| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # adapted from https://github.com/huggingface/transformers/blob/v4.39.3/src/transformers/models/ovis/modeling_ovis.py |
| 5 | # Copyright 2023 The vLLM team. |
| 6 | # Copyright 2023 HuggingFace Inc. team. All rights reserved. |
| 7 | # |
| 8 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 9 | # you may not use this file except in compliance with the License. |
| 10 | # You may obtain a copy of the License at |
| 11 | # |
| 12 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 13 | # |
| 14 | # Unless required by applicable law or agreed to in writing, software |
| 15 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 16 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 17 | # See the License for the specific language governing permissions and |
| 18 | # limitations under the License. |
| 19 | """PyTorch Ovis model.""" |
| 20 | |
| 21 | import math |
| 22 | from collections.abc import Iterable, Mapping |
| 23 | from typing import Annotated, Literal |
| 24 | |
| 25 | import torch |
| 26 | import torch.nn as nn |
| 27 | from torch import Tensor |
| 28 | from torch.nn.functional import gumbel_softmax, pad, softmax |
| 29 | from transformers import BatchFeature, PreTrainedConfig |
| 30 | |
| 31 | from vllm.config import VllmConfig |
| 32 | from vllm.config.multimodal import MultiModalDummyOptions |
| 33 | from vllm.inputs import MultiModalDataDict |
| 34 | from vllm.model_executor.layers.linear import ReplicatedLinear |
| 35 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 36 | from vllm.model_executor.models.aimv2 import AIMv2Model |
| 37 | from vllm.model_executor.models.siglip import SiglipVisionModel |
| 38 | from vllm.model_executor.models.utils import ( |
| 39 | AutoWeightsLoader, |
| 40 | flatten_bn, |
| 41 | init_vllm_registered_model, |
| 42 | maybe_prefix, |
| 43 | ) |
| 44 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 45 | from vllm.multimodal.inputs import ( |
| 46 | MultiModalFieldConfig, |
| 47 | MultiModalKwargsItems, |
| 48 | ) |
| 49 | from vllm.multimodal.parse import ImageSize, MultiModalDataItems |
| 50 | from vllm.multimodal.processing import ( |
| 51 | BaseDummyInputsBuilder, |
| 52 | BaseMultiModalProcessor, |
| 53 | BaseProcessingInfo, |
| 54 | PromptReplacement, |
| 55 | cached_encode, |
| 56 | ) |
| 57 | from vllm.renderers import TokenizeParams |
| 58 | from vllm.sequence import IntermediateTensors |
| 59 | from vllm.transformers_utils.processors.ovis import OvisProcessor |
| 60 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 61 | |
| 62 | from .interfaces import ( |
| 63 | MultiModalEmbeddings, |
| 64 | SupportsMultiModal, |
| 65 | SupportsPP, |
| 66 | supports_pp, |
| 67 | ) |
| 68 | |
| 69 | # Cannot find the following number from hf config. |
| 70 | IMAGE_TOKEN = "<image>" |
| 71 | IMAGE_INDICATOR_IDS = [-301, -302, -303, -304, -305] |
| 72 | |
| 73 | IMAGE_PAD_TOKEN_MAP = { |
| 74 | "gemma2": "<unused0>", |
| 75 | "llama": "<|reserved_special_token_0|>", |
| 76 | "qwen2": "<|image_pad|>", |
| 77 | } |
| 78 | IMAGE_PAD_TOKEN_ID_MAP = { |
| 79 | "gemma2": 7, |
| 80 | "llama": 128002, |
| 81 | "qwen2": 151655, |
| 82 | } |
| 83 | |
| 84 | |
| 85 | def st_argmax(y_soft: torch.Tensor, dim: int): # straight-through softmax |
| 86 | index = y_soft.argmax(dim, keepdim=True) |
| 87 | return torch.zeros_like( |
| 88 | y_soft, |
| 89 | memory_format=torch.legacy_contiguous_format, |
| 90 | ).scatter_(dim, index, 1.0) |
| 91 | |
| 92 | |
| 93 | class VisualTokenizer(torch.nn.Module): |
| 94 | def __init__( |
| 95 | self, |
| 96 | config: PreTrainedConfig, |
| 97 | quant_config: QuantizationConfig | None = None, |
| 98 | prefix: str = "", |
| 99 | ): |
| 100 | super().__init__() |
| 101 | self.config = config |
| 102 | self.backbone = self._init_backbone( |
| 103 | config=config, |
| 104 | quant_config=quant_config, |
| 105 | prefix=maybe_prefix(prefix, "backbone"), |
| 106 | ) |
| 107 | # reserved tokens for IMAGE_INDICATORS |
| 108 | head_dim = config.vocab_size - len(IMAGE_INDICATOR_IDS) |
| 109 | self.head = torch.nn.Sequential( |
| 110 | ReplicatedLinear( |
| 111 | config.backbone_config.hidden_size |
| 112 | * config.hidden_stride |
| 113 | * config.hidden_stride, |
| 114 | head_dim, |
| 115 | bias=False, |
| 116 | return_bias=False, |
| 117 | ), |
| 118 | torch.nn.LayerNorm(head_dim), |
| 119 | ) |
| 120 | |
| 121 | def _init_backbone( |
| 122 | self, |
| 123 | config: PreTrainedConfig, |
| 124 | quant_config: QuantizationConfig | None = None, |
| 125 | prefix: str = "", |
| 126 | ) -> nn.Module: |
| 127 | model_type = config.backbone_config.model_type |
| 128 | if model_type == "aimv2": |
| 129 | # No post rms_norm in Ovis2's AIMv2 ViT. |
| 130 | return AIMv2Model( |
| 131 | config=config.backbone_config, |
| 132 | quant_config=quant_config, |
| 133 | require_post_norm=False, |
| 134 | prefix=prefix, |
| 135 | ) |
| 136 | elif model_type == "siglip_vision_model": |
| 137 | return SiglipVisionModel( |
| 138 | config=config.backbone_config, |
| 139 | quant_config=quant_config, |
| 140 | prefix=prefix, |
| 141 | ) |
| 142 | raise ValueError(f"Unsupported visual tokenizer model_type: {model_type}") |
| 143 | |
| 144 | @property |
| 145 | def dtype(self) -> torch.dtype: |
| 146 | return next(self.head.parameters()).dtype |
| 147 | |
| 148 | @property |
| 149 | def device(self) -> torch.device: |
| 150 | return next(self.head.parameters()).device |
| 151 | |
| 152 | def tokenize(self, logits: torch.Tensor) -> torch.Tensor: |
| 153 | if self.config.tokenize_function == "softmax": |
| 154 | tokens = softmax(logits, dim=-1) |
| 155 | elif self.config.tokenize_function == "gumbel_argmax": |
| 156 | tokens = gumbel_softmax(logits, tau=self.config.tau, hard=True) |
| 157 | elif self.config.tokenize_function == "st_argmax": |
| 158 | tokens = st_argmax(logits, dim=-1) |
| 159 | else: |
| 160 | raise ValueError( |
| 161 | "Invalid `max_type`, expected softmax or gumbel_argmax " |
| 162 | f"or st_argmax, but got {self.config.tokenize_function}" |
| 163 | ) |
| 164 | return tokens |
| 165 | |
| 166 | def encode(self, pixel_values: torch.Tensor) -> torch.Tensor: |
| 167 | features = self.backbone(pixel_values) |
| 168 | if self.config.drop_cls_token: |
| 169 | features = features[:, 1:, :] |
| 170 | |
| 171 | # merge number of `hidden_stride * hidden_stride` hidden states together |
| 172 | # to reduce token sequence length |
| 173 | # e.g., for hidden_stride=2, this leads to a token length reduction: |
| 174 | # 1024 -> 256 for aimv2 |
| 175 | if self.config.hidden_stride > 1: |
| 176 | # this `d` maybe different from the above `d` |
| 177 | n, L, d = features.shape |
| 178 | sqrt_l = int(L**0.5) |
| 179 | assert sqrt_l**2 == L, ( |
| 180 | "The token sequence length should be a perfect square." |
| 181 | ) |
| 182 | features = features.reshape(n, sqrt_l, sqrt_l, d) |
| 183 | pl = ( |
| 184 | self.config.hidden_stride - (sqrt_l % self.config.hidden_stride) |
| 185 | ) % self.config.hidden_stride |
| 186 | features = pad(features, (0, 0, 0, pl, 0, pl), "constant", 0) |
| 187 | sqrt_l += pl |
| 188 | features = features.reshape( |
| 189 | n, |
| 190 | sqrt_l // self.config.hidden_stride, |
| 191 | self.config.hidden_stride, |
| 192 | sqrt_l // self.config.hidden_stride, |
| 193 | self.config.hidden_stride, |
| 194 | d, |
| 195 | ) |
| 196 | # [n, sqrt_l/hs, sqrt_l/hs, hs, hs, d] |
| 197 | features = features.permute(0, 1, 3, 2, 4, 5) |
| 198 | # [n, sqrt_l/hs, sqrt_l/hs, hs*hs*d] |
| 199 | features = features.flatten(3) |
| 200 | # [n, sqrt_l/hs*sqrt_l/hs, hs*hs*d] |
| 201 | features = features.reshape( |
| 202 | n, -1, self.config.hidden_stride * self.config.hidden_stride * d |
| 203 | ) |
| 204 | |
| 205 | return features |
| 206 | |
| 207 | def forward(self, pixel_values: torch.Tensor) -> torch.Tensor: |
| 208 | """[BatchSize, ImageShape] -> [BatchSize, Token, VocabSize].""" |
| 209 | features = self.encode(pixel_values) |
| 210 | logits = self.head(features) |
| 211 | tokens = self.tokenize(logits) |
| 212 | # tokens' shape is [BatchSize, #Token, VocabSize-5], so padding with |
| 213 | # [BatchSize, #Token, 5], after which, tokens' shape should become |
| 214 | # [BatchSize, #Token, VocabSize] |
| 215 | tokens = torch.nn.functional.pad( |
| 216 | tokens, |
| 217 | (0, len(IMAGE_INDICATOR_IDS)), |
| 218 | mode="constant", |
| 219 | value=0, |
| 220 | ) |
| 221 | return tokens |
| 222 | |
| 223 | |
| 224 | class OvisImagePatchInputs(TensorSchema): |
| 225 | """Dimensions: |
| 226 | - bnp: Batch size * number of images * number of patches |
| 227 | - h: Height of each patch |
| 228 | - w: Width of each patch |
| 229 | - patch_indicators: Batch size * (number of patches + 1) |
| 230 | - bn: Batch size * number of images |
| 231 | """ |
| 232 | |
| 233 | type: Literal["image_patches"] |
| 234 | flat_data: Annotated[torch.Tensor, TensorShape("bnp", 3, "h", "w")] |
| 235 | indicator_tokens: Annotated[torch.Tensor, TensorShape("patch_indicators")] |
| 236 | patches_per_image: Annotated[list[int], TensorShape("bn")] |
| 237 | # This is used to restore the first two dimensions of `flat_data`. |
| 238 | |
| 239 | |
| 240 | class VisualEmbedding(torch.nn.Embedding): |
| 241 | def __init__(self, *args, **kwargs): |
| 242 | super().__init__(*args, **kwargs) |
| 243 | |
| 244 | def forward(self, visual_tokens: Tensor) -> Tensor: |
| 245 | if visual_tokens.dtype in [ |
| 246 | torch.int8, |
| 247 | torch.int16, |
| 248 | torch.int32, |
| 249 | torch.int64, |
| 250 | torch.long, |
| 251 | ]: |
| 252 | return super().forward(visual_tokens) |
| 253 | return torch.matmul(visual_tokens, self.weight) |
| 254 | |
| 255 | @property |
| 256 | def device(self): |
| 257 | return self.weight.device |
| 258 | |
| 259 | @property |
| 260 | def dtype(self): |
| 261 | return self.weight.dtype |
| 262 | |
| 263 | |
| 264 | class OvisProcessingInfo(BaseProcessingInfo): |
| 265 | def get_hf_processor(self, **kwargs: object): |
| 266 | return self.ctx.get_hf_processor( |
| 267 | OvisProcessor, |
| 268 | image_pad_token=self.get_image_pad_token(), |
| 269 | image_segment_len=self.get_image_segment_len(), |
| 270 | **kwargs, |
| 271 | ) |
| 272 | |
| 273 | def get_default_tok_params(self) -> TokenizeParams: |
| 274 | return super().get_default_tok_params().with_kwargs(add_special_tokens=False) |
| 275 | |
| 276 | def get_image_segment_len(self) -> int: |
| 277 | visual_tokenizer_config = self.get_hf_config().visual_tokenizer_config |
| 278 | image_size = visual_tokenizer_config.backbone_config.image_size |
| 279 | patch_size = visual_tokenizer_config.backbone_config.patch_size |
| 280 | hidden_stride = visual_tokenizer_config.hidden_stride |
| 281 | patch_grid_length = math.ceil(image_size / patch_size) |
| 282 | assert patch_grid_length % hidden_stride == 0, ( |
| 283 | f"patch_grid_length {patch_grid_length} is not divisible by " |
| 284 | f"hidden_stride {hidden_stride}" |
| 285 | ) |
| 286 | # minus 1 for presented image token |
| 287 | return (patch_grid_length // hidden_stride) ** 2 - 1 |
| 288 | |
| 289 | def get_image_pad_token(self) -> str | None: |
| 290 | hf_text_config = self.get_hf_config().get_text_config() |
| 291 | text_model_type = hf_text_config.model_type |
| 292 | return IMAGE_PAD_TOKEN_MAP.get(text_model_type) |
| 293 | |
| 294 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 295 | return {"image": None} |
| 296 | |
| 297 | def get_image_size_with_most_features(self) -> ImageSize: |
| 298 | height, width = self.get_hf_processor().get_image_size() |
| 299 | hs = self.get_hf_config().visual_tokenizer_config.hidden_stride |
| 300 | # NOTE(Isotr0py): 9 is `max_partition` hardcoded in original code |
| 301 | # https://huggingface.co/AIDC-AI/Ovis2-1B/blob/main/modeling_ovis.py#L96 |
| 302 | return ImageSize(width=width * hs * 9, height=height * hs * 9) |
| 303 | |
| 304 | |
| 305 | class OvisDummyInputsBuilder(BaseDummyInputsBuilder[OvisProcessingInfo]): |
| 306 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 307 | num_images = mm_counts.get("image", 0) |
| 308 | return IMAGE_TOKEN * num_images |
| 309 | |
| 310 | def get_dummy_mm_data( |
| 311 | self, |
| 312 | seq_len: int, |
| 313 | mm_counts: Mapping[str, int], |
| 314 | mm_options: MultiModalDummyOptions, |
| 315 | ) -> MultiModalDataDict: |
| 316 | target_width, target_height = self.info.get_image_size_with_most_features() |
| 317 | |
| 318 | mm_data = { |
| 319 | "image": self._get_dummy_images( |
| 320 | width=target_width, |
| 321 | height=target_height, |
| 322 | num_images=mm_counts.get("image", 0), |
| 323 | overrides=mm_options.get("image"), |
| 324 | ), |
| 325 | } |
| 326 | return mm_data |
| 327 | |
| 328 | |
| 329 | class OvisMultiModalProcessor(BaseMultiModalProcessor[OvisProcessingInfo]): |
| 330 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 331 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 332 | |
| 333 | def image_indicators_to_visual_tokens( |
| 334 | self, |
| 335 | image_indicators: list[int], |
| 336 | ) -> list[int]: |
| 337 | """Filter image indicators placeholders and convert them to corresponding |
| 338 | tokens in visual tokenizer. |
| 339 | For example, [-301, -300, -302, -300, -303, -300, -304, -300, -305] |
| 340 | should return [vocab_size-1, vocab_size-2, ..., vocab_size-5] |
| 341 | """ |
| 342 | hf_config = self.info.get_hf_config() |
| 343 | vte_vocab_size = hf_config.visual_tokenizer_config.vocab_size |
| 344 | # -300 is image_atom token, filter them out |
| 345 | return [vte_vocab_size + x + 300 for x in image_indicators if x < -300] |
| 346 | |
| 347 | def _postprocess_hf_mm_data( |
| 348 | self, |
| 349 | mm_data: Mapping[str, object], |
| 350 | hf_processor_mm_kwargs: Mapping[str, object], |
| 351 | processed_data: BatchFeature, |
| 352 | ) -> BatchFeature: |
| 353 | if not mm_data: |
| 354 | return processed_data |
| 355 | |
| 356 | hf_processor = self.info.get_hf_processor() |
| 357 | image_indicators = [ |
| 358 | hf_processor.construct_image_indicators(grid) |
| 359 | for grid in processed_data["grids"] |
| 360 | ] |
| 361 | indicator_tokens = [ |
| 362 | self.image_indicators_to_visual_tokens(indicator) |
| 363 | for indicator in image_indicators |
| 364 | ] |
| 365 | processed_data["indicator_tokens"] = torch.tensor(indicator_tokens) |
| 366 | |
| 367 | return processed_data |
| 368 | |
| 369 | def _get_mm_fields_config( |
| 370 | self, |
| 371 | hf_inputs: BatchFeature, |
| 372 | hf_processor_mm_kwargs: Mapping[str, object], |
| 373 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 374 | return dict( |
| 375 | pixel_values=MultiModalFieldConfig.batched("image"), |
| 376 | grids=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 377 | indicator_tokens=MultiModalFieldConfig.batched("image"), |
| 378 | ) |
| 379 | |
| 380 | def _get_prompt_updates( |
| 381 | self, |
| 382 | mm_items: MultiModalDataItems, |
| 383 | hf_processor_mm_kwargs: Mapping[str, object], |
| 384 | out_mm_kwargs: MultiModalKwargsItems, |
| 385 | ) -> list[PromptReplacement]: |
| 386 | tokenizer = self.info.get_tokenizer() |
| 387 | |
| 388 | def get_replacement_ovis(item_idx: int): |
| 389 | out_item = out_mm_kwargs["image"][item_idx] |
| 390 | grid = out_item["grids"].data |
| 391 | |
| 392 | hf_processor = self.info.get_hf_processor() |
| 393 | return hf_processor.construct_image_placeholders(grid) |
| 394 | |
| 395 | return [ |
| 396 | PromptReplacement( |
| 397 | modality="image", |
| 398 | target=cached_encode(tokenizer, IMAGE_TOKEN, add_special_tokens=False), |
| 399 | replacement=get_replacement_ovis, |
| 400 | ), |
| 401 | ] |
| 402 | |
| 403 | |
| 404 | @MULTIMODAL_REGISTRY.register_processor( |
| 405 | OvisMultiModalProcessor, |
| 406 | info=OvisProcessingInfo, |
| 407 | dummy_inputs=OvisDummyInputsBuilder, |
| 408 | ) |
| 409 | class Ovis(nn.Module, SupportsMultiModal, SupportsPP): |
| 410 | @classmethod |
| 411 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 412 | if modality.startswith("image"): |
| 413 | return IMAGE_TOKEN |
| 414 | |
| 415 | raise ValueError("Only image modality is supported") |
| 416 | |
| 417 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 418 | super().__init__() |
| 419 | config = vllm_config.model_config.hf_config |
| 420 | quant_config = vllm_config.quant_config |
| 421 | |
| 422 | self.config: PreTrainedConfig = config |
| 423 | |
| 424 | with self._mark_language_model(vllm_config): |
| 425 | self.llm = init_vllm_registered_model( |
| 426 | vllm_config=vllm_config.with_hf_config(config.get_text_config()), |
| 427 | prefix=maybe_prefix(prefix, "llm"), |
| 428 | ) |
| 429 | |
| 430 | with self._mark_tower_model(vllm_config, "image"): |
| 431 | self.visual_tokenizer = VisualTokenizer( |
| 432 | config=config.visual_tokenizer_config, |
| 433 | quant_config=quant_config, |
| 434 | prefix=maybe_prefix(prefix, "visual_tokenizer"), |
| 435 | ) |
| 436 | self.vte = VisualEmbedding( |
| 437 | self.config.visual_tokenizer_config.vocab_size, self.config.hidden_size |
| 438 | ) |
| 439 | |
| 440 | text_model_type = self.config.get_text_config().model_type |
| 441 | self.image_pad_token_id = IMAGE_PAD_TOKEN_ID_MAP[text_model_type] |
| 442 | |
| 443 | language_model = self.get_language_model() |
| 444 | assert supports_pp(language_model) |
| 445 | self.make_empty_intermediate_tensors = ( |
| 446 | language_model.make_empty_intermediate_tensors |
| 447 | ) |
| 448 | |
| 449 | def _parse_and_validate_image_input( |
| 450 | self, **kwargs: object |
| 451 | ) -> OvisImagePatchInputs | None: |
| 452 | pixel_values = kwargs.pop("pixel_values", None) |
| 453 | indicator_tokens = kwargs.pop("indicator_tokens", None) |
| 454 | |
| 455 | if pixel_values is None and indicator_tokens is None: |
| 456 | return None |
| 457 | |
| 458 | if pixel_values is not None and indicator_tokens is not None: |
| 459 | if not isinstance(pixel_values, (torch.Tensor, list)): |
| 460 | raise ValueError( |
| 461 | f"Incorrect type of pixel values. Got type: {type(pixel_values)}" |
| 462 | ) |
| 463 | |
| 464 | if not isinstance(indicator_tokens, (torch.Tensor, list)): |
| 465 | raise ValueError( |
| 466 | "Incorrect type of indicator_tokens. " |
| 467 | f"Got type: {type(pixel_values)}" |
| 468 | ) |
| 469 | |
| 470 | return OvisImagePatchInputs( |
| 471 | type="image_patches", |
| 472 | flat_data=flatten_bn(pixel_values, concat=True), |
| 473 | patches_per_image=[x.shape[0] for x in pixel_values], |
| 474 | indicator_tokens=flatten_bn(indicator_tokens, concat=True), |
| 475 | ) |
| 476 | |
| 477 | raise AssertionError("This line should be unreachable.") |
| 478 | |
| 479 | def _process_image_input( |
| 480 | self, image_input: OvisImagePatchInputs |
| 481 | ) -> MultiModalEmbeddings: |
| 482 | image_patches_flat = image_input["flat_data"] |
| 483 | patches_per_image = image_input["patches_per_image"] |
| 484 | indicator_tokens = image_input["indicator_tokens"] |
| 485 | |
| 486 | indicator_per_image = list( |
| 487 | map(lambda x: x + 1 if x > 1 else x + 2, patches_per_image) |
| 488 | ) |
| 489 | |
| 490 | target_dtype = self.visual_tokenizer.dtype |
| 491 | visual_tokens = self.visual_tokenizer(image_patches_flat.to(target_dtype)) |
| 492 | visual_embeds = self.vte(visual_tokens) # 1:1 numeric eq. |
| 493 | |
| 494 | indicator_embeds = self.vte(indicator_tokens) |
| 495 | indicator_embeds_per_image = indicator_embeds.split(indicator_per_image) |
| 496 | |
| 497 | visual_embeds_per_image = visual_embeds.split(patches_per_image, dim=0) |
| 498 | vision_embeddings = [] |
| 499 | for indicator, visual in zip( |
| 500 | indicator_embeds_per_image, visual_embeds_per_image |
| 501 | ): |
| 502 | vision_embeddings_per_image = [] |
| 503 | for i in range(visual.shape[0]): |
| 504 | vision_embeddings_per_image.append( |
| 505 | torch.cat([indicator[i : i + 1], visual[i]], dim=0) |
| 506 | ) |
| 507 | vision_embeddings_per_image.append(indicator[i + 1 :]) |
| 508 | vision_embeddings.append(torch.cat(vision_embeddings_per_image, dim=0)) |
| 509 | |
| 510 | return tuple(vision_embeddings) |
| 511 | |
| 512 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 513 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 514 | if image_input is None: |
| 515 | return [] |
| 516 | |
| 517 | image_features = self._process_image_input(image_input) |
| 518 | |
| 519 | return image_features |
| 520 | |
| 521 | def forward( |
| 522 | self, |
| 523 | input_ids: torch.Tensor | None, |
| 524 | positions: torch.Tensor, |
| 525 | intermediate_tensors: IntermediateTensors | None = None, |
| 526 | inputs_embeds: torch.Tensor | None = None, |
| 527 | **kwargs: object, |
| 528 | ) -> torch.Tensor | IntermediateTensors: |
| 529 | if intermediate_tensors is not None: |
| 530 | inputs_embeds = None |
| 531 | |
| 532 | # up until here we have an inputs_embeds 100% numerical identity |
| 533 | # between the OG HF Transformers implementation and ours |
| 534 | hidden_states = self.llm( |
| 535 | input_ids=input_ids, |
| 536 | positions=positions, |
| 537 | intermediate_tensors=intermediate_tensors, |
| 538 | inputs_embeds=inputs_embeds, |
| 539 | ) |
| 540 | return hidden_states |
| 541 | |
| 542 | def compute_logits( |
| 543 | self, |
| 544 | hidden_states: torch.Tensor, |
| 545 | ) -> torch.Tensor | None: |
| 546 | return self.llm.compute_logits(hidden_states) |
| 547 | |
| 548 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 549 | loader = AutoWeightsLoader(self) |
| 550 | return loader.load_weights(weights) |