| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # adapted from https://github.com/deepseek-ai/DeepSeek-VL2/blob/faf18023f24b962b32d9f0a2d89e402a8d383a78/deepseek_vl2/models/modeling_deepseek_vl_v2.py |
| 5 | """Inference-only Deepseek-VL2 model compatible with HuggingFace weights.""" |
| 6 | |
| 7 | import math |
| 8 | from collections.abc import Iterable, Mapping, Sequence |
| 9 | from typing import Annotated, Literal, TypeAlias |
| 10 | |
| 11 | import torch |
| 12 | import torch.nn as nn |
| 13 | import torch.nn.functional as F |
| 14 | from einops import rearrange, repeat |
| 15 | from transformers import BatchFeature |
| 16 | |
| 17 | from vllm.config import VllmConfig |
| 18 | from vllm.config.multimodal import MultiModalDummyOptions |
| 19 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 20 | from vllm.inputs import MultiModalDataDict |
| 21 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 22 | from vllm.model_executor.models.transformers.utils import replace_linear_class |
| 23 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 24 | from vllm.multimodal.inputs import ( |
| 25 | MultiModalFieldConfig, |
| 26 | MultiModalKwargsItems, |
| 27 | ) |
| 28 | from vllm.multimodal.parse import ( |
| 29 | ImageEmbeddingItems, |
| 30 | ImageProcessorItems, |
| 31 | ImageSize, |
| 32 | MultiModalDataItems, |
| 33 | ) |
| 34 | from vllm.multimodal.processing import BaseDummyInputsBuilder |
| 35 | from vllm.multimodal.processing.processor import ( |
| 36 | BaseMultiModalProcessor, |
| 37 | BaseProcessingInfo, |
| 38 | MultiModalProcessingResult, |
| 39 | ProcessorInputs, |
| 40 | PromptReplacement, |
| 41 | PromptUpdate, |
| 42 | TimingContext, |
| 43 | ) |
| 44 | from vllm.sequence import IntermediateTensors |
| 45 | from vllm.tokenizers import cached_tokenizer_from_config |
| 46 | from vllm.transformers_utils.configs.deepseek_vl2 import ( |
| 47 | DeepseekVLV2Config, |
| 48 | MlpProjectorConfig, |
| 49 | VisionEncoderConfig, |
| 50 | ) |
| 51 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 52 | from vllm.utils.torch_utils import set_default_torch_dtype |
| 53 | |
| 54 | from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP |
| 55 | from .utils import ( |
| 56 | AutoWeightsLoader, |
| 57 | WeightsMapper, |
| 58 | init_vllm_registered_model, |
| 59 | maybe_prefix, |
| 60 | ) |
| 61 | |
| 62 | # The image token id may be various |
| 63 | _IMAGE_TOKEN = "<image>" |
| 64 | |
| 65 | |
| 66 | class DeepseekVL2ImagePixelInputs(TensorSchema): |
| 67 | """Dimensions: |
| 68 | - bnp: Batch size * number of images * number of patches |
| 69 | - p: Number of patches |
| 70 | - c: Number of channels (3) |
| 71 | - h: Height of each image |
| 72 | - w: Width of each image |
| 73 | """ |
| 74 | |
| 75 | type: Literal["pixel_values"] |
| 76 | data: Annotated[torch.Tensor, TensorShape("bnp", 3, "h", "w", dynamic_dims={"bnp"})] |
| 77 | images_spatial_crop: Annotated[torch.Tensor, TensorShape("bn", 2)] |
| 78 | |
| 79 | |
| 80 | class DeepseekVL2VImageEmbeddingInputs(TensorSchema): |
| 81 | """Dimensions: |
| 82 | - bn: Batch size * number of images |
| 83 | - f: Image feature size |
| 84 | - h: Hidden size (must match language model backbone) |
| 85 | """ |
| 86 | |
| 87 | type: Literal["image_embeds"] |
| 88 | data: Annotated[torch.Tensor | list[torch.Tensor], TensorShape("bn", "f", "h")] |
| 89 | |
| 90 | |
| 91 | DeepseekVL2ImageInputs: TypeAlias = ( |
| 92 | DeepseekVL2ImagePixelInputs | DeepseekVL2VImageEmbeddingInputs |
| 93 | ) |
| 94 | |
| 95 | |
| 96 | class MlpProjector(nn.Module): |
| 97 | def __init__(self, cfg: MlpProjectorConfig): |
| 98 | super().__init__() |
| 99 | |
| 100 | self.cfg = cfg |
| 101 | self.projector_type = cfg.projector_type |
| 102 | assert not cfg.token_pooling, "Token pooling is not supported currently." |
| 103 | |
| 104 | if self.projector_type == "downsample_mlp_gelu": |
| 105 | mlp_depth = cfg.depth |
| 106 | mlp_ratio = cfg.mlp_ratio |
| 107 | modules = [ |
| 108 | nn.Linear( |
| 109 | cfg.input_dim * cfg.downsample_ratio * cfg.downsample_ratio, |
| 110 | cfg.n_embed * mlp_ratio, |
| 111 | ) |
| 112 | ] |
| 113 | for _ in range(1, mlp_depth - 1): |
| 114 | modules.append(nn.GELU()) |
| 115 | modules.append( |
| 116 | nn.Linear(cfg.n_embed * mlp_ratio, cfg.n_embed * mlp_ratio) |
| 117 | ) |
| 118 | modules.append(nn.GELU()) |
| 119 | modules.append(nn.Linear(cfg.n_embed * mlp_ratio, cfg.n_embed)) |
| 120 | modules = nn.Sequential(*modules) |
| 121 | elif self.projector_type == "linear": |
| 122 | modules = nn.Linear(cfg.input_dim, cfg.n_embed) |
| 123 | else: |
| 124 | raise NotImplementedError( |
| 125 | f"Unsupported projector type: {cfg.projector_type}" |
| 126 | ) |
| 127 | |
| 128 | self.layers = modules |
| 129 | |
| 130 | def forward(self, x): |
| 131 | bs, hw, input_dim = x.shape |
| 132 | if self.projector_type == "downsample_mlp_gelu": |
| 133 | h = w = int((hw) ** 0.5) |
| 134 | """compute padding""" |
| 135 | if h % self.cfg.downsample_ratio: |
| 136 | pad = self.cfg.downsample_ratio - h % self.cfg.downsample_ratio |
| 137 | else: |
| 138 | pad = 0 |
| 139 | x = x.reshape(bs, h, w, input_dim) |
| 140 | if pad > 0: |
| 141 | x = F.pad(x, (0, 0, 0, pad, 0, pad), "constant", 0) |
| 142 | """4 to 1 concat""" |
| 143 | x = x.permute(0, 3, 1, 2) # B, C, H, W |
| 144 | x = F.unfold( |
| 145 | x, |
| 146 | kernel_size=self.cfg.downsample_ratio, |
| 147 | stride=self.cfg.downsample_ratio, |
| 148 | padding=0, |
| 149 | ) # B, C*4, HW // 4 |
| 150 | x = x.permute(0, 2, 1) |
| 151 | |
| 152 | return self.layers(x) |
| 153 | |
| 154 | |
| 155 | class DeepseekVL2ProcessingInfo(BaseProcessingInfo): |
| 156 | def get_hf_config(self): |
| 157 | return self.ctx.get_hf_config(DeepseekVLV2Config) |
| 158 | |
| 159 | def get_hf_processor(self, **kwargs: object): |
| 160 | return self.ctx.get_hf_processor(**kwargs) |
| 161 | |
| 162 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 163 | return {"image": None} |
| 164 | |
| 165 | def get_num_image_tokens( |
| 166 | self, *, image_width: int, image_height: int, cropping: bool = True |
| 167 | ) -> int: |
| 168 | hf_processor = self.get_hf_processor() |
| 169 | image_size = hf_processor.image_size |
| 170 | patch_size = hf_processor.patch_size |
| 171 | downsample_ratio = hf_processor.downsample_ratio |
| 172 | |
| 173 | if cropping: |
| 174 | best_width, best_height = hf_processor.select_best_resolution( |
| 175 | (image_width, image_height) |
| 176 | ) |
| 177 | num_width_tiles, num_height_tiles = ( |
| 178 | best_width // image_size, |
| 179 | best_height // image_size, |
| 180 | ) |
| 181 | else: |
| 182 | num_width_tiles = num_height_tiles = 1 |
| 183 | |
| 184 | h = w = math.ceil((image_size // patch_size) / downsample_ratio) |
| 185 | |
| 186 | global_views_tokens = h * (w + 1) |
| 187 | local_views_tokens = (num_height_tiles * h) * (num_width_tiles * w + 1) |
| 188 | return global_views_tokens + local_views_tokens + 1 |
| 189 | |
| 190 | def get_image_size_with_most_features(self) -> ImageSize: |
| 191 | hf_config = self.get_hf_config() |
| 192 | candidate_resolutions = hf_config.candidate_resolutions |
| 193 | height, width = max( |
| 194 | candidate_resolutions, |
| 195 | key=lambda x: self.get_num_image_tokens( |
| 196 | image_width=x[1], image_height=x[0] |
| 197 | ), |
| 198 | ) |
| 199 | return ImageSize(width=width, height=height) |
| 200 | |
| 201 | |
| 202 | class DeepseekVL2DummyInputsBuilder(BaseDummyInputsBuilder[DeepseekVL2ProcessingInfo]): |
| 203 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 204 | num_images = mm_counts.get("image", 0) |
| 205 | |
| 206 | processor = self.info.get_hf_processor() |
| 207 | image_token = processor.image_token |
| 208 | |
| 209 | return image_token * num_images |
| 210 | |
| 211 | def get_dummy_mm_data( |
| 212 | self, |
| 213 | seq_len: int, |
| 214 | mm_counts: Mapping[str, int], |
| 215 | mm_options: MultiModalDummyOptions, |
| 216 | ) -> MultiModalDataDict: |
| 217 | max_image_size = self.info.get_image_size_with_most_features() |
| 218 | |
| 219 | return { |
| 220 | "image": self._get_dummy_images( |
| 221 | width=max_image_size.width, |
| 222 | height=max_image_size.height, |
| 223 | num_images=mm_counts.get("image", 0), |
| 224 | overrides=mm_options.get("image"), |
| 225 | ) |
| 226 | } |
| 227 | |
| 228 | |
| 229 | class DeepseekVL2MultiModalProcessor( |
| 230 | BaseMultiModalProcessor[DeepseekVL2ProcessingInfo] |
| 231 | ): |
| 232 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 233 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 234 | |
| 235 | def _postprocess_hf_mm_data( |
| 236 | self, |
| 237 | mm_data: Mapping[str, object], |
| 238 | hf_processor_mm_kwargs: Mapping[str, object], |
| 239 | processed_data: BatchFeature, |
| 240 | ) -> BatchFeature: |
| 241 | if not mm_data: |
| 242 | return processed_data |
| 243 | |
| 244 | processed_data["num_patches"] = ( |
| 245 | processed_data["images_spatial_crop"].prod(-1) + 1 |
| 246 | ) |
| 247 | |
| 248 | return processed_data |
| 249 | |
| 250 | def _get_mm_fields_config( |
| 251 | self, |
| 252 | hf_inputs: BatchFeature, |
| 253 | hf_processor_mm_kwargs: Mapping[str, object], |
| 254 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 255 | num_patches = hf_inputs.get("num_patches", torch.empty(0)) |
| 256 | |
| 257 | return dict( |
| 258 | pixel_values=MultiModalFieldConfig.flat_from_sizes("image", num_patches), |
| 259 | images_spatial_crop=MultiModalFieldConfig.batched( |
| 260 | "image", keep_on_cpu=True |
| 261 | ), |
| 262 | image_embeds=MultiModalFieldConfig.batched("image"), |
| 263 | ) |
| 264 | |
| 265 | def _get_prompt_updates( |
| 266 | self, |
| 267 | mm_items: MultiModalDataItems, |
| 268 | hf_processor_mm_kwargs: Mapping[str, object], |
| 269 | out_mm_kwargs: MultiModalKwargsItems, |
| 270 | ) -> Sequence[PromptUpdate]: |
| 271 | hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 272 | |
| 273 | image_token_id = hf_processor.image_token_id |
| 274 | assert isinstance(image_token_id, int) |
| 275 | |
| 276 | def get_replacement_deepseek_vl2(item_idx: int): |
| 277 | images = mm_items.get_items( |
| 278 | "image", (ImageEmbeddingItems, ImageProcessorItems) |
| 279 | ) |
| 280 | |
| 281 | if isinstance(images, ImageEmbeddingItems): |
| 282 | num_image_tokens = images.get_feature_size(item_idx) |
| 283 | else: |
| 284 | assert isinstance(images, ImageProcessorItems) |
| 285 | image_size = images.get_image_size(item_idx) |
| 286 | |
| 287 | num_image_tokens = self.info.get_num_image_tokens( |
| 288 | image_width=image_size.width, |
| 289 | image_height=image_size.height, |
| 290 | cropping=len(images) <= 2, |
| 291 | ) |
| 292 | return [image_token_id] * num_image_tokens |
| 293 | |
| 294 | return [ |
| 295 | PromptReplacement( |
| 296 | modality="image", |
| 297 | target=[image_token_id], |
| 298 | replacement=get_replacement_deepseek_vl2, |
| 299 | ) |
| 300 | ] |
| 301 | |
| 302 | def _cached_apply_hf_processor( |
| 303 | self, |
| 304 | inputs: ProcessorInputs, |
| 305 | timing_ctx: TimingContext, |
| 306 | ) -> MultiModalProcessingResult: |
| 307 | # The processor logic is different for len(images) <= 2 vs > 2 |
| 308 | # Since the processing cache assumes that the processor output is |
| 309 | # invariant of how many images are passed per prompt, we only |
| 310 | # perform caching for the most common case |
| 311 | if inputs.mm_data_items.get_count("image", strict=False) > 2: |
| 312 | return self._apply_hf_processor(inputs, timing_ctx) |
| 313 | |
| 314 | return super()._cached_apply_hf_processor(inputs, timing_ctx) |
| 315 | |
| 316 | |
| 317 | @MULTIMODAL_REGISTRY.register_processor( |
| 318 | DeepseekVL2MultiModalProcessor, |
| 319 | info=DeepseekVL2ProcessingInfo, |
| 320 | dummy_inputs=DeepseekVL2DummyInputsBuilder, |
| 321 | ) |
| 322 | class DeepseekVLV2ForCausalLM(nn.Module, SupportsMultiModal, SupportsPP): |
| 323 | hf_to_vllm_mapper = WeightsMapper( |
| 324 | orig_to_new_prefix={ |
| 325 | "language.": "language_model.", |
| 326 | } |
| 327 | ) |
| 328 | |
| 329 | @classmethod |
| 330 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 331 | if modality.startswith("image"): |
| 332 | return "<image>" |
| 333 | |
| 334 | raise ValueError("Only image modality is supported") |
| 335 | |
| 336 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 337 | super().__init__() |
| 338 | config: DeepseekVLV2Config = vllm_config.model_config.hf_config |
| 339 | quant_config = vllm_config.quant_config |
| 340 | multimodal_config = vllm_config.model_config.multimodal_config |
| 341 | |
| 342 | self.config = config |
| 343 | self.multimodal_config = multimodal_config |
| 344 | |
| 345 | self.vision_config = config.vision_config |
| 346 | self.projector_config = config.projector_config |
| 347 | self.text_config = config.text_config |
| 348 | |
| 349 | model_config = vllm_config.model_config |
| 350 | tokenizer = cached_tokenizer_from_config(model_config) |
| 351 | self.image_token_id: int = tokenizer.vocab[_IMAGE_TOKEN] |
| 352 | |
| 353 | with self._mark_tower_model(vllm_config, "image"): |
| 354 | self.vision = self._init_vision_module( |
| 355 | self.vision_config, quant_config, maybe_prefix(prefix, "vision") |
| 356 | ) |
| 357 | |
| 358 | self.projector = MlpProjector(self.projector_config) |
| 359 | self.tile_tag = config.tile_tag |
| 360 | self.global_view_pos = config.global_view_pos |
| 361 | |
| 362 | # special token for image token sequence format |
| 363 | embed_std = 1 / torch.sqrt( |
| 364 | torch.tensor(self.projector_config.n_embed, dtype=torch.float32) |
| 365 | ) |
| 366 | if self.tile_tag == "2D": |
| 367 | # <|view_seperator|>, <|\n|> |
| 368 | self.image_newline = nn.Parameter( |
| 369 | torch.randn(self.projector_config.n_embed) * embed_std |
| 370 | ) |
| 371 | # This is a typo in original implementation |
| 372 | self.view_seperator = nn.Parameter( |
| 373 | torch.randn(self.projector_config.n_embed) * embed_std |
| 374 | ) |
| 375 | else: |
| 376 | raise ValueError( |
| 377 | f"Only 2D tile_tag is supported currently, got: {self.tile_tag}" |
| 378 | ) |
| 379 | |
| 380 | with self._mark_language_model(vllm_config): |
| 381 | self.language_model = init_vllm_registered_model( |
| 382 | vllm_config=vllm_config, |
| 383 | hf_config=self.text_config, |
| 384 | prefix=maybe_prefix(prefix, "language"), |
| 385 | ) |
| 386 | |
| 387 | self.make_empty_intermediate_tensors = ( |
| 388 | self.language_model.make_empty_intermediate_tensors |
| 389 | ) |
| 390 | |
| 391 | def _get_parent_and_attr(self, root: torch.nn.Module, dotted_name: str): |
| 392 | """Return (parent_module, final_attr_name) for a dotted module path.""" |
| 393 | names = dotted_name.split(".") |
| 394 | parent = root |
| 395 | for n in names[:-1]: |
| 396 | parent = getattr(parent, n) |
| 397 | return parent, names[-1] |
| 398 | |
| 399 | # patch for timm ViT instance to support tensor parallel |
| 400 | def patch_vit_for_tp( |
| 401 | self, vit: torch.nn.Module, quant_config: QuantizationConfig | None |
| 402 | ): |
| 403 | try: |
| 404 | import timm |
| 405 | except ImportError as e: |
| 406 | raise ImportError("Please install timm") from e |
| 407 | |
| 408 | for name, module in vit.named_modules(): |
| 409 | if isinstance(module, nn.Linear): |
| 410 | parent, attr_name = self._get_parent_and_attr(vit, name) |
| 411 | if isinstance(parent, timm.layers.Mlp) and attr_name == "fc1": |
| 412 | new_linear = replace_linear_class( |
| 413 | module, "colwise", quant_config, prefix=name |
| 414 | ) |
| 415 | setattr(parent, attr_name, new_linear) |
| 416 | elif isinstance(parent, timm.layers.Mlp) and attr_name == "fc2": |
| 417 | new_linear = replace_linear_class( |
| 418 | module, "rowwise", quant_config, prefix=name |
| 419 | ) |
| 420 | setattr(parent, attr_name, new_linear) |
| 421 | |
| 422 | return vit |
| 423 | |
| 424 | def _init_vision_module( |
| 425 | self, |
| 426 | vision_config: VisionEncoderConfig, |
| 427 | quant_config: QuantizationConfig | None, |
| 428 | prefix: str = "", |
| 429 | ) -> nn.Module: |
| 430 | # TODO: refactor vision model through timm wrapper from transformers |
| 431 | try: |
| 432 | import timm |
| 433 | except ImportError as e: |
| 434 | raise ImportError("Please install timm") from e |
| 435 | |
| 436 | with set_default_torch_dtype(torch.float16): |
| 437 | model = timm.create_model( |
| 438 | "vit_so400m_patch14_siglip_384.webli", |
| 439 | pretrained=False, |
| 440 | num_classes=0, |
| 441 | dynamic_img_size=True, |
| 442 | dynamic_img_pad=True, |
| 443 | ) |
| 444 | |
| 445 | if get_tensor_model_parallel_world_size() > 1: |
| 446 | model = self.patch_vit_for_tp(model, quant_config) |
| 447 | |
| 448 | model = model.to(dtype=torch.get_default_dtype()) |
| 449 | return model |
| 450 | |
| 451 | def _parse_and_validate_image_input( |
| 452 | self, **kwargs: object |
| 453 | ) -> DeepseekVL2ImageInputs | None: |
| 454 | pixel_values = kwargs.pop("pixel_values", None) |
| 455 | images_spatial_crop = kwargs.pop("images_spatial_crop", None) |
| 456 | image_embeds = kwargs.pop("image_embeds", None) |
| 457 | |
| 458 | if pixel_values is None and image_embeds is None: |
| 459 | return None |
| 460 | |
| 461 | if pixel_values is not None: |
| 462 | expected_h = expected_w = self.vision_config.image_size |
| 463 | return DeepseekVL2ImagePixelInputs( |
| 464 | type="pixel_values", |
| 465 | data=pixel_values, |
| 466 | images_spatial_crop=images_spatial_crop, |
| 467 | resolve_bindings={ |
| 468 | "h": expected_h, |
| 469 | "w": expected_w, |
| 470 | }, |
| 471 | ) |
| 472 | |
| 473 | if image_embeds is not None: |
| 474 | return DeepseekVL2VImageEmbeddingInputs( |
| 475 | type="image_embeds", |
| 476 | data=image_embeds, |
| 477 | ) |
| 478 | |
| 479 | raise AssertionError("This line should be unreachable.") |
| 480 | |
| 481 | def _pixel_values_to_embedding( |
| 482 | self, |
| 483 | pixel_values: torch.Tensor, |
| 484 | images_spatial_crop: torch.Tensor, |
| 485 | ) -> list[torch.Tensor]: |
| 486 | # [batch_all_tiles, vit_seq_len, c] |
| 487 | images_feature = self.vision.forward_features(pixel_values) |
| 488 | |
| 489 | # [batch_all_tiles, hw, D] |
| 490 | images_embeds = self.projector(images_feature) |
| 491 | |
| 492 | _, hw, n_dim = images_embeds.shape |
| 493 | h = w = int(hw**0.5) |
| 494 | |
| 495 | # fill image token based on self.tile_tag & self.global_view_pos |
| 496 | tile_index = 0 |
| 497 | vision_embeddings = [] |
| 498 | for jdx in range(images_spatial_crop.size(0)): |
| 499 | # extra global & local features |
| 500 | num_width_tiles, num_height_tiles = images_spatial_crop[jdx] |
| 501 | if num_width_tiles == 0 or num_height_tiles == 0: |
| 502 | break |
| 503 | num_tiles_in_image = num_width_tiles * num_height_tiles |
| 504 | |
| 505 | # [hw, D] |
| 506 | global_features = images_embeds[tile_index] |
| 507 | |
| 508 | # [num_height_tiles * num_width_tiles, hw, D] |
| 509 | local_features = images_embeds[ |
| 510 | tile_index + 1 : tile_index + 1 + num_tiles_in_image |
| 511 | ] |
| 512 | tile_index += num_tiles_in_image + 1 |
| 513 | |
| 514 | # format global and local features |
| 515 | # ----------------- global view add newline ----------------- |
| 516 | # [hw, D] -> [h, w, D] |
| 517 | global_features = global_features.view(h, w, n_dim) |
| 518 | |
| 519 | # [D] -> [h, 1, D] |
| 520 | new_lines_in_global = repeat(self.image_newline, "d -> h 1 d", h=h) |
| 521 | |
| 522 | # cat([h, w, D], [h, 1, D], dim=1) -> [h, w + 1, D] |
| 523 | global_features = torch.cat([global_features, new_lines_in_global], dim=1) |
| 524 | |
| 525 | # [h, w + 1, D] -> [h * (w + 1), D] |
| 526 | global_features = global_features.view(-1, n_dim) |
| 527 | |
| 528 | # ----------------- local view add newline ----------------- |
| 529 | # [num_height_tiles * num_width_tiles, h * w, D] -> |
| 530 | # [num_height_tiles * h, num_width_tiles * w, D] |
| 531 | local_features = rearrange( |
| 532 | local_features, |
| 533 | "(th tw) (h w) d -> (th h) (tw w) d", |
| 534 | th=num_height_tiles, |
| 535 | tw=num_width_tiles, |
| 536 | h=h, |
| 537 | w=w, |
| 538 | ) |
| 539 | |
| 540 | # [D] -> [num_height_tiles * h, 1, D] |
| 541 | new_lines_in_local = repeat( |
| 542 | self.image_newline, "d -> (th h) 1 d", th=num_height_tiles, h=h |
| 543 | ) |
| 544 | |
| 545 | # [num_height_tiles * h, num_width_tiles * w + 1, D] |
| 546 | local_features = torch.cat([local_features, new_lines_in_local], dim=1) |
| 547 | |
| 548 | # [num_height_tiles * h, num_width_tiles * w + 1, D] |
| 549 | # --> [(num_height_tiles * h) * (num_width_tiles * w + 1), D] |
| 550 | local_features = local_features.view(-1, n_dim) |
| 551 | |
| 552 | # merge global and local tiles |
| 553 | if self.global_view_pos == "head": |
| 554 | global_local_features = torch.cat( |
| 555 | [ |
| 556 | global_features, |
| 557 | self.view_seperator[None, :], |
| 558 | local_features, |
| 559 | ] |
| 560 | ) |
| 561 | else: |
| 562 | global_local_features = torch.cat( |
| 563 | [ |
| 564 | local_features, |
| 565 | self.view_seperator[None, :], |
| 566 | global_features, |
| 567 | ] |
| 568 | ) |
| 569 | |
| 570 | vision_embeddings.append(global_local_features) |
| 571 | return vision_embeddings |
| 572 | |
| 573 | def _process_image_input( |
| 574 | self, image_input: DeepseekVL2ImageInputs |
| 575 | ) -> torch.Tensor | list[torch.Tensor]: |
| 576 | if image_input["type"] == "image_embeds": |
| 577 | return image_input["data"] |
| 578 | |
| 579 | pixel_values = image_input["data"] |
| 580 | images_spatial_crop = image_input["images_spatial_crop"] |
| 581 | |
| 582 | return self._pixel_values_to_embedding( |
| 583 | pixel_values=pixel_values, images_spatial_crop=images_spatial_crop |
| 584 | ) |
| 585 | |
| 586 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 587 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 588 | if image_input is None: |
| 589 | return [] |
| 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 |