| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # adapted from https://huggingface.co/OpenGVLab/InternVL2-4B/blob/main/modeling_internvl_chat.py |
| 5 | # -------------------------------------------------------- |
| 6 | # InternVL |
| 7 | # Copyright (c) 2023 OpenGVLab |
| 8 | # Licensed under The MIT License [see LICENSE for details] |
| 9 | # -------------------------------------------------------- |
| 10 | from abc import abstractmethod |
| 11 | from collections.abc import Hashable, Iterable, Mapping, Sequence |
| 12 | from functools import cached_property |
| 13 | from typing import Annotated, Any, Literal, TypeAlias, TypedDict, TypeVar |
| 14 | |
| 15 | import torch |
| 16 | import torch.nn as nn |
| 17 | from transformers import BatchFeature, PreTrainedConfig |
| 18 | |
| 19 | from vllm.config import VllmConfig |
| 20 | from vllm.config.multimodal import MultiModalDummyOptions |
| 21 | from vllm.inputs import MultiModalDataDict |
| 22 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 23 | from vllm.model_executor.layers.quantization.auto_awq import AutoAWQConfig |
| 24 | from vllm.model_executor.models.intern_vit import ( |
| 25 | InternVisionModel, |
| 26 | ) |
| 27 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 28 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 29 | from vllm.multimodal.inputs import ( |
| 30 | BatchedTensorInputs, |
| 31 | MultiModalFieldConfig, |
| 32 | MultiModalKwargsItem, |
| 33 | MultiModalKwargsItems, |
| 34 | ) |
| 35 | from vllm.multimodal.parse import ( |
| 36 | ImageEmbeddingItems, |
| 37 | ImageProcessorItems, |
| 38 | ImageSize, |
| 39 | MultiModalDataItems, |
| 40 | ) |
| 41 | from vllm.multimodal.processing import ( |
| 42 | BaseDummyInputsBuilder, |
| 43 | BaseMultiModalProcessor, |
| 44 | BaseProcessingInfo, |
| 45 | PromptReplacement, |
| 46 | PromptUpdate, |
| 47 | cached_encode, |
| 48 | ) |
| 49 | from vllm.sequence import IntermediateTensors |
| 50 | from vllm.transformers_utils.processors.internvl import ( |
| 51 | InternVLImageProcessor, |
| 52 | InternVLProcessor, |
| 53 | InternVLVideoProcessor, |
| 54 | ) |
| 55 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 56 | |
| 57 | from .interfaces import ( |
| 58 | MultiModalEmbeddings, |
| 59 | SupportsEncoderCudaGraph, |
| 60 | SupportsLoRA, |
| 61 | SupportsMultiModal, |
| 62 | SupportsPP, |
| 63 | ) |
| 64 | from .utils import ( |
| 65 | AutoWeightsLoader, |
| 66 | WeightsMapper, |
| 67 | init_vllm_registered_model, |
| 68 | maybe_prefix, |
| 69 | ) |
| 70 | |
| 71 | |
| 72 | class InternVLImagePixelInputs(TensorSchema): |
| 73 | """Dimensions: |
| 74 | - bn: Batch size * number of images |
| 75 | - bnp: Batch size * number of images * (1 + num_patches) |
| 76 | - c: Number of channels (3) |
| 77 | - h: Height of each image patch |
| 78 | - w: Width of each image patch |
| 79 | """ |
| 80 | |
| 81 | type: Literal["pixel_values"] |
| 82 | pixel_values_flat: Annotated[torch.Tensor, TensorShape("bnp", 3, "h", "w")] |
| 83 | num_patches: Annotated[torch.Tensor, TensorShape("bn")] |
| 84 | |
| 85 | |
| 86 | class InternVLImageEmbeddingInputs(TensorSchema): |
| 87 | """Dimensions: |
| 88 | - n: Number of images |
| 89 | - f: Total image feature size |
| 90 | - h: Hidden size (must match the hidden size of language model backbone) |
| 91 | """ |
| 92 | |
| 93 | type: Literal["image_embeds"] |
| 94 | data: Annotated[torch.Tensor | list[torch.Tensor], TensorShape("n", "f", "h")] |
| 95 | |
| 96 | |
| 97 | InternVLImageInputs: TypeAlias = InternVLImagePixelInputs | InternVLImageEmbeddingInputs |
| 98 | |
| 99 | |
| 100 | class InternVLVideoPixelInputs(TensorSchema): |
| 101 | """Dimensions: |
| 102 | - bvf: Batch size * number of videos * num_frames |
| 103 | - bn: Batch size * number of images |
| 104 | - c: Number of channels (3) |
| 105 | - h: Height of each video frame |
| 106 | - w: Width of each video frame |
| 107 | """ |
| 108 | |
| 109 | type: Literal["pixel_values_videos"] |
| 110 | pixel_values_flat: Annotated[torch.Tensor, TensorShape("bvf", 3, "h", "w")] |
| 111 | num_patches: Annotated[torch.Tensor, TensorShape("bn")] |
| 112 | |
| 113 | |
| 114 | class InternVLVideoEmbeddingInputs(TensorSchema): |
| 115 | """Dimensions: |
| 116 | - n: Number of videos |
| 117 | - f: Total video feature size |
| 118 | - h: Hidden size (must match the hidden size of language model backbone) |
| 119 | """ |
| 120 | |
| 121 | type: Literal["video_embeds"] |
| 122 | data: Annotated[torch.Tensor | list[torch.Tensor], TensorShape("n", "f", "h")] |
| 123 | |
| 124 | |
| 125 | InternVLVideoInputs: TypeAlias = InternVLVideoPixelInputs | InternVLVideoEmbeddingInputs |
| 126 | |
| 127 | |
| 128 | class InternVLMultiModalInputs(TypedDict, total=False): |
| 129 | images: InternVLImageInputs | None |
| 130 | videos: InternVLVideoInputs | None |
| 131 | |
| 132 | |
| 133 | class BaseInternVLProcessingInfo(BaseProcessingInfo): |
| 134 | """Basic image-only ProcessingInfo for InternVL-style models.""" |
| 135 | |
| 136 | @abstractmethod |
| 137 | def get_hf_processor(self, **kwargs: object) -> InternVLProcessor: |
| 138 | raise NotImplementedError |
| 139 | |
| 140 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 141 | return {"image": None} |
| 142 | |
| 143 | def get_num_image_tokens( |
| 144 | self, |
| 145 | *, |
| 146 | image_width: int, |
| 147 | image_height: int, |
| 148 | processor: InternVLProcessor, |
| 149 | ) -> int: |
| 150 | return processor.get_num_image_tokens( |
| 151 | image_width=image_width, |
| 152 | image_height=image_height, |
| 153 | ) |
| 154 | |
| 155 | def get_image_size_with_most_features(self) -> ImageSize: |
| 156 | processor = self.get_hf_processor() |
| 157 | image_processor = processor.image_processor |
| 158 | |
| 159 | base_size = image_processor.image_size |
| 160 | target_ratios = processor.resolve_target_ratios() |
| 161 | |
| 162 | largest_feature_size, largest_feature_pinpoint = 0, None |
| 163 | for wr, hr in target_ratios: |
| 164 | width, height = base_size * wr, base_size * hr |
| 165 | |
| 166 | feat_size = self.get_num_image_tokens( |
| 167 | image_width=width, |
| 168 | image_height=height, |
| 169 | processor=processor, |
| 170 | ) |
| 171 | if feat_size > largest_feature_size: |
| 172 | largest_feature_size = feat_size |
| 173 | largest_feature_pinpoint = ImageSize(width=width, height=height) |
| 174 | |
| 175 | if largest_feature_size == 0 or largest_feature_pinpoint is None: |
| 176 | raise ValueError("Cannot have a largest feature size of 0!") |
| 177 | |
| 178 | return largest_feature_pinpoint |
| 179 | |
| 180 | def get_max_image_tokens(self) -> int: |
| 181 | processor = self.get_hf_processor() |
| 182 | target_width, target_height = self.get_image_size_with_most_features() |
| 183 | |
| 184 | return self.get_num_image_tokens( |
| 185 | image_width=target_width, |
| 186 | image_height=target_height, |
| 187 | processor=processor, |
| 188 | ) |
| 189 | |
| 190 | |
| 191 | _I = TypeVar("_I", bound=BaseInternVLProcessingInfo) |
| 192 | |
| 193 | |
| 194 | class BaseInternVLDummyInputsBuilder(BaseDummyInputsBuilder[_I]): |
| 195 | """Basic image-only DummyInputsBuilder for InternVL-style models.""" |
| 196 | |
| 197 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 198 | num_images = mm_counts.get("image", 0) |
| 199 | |
| 200 | return "<image>" * num_images |
| 201 | |
| 202 | def get_dummy_mm_data( |
| 203 | self, |
| 204 | seq_len: int, |
| 205 | mm_counts: Mapping[str, int], |
| 206 | mm_options: MultiModalDummyOptions, |
| 207 | ) -> MultiModalDataDict: |
| 208 | target_width, target_height = self.info.get_image_size_with_most_features() |
| 209 | |
| 210 | return { |
| 211 | "image": self._get_dummy_images( |
| 212 | width=target_width, |
| 213 | height=target_height, |
| 214 | num_images=mm_counts.get("image", 0), |
| 215 | overrides=mm_options.get("image"), |
| 216 | ) |
| 217 | } |
| 218 | |
| 219 | |
| 220 | class BaseInternVLMultiModalProcessor(BaseMultiModalProcessor[_I]): |
| 221 | """Basic image-only MultiModalProcessor for InternVL-style models.""" |
| 222 | |
| 223 | def _postprocess_hf_mm_data( |
| 224 | self, |
| 225 | mm_data: Mapping[str, object], |
| 226 | hf_processor_mm_kwargs: Mapping[str, object], |
| 227 | processed_data: BatchFeature, |
| 228 | ) -> BatchFeature: |
| 229 | hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 230 | image_token_id = hf_processor.ctx_image_token_id |
| 231 | |
| 232 | # Since there may be extra tokens in the feature placeholders, |
| 233 | # we need to pass the image token ID to the model to select the |
| 234 | # tokens to merge from the vision encoder outputs |
| 235 | processed_data["image_token_id"] = torch.tensor(image_token_id) |
| 236 | |
| 237 | return processed_data |
| 238 | |
| 239 | def _get_image_fields_config(self, hf_inputs: BatchFeature): |
| 240 | image_num_patches = hf_inputs.get("image_num_patches", torch.empty(0)) |
| 241 | num_images = len(image_num_patches) |
| 242 | |
| 243 | return dict( |
| 244 | pixel_values_flat=MultiModalFieldConfig.flat_from_sizes( |
| 245 | "image", image_num_patches |
| 246 | ), |
| 247 | image_num_patches=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 248 | image_embeds=MultiModalFieldConfig.batched("image"), |
| 249 | image_token_id=MultiModalFieldConfig.shared( |
| 250 | "image", num_images, keep_on_cpu=True |
| 251 | ), |
| 252 | ) |
| 253 | |
| 254 | def _get_mm_fields_config( |
| 255 | self, |
| 256 | hf_inputs: BatchFeature, |
| 257 | hf_processor_mm_kwargs: Mapping[str, object], |
| 258 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 259 | return self._get_image_fields_config(hf_inputs) |
| 260 | |
| 261 | def _get_prompt_repl_image( |
| 262 | self, |
| 263 | mm_items: MultiModalDataItems, |
| 264 | hf_processor: InternVLProcessor, |
| 265 | out_mm_data: BatchedTensorInputs, |
| 266 | ): |
| 267 | tokenizer = self.info.get_tokenizer() |
| 268 | |
| 269 | if "image_num_patches" in out_mm_data: |
| 270 | image_num_patches = out_mm_data["image_num_patches"] |
| 271 | assert isinstance(image_num_patches, torch.Tensor) |
| 272 | image_num_patches = image_num_patches.tolist() |
| 273 | elif "image_embeds" in out_mm_data: |
| 274 | # TODO: Use image size information in dictionary embedding inputs |
| 275 | # to compute num_patches (similar to Qwen2-VL) |
| 276 | image_num_patches = [None] * len(out_mm_data["image_embeds"]) |
| 277 | else: |
| 278 | image_num_patches = [] |
| 279 | |
| 280 | def get_replacement_internvl(item_idx: int): |
| 281 | images = mm_items.get_items( |
| 282 | "image", (ImageEmbeddingItems, ImageProcessorItems) |
| 283 | ) |
| 284 | |
| 285 | if isinstance(images, ImageEmbeddingItems): |
| 286 | feature_size = images.get_feature_size(item_idx) |
| 287 | else: |
| 288 | assert isinstance(images, ImageProcessorItems) |
| 289 | image_size = images.get_image_size(item_idx) |
| 290 | feature_size = self.info.get_num_image_tokens( |
| 291 | image_width=image_size.width, |
| 292 | image_height=image_size.height, |
| 293 | processor=hf_processor, |
| 294 | ) |
| 295 | |
| 296 | num_patches = image_num_patches[item_idx] |
| 297 | if num_patches is not None: |
| 298 | assert isinstance(num_patches, int) |
| 299 | |
| 300 | return hf_processor.get_image_repl(num_patches, num_features=feature_size) |
| 301 | |
| 302 | return PromptReplacement( |
| 303 | modality="image", |
| 304 | target=cached_encode(tokenizer, "<image>", add_special_tokens=False), |
| 305 | replacement=get_replacement_internvl, |
| 306 | ) |
| 307 | |
| 308 | def _get_prompt_updates( |
| 309 | self, |
| 310 | mm_items: MultiModalDataItems, |
| 311 | hf_processor_mm_kwargs: Mapping[str, object], |
| 312 | out_mm_kwargs: MultiModalKwargsItems, |
| 313 | ) -> Sequence[PromptUpdate]: |
| 314 | hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 315 | out_mm_data = out_mm_kwargs.get_data() |
| 316 | |
| 317 | return [ |
| 318 | self._get_prompt_repl_image(mm_items, hf_processor, out_mm_data), |
| 319 | ] |
| 320 | |
| 321 | |
| 322 | class InternVLProcessingInfo(BaseInternVLProcessingInfo): |
| 323 | """InternVL ProcessingInfo extended for video processing.""" |
| 324 | |
| 325 | def get_image_processor(self, **kwargs): |
| 326 | config = self.get_hf_config() |
| 327 | vision_config = config.vision_config |
| 328 | |
| 329 | kwargs = self.ctx.get_merged_mm_kwargs(kwargs) |
| 330 | kwargs.setdefault("image_size", vision_config.image_size) |
| 331 | kwargs.setdefault("min_dynamic_patch", config.min_dynamic_patch) |
| 332 | kwargs.setdefault("max_dynamic_patch", config.max_dynamic_patch) |
| 333 | kwargs.setdefault("dynamic_image_size", config.dynamic_image_size) |
| 334 | kwargs.setdefault("use_thumbnail", config.use_thumbnail) |
| 335 | |
| 336 | return InternVLImageProcessor(**kwargs) |
| 337 | |
| 338 | def get_video_processor(self, **kwargs): |
| 339 | config = self.get_hf_config() |
| 340 | vision_config = config.vision_config |
| 341 | |
| 342 | kwargs = self.ctx.get_merged_mm_kwargs(kwargs) |
| 343 | kwargs.setdefault("image_size", vision_config.image_size) |
| 344 | |
| 345 | return InternVLVideoProcessor(**kwargs) |
| 346 | |
| 347 | @cached_property |
| 348 | def ctx_video_token(self): |
| 349 | text_model_type = self.get_hf_config().get_text_config().model_type |
| 350 | ctx_video_token_map = { |
| 351 | "qwen2": "<|video_pad|>", |
| 352 | "qwen3": "<|video_pad|>", |
| 353 | "qwen3_moe": "<|video_pad|>", |
| 354 | "gpt_oss": "<|reserved_200000|>", |
| 355 | } |
| 356 | |
| 357 | if text_model_type not in ctx_video_token_map: |
| 358 | return None |
| 359 | |
| 360 | ctx_video_token = ctx_video_token_map[text_model_type] |
| 361 | if ctx_video_token not in self.get_tokenizer().get_vocab(): |
| 362 | return None |
| 363 | |
| 364 | return ctx_video_token |
| 365 | |
| 366 | def get_hf_processor(self, **kwargs: object) -> InternVLProcessor: |
| 367 | config = self.get_hf_config() |
| 368 | vision_config = config.vision_config |
| 369 | |
| 370 | image_processor = self.get_image_processor(**kwargs) |
| 371 | image_size = image_processor.image_size |
| 372 | patch_size = vision_config.patch_size |
| 373 | downsample_ratio = config.downsample_ratio |
| 374 | image_seq_length = int((image_size // patch_size) ** 2 * (downsample_ratio**2)) |
| 375 | |
| 376 | ctx_video_token = self.ctx_video_token |
| 377 | video_processor = ( |
| 378 | self.get_video_processor(**kwargs) if ctx_video_token else None |
| 379 | ) |
| 380 | |
| 381 | return InternVLProcessor( |
| 382 | tokenizer=self.get_tokenizer(), |
| 383 | image_processor=image_processor, |
| 384 | video_processor=video_processor, |
| 385 | image_seq_length=image_seq_length, |
| 386 | ctx_video_token=ctx_video_token, |
| 387 | ) |
| 388 | |
| 389 | def get_supported_mm_limits(self): |
| 390 | video_limit = {"video": None} if self.ctx_video_token else {} |
| 391 | return {**super().get_supported_mm_limits(), **video_limit} |
| 392 | |
| 393 | def get_num_frames_with_most_features( |
| 394 | self, |
| 395 | seq_len: int, |
| 396 | mm_counts: Mapping[str, int], |
| 397 | ) -> int: |
| 398 | max_images = mm_counts.get("image", 0) |
| 399 | max_videos = mm_counts.get("video", 0) |
| 400 | |
| 401 | processor = self.get_hf_processor() |
| 402 | num_image_token = processor.image_seq_length |
| 403 | |
| 404 | max_image_tokens = self.get_max_image_tokens() * max_images |
| 405 | max_total_frames = (seq_len - max_image_tokens) // num_image_token |
| 406 | max_frames_per_video = max_total_frames // max(max_videos, 1) |
| 407 | |
| 408 | return max(max_frames_per_video, 1) |
| 409 | |
| 410 | |
| 411 | class InternVLDummyInputsBuilder( |
| 412 | BaseInternVLDummyInputsBuilder[InternVLProcessingInfo] |
| 413 | ): |
| 414 | """InternVL DummyInputsBuilder extended for video support.""" |
| 415 | |
| 416 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 417 | num_videos = mm_counts.get("video", 0) |
| 418 | |
| 419 | return super().get_dummy_text(mm_counts) + "<video>" * num_videos |
| 420 | |
| 421 | def get_dummy_mm_data( |
| 422 | self, |
| 423 | seq_len: int, |
| 424 | mm_counts: Mapping[str, int], |
| 425 | mm_options: MultiModalDummyOptions, |
| 426 | ) -> MultiModalDataDict: |
| 427 | dummy_image = super().get_dummy_mm_data(seq_len, mm_counts, mm_options) |
| 428 | if self.info.ctx_video_token: |
| 429 | config = self.info.get_hf_config() |
| 430 | image_size: int = config.vision_config.image_size |
| 431 | target_num_frames = self.info.get_num_frames_with_most_features( |
| 432 | seq_len, mm_counts |
| 433 | ) |
| 434 | num_videos = mm_counts.get("video", 0) |
| 435 | video_overrides = mm_options.get("video") |
| 436 | dummy_video = { |
| 437 | "video": self._get_dummy_videos( |
| 438 | width=image_size, |
| 439 | height=image_size, |
| 440 | num_frames=target_num_frames, |
| 441 | num_videos=num_videos, |
| 442 | overrides=video_overrides, |
| 443 | ) |
| 444 | } |
| 445 | else: |
| 446 | dummy_video = {} |
| 447 | return {**dummy_image, **dummy_video} |
| 448 | |
| 449 | |
| 450 | class InternVLMultiModalProcessor( |
| 451 | BaseInternVLMultiModalProcessor[InternVLProcessingInfo] |
| 452 | ): |
| 453 | """InternVL MultiModalProcessor extended for video support.""" |
| 454 | |
| 455 | def _postprocess_hf_mm_data( |
| 456 | self, |
| 457 | mm_data: Mapping[str, object], |
| 458 | hf_processor_mm_kwargs: Mapping[str, object], |
| 459 | processed_data: BatchFeature, |
| 460 | ) -> BatchFeature: |
| 461 | processed_data = super()._postprocess_hf_mm_data( |
| 462 | mm_data, |
| 463 | hf_processor_mm_kwargs, |
| 464 | processed_data, |
| 465 | ) |
| 466 | |
| 467 | hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 468 | if (video_token_id := hf_processor.ctx_video_token_id) is not None: |
| 469 | processed_data["video_token_id"] = torch.tensor(video_token_id) |
| 470 | |
| 471 | return processed_data |
| 472 | |
| 473 | def _get_video_fields_config(self, hf_inputs: BatchFeature): |
| 474 | video_num_patches = hf_inputs.get("video_num_patches", torch.empty(0)) |
| 475 | num_videos = len(video_num_patches) |
| 476 | |
| 477 | return dict( |
| 478 | pixel_values_flat_video=MultiModalFieldConfig.flat_from_sizes( |
| 479 | "video", video_num_patches |
| 480 | ), |
| 481 | video_num_patches=MultiModalFieldConfig.batched("video", keep_on_cpu=True), |
| 482 | video_token_id=MultiModalFieldConfig.shared( |
| 483 | "video", num_videos, keep_on_cpu=True |
| 484 | ), |
| 485 | ) |
| 486 | |
| 487 | def _get_mm_fields_config( |
| 488 | self, |
| 489 | hf_inputs: BatchFeature, |
| 490 | hf_processor_mm_kwargs: Mapping[str, object], |
| 491 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 492 | fields = self._get_image_fields_config(hf_inputs) |
| 493 | if self.info.ctx_video_token: |
| 494 | fields |= self._get_video_fields_config(hf_inputs) |
| 495 | |
| 496 | return fields |
| 497 | |
| 498 | def _get_prompt_repl_video( |
| 499 | self, |
| 500 | mm_items: MultiModalDataItems, |
| 501 | hf_processor: InternVLProcessor, |
| 502 | out_mm_data: BatchedTensorInputs, |
| 503 | ): |
| 504 | tokenizer = self.info.get_tokenizer() |
| 505 | |
| 506 | if "video_num_patches" in out_mm_data: |
| 507 | video_num_patches_tensor = out_mm_data["video_num_patches"] |
| 508 | assert isinstance(video_num_patches_tensor, torch.Tensor) |
| 509 | video_num_patches: list[int] = video_num_patches_tensor.tolist() |
| 510 | else: |
| 511 | video_num_patches = [] |
| 512 | |
| 513 | def get_video_replacement_internvl(item_idx: int): |
| 514 | num_patches = video_num_patches[item_idx] |
| 515 | if num_patches is not None: |
| 516 | assert isinstance(num_patches, int) |
| 517 | |
| 518 | return hf_processor.get_video_repl(num_patches) |
| 519 | |
| 520 | return PromptReplacement( |
| 521 | modality="video", |
| 522 | target=cached_encode(tokenizer, "<video>", add_special_tokens=False), |
| 523 | replacement=get_video_replacement_internvl, |
| 524 | ) |
| 525 | |
| 526 | def _get_prompt_updates( |
| 527 | self, |
| 528 | mm_items: MultiModalDataItems, |
| 529 | hf_processor_mm_kwargs: Mapping[str, object], |
| 530 | out_mm_kwargs: MultiModalKwargsItems, |
| 531 | ) -> Sequence[PromptUpdate]: |
| 532 | hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 533 | out_mm_data = out_mm_kwargs.get_data() |
| 534 | |
| 535 | prompt_repls = [ |
| 536 | self._get_prompt_repl_image(mm_items, hf_processor, out_mm_data), |
| 537 | ] |
| 538 | if self.info.ctx_video_token is not None: |
| 539 | prompt_repls.append( |
| 540 | self._get_prompt_repl_video(mm_items, hf_processor, out_mm_data) |
| 541 | ) |
| 542 | |
| 543 | return prompt_repls |
| 544 | |
| 545 | |
| 546 | @MULTIMODAL_REGISTRY.register_processor( |
| 547 | InternVLMultiModalProcessor, |
| 548 | info=InternVLProcessingInfo, |
| 549 | dummy_inputs=InternVLDummyInputsBuilder, |
| 550 | ) |
| 551 | class InternVLChatModel( |
| 552 | nn.Module, |
| 553 | SupportsMultiModal, |
| 554 | SupportsPP, |
| 555 | SupportsLoRA, |
| 556 | SupportsEncoderCudaGraph, |
| 557 | ): |
| 558 | supports_encoder_tp_data = True |
| 559 | supports_tower_connector_lora = True |
| 560 | |
| 561 | hf_to_vllm_mapper = WeightsMapper( |
| 562 | orig_to_new_prefix=dict.fromkeys( |
| 563 | [ |
| 564 | "action_embed", |
| 565 | "temporal_embed", |
| 566 | "track_embed", |
| 567 | "track_embed_decoder", |
| 568 | "box_token", |
| 569 | "cg_criterion", |
| 570 | "cg_model", |
| 571 | "loc_encoder", |
| 572 | "loc_decoder", |
| 573 | "sam", |
| 574 | "temporal_token", |
| 575 | "track_token", |
| 576 | ] |
| 577 | ) |
| 578 | ) |
| 579 | |
| 580 | @classmethod |
| 581 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 582 | if modality.startswith("image"): |
| 583 | return "<image>" |
| 584 | if modality.startswith("video"): |
| 585 | return "<video>" |
| 586 | |
| 587 | raise ValueError("Only image or video modality is supported") |
| 588 | |
| 589 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 590 | super().__init__() |
| 591 | |
| 592 | config = vllm_config.model_config.hf_config |
| 593 | quant_config = vllm_config.quant_config |
| 594 | multimodal_config = vllm_config.model_config.get_multimodal_config() |
| 595 | |
| 596 | self.config = config |
| 597 | self.multimodal_config = multimodal_config |
| 598 | self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data" |
| 599 | self._patch_quant_config(config, quant_config) |
| 600 | |
| 601 | image_size = config.force_image_size or config.vision_config.image_size |
| 602 | patch_size = config.vision_config.patch_size |
| 603 | self.patch_size = patch_size |
| 604 | self.patch_tokens = (image_size // patch_size) ** 2 |
| 605 | self.num_image_token = int(self.patch_tokens * (config.downsample_ratio**2)) |
| 606 | self.downsample_ratio = config.downsample_ratio |
| 607 | self.ps_version = config.ps_version |
| 608 | |
| 609 | with self._mark_tower_model(vllm_config, {"image", "video"}): |
| 610 | self.vision_model = self._init_vision_model( |
| 611 | config, |
| 612 | quant_config=quant_config, |
| 613 | prefix=maybe_prefix(prefix, "vision_model"), |
| 614 | ) |
| 615 | self.mlp1 = self._init_mlp1(config) |
| 616 | |
| 617 | with self._mark_language_model(vllm_config): |
| 618 | self.language_model = init_vllm_registered_model( |
| 619 | vllm_config=vllm_config, |
| 620 | hf_config=config.text_config, |
| 621 | prefix=maybe_prefix(prefix, "language_model"), |
| 622 | ) |
| 623 | |
| 624 | self.img_context_token_id: int | None = None |
| 625 | self.video_context_token_id: int | None = None |
| 626 | |
| 627 | self.make_empty_intermediate_tensors = ( |
| 628 | self.language_model.make_empty_intermediate_tensors |
| 629 | ) |
| 630 | |
| 631 | def _patch_quant_config( |
| 632 | self, config: PreTrainedConfig, quant_config: QuantizationConfig | None |
| 633 | ): |
| 634 | # the awq models from OpenGVLab missing `modules_to_not_convert` |
| 635 | # patch the quant_config to add `modules_to_not_convert` back |
| 636 | if isinstance(quant_config, AutoAWQConfig): |
| 637 | text_config = config.text_config |
| 638 | llm_quant_config = getattr(text_config, "quantization_config", None) |
| 639 | if (not quant_config.modules_to_not_convert) and ( |
| 640 | llm_quant_config is not None |
| 641 | ): |
| 642 | quant_config.modules_to_not_convert.append("vision_model") |
| 643 | |
| 644 | def _init_vision_model( |
| 645 | self, |
| 646 | config: PreTrainedConfig, |
| 647 | quant_config: QuantizationConfig | None, |
| 648 | *, |
| 649 | prefix: str, |
| 650 | ): |
| 651 | vision_feature_layer = config.select_layer |
| 652 | if vision_feature_layer < 0: |
| 653 | num_hidden_layers = ( |
| 654 | config.vision_config.num_hidden_layers + vision_feature_layer + 1 |
| 655 | ) |
| 656 | else: |
| 657 | num_hidden_layers = vision_feature_layer + 1 |
| 658 | |
| 659 | return InternVisionModel( |
| 660 | config.vision_config, |
| 661 | quant_config=quant_config, |
| 662 | num_hidden_layers_override=num_hidden_layers, |
| 663 | prefix=prefix, |
| 664 | ) |
| 665 | |
| 666 | def _init_mlp1(self, config: PreTrainedConfig) -> nn.Module: |
| 667 | vit_hidden_size = config.vision_config.hidden_size |
| 668 | llm_hidden_size = config.text_config.hidden_size |
| 669 | |
| 670 | return nn.Sequential( |
| 671 | nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2), |
| 672 | nn.Linear( |
| 673 | vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size |
| 674 | ), |
| 675 | nn.GELU(), |
| 676 | nn.Linear(llm_hidden_size, llm_hidden_size), |
| 677 | ) |
| 678 | |
| 679 | def pixel_shuffle(self, x, scale_factor=0.5): |
| 680 | n, w, h, c = x.size() |
| 681 | # N, W, H, C --> N, W, H * scale, C // scale |
| 682 | x = x.view(n, w, int(h * scale_factor), int(c / scale_factor)) |
| 683 | # N, W, H * scale, C // scale --> N, H * scale, W, C // scale |
| 684 | x = x.permute(0, 2, 1, 3).contiguous() |
| 685 | x = x.view( |
| 686 | n, |
| 687 | int(h * scale_factor), |
| 688 | int(w * scale_factor), |
| 689 | int(c / (scale_factor * scale_factor)), |
| 690 | ) |
| 691 | if self.ps_version == "v1": |
| 692 | pass |
| 693 | else: |
| 694 | x = x.permute(0, 2, 1, 3).contiguous() |
| 695 | return x |
| 696 | |
| 697 | def extract_feature(self, pixel_values: torch.Tensor) -> torch.Tensor: |
| 698 | vit_embeds = self.vision_model(pixel_values=pixel_values) |
| 699 | vit_embeds = vit_embeds[:, 1:, :] |
| 700 | |
| 701 | h = w = int(vit_embeds.shape[1] ** 0.5) |
| 702 | vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1) |
| 703 | vit_embeds = self.pixel_shuffle(vit_embeds, scale_factor=self.downsample_ratio) |
| 704 | vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1]) |
| 705 | vit_embeds = self.mlp1(vit_embeds) |
| 706 | return vit_embeds |
| 707 | |
| 708 | def _parse_and_validate_image_input( |
| 709 | self, **kwargs: object |
| 710 | ) -> InternVLImageInputs | None: |
| 711 | pixel_values_flat = kwargs.pop("pixel_values_flat", None) |
| 712 | image_num_patches = kwargs.pop("image_num_patches", None) |
| 713 | image_embeds = kwargs.pop("image_embeds", None) |
| 714 | |
| 715 | if pixel_values_flat is None and image_embeds is None: |
| 716 | return None |
| 717 | |
| 718 | if image_embeds is not None: |
| 719 | return InternVLImageEmbeddingInputs( |
| 720 | type="image_embeds", |
| 721 | data=image_embeds, |
| 722 | ) |
| 723 | |
| 724 | image_token_id = kwargs["image_token_id"] |
| 725 | if isinstance(image_token_id, torch.Tensor): |
| 726 | image_token_id = image_token_id.flatten().unique().item() |
| 727 | |
| 728 | assert isinstance(image_token_id, int) |
| 729 | self.img_context_token_id = image_token_id |
| 730 | |
| 731 | if pixel_values_flat is not None: |
| 732 | expected_h = expected_w = self.config.vision_config.image_size |
| 733 | resolve_bindings = {"h": expected_h, "w": expected_w} |
| 734 | |
| 735 | return InternVLImagePixelInputs( |
| 736 | type="pixel_values", |
| 737 | pixel_values_flat=pixel_values_flat, |
| 738 | num_patches=image_num_patches, |
| 739 | resolve_bindings=resolve_bindings, |
| 740 | ) |
| 741 | |
| 742 | raise AssertionError("This line should be unreachable.") |
| 743 | |
| 744 | def _parse_and_validate_video_input( |
| 745 | self, **kwargs: object |
| 746 | ) -> InternVLVideoInputs | None: |
| 747 | pixel_values_flat_video = kwargs.pop("pixel_values_flat_video", None) |
| 748 | video_num_patches = kwargs.pop("video_num_patches", None) |
| 749 | video_embeds = kwargs.pop("video_embeds", None) |
| 750 | |
| 751 | if pixel_values_flat_video is None and video_embeds is None: |
| 752 | return None |
| 753 | |
| 754 | if video_embeds is not None: |
| 755 | return InternVLVideoEmbeddingInputs( |
| 756 | type="video_embeds", |
| 757 | data=video_embeds, |
| 758 | ) |
| 759 | |
| 760 | video_token_id = kwargs["video_token_id"] |
| 761 | if isinstance(video_token_id, torch.Tensor): |
| 762 | video_token_id = video_token_id.flatten().unique().item() |
| 763 | |
| 764 | assert isinstance(video_token_id, int) |
| 765 | self.video_context_token_id = video_token_id |
| 766 | |
| 767 | if pixel_values_flat_video is not None: |
| 768 | expected_h = expected_w = self.config.vision_config.image_size |
| 769 | resolve_bindings = {"h": expected_h, "w": expected_w} |
| 770 | |
| 771 | return InternVLVideoPixelInputs( |
| 772 | type="pixel_values_videos", |
| 773 | pixel_values_flat=pixel_values_flat_video, |
| 774 | num_patches=video_num_patches, |
| 775 | resolve_bindings=resolve_bindings, |
| 776 | ) |
| 777 | |
| 778 | raise AssertionError("This line should be unreachable.") |
| 779 | |
| 780 | def _process_vision_input( |
| 781 | self, |
| 782 | image_input: InternVLImageInputs | InternVLVideoInputs, |
| 783 | ) -> tuple[torch.Tensor, ...]: |
| 784 | if ( |
| 785 | image_input["type"] == "image_embeds" |
| 786 | or image_input["type"] == "video_embeds" |
| 787 | ): |
| 788 | return image_input["data"] |
| 789 | |
| 790 | image_embeds = self.extract_feature(image_input["pixel_values_flat"]) |
| 791 | |
| 792 | num_patches = image_input["num_patches"] |
| 793 | |
| 794 | # Only one image in the current batch |
| 795 | if len(num_patches) == 1: |
| 796 | return (image_embeds.view(-1, self.config.text_config.hidden_size),) |
| 797 | |
| 798 | # NOTE: Image embeddings are split into separate tensors for each image |
| 799 | # by the size of each embedding. |
| 800 | feature_size = image_embeds.shape[1] |
| 801 | image_embeds = image_embeds.view(-1, self.config.text_config.hidden_size) |
| 802 | image_feature_sizes = [ |
| 803 | num_patches * feature_size for num_patches in num_patches |
| 804 | ] |
| 805 | return image_embeds.split(image_feature_sizes) |
| 806 | |
| 807 | def _parse_and_validate_multimodal_inputs( |
| 808 | self, **kwargs: object |
| 809 | ) -> InternVLMultiModalInputs: |
| 810 | modalities: InternVLMultiModalInputs = {} |
| 811 | |
| 812 | # Preserve the order of modalities if there are multiple of them |
| 813 | # from the order of kwargs. |
| 814 | for input_key in kwargs: |
| 815 | if ( |
| 816 | input_key in ("pixel_values_flat", "image_embeds") |
| 817 | and "images" not in modalities |
| 818 | ): |
| 819 | modalities["images"] = self._parse_and_validate_image_input(**kwargs) |
| 820 | if input_key in ("pixel_values_flat_video",) and "videos" not in modalities: |
| 821 | modalities["videos"] = self._parse_and_validate_video_input(**kwargs) |
| 822 | |
| 823 | return modalities |
| 824 | |
| 825 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 826 | modalities = self._parse_and_validate_multimodal_inputs(**kwargs) |
| 827 | if not modalities: |
| 828 | return [] |
| 829 | |
| 830 | # The result multimodal_embeddings is tuple of tensors, with each |
| 831 | # tensor correspoending to a multimodal data item (image or video). |
| 832 | multimodal_embeddings: tuple[torch.Tensor, ...] = () |
| 833 | |
| 834 | # NOTE: It is important to iterate over the keys in this dictionary |
| 835 | # to preserve the order of the modalities. |
| 836 | for modality in modalities: |
| 837 | if modality == "images": |
| 838 | image_input = modalities["images"] |
| 839 | if image_input is not None: |
| 840 | image_embeddings = self._process_vision_input(image_input) |
| 841 | multimodal_embeddings += tuple(image_embeddings) |
| 842 | if modality == "videos": |
| 843 | video_input = modalities["videos"] |
| 844 | if video_input is not None: |
| 845 | video_embeddings = self._process_vision_input(video_input) |
| 846 | multimodal_embeddings += tuple(video_embeddings) |
| 847 | |
| 848 | return multimodal_embeddings |
| 849 | |
| 850 | def embed_input_ids( |
| 851 | self, |
| 852 | input_ids: torch.Tensor, |
| 853 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 854 | *, |
| 855 | is_multimodal: torch.Tensor | None = None, |
| 856 | ) -> torch.Tensor: |
| 857 | # This is to satisfy the type checker for each overload |
| 858 | if multimodal_embeddings is None or is_multimodal is None: |
| 859 | return super().embed_input_ids(input_ids) |
| 860 | |
| 861 | return super().embed_input_ids( |
| 862 | input_ids, |
| 863 | multimodal_embeddings=multimodal_embeddings, |
| 864 | is_multimodal=is_multimodal, |
| 865 | ) |
| 866 | |
| 867 | def forward( |
| 868 | self, |
| 869 | input_ids: torch.Tensor | None, |
| 870 | positions: torch.Tensor, |
| 871 | intermediate_tensors: IntermediateTensors | None = None, |
| 872 | inputs_embeds: torch.Tensor | None = None, |
| 873 | **kwargs: object, |
| 874 | ) -> IntermediateTensors: |
| 875 | if intermediate_tensors is not None: |
| 876 | inputs_embeds = None |
| 877 | |
| 878 | forward_kwargs = { |
| 879 | "input_ids": input_ids, |
| 880 | "positions": positions, |
| 881 | "intermediate_tensors": intermediate_tensors, |
| 882 | "inputs_embeds": inputs_embeds, |
| 883 | } |
| 884 | |
| 885 | hidden_states = self.language_model.model(**forward_kwargs) |
| 886 | return hidden_states |
| 887 | |
| 888 | def compute_logits( |
| 889 | self, |
| 890 | hidden_states: torch.Tensor, |
| 891 | ) -> torch.Tensor | None: |
| 892 | return self.language_model.compute_logits(hidden_states) |
| 893 | |
| 894 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 895 | loader = AutoWeightsLoader(self) |
| 896 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 897 | |
| 898 | def get_mm_mapping(self) -> MultiModelKeys: |
| 899 | """Get the module prefix in multimodal models.""" |
| 900 | return MultiModelKeys.from_string_field( |
| 901 | language_model="language_model", |
| 902 | connector="mlp1", |
| 903 | tower_model="vision_model", |
| 904 | ) |
| 905 | |
| 906 | def get_mm_lora_token_counts( |
| 907 | self, |
| 908 | *, |
| 909 | modality: str, |
| 910 | mm_kwargs: MultiModalKwargsItem | None, |
| 911 | num_mm_embeds: int, |
| 912 | ) -> tuple[int, int | None]: |
| 913 | del modality, mm_kwargs |
| 914 | if num_mm_embeds <= 0 or self.num_image_token <= 0: |
| 915 | return 0, 0 |
| 916 | |
| 917 | num_patches = num_mm_embeds // self.num_image_token |
| 918 | tower_tokens = num_patches * (self.patch_tokens + 1) |
| 919 | return tower_tokens, num_patches * self.num_image_token |
| 920 | |
| 921 | # -- SupportsEncoderCudaGraph protocol methods -- |
| 922 | |
| 923 | def get_encoder_cudagraph_config(self): |
| 924 | from vllm.v1.worker.encoder_cudagraph_defs import EncoderCudaGraphConfig |
| 925 | |
| 926 | return EncoderCudaGraphConfig( |
| 927 | modalities=["image", "video"], |
| 928 | # InternVision uses standard ViT attention (no rotary embeddings, |
| 929 | # no variable-length sequence metadata), so the only graph-recorded |
| 930 | # buffer is pixel_values_flat itself. |
| 931 | buffer_keys=["pixel_values_flat"], |
| 932 | out_hidden_size=self.config.text_config.hidden_size, |
| 933 | ) |
| 934 | |
| 935 | def get_input_modality( |
| 936 | self, |
| 937 | mm_kwargs: dict[str, Any], |
| 938 | ) -> str: |
| 939 | if "pixel_values_flat" in mm_kwargs: |
| 940 | return "image" |
| 941 | return "video" |
| 942 | |
| 943 | def get_encoder_cudagraph_budget_range( |
| 944 | self, |
| 945 | vllm_config: "VllmConfig", |
| 946 | ) -> tuple[int, int]: |
| 947 | # Min: 1 tile → num_image_token output tokens. |
| 948 | min_budget = self.num_image_token |
| 949 | max_budget = min( |
| 950 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 951 | vllm_config.model_config.max_model_len, |
| 952 | ) |
| 953 | return (min_budget, max_budget) |
| 954 | |
| 955 | def _get_internvl_patches_list( |
| 956 | self, |
| 957 | mm_kwargs: dict[str, Any], |
| 958 | ) -> list[int]: |
| 959 | """Return per-item tile counts as a plain list of ints.""" |
| 960 | if self.get_input_modality(mm_kwargs) == "image": |
| 961 | patches = mm_kwargs.get("image_num_patches", []) |
| 962 | else: |
| 963 | patches = mm_kwargs.get("video_num_patches", []) |
| 964 | if isinstance(patches, torch.Tensor): |
| 965 | return patches.tolist() |
| 966 | return [int(n) for n in patches] |
| 967 | |
| 968 | def get_encoder_cudagraph_item_specs( |
| 969 | self, |
| 970 | mm_kwargs: dict[str, Any], |
| 971 | ): |
| 972 | from vllm.v1.worker.encoder_cudagraph_defs import EncoderItemSpec |
| 973 | |
| 974 | return [ |
| 975 | EncoderItemSpec( |
| 976 | input_size=n, |
| 977 | output_tokens=n * self.num_image_token, |
| 978 | ) |
| 979 | for n in self._get_internvl_patches_list(mm_kwargs) |
| 980 | ] |
| 981 | |
| 982 | def select_encoder_cudagraph_items( |
| 983 | self, |
| 984 | mm_kwargs: dict[str, Any], |
| 985 | indices: list[int], |
| 986 | ) -> dict[str, Any]: |
| 987 | modality = self.get_input_modality(mm_kwargs) |
| 988 | pv_key = ( |
| 989 | "pixel_values_flat" if modality == "image" else "pixel_values_flat_video" |
| 990 | ) |
| 991 | patches_key = ( |
| 992 | "image_num_patches" if modality == "image" else "video_num_patches" |
| 993 | ) |
| 994 | |
| 995 | pixel_values = mm_kwargs[pv_key] |
| 996 | patches_list = self._get_internvl_patches_list(mm_kwargs) |
| 997 | |
| 998 | if len(indices) == 0: |
| 999 | return {pv_key: pixel_values[:0], patches_key: []} |
| 1000 | |
| 1001 | # Compute cumulative tile offsets for slicing pixel_values. |
| 1002 | cum_patches = [0] |
| 1003 | for n in patches_list: |
| 1004 | cum_patches.append(cum_patches[-1] + n) |
| 1005 | |
| 1006 | selected_pv = torch.cat( |
| 1007 | [pixel_values[cum_patches[i] : cum_patches[i + 1]] for i in indices] |
| 1008 | ) |
| 1009 | selected_patches = [patches_list[i] for i in indices] |
| 1010 | |
| 1011 | return {pv_key: selected_pv, patches_key: selected_patches} |
| 1012 | |
| 1013 | def prepare_encoder_cudagraph_capture_inputs( |
| 1014 | self, |
| 1015 | token_budget: int, |
| 1016 | max_batch_size: int, |
| 1017 | max_frames_per_batch: int, |
| 1018 | device: torch.device, |
| 1019 | dtype: torch.dtype, |
| 1020 | path: str = "default", |
| 1021 | axis_keys: tuple[Hashable, ...] | None = None, |
| 1022 | ): |
| 1023 | from vllm.v1.worker.encoder_cudagraph_defs import ( |
| 1024 | EncoderCudaGraphCaptureInputs, |
| 1025 | ) |
| 1026 | |
| 1027 | # Size the buffer to hold the maximum possible tiles for this budget. |
| 1028 | total_tiles = max(token_budget // self.num_image_token, 1) |
| 1029 | image_size = self.config.vision_config.image_size |
| 1030 | |
| 1031 | dummy_pixel_values = torch.randn( |
| 1032 | total_tiles, 3, image_size, image_size, device=device, dtype=dtype |
| 1033 | ) |
| 1034 | |
| 1035 | return EncoderCudaGraphCaptureInputs( |
| 1036 | values={"pixel_values_flat": dummy_pixel_values}, |
| 1037 | ) |
| 1038 | |
| 1039 | def prepare_encoder_cudagraph_replay_buffers( |
| 1040 | self, |
| 1041 | mm_kwargs: dict[str, Any], |
| 1042 | max_batch_size: int, |
| 1043 | max_frames_per_batch: int, |
| 1044 | path: str = "default", |
| 1045 | ): |
| 1046 | from vllm.v1.worker.encoder_cudagraph_defs import ( |
| 1047 | EncoderCudaGraphReplayBuffers, |
| 1048 | ) |
| 1049 | |
| 1050 | modality = self.get_input_modality(mm_kwargs) |
| 1051 | pv_key = ( |
| 1052 | "pixel_values_flat" if modality == "image" else "pixel_values_flat_video" |
| 1053 | ) |
| 1054 | return EncoderCudaGraphReplayBuffers( |
| 1055 | values={"pixel_values_flat": mm_kwargs[pv_key]}, |
| 1056 | ) |
| 1057 | |
| 1058 | def encoder_cudagraph_forward( |
| 1059 | self, |
| 1060 | values: dict[str, torch.Tensor], |
| 1061 | path: str = "default", |
| 1062 | ) -> torch.Tensor: |
| 1063 | # The graph is always captured with pixel_values_flat as the input |
| 1064 | # buffer. During video replay the manager copies video tiles into |
| 1065 | # this same buffer before calling graph.replay(), so we always read |
| 1066 | # from pixel_values_flat here. |
| 1067 | pixel_values = values["pixel_values_flat"] |
| 1068 | out = self.extract_feature(pixel_values) # [N, num_image_token, H] |
| 1069 | return out.view(-1, self.config.text_config.hidden_size) |
| 1070 | |
| 1071 | def encoder_eager_forward( |
| 1072 | self, |
| 1073 | mm_kwargs: dict[str, Any], |
| 1074 | path: str = "default", |
| 1075 | ) -> torch.Tensor: |
| 1076 | if self.get_input_modality(mm_kwargs) == "image": |
| 1077 | pixel_values = mm_kwargs["pixel_values_flat"] |
| 1078 | else: |
| 1079 | pixel_values = mm_kwargs["pixel_values_flat_video"] |
| 1080 | out = self.extract_feature(pixel_values) # [N, num_image_token, H] |
| 1081 | return out.view(-1, self.config.text_config.hidden_size) |