| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inference-only Deepseek-OCR model compatible with HuggingFace weights.""" |
| 4 | |
| 5 | import math |
| 6 | from collections.abc import Hashable, Iterable, Mapping, Sequence |
| 7 | from typing import Annotated, Any, Literal |
| 8 | |
| 9 | import torch |
| 10 | import torch.nn as nn |
| 11 | from transformers import BatchFeature, CLIPVisionConfig |
| 12 | |
| 13 | from vllm.config import VllmConfig |
| 14 | from vllm.config.multimodal import MultiModalDummyOptions |
| 15 | from vllm.inputs import MultiModalDataDict |
| 16 | from vllm.model_executor.models.interfaces import ( |
| 17 | MultiModalEmbeddings, |
| 18 | SupportsEncoderCudaGraph, |
| 19 | SupportsLoRA, |
| 20 | SupportsMultiModal, |
| 21 | SupportsPP, |
| 22 | ) |
| 23 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 24 | from vllm.model_executor.models.utils import ( |
| 25 | AutoWeightsLoader, |
| 26 | WeightsMapper, |
| 27 | init_vllm_registered_model, |
| 28 | maybe_prefix, |
| 29 | ) |
| 30 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 31 | from vllm.multimodal.inputs import ( |
| 32 | MultiModalFieldConfig, |
| 33 | MultiModalKwargsItems, |
| 34 | NestedTensors, |
| 35 | ) |
| 36 | from vllm.multimodal.parse import ( |
| 37 | ImageEmbeddingItems, |
| 38 | ImageProcessorItems, |
| 39 | ImageSize, |
| 40 | MultiModalDataItems, |
| 41 | ) |
| 42 | from vllm.multimodal.processing import ( |
| 43 | BaseDummyInputsBuilder, |
| 44 | BaseMultiModalProcessor, |
| 45 | BaseProcessingInfo, |
| 46 | PromptReplacement, |
| 47 | PromptUpdate, |
| 48 | ) |
| 49 | from vllm.multimodal.processing.processor import HFMultiModalInputs |
| 50 | from vllm.sampling_params import SamplingParams |
| 51 | from vllm.sequence import IntermediateTensors |
| 52 | from vllm.tokenizers import cached_tokenizer_from_config |
| 53 | from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekVLV2Config |
| 54 | from vllm.transformers_utils.processors.deepseek_ocr import ( |
| 55 | BASE_SIZE, |
| 56 | CROP_MODE, |
| 57 | IMAGE_SIZE, |
| 58 | DeepseekOCRProcessor, |
| 59 | count_tiles, |
| 60 | ) |
| 61 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 62 | from vllm.v1.sample.logits_processor import ( |
| 63 | AdapterLogitsProcessor, |
| 64 | RequestLogitsProcessor, |
| 65 | ) |
| 66 | from vllm.v1.worker.encoder_cudagraph_defs import ( |
| 67 | EncoderCudaGraphCaptureInputs, |
| 68 | EncoderCudaGraphConfig, |
| 69 | EncoderCudaGraphPathConfig, |
| 70 | EncoderCudaGraphReplayBuffers, |
| 71 | EncoderItemSpec, |
| 72 | ) |
| 73 | |
| 74 | from .deepencoder import DeepCLIPVisionTransformer, build_sam_vit_b |
| 75 | from .deepseek_vl2 import MlpProjector |
| 76 | |
| 77 | # The image token id may be various |
| 78 | _IMAGE_TOKEN = "<image>" |
| 79 | |
| 80 | |
| 81 | class DeepseekOCRImagePixelInputs(TensorSchema): |
| 82 | """Dimensions: |
| 83 | - b: Batch size |
| 84 | - n: Number of images |
| 85 | - p: Number of patches |
| 86 | - base_size: Base size of the processor |
| 87 | - image_size: Image size of the processor |
| 88 | """ |
| 89 | |
| 90 | type: Literal["pixel_values"] |
| 91 | data: Annotated[ |
| 92 | torch.Tensor, |
| 93 | TensorShape("bn", 3, "base_size", "base_size", dynamic_dims={"bnp"}), |
| 94 | ] |
| 95 | images_crop: Annotated[ |
| 96 | torch.Tensor, |
| 97 | TensorShape("bnp", 3, "image_size", "image_size", dynamic_dims={"bnp"}), |
| 98 | ] |
| 99 | images_spatial_crop: Annotated[torch.Tensor, TensorShape("bn", 2)] |
| 100 | |
| 101 | |
| 102 | class NoRepeatNGramLogitsProcessor: |
| 103 | def __init__( |
| 104 | self, |
| 105 | ngram_size: int, |
| 106 | window_size: int, |
| 107 | whitelist_token_ids: set[int] | None = None, |
| 108 | ): |
| 109 | self.ngram_size = ngram_size |
| 110 | self.window_size = window_size |
| 111 | self.whitelist_token_ids = whitelist_token_ids or set() |
| 112 | |
| 113 | def __call__( |
| 114 | self, |
| 115 | output_ids: list[int], |
| 116 | logits: torch.Tensor, |
| 117 | ) -> torch.Tensor: |
| 118 | if len(output_ids) < self.ngram_size: |
| 119 | return logits |
| 120 | |
| 121 | current_prefix = tuple(output_ids[-(self.ngram_size - 1) :]) |
| 122 | |
| 123 | search_start = max(0, len(output_ids) - self.window_size) |
| 124 | search_end = len(output_ids) - self.ngram_size + 1 |
| 125 | |
| 126 | banned_tokens = set() |
| 127 | for i in range(search_start, search_end): |
| 128 | ngram = tuple(output_ids[i : i + self.ngram_size]) |
| 129 | if ngram[:-1] == current_prefix: |
| 130 | banned_tokens.add(ngram[-1]) |
| 131 | |
| 132 | banned_tokens = banned_tokens - self.whitelist_token_ids |
| 133 | |
| 134 | if banned_tokens: |
| 135 | logits[list(banned_tokens)] = -float("inf") |
| 136 | |
| 137 | return logits |
| 138 | |
| 139 | |
| 140 | class NGramPerReqLogitsProcessor(AdapterLogitsProcessor): |
| 141 | """Example of overriding the wrapper class `__init__()` in order to utilize |
| 142 | info about the device type""" |
| 143 | |
| 144 | @classmethod |
| 145 | def validate_params(cls, params: SamplingParams): |
| 146 | ngram_size = params.extra_args and params.extra_args.get("ngram_size") |
| 147 | window_size = params.extra_args and params.extra_args.get("window_size", 100) |
| 148 | whitelist_token_ids = params.extra_args and params.extra_args.get( |
| 149 | "whitelist_token_ids", None |
| 150 | ) |
| 151 | # if ngram_size is not provided, skip validation because the processor |
| 152 | # will not be used. |
| 153 | if ngram_size is None: |
| 154 | return None |
| 155 | |
| 156 | if not isinstance(ngram_size, int) or ngram_size <= 0: |
| 157 | raise ValueError( |
| 158 | f"`ngram_size` has to be a strictly positive integer, got {ngram_size}." |
| 159 | ) |
| 160 | if not isinstance(window_size, int) or window_size <= 0: |
| 161 | raise ValueError( |
| 162 | "`window_size` has to be a strictly positive integer, " |
| 163 | f"got {window_size}." |
| 164 | ) |
| 165 | if whitelist_token_ids is not None and not isinstance( |
| 166 | whitelist_token_ids, Iterable |
| 167 | ): |
| 168 | raise ValueError( |
| 169 | "`whitelist_token_ids` has to be a sequence of integers, " |
| 170 | f"got {whitelist_token_ids}." |
| 171 | ) |
| 172 | |
| 173 | def is_argmax_invariant(self) -> bool: |
| 174 | return False |
| 175 | |
| 176 | def new_req_logits_processor( |
| 177 | self, |
| 178 | params: SamplingParams, |
| 179 | ) -> RequestLogitsProcessor | None: |
| 180 | ngram_size = params.extra_args and params.extra_args.get("ngram_size") |
| 181 | window_size = params.extra_args and params.extra_args.get("window_size", 100) |
| 182 | whitelist_token_ids = params.extra_args and params.extra_args.get( |
| 183 | "whitelist_token_ids", None |
| 184 | ) |
| 185 | if ngram_size is None: |
| 186 | return None |
| 187 | assert isinstance(ngram_size, int) and isinstance(window_size, int) |
| 188 | |
| 189 | whitelist_token_ids = set(whitelist_token_ids) if whitelist_token_ids else None |
| 190 | return NoRepeatNGramLogitsProcessor( |
| 191 | ngram_size=ngram_size, |
| 192 | window_size=window_size, |
| 193 | whitelist_token_ids=whitelist_token_ids, |
| 194 | ) |
| 195 | |
| 196 | |
| 197 | class DeepseekOCRProcessingInfo(BaseProcessingInfo): |
| 198 | def get_hf_config(self): |
| 199 | return self.ctx.get_hf_config(DeepseekVLV2Config) |
| 200 | |
| 201 | def get_hf_processor(self, **kwargs: object): |
| 202 | v1_processor_config = dict( |
| 203 | image_size=IMAGE_SIZE, |
| 204 | base_size=BASE_SIZE, |
| 205 | crop_mode=CROP_MODE, |
| 206 | strategy="v1", |
| 207 | ) |
| 208 | |
| 209 | return self.ctx.get_hf_processor( |
| 210 | DeepseekOCRProcessor, |
| 211 | **{**v1_processor_config, **kwargs}, |
| 212 | ) |
| 213 | |
| 214 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 215 | return {"image": None} |
| 216 | |
| 217 | def get_num_image_tokens( |
| 218 | self, *, image_width: int, image_height: int, cropping: bool = True |
| 219 | ) -> int: |
| 220 | image_size = IMAGE_SIZE |
| 221 | base_size = BASE_SIZE |
| 222 | patch_size = 16 |
| 223 | downsample_ratio = 4 |
| 224 | |
| 225 | # Use the caller-supplied `cropping` flag so that callers that disable |
| 226 | # crop mode for multi-image requests get a consistent token count. |
| 227 | if cropping: |
| 228 | if image_width <= IMAGE_SIZE and image_height <= IMAGE_SIZE: |
| 229 | crop_ratio = [1, 1] |
| 230 | else: |
| 231 | # find the closest aspect ratio to the target |
| 232 | crop_ratio = count_tiles( |
| 233 | image_width, image_height, image_size=IMAGE_SIZE |
| 234 | ) |
| 235 | |
| 236 | num_width_tiles, num_height_tiles = crop_ratio |
| 237 | else: |
| 238 | num_width_tiles = num_height_tiles = 1 |
| 239 | |
| 240 | h = w = math.ceil((base_size // patch_size) / downsample_ratio) |
| 241 | |
| 242 | h2 = w2 = math.ceil((image_size // patch_size) / downsample_ratio) |
| 243 | |
| 244 | global_views_tokens = h * (w + 1) |
| 245 | if num_width_tiles > 1 or num_height_tiles > 1: |
| 246 | local_views_tokens = (num_height_tiles * h2) * (num_width_tiles * w2 + 1) |
| 247 | else: |
| 248 | local_views_tokens = 0 |
| 249 | |
| 250 | return global_views_tokens + local_views_tokens + 1 |
| 251 | |
| 252 | def get_image_size_with_most_features(self) -> ImageSize: |
| 253 | if IMAGE_SIZE == 1024 and BASE_SIZE == 1280: |
| 254 | return ImageSize(width=1024 * 2, height=1024 * 2) |
| 255 | return ImageSize(width=640 * 2, height=640 * 2) |
| 256 | |
| 257 | |
| 258 | class DeepseekOCRDummyInputsBuilder(BaseDummyInputsBuilder[DeepseekOCRProcessingInfo]): |
| 259 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 260 | num_images = mm_counts.get("image", 0) |
| 261 | |
| 262 | processor = self.info.get_hf_processor() |
| 263 | image_token = processor.image_token |
| 264 | |
| 265 | return image_token * num_images |
| 266 | |
| 267 | def get_dummy_mm_data( |
| 268 | self, |
| 269 | seq_len: int, |
| 270 | mm_counts: Mapping[str, int], |
| 271 | mm_options: MultiModalDummyOptions, |
| 272 | ) -> MultiModalDataDict: |
| 273 | max_image_size = self.info.get_image_size_with_most_features() |
| 274 | |
| 275 | return { |
| 276 | "image": self._get_dummy_images( |
| 277 | width=max_image_size.width, |
| 278 | height=max_image_size.height, |
| 279 | num_images=mm_counts.get("image", 0), |
| 280 | ) |
| 281 | } |
| 282 | |
| 283 | |
| 284 | class DeepseekOCRMultiModalProcessor( |
| 285 | BaseMultiModalProcessor[DeepseekOCRProcessingInfo] |
| 286 | ): |
| 287 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 288 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 289 | |
| 290 | def _get_hf_mm_inputs( |
| 291 | self, |
| 292 | mm_items: MultiModalDataItems, |
| 293 | hf_kwargs: Mapping[str, object], |
| 294 | ) -> HFMultiModalInputs: |
| 295 | hf_inputs = super()._get_hf_mm_inputs(mm_items, hf_kwargs) |
| 296 | if "text" in hf_inputs.hf_data: |
| 297 | hf_inputs.hf_data["prompt"] = hf_inputs.hf_data.pop("text") |
| 298 | |
| 299 | return hf_inputs |
| 300 | |
| 301 | def _get_mm_fields_config( |
| 302 | self, |
| 303 | hf_inputs: BatchFeature, |
| 304 | hf_processor_mm_kwargs: Mapping[str, object], |
| 305 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 306 | images_spatial_crop = hf_inputs.get("images_spatial_crop", torch.empty((0, 2))) |
| 307 | is_tiled = (images_spatial_crop[:, 0] > 1) | (images_spatial_crop[:, 1] > 1) |
| 308 | patches_per_image = torch.where(is_tiled, images_spatial_crop.prod(dim=-1), 0) |
| 309 | return dict( |
| 310 | pixel_values=MultiModalFieldConfig.batched("image"), |
| 311 | images_spatial_crop=MultiModalFieldConfig.batched( |
| 312 | "image", keep_on_cpu=True |
| 313 | ), |
| 314 | images_crop=MultiModalFieldConfig.flat_from_sizes( |
| 315 | "image", patches_per_image |
| 316 | ), |
| 317 | ) |
| 318 | |
| 319 | def _get_prompt_updates( |
| 320 | self, |
| 321 | mm_items: MultiModalDataItems, |
| 322 | hf_processor_mm_kwargs: Mapping[str, object], |
| 323 | out_mm_kwargs: MultiModalKwargsItems, |
| 324 | ) -> Sequence[PromptUpdate]: |
| 325 | hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 326 | |
| 327 | image_token_id = hf_processor.image_token_id |
| 328 | assert isinstance(image_token_id, int) |
| 329 | |
| 330 | def get_replacement_deepseek_vl2(item_idx: int): |
| 331 | images = mm_items.get_items( |
| 332 | "image", (ImageEmbeddingItems, ImageProcessorItems) |
| 333 | ) |
| 334 | |
| 335 | if isinstance(images, ImageEmbeddingItems): |
| 336 | num_image_tokens = images.get_feature_size(item_idx) |
| 337 | else: |
| 338 | assert isinstance(images, ImageProcessorItems) |
| 339 | size = images.get_image_size(item_idx) |
| 340 | |
| 341 | num_image_tokens = self.info.get_num_image_tokens( |
| 342 | image_width=size.width, |
| 343 | image_height=size.height, |
| 344 | cropping=CROP_MODE, |
| 345 | ) |
| 346 | return [image_token_id] * num_image_tokens |
| 347 | |
| 348 | return [ |
| 349 | PromptReplacement( |
| 350 | modality="image", |
| 351 | target=[image_token_id], |
| 352 | replacement=get_replacement_deepseek_vl2, |
| 353 | ) |
| 354 | ] |
| 355 | |
| 356 | |
| 357 | @MULTIMODAL_REGISTRY.register_processor( |
| 358 | DeepseekOCRMultiModalProcessor, |
| 359 | info=DeepseekOCRProcessingInfo, |
| 360 | dummy_inputs=DeepseekOCRDummyInputsBuilder, |
| 361 | ) |
| 362 | class DeepseekOCRForCausalLM( |
| 363 | nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA, SupportsEncoderCudaGraph |
| 364 | ): |
| 365 | hf_to_vllm_mapper = WeightsMapper( |
| 366 | orig_to_new_prefix={ |
| 367 | # map prefix for language backbone |
| 368 | "model.embed_tokens.": "language_model.model.embed_tokens.", |
| 369 | "model.layers.": "language_model.model.layers.", |
| 370 | "model.norm.": "language_model.model.norm.", |
| 371 | "lm_head.": "language_model.lm_head.", |
| 372 | # remove "model." prefix for other components |
| 373 | "model.": "", |
| 374 | } |
| 375 | ) |
| 376 | |
| 377 | @classmethod |
| 378 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 379 | if modality.startswith("image"): |
| 380 | return "<image>" |
| 381 | |
| 382 | raise ValueError("Only image modality is supported") |
| 383 | |
| 384 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 385 | super().__init__() |
| 386 | |
| 387 | config: DeepseekVLV2Config = vllm_config.model_config.hf_config |
| 388 | quant_config = vllm_config.quant_config |
| 389 | multimodal_config = vllm_config.model_config.multimodal_config |
| 390 | |
| 391 | self.config = config |
| 392 | self.model_config = vllm_config.model_config |
| 393 | self.multimodal_config = multimodal_config |
| 394 | |
| 395 | self.vision_config = config.vision_config |
| 396 | self.projector_config = config.projector_config |
| 397 | self.text_config = config.text_config |
| 398 | |
| 399 | model_config = vllm_config.model_config |
| 400 | tokenizer = cached_tokenizer_from_config(model_config) |
| 401 | self.image_token_id = tokenizer.vocab[_IMAGE_TOKEN] |
| 402 | |
| 403 | with self._mark_tower_model(vllm_config, "image"): |
| 404 | self.sam_model = build_sam_vit_b() |
| 405 | clip_vision_config = CLIPVisionConfig( |
| 406 | hidden_size=1024, |
| 407 | intermediate_size=4096, |
| 408 | num_attention_heads=16, |
| 409 | num_hidden_layers=24, |
| 410 | image_size=224, |
| 411 | patch_size=14, |
| 412 | projection_dim=512, |
| 413 | layer_norm_eps=1e-5, |
| 414 | ) |
| 415 | self.vision_model = DeepCLIPVisionTransformer( |
| 416 | config=clip_vision_config, |
| 417 | quant_config=quant_config, |
| 418 | prefix=maybe_prefix(prefix, "vision_model"), |
| 419 | ) |
| 420 | |
| 421 | self.projector = MlpProjector(self.projector_config) |
| 422 | self.tile_tag = config.tile_tag |
| 423 | self.global_view_pos = config.global_view_pos |
| 424 | |
| 425 | # special token for image token sequence format |
| 426 | n_embed = self.projector_config.n_embed |
| 427 | embed_std = 1 / torch.sqrt(torch.tensor(n_embed, dtype=torch.float32)) |
| 428 | if self.tile_tag == "2D": |
| 429 | # <|view_separator|>, <|\n|> |
| 430 | self.image_newline = nn.Parameter(torch.randn(n_embed) * embed_std) |
| 431 | # This is a typo in original implementation |
| 432 | self.view_seperator = nn.Parameter(torch.randn(n_embed) * embed_std) |
| 433 | else: |
| 434 | raise ValueError( |
| 435 | f"Only 2D tile_tag is supported currently, got: {self.tile_tag}" |
| 436 | ) |
| 437 | |
| 438 | with self._mark_language_model(vllm_config): |
| 439 | self.language_model = init_vllm_registered_model( |
| 440 | vllm_config=vllm_config, |
| 441 | hf_config=self.text_config, |
| 442 | prefix=maybe_prefix(prefix, "language_model"), |
| 443 | ) |
| 444 | |
| 445 | self.make_empty_intermediate_tensors = ( |
| 446 | self.language_model.make_empty_intermediate_tensors |
| 447 | ) |
| 448 | |
| 449 | def _parse_and_validate_image_input( |
| 450 | self, **kwargs: object |
| 451 | ) -> DeepseekOCRImagePixelInputs | None: |
| 452 | pixel_values = kwargs.pop("pixel_values", None) |
| 453 | images_spatial_crop = kwargs.pop("images_spatial_crop", None) |
| 454 | images_crop = kwargs.pop("images_crop", None) |
| 455 | |
| 456 | if pixel_values is None: |
| 457 | return None |
| 458 | assert isinstance(pixel_values, torch.Tensor) |
| 459 | assert images_crop is None or isinstance(images_crop, torch.Tensor) |
| 460 | |
| 461 | # Use actual tensor spatial dim instead of hardcoded |
| 462 | # vision_config.image_size (1024). The vision encoders (SAM & CLIP) |
| 463 | # support arbitrary resolutions via pos-encoding interpolation, |
| 464 | # so Tiny/Small/Base/Large variants all work with the same weights. |
| 465 | base_size = pixel_values.shape[-1] |
| 466 | image_size = images_crop.shape[-1] if images_crop is not None else base_size |
| 467 | |
| 468 | return DeepseekOCRImagePixelInputs( |
| 469 | type="pixel_values", |
| 470 | data=pixel_values, |
| 471 | images_crop=images_crop, |
| 472 | images_spatial_crop=images_spatial_crop, |
| 473 | resolve_bindings={ |
| 474 | "base_size": base_size, |
| 475 | "image_size": image_size, |
| 476 | }, |
| 477 | ) |
| 478 | |
| 479 | def _encode_global_features(self, image_tensor: torch.Tensor) -> torch.Tensor: |
| 480 | global_features_1 = self.sam_model(image_tensor) |
| 481 | global_features_2 = self.vision_model(image_tensor, global_features_1) |
| 482 | features = torch.cat( |
| 483 | ( |
| 484 | global_features_2[:, 1:], |
| 485 | global_features_1.flatten(2).permute(0, 2, 1), |
| 486 | ), |
| 487 | dim=-1, |
| 488 | ) |
| 489 | features = self.projector(features) |
| 490 | |
| 491 | _, hw, dim = features.shape |
| 492 | side = int(hw**0.5) |
| 493 | |
| 494 | features = features.view(side, side, dim) |
| 495 | newline = self.image_newline[None, None, :].expand(side, 1, dim) |
| 496 | features = torch.cat([features, newline], dim=1) |
| 497 | return features.view(-1, dim) |
| 498 | |
| 499 | def _encode_local_features( |
| 500 | self, patches: torch.Tensor, crop_shape: torch.Tensor |
| 501 | ) -> torch.Tensor | None: |
| 502 | if torch.sum(patches).item() == 0: |
| 503 | return None |
| 504 | |
| 505 | local_features_1 = self.sam_model(patches) |
| 506 | local_features_2 = self.vision_model(patches, local_features_1) |
| 507 | features = torch.cat( |
| 508 | ( |
| 509 | local_features_2[:, 1:], |
| 510 | local_features_1.flatten(2).permute(0, 2, 1), |
| 511 | ), |
| 512 | dim=-1, |
| 513 | ) |
| 514 | features = self.projector(features) |
| 515 | |
| 516 | return self._assemble_patch_grid(features, crop_shape) |
| 517 | |
| 518 | def _assemble_patch_grid( |
| 519 | self, features: torch.Tensor, crop_shape: torch.Tensor |
| 520 | ) -> torch.Tensor: |
| 521 | """Assemble projected patches into a 2-D tile grid with newline columns.""" |
| 522 | _, hw, dim = features.shape |
| 523 | patch_side = int(hw**0.5) |
| 524 | width_tiles = int(crop_shape[0].item()) |
| 525 | height_tiles = int(crop_shape[1].item()) |
| 526 | |
| 527 | features = ( |
| 528 | features.view(height_tiles, width_tiles, patch_side, patch_side, dim) |
| 529 | .permute(0, 2, 1, 3, 4) |
| 530 | .reshape(height_tiles * patch_side, width_tiles * patch_side, dim) |
| 531 | ) |
| 532 | newline = self.image_newline[None, None, :].expand( |
| 533 | height_tiles * patch_side, 1, dim |
| 534 | ) |
| 535 | features = torch.cat([features, newline], dim=1) |
| 536 | return features.view(-1, dim) |
| 537 | |
| 538 | def _pixel_values_to_embedding( |
| 539 | self, |
| 540 | pixel_values: torch.Tensor, |
| 541 | images_crop: torch.Tensor, |
| 542 | images_spatial_crop: torch.Tensor, |
| 543 | ) -> NestedTensors: |
| 544 | images_in_this_batch = [] |
| 545 | |
| 546 | is_tiled = (images_spatial_crop[:, 0] > 1) | (images_spatial_crop[:, 1] > 1) |
| 547 | patches_per_image = torch.where(is_tiled, images_spatial_crop.prod(dim=-1), 0) |
| 548 | images_crop = images_crop.split(patches_per_image.tolist()) |
| 549 | for jdx in range(images_spatial_crop.size(0)): |
| 550 | patches = images_crop[jdx] |
| 551 | image_ori = pixel_values[[jdx]] |
| 552 | crop_shape = images_spatial_crop[jdx] |
| 553 | |
| 554 | global_features = self._encode_global_features(image_ori) |
| 555 | local_features = self._encode_local_features(patches, crop_shape) |
| 556 | |
| 557 | if local_features is not None: |
| 558 | combined = torch.cat( |
| 559 | [local_features, global_features, self.view_seperator[None, :]], |
| 560 | dim=0, |
| 561 | ) |
| 562 | else: |
| 563 | combined = torch.cat( |
| 564 | [global_features, self.view_seperator[None, :]], dim=0 |
| 565 | ) |
| 566 | |
| 567 | images_in_this_batch.append(combined) |
| 568 | |
| 569 | return images_in_this_batch |
| 570 | |
| 571 | def _process_image_input( |
| 572 | self, image_input: DeepseekOCRImagePixelInputs |
| 573 | ) -> torch.Tensor: |
| 574 | pixel_values = image_input.data |
| 575 | images_crop = image_input.images_crop |
| 576 | images_spatial_crop = image_input.images_spatial_crop.to(dtype=torch.long) |
| 577 | |
| 578 | vision_features = self._pixel_values_to_embedding( |
| 579 | pixel_values=pixel_values, |
| 580 | images_crop=images_crop, |
| 581 | images_spatial_crop=images_spatial_crop, |
| 582 | ) |
| 583 | |
| 584 | return vision_features |
| 585 | |
| 586 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings | None: |
| 587 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 588 | if image_input is None: |
| 589 | return None |
| 590 | vision_embeddings = self._process_image_input(image_input) |
| 591 | return vision_embeddings |
| 592 | |
| 593 | def forward( |
| 594 | self, |
| 595 | input_ids: torch.Tensor | None, |
| 596 | positions: torch.Tensor, |
| 597 | intermediate_tensors: IntermediateTensors | None = None, |
| 598 | inputs_embeds: torch.Tensor | None = None, |
| 599 | **kwargs: object, |
| 600 | ): |
| 601 | if intermediate_tensors is not None: |
| 602 | inputs_embeds = None |
| 603 | |
| 604 | hidden_states = self.language_model( |
| 605 | input_ids, positions, intermediate_tensors, inputs_embeds=inputs_embeds |
| 606 | ) |
| 607 | |
| 608 | return hidden_states |
| 609 | |
| 610 | def compute_logits( |
| 611 | self, |
| 612 | hidden_states: torch.Tensor, |
| 613 | ) -> torch.Tensor | None: |
| 614 | return self.language_model.compute_logits(hidden_states) |
| 615 | |
| 616 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 617 | loader = AutoWeightsLoader(self) |
| 618 | autoloaded_weights = loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 619 | return autoloaded_weights |
| 620 | |
| 621 | def get_mm_mapping(self) -> MultiModelKeys: |
| 622 | """Get the module prefix in multimodal models.""" |
| 623 | return MultiModelKeys.from_string_field( |
| 624 | language_model="language_model", |
| 625 | connector="projector", |
| 626 | tower_model=["sam_model", "vision_model"], |
| 627 | ) |
| 628 | |
| 629 | # -- Fixed spatial constants (computed from BASE_SIZE / IMAGE_SIZE) -- |
| 630 | |
| 631 | @property |
| 632 | def image_side(self) -> int: |
| 633 | """Number of output grid cells per spatial dim for a global image.""" |
| 634 | return math.ceil((BASE_SIZE // 16) / 4) # 16 |
| 635 | |
| 636 | @property |
| 637 | def global_image_output_token(self) -> int: |
| 638 | """Tokens per global image (grid + one newline per row).""" |
| 639 | return self.image_side * (self.image_side + 1) # 272 |
| 640 | |
| 641 | @property |
| 642 | def patch_side(self) -> int: |
| 643 | """Number of output grid cells per spatial dim for a local patch.""" |
| 644 | return math.ceil((IMAGE_SIZE // 16) / 4) # 10 |
| 645 | |
| 646 | @property |
| 647 | def single_patch_output_token(self) -> int: |
| 648 | """Tokens per local patch (square grid, no newlines).""" |
| 649 | return self.patch_side * self.patch_side # 100 |
| 650 | |
| 651 | # -- SupportsEncoderCudaGraph protocol methods -- |
| 652 | |
| 653 | def _get_num_input_output_tokens( |
| 654 | self, |
| 655 | image_spatial_crop: torch.Tensor | None = None, |
| 656 | ) -> tuple[int, int, int, int]: |
| 657 | """Return (num_input_tokens, num_output_tokens, global_output_token, |
| 658 | local_output_token) for a single image described by |
| 659 | ``image_spatial_crop``. |
| 660 | """ |
| 661 | is_tiled = False |
| 662 | if image_spatial_crop is not None: |
| 663 | is_tiled = image_spatial_crop[0] > 1 or image_spatial_crop[1] > 1 |
| 664 | |
| 665 | # Compute input size: |
| 666 | global_input_side = BASE_SIZE // 16 # 64 |
| 667 | local_input_side = IMAGE_SIZE // 16 # 40 |
| 668 | num_input_tokens = global_input_side**2 |
| 669 | |
| 670 | if is_tiled: |
| 671 | assert image_spatial_crop is not None |
| 672 | num_patches = int(image_spatial_crop.prod().item()) |
| 673 | num_input_tokens += num_patches * (local_input_side**2) |
| 674 | |
| 675 | global_output_token = self.global_image_output_token |
| 676 | num_output_tokens = global_output_token |
| 677 | |
| 678 | local_output_token = 0 |
| 679 | if is_tiled: |
| 680 | local_output_token = num_patches * self.single_patch_output_token |
| 681 | num_output_tokens += local_output_token |
| 682 | |
| 683 | return ( |
| 684 | num_input_tokens, |
| 685 | num_output_tokens, |
| 686 | global_output_token, |
| 687 | local_output_token, |
| 688 | ) |
| 689 | |
| 690 | def get_encoder_cudagraph_config(self): |
| 691 | return EncoderCudaGraphConfig( |
| 692 | modalities=["image"], |
| 693 | buffer_keys=["pixel_values"], |
| 694 | out_hidden_size=self.projector_config.n_embed, |
| 695 | paths={ |
| 696 | "global": EncoderCudaGraphPathConfig( |
| 697 | min_token_budget=self.global_image_output_token |
| 698 | ), |
| 699 | "local": EncoderCudaGraphPathConfig( |
| 700 | min_token_budget=self.single_patch_output_token, |
| 701 | allow_zero_tokens=True, |
| 702 | ), |
| 703 | }, |
| 704 | ) |
| 705 | |
| 706 | def get_encoder_cudagraph_budget_range( |
| 707 | self, |
| 708 | vllm_config, |
| 709 | ) -> tuple[int, int]: |
| 710 | # Min budget: at least one global image with newline tokens (without patches). |
| 711 | min_budget = self.global_image_output_token |
| 712 | max_budget = min( |
| 713 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 714 | self.model_config.max_model_len, |
| 715 | ) |
| 716 | return (min_budget, max_budget) |
| 717 | |
| 718 | def get_encoder_cudagraph_item_specs( |
| 719 | self, |
| 720 | mm_kwargs: dict[str, Any], |
| 721 | ) -> list[EncoderItemSpec]: |
| 722 | item_specs = [] |
| 723 | for image_spatial_crop in mm_kwargs["images_spatial_crop"]: |
| 724 | ( |
| 725 | num_input_tokens, |
| 726 | num_output_tokens, |
| 727 | global_output_token, |
| 728 | local_output_token, |
| 729 | ) = self._get_num_input_output_tokens(image_spatial_crop) |
| 730 | item_specs.append( |
| 731 | EncoderItemSpec( |
| 732 | input_size=num_input_tokens, |
| 733 | output_tokens=num_output_tokens, |
| 734 | path_output_tokens={ |
| 735 | "global": global_output_token, |
| 736 | "local": local_output_token, |
| 737 | }, |
| 738 | ) |
| 739 | ) |
| 740 | return item_specs |
| 741 | |
| 742 | def select_encoder_cudagraph_items( |
| 743 | self, |
| 744 | mm_kwargs: dict[str, Any], |
| 745 | indices: list[int], |
| 746 | ) -> dict[str, Any]: |
| 747 | pixel_values = mm_kwargs["pixel_values"] |
| 748 | images_crop = mm_kwargs["images_crop"] |
| 749 | images_spatial_crop = mm_kwargs["images_spatial_crop"] |
| 750 | |
| 751 | if len(indices) == 0: |
| 752 | return { |
| 753 | "pixel_values": pixel_values[:0], |
| 754 | "images_crop": images_crop[:0], |
| 755 | "images_spatial_crop": images_spatial_crop[:0], |
| 756 | } |
| 757 | |
| 758 | is_tiled = (images_spatial_crop[:, 0] > 1) | (images_spatial_crop[:, 1] > 1) |
| 759 | patches_per_image = torch.where(is_tiled, images_spatial_crop.prod(dim=-1), 0) |
| 760 | cum_patches = [0] |
| 761 | for num_patches in patches_per_image: |
| 762 | cum_patches.append(cum_patches[-1] + int(num_patches)) |
| 763 | |
| 764 | selected_pv = pixel_values[indices] |
| 765 | selected_ic = torch.cat( |
| 766 | [images_crop[cum_patches[i] : cum_patches[i + 1]] for i in indices] |
| 767 | ) |
| 768 | selected_sp = images_spatial_crop[indices] |
| 769 | |
| 770 | return { |
| 771 | "pixel_values": selected_pv, |
| 772 | "images_crop": selected_ic, |
| 773 | "images_spatial_crop": selected_sp, |
| 774 | } |
| 775 | |
| 776 | def prepare_encoder_cudagraph_capture_inputs( |
| 777 | self, |
| 778 | token_budget: int, |
| 779 | max_batch_size: int, |
| 780 | max_frames_per_batch: int, |
| 781 | device: torch.device, |
| 782 | dtype: torch.dtype, |
| 783 | path: str = "default", |
| 784 | axis_keys: tuple[Hashable, ...] | None = None, |
| 785 | ): |
| 786 | assert path in ("global", "local") |
| 787 | |
| 788 | if path == "global": |
| 789 | max_num_images = token_budget // self.global_image_output_token |
| 790 | max_batch_size = min(max_batch_size, max_num_images) |
| 791 | dummy_pixel_values = torch.randn( |
| 792 | max_batch_size, |
| 793 | 3, |
| 794 | BASE_SIZE, |
| 795 | BASE_SIZE, |
| 796 | device=device, |
| 797 | dtype=dtype, |
| 798 | ) |
| 799 | values = {"pixel_values": dummy_pixel_values} |
| 800 | else: |
| 801 | max_num_patches = token_budget // self.single_patch_output_token |
| 802 | dummy_images_crop = torch.randn( |
| 803 | max_num_patches, |
| 804 | 3, |
| 805 | IMAGE_SIZE, |
| 806 | IMAGE_SIZE, |
| 807 | device=device, |
| 808 | dtype=dtype, |
| 809 | ) |
| 810 | values = {"images_crop": dummy_images_crop} |
| 811 | |
| 812 | return EncoderCudaGraphCaptureInputs(values=values) |
| 813 | |
| 814 | def prepare_encoder_cudagraph_replay_buffers( |
| 815 | self, |
| 816 | mm_kwargs: dict[str, Any], |
| 817 | max_batch_size: int, |
| 818 | max_frames_per_batch: int, |
| 819 | path: str = "default", |
| 820 | ): |
| 821 | assert path in ("global", "local") |
| 822 | |
| 823 | if path == "global": |
| 824 | values = {"pixel_values": mm_kwargs["pixel_values"]} |
| 825 | else: |
| 826 | values = {"images_crop": mm_kwargs["images_crop"]} |
| 827 | |
| 828 | return EncoderCudaGraphReplayBuffers(values=values) |
| 829 | |
| 830 | def _batched_encoder_forward_global_path( |
| 831 | self, |
| 832 | pixel_values: torch.Tensor, |
| 833 | ) -> torch.Tensor: |
| 834 | """Encode batched global images with newline tokens inserted. |
| 835 | Output shape: ``[B * 272, n_embed]``. |
| 836 | """ |
| 837 | bsz = pixel_values.shape[0] |
| 838 | global_features_1 = self.sam_model(pixel_values) |
| 839 | global_features_2 = self.vision_model(pixel_values, global_features_1) |
| 840 | features = torch.cat( |
| 841 | ( |
| 842 | global_features_2[:, 1:], |
| 843 | global_features_1.flatten(2).permute(0, 2, 1), |
| 844 | ), |
| 845 | dim=-1, |
| 846 | ) |
| 847 | features = self.projector(features) |
| 848 | side = self.image_side |
| 849 | dim = features.shape[-1] |
| 850 | features = features.view(bsz, side, side, dim) |
| 851 | newline = self.image_newline.view(1, 1, 1, dim).expand(bsz, side, 1, dim) |
| 852 | features = torch.cat([features, newline], dim=2) |
| 853 | return features.view(-1, dim) |
| 854 | |
| 855 | def _batched_encoder_forward_local_path( |
| 856 | self, |
| 857 | images_crop: torch.Tensor, |
| 858 | ) -> torch.Tensor: |
| 859 | """Encode local patches without newline insertion (newlines are added later |
| 860 | in ``postprocess_encoder_output`` via ``_assemble_patch_grid``). |
| 861 | Output shape: ``[P * 100, n_embed]``. |
| 862 | """ |
| 863 | features_1 = self.sam_model(images_crop) |
| 864 | features_2 = self.vision_model(images_crop, features_1) |
| 865 | features = torch.cat( |
| 866 | ( |
| 867 | features_2[:, 1:], |
| 868 | features_1.flatten(2).permute(0, 2, 1), |
| 869 | ), |
| 870 | dim=-1, |
| 871 | ) |
| 872 | features = self.projector(features) |
| 873 | return features.view(-1, features.shape[-1]) |
| 874 | |
| 875 | def encoder_cudagraph_forward( |
| 876 | self, |
| 877 | values: dict[str, torch.Tensor], |
| 878 | path: str = "default", |
| 879 | ) -> torch.Tensor: |
| 880 | assert path in ("global", "local") |
| 881 | |
| 882 | if path == "global": |
| 883 | pixel_values = values["pixel_values"] |
| 884 | return self._batched_encoder_forward_global_path(pixel_values) |
| 885 | else: |
| 886 | images_crop = values["images_crop"] |
| 887 | return self._batched_encoder_forward_local_path(images_crop) |
| 888 | |
| 889 | def encoder_eager_forward( |
| 890 | self, |
| 891 | mm_kwargs: dict[str, Any], |
| 892 | path: str = "default", |
| 893 | ) -> torch.Tensor: |
| 894 | """Eager encoder forward with optional per-path execution. |
| 895 | |
| 896 | ``path="default"``: full forward (global + local + assembly). |
| 897 | ``path="global"``: global-only batched forward with newlines. |
| 898 | ``path="local"``: local-only batched forward without newlines. |
| 899 | """ |
| 900 | if path == "default": |
| 901 | # Original eager implementation: process each image one by one |
| 902 | # (with both global and local paths) and concatenate results. |
| 903 | image_input = DeepseekOCRImagePixelInputs( |
| 904 | type="pixel_values", |
| 905 | data=mm_kwargs["pixel_values"], |
| 906 | images_crop=mm_kwargs["images_crop"], |
| 907 | images_spatial_crop=mm_kwargs["images_spatial_crop"], |
| 908 | ) |
| 909 | vision_embeddings = self._process_image_input(image_input) |
| 910 | return torch.cat(vision_embeddings, dim=0) |
| 911 | |
| 912 | assert path in ("global", "local") |
| 913 | if path == "global": |
| 914 | pixel_values = mm_kwargs["pixel_values"] |
| 915 | return self._batched_encoder_forward_global_path(pixel_values) |
| 916 | else: |
| 917 | images_crop = mm_kwargs["images_crop"] |
| 918 | return self._batched_encoder_forward_local_path(images_crop) |
| 919 | |
| 920 | def postprocess_encoder_output( |
| 921 | self, |
| 922 | outputs: dict[str, torch.Tensor], |
| 923 | indices: list[int], |
| 924 | per_item_out_tokens: list[int], |
| 925 | dest: dict[int, torch.Tensor] | list[torch.Tensor | None], |
| 926 | clone: bool = False, |
| 927 | batch_mm_kwargs: dict[str, Any] | None = None, |
| 928 | ) -> None: |
| 929 | """Assemble per-image embeddings from global and local encoder outputs. |
| 930 | |
| 931 | ``output['global']`` contains global-image features with newlines already |
| 932 | inserted (from CUDA graph replay or eager fallback): |
| 933 | ``[B * 272, n_embed]``. |
| 934 | |
| 935 | ``output['local']`` contains local-patch features without |
| 936 | newlines (from CUDA graph replay or eager fallback): |
| 937 | ``[P * 100, n_embed]``. May be ``None`` if no patches in batch. |
| 938 | |
| 939 | This method: |
| 940 | 1. Splits ``output['global']`` into per-image global portions. |
| 941 | 2. Splits ``output['local']`` into per-image patch groups. |
| 942 | 3. For each image: assembles patch grid with newlines via |
| 943 | ``_assemble_patch_grid``, then concatenates |
| 944 | ``[local_tiled, global, view_seperator]``. |
| 945 | """ |
| 946 | output = outputs["global"] |
| 947 | local_output = outputs.get("local") |
| 948 | assert batch_mm_kwargs is not None |
| 949 | bsz = len(indices) |
| 950 | n_embed = output.shape[-1] |
| 951 | |
| 952 | images_spatial_crop = batch_mm_kwargs["images_spatial_crop"] |
| 953 | is_tiled = (images_spatial_crop[:, 0] > 1) | (images_spatial_crop[:, 1] > 1) |
| 954 | num_patches = [ |
| 955 | int(np) for np in torch.where(is_tiled, images_spatial_crop.prod(dim=-1), 0) |
| 956 | ] |
| 957 | total_patches = sum(num_patches) |
| 958 | |
| 959 | global_part = output[: bsz * self.global_image_output_token].reshape( |
| 960 | bsz, self.global_image_output_token, n_embed |
| 961 | ) |
| 962 | |
| 963 | # Split local output into per-patch groups. |
| 964 | local_flat = None |
| 965 | if total_patches > 0 and local_output is not None: |
| 966 | local_flat = local_output[: total_patches * self.single_patch_output_token] |
| 967 | local_flat = local_flat.reshape( |
| 968 | total_patches, self.single_patch_output_token, n_embed |
| 969 | ) |
| 970 | |
| 971 | cur_patch = 0 |
| 972 | for i, idx in enumerate(indices): |
| 973 | num_patch = num_patches[i] |
| 974 | single_image_output: list[torch.Tensor] = [] |
| 975 | |
| 976 | # 1. Process local patches: assemble tile grid, add 1 newline per row. |
| 977 | if num_patch > 0 and local_flat is not None: |
| 978 | patches = local_flat[cur_patch : cur_patch + num_patch] |
| 979 | cur_patch += num_patch |
| 980 | single_image_output.append( |
| 981 | self._assemble_patch_grid(patches, images_spatial_crop[i]) |
| 982 | ) |
| 983 | |
| 984 | # 2. Global image: newlines already inserted. |
| 985 | single_image_output.append(global_part[i]) |
| 986 | |
| 987 | # 3. Add view separator for each image. |
| 988 | single_image_output.append(self.view_seperator[None, :]) |
| 989 | |
| 990 | # 4. Save final outputs for each image. |
| 991 | dest[idx] = torch.cat(single_image_output, dim=0) |