_ids = token_type_ids.unsqueeze(0)
if mm_token_type_ids.dim() == 1:
mm_token_type_ids = mm_token_type_ids.unsqueeze(0)
output = {
"input_ids": input_ids,
"attention_mask": attention_mask,
"token_type_ids": token_type_ids,
"mm_token_type_ids": mm_token_type_ids,
"pixel_values": pixel_values,
"patch_attention_mask": patch_attention_mask,
}
if return_prompt:
output["prompt"] = prompt
return output if return_dict else output
def __call__(
self,
images: Optional[Any] = None,
text: Optional[Union[str, List[str]]] = None,
videos: Optional[Any] = None,
audio: Optional[Any] = None,
**kwargs: Any,
) -> Any:
"""Run the processor and ensure multimodal token identifiers are present.
Args:
images: Optional image inputs.
text: Optional textual inputs.
videos: Optional video inputs.
audio: Optional audio inputs.
**kwargs: Additional keyword arguments forwarded to the base processor.
Returns:
Any: Processor outputs augmented with token type ids when needed.
"""
kwargs.pop("return_mm_token_type_ids", None)
image_processor = getattr(self, "image_processor", None)
img_tiling = kwargs.get("img_tiling", True)
tiling_method = kwargs.get(
"tiling_method",
getattr(image_processor, "tiling_method", "llava-uhd"),
)
tiling_size = kwargs.get("tiling_size")
if tiling_size is None and image_processor is not None:
size = getattr(image_processor, "size", None)
if isinstance(size, dict) and "shortest_edge" in size:
tiling_size = int(size["shortest_edge"])
elif isinstance(size, int):
tiling_size = size
tiling_size = tiling_size or 512
grid_pinpoints = kwargs.get("grid_pinpoints")
if grid_pinpoints is None:
grid_pinpoints = [
(2, 2),
(1, 2),
(2, 1),
(1, 3),
(3, 1),
(1, 4),
(4, 1),
]
max_tiles_num = kwargs.get(
"max_tiles_num", getattr(image_processor, "max_tiles_num", 4)
)
patch_size = kwargs.get(
"patch_size", getattr(image_processor, "patch_size", 14)
)
# vLLM infers placeholder splits from the tokenized prompt, so expand
# image placeholders to the exact tile count before tokenization.
if isinstance(text, str) and (
images is not None or videos is not None
):
if (
REKA_IMG_TOKEN not in text
and IMAGE_START not in text
and VIDEO_START not in text
):
text = self._prepend_mm_placeholders(
text=text, images=images, videos=videos, **kwargs
)
else:
text = self._expand_image_placeholders(
text=text, images=images, **kwargs
)
# vLLM derives placeholder lengths from processor outputs; tile before tokenization.
if images is not None and img_tiling:
images = self._tile_images(
images=images,
tiling_method=tiling_method,
tiling_size=tiling_size,
grid_pinpoints=grid_pinpoints,
max_tiles_num=max_tiles_num,
patch_size=patch_size,
)
# vLLM should treat tiled images as one prompt with multiple images.
if isinstance(text, str) and isinstance(images, list):
text = [text]
images = [images]
outputs = super().__call__(
images=images, text=text, videos=videos, audio=audio, **kwargs
)
if "input_ids" in outputs and "token_type_ids" not in outputs:
token_type_ids, mm_token_type_ids = self._build_mm_token_type_ids(
outputs["input_ids"]
)
outputs["token_type_ids"] = token_type_ids
outputs["mm_token_type_ids"] = mm_token_type_ids
return outputs
def _expand_image_placeholders(
self,
text: str,
images: Optional[Any],
**kwargs: Any,
) -> str:
if images is None or IMAGE_START not in text or IMAGE_END not in text:
return text
image_list = (
list(images) if isinstance(images, (list, tuple)) else [images]
)
image_processor = getattr(self, "image_processor", None)
img_tiling = kwargs.get("img_tiling", True)
tiling_method = kwargs.get(
"tiling_method",
getattr(image_processor, "tiling_method", "llava-uhd"),
)
tiling_size = kwargs.get("tiling_size")
if tiling_size is None and image_processor is not None:
size = getattr(image_processor, "size", None)
if isinstance(size, dict) and "shortest_edge" in size:
tiling_size = int(size["shortest_edge"])
elif isinstance(size, int):
tiling_size = size
tiling_size = tiling_size or 512
grid_pinpoints = kwargs.get("grid_pinpoints")
if grid_pinpoints is None:
grid_pinpoints = [
(2, 2),
(1, 2),
(2, 1),
(1, 3),
(3, 1),
(1, 4),
(4, 1),
]
max_tiles_num = kwargs.get(
"max_tiles_num", getattr(image_processor, "max_tiles_num", 4)
)
patch_size = kwargs.get(
"patch_size", getattr(image_processor, "patch_size", 14)
)
expected_tokens = []
for image in image_list:
width = height = 0
if hasattr(image, "size"):
width, height = image.size
elif isinstance(image, (list, tuple)) and len(image) >= 2:
height, width = int(image[0]), int(image[1])
if img_tiling and width > 0 and height > 0:
if str(tiling_method).lower() == "llava-next":
tiles = estimate_num_tiles_llava_next(
(width, height),
size=tiling_size,
grid_pinpoints=grid_pinpoints,
)
else:
tiles = estimate_num_tiles_llava_uhd(
(width, height),
max_tiles_num=max_tiles_num,
scale_resolution=tiling_size,
patch_size=patch_size,
never_split=False,
)
else:
tiles = 1
expected_tokens.append(self.num_img_tokens * tiles)
parts = []
remaining = text
for tokens in expected_tokens:
start = remaining.find(IMAGE_START)
end = remaining.find(IMAGE_END, start + len(IMAGE_START))
if start == -1 or end == -1:
return text
parts.append(remaining[:start])
parts.append(IMAGE_START + (REKA_IMG_TOKEN * tokens) + IMAGE_END)
remaining = remaining[end + len(IMAGE_END) :]
parts.append(remaining)
new_text = "".join(parts)
return new_text
def _tile_images(
self,
images: Any,
tiling_method: str,
tiling_size: int,
grid_pinpoints: List[Tuple[int, int]],
max_tiles_num: int,
patch_size: int,
) -> Any:
# vLLM expects one image entry per tile so it can emit per-tile embeddings.
image_list = (
list(images) if isinstance(images, (list, tuple)) else [images]
)
tiled_images: List[Any] = []
for image in image_list:
if image is None:
continue
if isinstance(image, torch.Tensor):
tiled_images.append(image)
continue
if isinstance(image, np.ndarray):
image = Image.fromarray(image)
if isinstance(image, Image.Image):
# Match the tiling logic used for placeholder expansion.
if str(tiling_method).lower() == "llava-next":
tiles = process_anyres_image(
image, size=tiling_size, grid_pinpoints=grid_pinpoints
)
else:
tiles = process_anyres_image_uhd(
image,
max_tiles_num=max_tiles_num,
scale_resolution=tiling_size,
patch_size=patch_size,
never_split=False,
)
tiled_images.extend(tiles)
continue
tiled_images.append(image)
return (
tiled_images
if isinstance(images, (list, tuple))
else tiled_images[0]
)
def _prepend_mm_placeholders(
self,
text: str,
images: Optional[Any],
videos: Optional[Any],
**kwargs: Any,
) -> str:
"""Prepend placeholder tokens when media is provided without markers."""
# Keep placeholders aligned with tiling so vLLM doesn't under/over-allocate.
image_list = (
list(images)
if isinstance(images, (list, tuple))
else ([images] if images is not None else [])
)
num_images = len(image_list)
num_videos = self._count_media_items(videos)
if num_images == 0 and num_videos == 0:
return text
image_processor = getattr(self, "image_processor", None)
img_tiling = kwargs.get("img_tiling", True)
tiling_method = kwargs.get(
"tiling_method",
getattr(image_processor, "tiling_method", "llava-uhd"),
)
tiling_size = kwargs.get("tiling_size")
if tiling_size is None and image_processor is not None:
size = getattr(image_processor, "size", None)
if isinstance(size, dict) and "shortest_edge" in size:
tiling_size = int(size["shortest_edge"])
elif isinstance(size, int):
tiling_size = size
tiling_size = tiling_size or 512
grid_pinpoints = kwargs.get("grid_pinpoints")
if grid_pinpoints is None:
grid_pinpoints = [
(2, 2),
(1, 2),
(2, 1),
(1, 3),
(3, 1),
(1, 4),
(4, 1),
]
max_tiles_num = kwargs.get(
"max_tiles_num", getattr(image_processor, "max_tiles_num", 4)
)
patch_size = kwargs.get(
"patch_size", getattr(image_processor, "patch_size", 14)
)
def _get_image_size(image: Any) -> Tuple[int, int]:
if hasattr(image, "size"):
size = image.size
if isinstance(size, (list, tuple)) and len(size) >= 2:
return int(size[0]), int(size[1])
if hasattr(image, "shape"):
shape = image.shape
if isinstance(shape, (list, tuple)) and len(shape) >= 2:
return int(shape[1]), int(shape[0])
if isinstance(image, (list, tuple)) and len(image) >= 2:
return int(image[1]), int(image[0])
return 0, 0
placeholder = ""
for image in image_list:
tiles = 1
if img_tiling:
width, height = _get_image_size(image)
if width > 0 and height > 0:
if str(tiling_method).lower() == "llava-next":
tiles = estimate_num_tiles_llava_next(
(width, height),
size=tiling_size,
grid_pinpoints=grid_pinpoints,
)
else:
tiles = estimate_num_tiles_llava_uhd(
(width, height),
max_tiles_num=max_tiles_num,
scale_resolution=tiling_size,
patch_size=patch_size,
never_split=False,
)
placeholder += IMAGE_START
placeholder += REKA_IMG_TOKEN * (self.num_img_tokens * tiles)
placeholder += IMAGE_END
for _ in range(num_videos):
placeholder += VIDEO_START
placeholder += REKA_IMG_TOKEN * (
self.num_img_tokens * self.num_video_frames
)
placeholder += VIDEO_END
return f"{placeholder}{text}"
@staticmethod
def _count_media_items(payload: Optional[Any]) -> int:
"""Best-effort count of media items for placeholder insertion."""
if payload is None:
return 0
if isinstance(payload, (list, tuple)):
return len(payload)
return 1
def _build_mm_token_type_ids(self, input_ids: Any) -> Tuple[Any, Any]:
"""Compute token_type_ids that mark multimodal placeholders.
Args:
input_ids: Input IDs or sequences containing tokenizer ids.
Returns:
Tuple[Any, Any]: Regular and multimodal token type ids detected from placeholders.
"""
if self.tokenizer is None:
return input_ids, input_ids
img_token_id = self.image_token_id
if isinstance(input_ids, torch.Tensor):
mm_token_type_ids = (input_ids == img_token_id).long()
token_type_ids = mm_token_type_ids.clone()
return token_type_ids, mm_token_type_ids
if isinstance(input_ids, (list, tuple)):
if input_ids and isinstance(input_ids[0], (list, tuple)):
mm_token_type_ids = [
[1 if token_id == img_token_id else 0 for token_id in seq]
for seq in input_ids
]
else:
mm_token_type_ids = [
1 if token_id == img_token_id else 0
for token_id in input_ids
]
token_type_ids = list(mm_token_type_ids)
return token_type_ids, mm_token_type_ids
if hasattr(input_ids, "tolist"):
ids = input_ids.tolist()
token_type_ids, mm_token_type_ids = self._build_mm_token_type_ids(
ids
)
return token_type_ids, mm_token_type_ids
return input_ids, input_ids
def _get_num_multimodal_tokens(
self,
image_sizes: Optional[List[List[int]]] = None,
video_sizes: Optional[List[List[int]]] = None,
**kwargs: Any,
) -> MultiModalData:
"""Estimate the count of multimodal tokens from provided media sizes.
Args:
image_sizes: Per-image sizes as (height, width) tuples.
video_sizes: Per-video sizes as (num_frames, height, width) tuples.
**kwargs: Ignored compatibility arguments accepted by parent helpers.
Returns:
MultiModalData: Token counts for the vision modalities.
"""
vision_data: Dict[str, List[int]] = {}
if image_sizes is not None:
image_processor = getattr(self, "image_processor", None)
img_tiling = kwargs.get("img_tiling", True)
tiling_method = kwargs.get(
"tiling_method",
getattr(image_processor, "tiling_method", "llava-uhd"),
)
tiling_size = kwargs.get("tiling_size")
if tiling_size is None and image_processor is not None:
size = getattr(image_processor, "size", None)
if isinstance(size, dict) and "shortest_edge" in size:
tiling_size = int(size["shortest_edge"])
elif isinstance(size, int):
tiling_size = size
tiling_size = tiling_size or 512
grid_pinpoints = kwargs.get("grid_pinpoints")
if grid_pinpoints is None:
grid_pinpoints = [
(2, 2),
(1, 2),
(2, 1),
(1, 3),
(3, 1),
(1, 4),
(4, 1),
]
max_tiles_num = kwargs.get(
"max_tiles_num", getattr(image_processor, "max_tiles_num", 4)
)
patch_size = kwargs.get(
"patch_size", getattr(image_processor, "patch_size", 14)
)
# vLLM splits placeholder positions using per-image token/patch counts.
num_image_tokens: List[int] = []
num_image_patches: List[int] = []
for image_size in image_sizes:
height = width = 0
if image_size and len(image_size) >= 2:
height, width = int(image_size[0]), int(image_size[1])
tiles = 1
if img_tiling and width > 0 and height > 0:
if str(tiling_method).lower() == "llava-next":
tiles = estimate_num_tiles_llava_next(
(width, height),
size=tiling_size,
grid_pinpoints=grid_pinpoints,
)
else:
tiles = estimate_num_tiles_llava_uhd(
(width, height),
max_tiles_num=max_tiles_num,
scale_resolution=tiling_size,
patch_size=patch_size,
never_split=False,
)
num_image_tokens.append(self.num_img_tokens * tiles)
num_image_patches.append(tiles)
vision_data["num_image_tokens"] = num_image_tokens
vision_data["num_image_patches"] = num_image_patches
else:
vision_data["num_image_tokens"] = []
vision_data["num_image_patches"] = []
if video_sizes is not None:
video_tokens: List[int] = []
for video_size in video_sizes:
num_frames = video_size[0] if video_size else 0
num_frames = min(
num_frames or self.num_video_frames, self.num_video_frames
)
video_tokens.append(self.num_img_tokens * num_frames)
vision_data["num_video_tokens"] = video_tokens
return MultiModalData(**vision_data)
Yasa2Processor.register_for_auto_class()
|