| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import math |
| 5 | from collections.abc import Iterable, Mapping, Sequence |
| 6 | from typing import Annotated, Any, Literal, cast |
| 7 | |
| 8 | import torch |
| 9 | import torch.nn as nn |
| 10 | from transformers import BatchFeature, Mistral3Config, PixtralVisionConfig |
| 11 | from transformers.models.pixtral import PixtralProcessor |
| 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.layers.activation import get_act_fn |
| 17 | from vllm.model_executor.layers.fusion.mm_input_norm import build_mm_input_norm |
| 18 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 19 | from vllm.model_executor.layers.linear import ColumnParallelLinear, RowParallelLinear |
| 20 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 21 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 22 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 23 | from vllm.multimodal.inputs import ( |
| 24 | MultiModalFieldConfig, |
| 25 | MultiModalKwargsItems, |
| 26 | ) |
| 27 | from vllm.multimodal.parse import ( |
| 28 | ImageProcessorItems, |
| 29 | ImageSize, |
| 30 | MultiModalDataItems, |
| 31 | ) |
| 32 | from vllm.multimodal.processing import ( |
| 33 | BaseDummyInputsBuilder, |
| 34 | BaseMultiModalProcessor, |
| 35 | BaseProcessingInfo, |
| 36 | PromptReplacement, |
| 37 | PromptUpdate, |
| 38 | PromptUpdateDetails, |
| 39 | ) |
| 40 | from vllm.sequence import IntermediateTensors |
| 41 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 42 | |
| 43 | from .interfaces import ( |
| 44 | MultiModalEmbeddings, |
| 45 | SupportsEagle, |
| 46 | SupportsEagle3, |
| 47 | SupportsLoRA, |
| 48 | SupportsMultiModal, |
| 49 | SupportsPP, |
| 50 | ) |
| 51 | from .pixtral import PixtralHFEncoderInfo, PixtralHFVisionModel |
| 52 | from .utils import ( |
| 53 | AutoWeightsLoader, |
| 54 | WeightsMapper, |
| 55 | get_layer_index, |
| 56 | init_vllm_registered_model, |
| 57 | maybe_prefix, |
| 58 | ) |
| 59 | |
| 60 | |
| 61 | class Mistral3ImagePixelInputs(TensorSchema): |
| 62 | """Dimensions: |
| 63 | - bn: Batch size * number of images |
| 64 | - c: Number of channels (3) |
| 65 | - h: Height of each image |
| 66 | - w: Width of each image |
| 67 | """ |
| 68 | |
| 69 | type: Literal["pixel_values_pixtral"] = "pixel_values_pixtral" |
| 70 | |
| 71 | # Note that `height` or `width` may be different per batch and image, |
| 72 | # in which case the data is passed as a list instead of a batched tensor. |
| 73 | pixel_values: Annotated[ |
| 74 | torch.Tensor | list[torch.Tensor], |
| 75 | TensorShape("bn", 3, "h", "w", dynamic_dims={"h", "w"}), |
| 76 | ] |
| 77 | |
| 78 | |
| 79 | class Mistral3PatchMerger(nn.Module): |
| 80 | """Learned merging of spatial_merge_size ** 2 patches.""" |
| 81 | |
| 82 | def __init__( |
| 83 | self, vision_hidden_size: int, spatial_merge_size: int, patch_size: int |
| 84 | ): |
| 85 | super().__init__() |
| 86 | |
| 87 | self.vision_hidden_size = vision_hidden_size |
| 88 | self.spatial_merge_size = spatial_merge_size |
| 89 | self.patch_size = patch_size |
| 90 | self.merging_layer = nn.Linear( |
| 91 | vision_hidden_size * self.spatial_merge_size**2, |
| 92 | vision_hidden_size, |
| 93 | bias=False, |
| 94 | ) |
| 95 | |
| 96 | def forward( |
| 97 | self, image_features: torch.Tensor, image_sizes: torch.Tensor |
| 98 | ) -> torch.Tensor: |
| 99 | image_sizes = [ |
| 100 | (image_size[0] // self.patch_size, image_size[1] // self.patch_size) |
| 101 | for image_size in image_sizes |
| 102 | ] |
| 103 | |
| 104 | tokens_per_image = [h * w for h, w in image_sizes] |
| 105 | d = image_features.shape[-1] |
| 106 | |
| 107 | permuted_tensor = [] |
| 108 | for image_index, image_tokens in enumerate( |
| 109 | image_features.split(tokens_per_image) |
| 110 | ): |
| 111 | # Reshape image_tokens into a 2D grid |
| 112 | h, w = image_sizes[image_index] |
| 113 | image_grid = image_tokens.view(h, w, d).permute(2, 0, 1).unsqueeze(0) |
| 114 | grid = torch.nn.functional.unfold( |
| 115 | image_grid, |
| 116 | kernel_size=self.spatial_merge_size, |
| 117 | stride=self.spatial_merge_size, |
| 118 | ) |
| 119 | grid = grid.view(d * self.spatial_merge_size**2, -1).t() |
| 120 | permuted_tensor.append(grid) |
| 121 | |
| 122 | image_features = torch.cat(permuted_tensor, dim=0) |
| 123 | image_features = self.merging_layer(image_features) |
| 124 | return image_features |
| 125 | |
| 126 | |
| 127 | class Mistral3MultiModalProjector(nn.Module): |
| 128 | def __init__( |
| 129 | self, |
| 130 | vision_hidden_size: int, |
| 131 | text_hidden_size: int, |
| 132 | spatial_merge_size: int, |
| 133 | patch_size: int, |
| 134 | projector_hidden_act: str, |
| 135 | multimodal_projector_bias: bool, |
| 136 | quant_config: QuantizationConfig | None = None, |
| 137 | prefix: str = "", |
| 138 | ): |
| 139 | super().__init__() |
| 140 | |
| 141 | self.norm = RMSNorm(vision_hidden_size, eps=1e-5) |
| 142 | self.patch_merger = Mistral3PatchMerger( |
| 143 | vision_hidden_size=vision_hidden_size, |
| 144 | spatial_merge_size=spatial_merge_size, |
| 145 | patch_size=patch_size, |
| 146 | ) |
| 147 | |
| 148 | self.linear_1 = ColumnParallelLinear( |
| 149 | vision_hidden_size, |
| 150 | text_hidden_size, |
| 151 | bias=multimodal_projector_bias, |
| 152 | quant_config=quant_config, |
| 153 | prefix=f"{prefix}.linear_1", |
| 154 | ) |
| 155 | self.act = get_act_fn(projector_hidden_act) |
| 156 | self.linear_2 = RowParallelLinear( |
| 157 | text_hidden_size, |
| 158 | text_hidden_size, |
| 159 | bias=multimodal_projector_bias, |
| 160 | quant_config=quant_config, |
| 161 | prefix=f"{prefix}.linear_2", |
| 162 | ) |
| 163 | |
| 164 | def forward( |
| 165 | self, image_features: torch.Tensor, image_sizes: torch.Tensor |
| 166 | ) -> torch.Tensor: |
| 167 | image_features = self.norm(image_features) |
| 168 | image_features = self.patch_merger(image_features, image_sizes) |
| 169 | hidden_states, _ = self.linear_1(image_features) |
| 170 | hidden_states = self.act(hidden_states) |
| 171 | hidden_states, _ = self.linear_2(hidden_states) |
| 172 | return hidden_states |
| 173 | |
| 174 | |
| 175 | class Mistral3HFEncoderInfo(PixtralHFEncoderInfo): |
| 176 | def __init__(self, hf_config: Mistral3Config, image_size: int) -> None: |
| 177 | super().__init__(hf_config) |
| 178 | self.image_size = image_size |
| 179 | |
| 180 | def get_image_size(self) -> int: |
| 181 | return self.image_size |
| 182 | |
| 183 | def get_patch_grid_size( |
| 184 | self, |
| 185 | *, |
| 186 | image_width: int, |
| 187 | image_height: int, |
| 188 | ) -> tuple[int, int]: |
| 189 | max_size = self.get_image_size() |
| 190 | ratio = max(image_width / max_size, image_height / max_size) |
| 191 | |
| 192 | if ratio > 1: |
| 193 | image_width = math.floor(image_width / ratio) |
| 194 | image_height = math.floor(image_height / ratio) |
| 195 | |
| 196 | assert isinstance(self.hf_config, Mistral3Config) |
| 197 | merged_patch_size = ( |
| 198 | self.vision_config.patch_size * self.hf_config.spatial_merge_size |
| 199 | ) |
| 200 | |
| 201 | ncols = (image_width - 1) // merged_patch_size + 1 |
| 202 | nrows = (image_height - 1) // merged_patch_size + 1 |
| 203 | return ncols, nrows |
| 204 | |
| 205 | |
| 206 | class Mistral3ProcessingInfo(BaseProcessingInfo): |
| 207 | def get_hf_config(self) -> Mistral3Config: |
| 208 | return self.ctx.get_hf_config(Mistral3Config) |
| 209 | |
| 210 | def get_vision_encoder_info( |
| 211 | self, |
| 212 | mm_processor_kwargs: Mapping[str, object] | None = None, |
| 213 | ) -> PixtralHFEncoderInfo: |
| 214 | processor = self.get_hf_processor() |
| 215 | size = processor.image_processor.size |
| 216 | merged_mm_kwargs = self._merge_and_resolve_mm_processor_kwargs( |
| 217 | mm_processor_kwargs or {} |
| 218 | ) |
| 219 | image_mm_kwargs = cast( |
| 220 | Mapping[str, Any], |
| 221 | merged_mm_kwargs.get("images_kwargs", {}), |
| 222 | ) |
| 223 | if override_size := image_mm_kwargs.get("size"): |
| 224 | size = size | override_size |
| 225 | |
| 226 | image_size = size["longest_edge"] |
| 227 | return Mistral3HFEncoderInfo(self.get_hf_config(), image_size) |
| 228 | |
| 229 | def get_hf_processor(self, **kwargs: object) -> PixtralProcessor: |
| 230 | return self.ctx.get_hf_processor(PixtralProcessor, **kwargs) |
| 231 | |
| 232 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 233 | return {"image": None} |
| 234 | |
| 235 | def get_num_image_tokens( |
| 236 | self, |
| 237 | *, |
| 238 | image_width: int, |
| 239 | image_height: int, |
| 240 | ) -> int: |
| 241 | vision_encoder_info = self.get_vision_encoder_info() |
| 242 | return vision_encoder_info.get_num_image_tokens( |
| 243 | image_width=image_width, |
| 244 | image_height=image_height, |
| 245 | ) |
| 246 | |
| 247 | def get_image_size_with_most_features(self) -> ImageSize: |
| 248 | vision_encoder_info = self.get_vision_encoder_info() |
| 249 | width = height = vision_encoder_info.get_image_size() |
| 250 | return ImageSize(width=width, height=height) |
| 251 | |
| 252 | |
| 253 | class Mistral3DummyInputsBuilder(BaseDummyInputsBuilder): |
| 254 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 255 | num_images = mm_counts.get("image", 0) |
| 256 | |
| 257 | processor = self.info.get_hf_processor() |
| 258 | image_token = processor.image_token |
| 259 | |
| 260 | return image_token * num_images |
| 261 | |
| 262 | def get_dummy_mm_data( |
| 263 | self, |
| 264 | seq_len: int, |
| 265 | mm_counts: Mapping[str, int], |
| 266 | mm_options: MultiModalDummyOptions, |
| 267 | ) -> MultiModalDataDict: |
| 268 | target_width, target_height = self.info.get_image_size_with_most_features() |
| 269 | |
| 270 | return { |
| 271 | "image": self._get_dummy_images( |
| 272 | width=target_width, |
| 273 | height=target_height, |
| 274 | num_images=mm_counts.get("image", 0), |
| 275 | overrides=mm_options.get("image"), |
| 276 | ) |
| 277 | } |
| 278 | |
| 279 | |
| 280 | class Mistral3MultiModalProcessor(BaseMultiModalProcessor[Mistral3ProcessingInfo]): |
| 281 | def _postprocess_hf_mm_data( |
| 282 | self, |
| 283 | mm_data: Mapping[str, object], |
| 284 | hf_processor_mm_kwargs: Mapping[str, object], |
| 285 | processed_data: BatchFeature, |
| 286 | ) -> BatchFeature: |
| 287 | if not mm_data: |
| 288 | return processed_data |
| 289 | |
| 290 | pixel_values = processed_data.get("pixel_values") |
| 291 | if pixel_values is not None: |
| 292 | # Avoid padding since we need the output for each image to be |
| 293 | # independent of other images for the cache to work correctly |
| 294 | image_sizes = processed_data["image_sizes"] |
| 295 | assert len(pixel_values) == len(image_sizes) |
| 296 | |
| 297 | processed_data["pixel_values"] = [ |
| 298 | p[:, :h, :w] for p, (h, w) in zip(pixel_values, image_sizes) |
| 299 | ] |
| 300 | |
| 301 | return processed_data |
| 302 | |
| 303 | def _get_mm_fields_config( |
| 304 | self, |
| 305 | hf_inputs: BatchFeature, |
| 306 | hf_processor_mm_kwargs: Mapping[str, object], |
| 307 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 308 | return dict( |
| 309 | pixel_values=MultiModalFieldConfig.batched("image"), |
| 310 | image_embeds=MultiModalFieldConfig.batched("image"), |
| 311 | ) |
| 312 | |
| 313 | def _get_prompt_updates( |
| 314 | self, |
| 315 | mm_items: MultiModalDataItems, |
| 316 | hf_processor_mm_kwargs: Mapping[str, object], |
| 317 | out_mm_kwargs: MultiModalKwargsItems, |
| 318 | ) -> Sequence[PromptUpdate]: |
| 319 | processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 320 | hf_config = self.info.get_hf_config() |
| 321 | tokenizer = self.info.get_tokenizer() |
| 322 | vocab = tokenizer.get_vocab() |
| 323 | |
| 324 | image_break_id = vocab[processor.image_break_token] |
| 325 | image_token_id = hf_config.image_token_index |
| 326 | image_end_id = vocab[processor.image_end_token] |
| 327 | |
| 328 | assert isinstance(hf_config.vision_config, PixtralVisionConfig) |
| 329 | encoder_info = self.info.get_vision_encoder_info(hf_processor_mm_kwargs) |
| 330 | |
| 331 | def get_replacement(item_idx: int): |
| 332 | images = mm_items.get_items("image", ImageProcessorItems) |
| 333 | image_size = images.get_image_size(item_idx) |
| 334 | |
| 335 | ncols, nrows = encoder_info.get_patch_grid_size( |
| 336 | image_width=image_size.width, |
| 337 | image_height=image_size.height, |
| 338 | ) |
| 339 | |
| 340 | tokens = ([image_token_id] * ncols + [image_break_id]) * nrows |
| 341 | tokens[-1] = image_end_id |
| 342 | |
| 343 | return PromptUpdateDetails.select_token_id(tokens, image_token_id) |
| 344 | |
| 345 | return [ |
| 346 | PromptReplacement( |
| 347 | modality="image", |
| 348 | target=[image_token_id], |
| 349 | replacement=get_replacement, |
| 350 | ), |
| 351 | ] |
| 352 | |
| 353 | |
| 354 | def _get_num_hidden_layers(hf_config: Mistral3Config) -> int: |
| 355 | """Determine the number of hidden layers to initialize up to in the |
| 356 | visual encoder. |
| 357 | |
| 358 | Args: |
| 359 | hf_config: Model config with vision feature layer(s). |
| 360 | |
| 361 | """ |
| 362 | feature_layers = hf_config.vision_feature_layer |
| 363 | num_hidden_layers = hf_config.vision_config.num_hidden_layers |
| 364 | # If we have one feature layer, initialize up to that layer |
| 365 | if isinstance(feature_layers, int): |
| 366 | return get_layer_index(feature_layers, num_hidden_layers) |
| 367 | # If we have multiple feature layers, initialize up to the deepest one |
| 368 | elif isinstance(feature_layers, (list, tuple)): |
| 369 | return max(get_layer_index(idx, num_hidden_layers) for idx in feature_layers) |
| 370 | raise TypeError( |
| 371 | f"vision_layer_feature type: {type(feature_layers)} is not supported" |
| 372 | ) |
| 373 | |
| 374 | |
| 375 | def init_vision_tower_for_mistral3( |
| 376 | hf_config: Mistral3Config, |
| 377 | quant_config: QuantizationConfig | None, |
| 378 | *, |
| 379 | input_norm: nn.Module | None = None, |
| 380 | require_post_norm: bool | None = None, |
| 381 | prefix: str = "", |
| 382 | ) -> PixtralHFVisionModel: |
| 383 | vision_config = hf_config.vision_config |
| 384 | |
| 385 | # Initialize the vision tower only up to the deepest required feature layer |
| 386 | num_hidden_layers = _get_num_hidden_layers(hf_config) |
| 387 | |
| 388 | assert isinstance(vision_config, PixtralVisionConfig) |
| 389 | |
| 390 | return PixtralHFVisionModel( |
| 391 | vision_config, |
| 392 | quant_config=quant_config, |
| 393 | input_norm=input_norm, |
| 394 | num_hidden_layers_override=num_hidden_layers, |
| 395 | require_post_norm=require_post_norm, |
| 396 | prefix=prefix, |
| 397 | ) |
| 398 | |
| 399 | |
| 400 | @MULTIMODAL_REGISTRY.register_processor( |
| 401 | Mistral3MultiModalProcessor, |
| 402 | info=Mistral3ProcessingInfo, |
| 403 | dummy_inputs=Mistral3DummyInputsBuilder, |
| 404 | ) |
| 405 | class Mistral3ForConditionalGeneration( |
| 406 | nn.Module, |
| 407 | SupportsLoRA, |
| 408 | SupportsMultiModal, |
| 409 | SupportsPP, |
| 410 | SupportsEagle, |
| 411 | SupportsEagle3, |
| 412 | ): |
| 413 | supports_mm_device_do_normalize = True |
| 414 | |
| 415 | packed_modules_mapping = { |
| 416 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 417 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 418 | } |
| 419 | |
| 420 | hf_to_vllm_mapper = WeightsMapper( |
| 421 | orig_to_new_prefix={ |
| 422 | # mapping for new names in checkpoint saved after transformers v4.52 |
| 423 | "model.language_model.": "language_model.model.", |
| 424 | "model.vision_tower.": "vision_tower.", |
| 425 | "model.multi_modal_projector.": "multi_modal_projector.", |
| 426 | "lm_head.": "language_model.lm_head.", |
| 427 | # Some PEFT LoRAs are trained against the text submodule directly |
| 428 | # and produce names like `base_model.model.model.layers.*`. |
| 429 | "model.": "language_model.model.", |
| 430 | }, |
| 431 | orig_to_new_suffix={ |
| 432 | # FP8 quantized HF checkpoints use "activation_scale" and |
| 433 | # "weight_scale_inv" but vLLM's FP8 linear layers register |
| 434 | # them as "input_scale" and "weight_scale" |
| 435 | ".activation_scale": ".input_scale", |
| 436 | ".weight_scale_inv": ".weight_scale", |
| 437 | }, |
| 438 | ) |
| 439 | |
| 440 | @classmethod |
| 441 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 442 | if modality.startswith("image"): |
| 443 | return None |
| 444 | |
| 445 | raise ValueError("Only image modality is supported") |
| 446 | |
| 447 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 448 | super().__init__() |
| 449 | |
| 450 | config = vllm_config.model_config.hf_config |
| 451 | quant_config = vllm_config.quant_config |
| 452 | multimodal_config = vllm_config.model_config.multimodal_config |
| 453 | |
| 454 | self.config = config |
| 455 | self.multimodal_config = multimodal_config |
| 456 | |
| 457 | # NOTE: This is a special case for Pixtral-12B in the HF-format |
| 458 | # https://huggingface.co/mistral-community/pixtral-12b/blob/main/config.json # noqa |
| 459 | if ( |
| 460 | config.projector_hidden_act is None |
| 461 | and config.vision_config.hidden_act == "gelu" |
| 462 | ): |
| 463 | config.projector_hidden_act = "gelu" |
| 464 | |
| 465 | with self._mark_tower_model(vllm_config, "image"): |
| 466 | self.vision_tower = init_vision_tower_for_mistral3( |
| 467 | config, |
| 468 | quant_config=quant_config, |
| 469 | input_norm=build_mm_input_norm(vllm_config.model_config), |
| 470 | require_post_norm=False, |
| 471 | prefix=maybe_prefix(prefix, "vision_tower"), |
| 472 | ) |
| 473 | self.multi_modal_projector = Mistral3MultiModalProjector( |
| 474 | vision_hidden_size=config.vision_config.hidden_size, |
| 475 | text_hidden_size=config.text_config.hidden_size, |
| 476 | projector_hidden_act=config.projector_hidden_act, |
| 477 | spatial_merge_size=config.spatial_merge_size, |
| 478 | patch_size=config.vision_config.patch_size, |
| 479 | multimodal_projector_bias=config.multimodal_projector_bias, |
| 480 | quant_config=quant_config, |
| 481 | prefix=maybe_prefix(prefix, "multi_modal_projector"), |
| 482 | ) |
| 483 | |
| 484 | with self._mark_language_model(vllm_config): |
| 485 | self.language_model = init_vllm_registered_model( |
| 486 | vllm_config=vllm_config, |
| 487 | hf_config=config.text_config, |
| 488 | prefix=maybe_prefix(prefix, "language_model"), |
| 489 | ) |
| 490 | |
| 491 | self.make_empty_intermediate_tensors = ( |
| 492 | self.language_model.make_empty_intermediate_tensors |
| 493 | ) |
| 494 | |
| 495 | def _parse_and_validate_image_input( |
| 496 | self, **kwargs: object |
| 497 | ) -> Mistral3ImagePixelInputs | None: |
| 498 | pixel_values = kwargs.pop("pixel_values", None) |
| 499 | image_embeds = kwargs.pop("image_embeds", None) |
| 500 | |
| 501 | if pixel_values is None and image_embeds is None: |
| 502 | return None |
| 503 | |
| 504 | return Mistral3ImagePixelInputs( |
| 505 | type="pixel_values_pixtral", |
| 506 | pixel_values=pixel_values, |
| 507 | ) |
| 508 | |
| 509 | def _process_image_input( |
| 510 | self, |
| 511 | image_input: Mistral3ImagePixelInputs, |
| 512 | ) -> torch.Tensor | tuple[torch.Tensor, ...]: |
| 513 | if image_input["type"] == "image_embeds": |
| 514 | return image_input["data"] |
| 515 | |
| 516 | image_sizes = [ |
| 517 | (img.shape[-2], img.shape[-1]) for img in image_input["pixel_values"] |
| 518 | ] |
| 519 | |
| 520 | image_features = self.vision_tower(image_input["pixel_values"]) |
| 521 | |
| 522 | if isinstance(image_features, torch.Tensor): |
| 523 | return self.multi_modal_projector(image_features, image_sizes) |
| 524 | |
| 525 | feature_sizes = [ |
| 526 | image_feature.shape[0] // self.config.spatial_merge_size**2 |
| 527 | for image_feature in image_features |
| 528 | ] |
| 529 | |
| 530 | image_embeds = self.multi_modal_projector( |
| 531 | torch.cat(image_features), image_sizes |
| 532 | ) |
| 533 | if len(feature_sizes) > 1: |
| 534 | image_embeds = torch.split(image_embeds, feature_sizes) |
| 535 | else: |
| 536 | image_embeds = (image_embeds,) |
| 537 | return image_embeds |
| 538 | |
| 539 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 540 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 541 | if image_input is None: |
| 542 | return [] |
| 543 | |
| 544 | vision_embeddings = self._process_image_input(image_input) |
| 545 | |
| 546 | return vision_embeddings |
| 547 | |
| 548 | def forward( |
| 549 | self, |
| 550 | input_ids: torch.Tensor | None, |
| 551 | positions: torch.Tensor, |
| 552 | intermediate_tensors: IntermediateTensors | None = None, |
| 553 | inputs_embeds: torch.Tensor | None = None, |
| 554 | **kwargs: object, |
| 555 | ) -> torch.Tensor | IntermediateTensors: |
| 556 | """Run forward pass for Mistral3. |
| 557 | |
| 558 | One key thing to understand is the `input_ids` already accounts for the |
| 559 | positions of the to-be-inserted image embeddings. |
| 560 | |
| 561 | Concretely, consider a text prompt: |
| 562 | `"USER: <image>\\nWhat's the content of the image?\\nASSISTANT:"`. |
| 563 | |
| 564 | Tokenizer outputs: |
| 565 | `[1, 3148, 1001, 29901, 29871, 32000, 29871, 13, 5618, 29915, 29879, |
| 566 | 278, 2793, 310, 278, 1967, 29973, 13, 22933, 9047, 13566, 29901]`. |
| 567 | |
| 568 | To reserve space in KV cache, we have to insert placeholder tokens |
| 569 | before they are inputted to the model, so the input processor prepends |
| 570 | additional image tokens (denoted as `32000`), resulting in: |
| 571 | `[1, 3148, 1001, 29901, 29871, 32000, ..., 32000, 29871, 13, 5618, |
| 572 | 29915, 29879, 278, 2793, 310, 278, 1967, 29973, 13, 22933, 9047, 13566, |
| 573 | 29901]`. |
| 574 | |
| 575 | We insert 575 tokens so that including the original image token in the |
| 576 | input, there are a total of 576 (24 * 24) image tokens, which |
| 577 | corresponds to the number of image tokens inputted to the language |
| 578 | model, i.e. the number of image tokens outputted by the visual encoder. |
| 579 | |
| 580 | This way, the `positions` and `attn_metadata` are consistent |
| 581 | with the `input_ids`. |
| 582 | |
| 583 | Args: |
| 584 | input_ids: Flattened (concatenated) input_ids corresponding to a |
| 585 | batch. |
| 586 | positions: Position indices for the input tokens. |
| 587 | intermediate_tensors: Intermediate tensors from prior forward pass. |
| 588 | inputs_embeds: Optional tensor of input embeddings. |
| 589 | **kwargs: Multimodal inputs for this batch, forwarded to the |
| 590 | multimodal embedding path. |
| 591 | |
| 592 | Info: |
| 593 | [`Mistral3ImagePixelInputs`][vllm.model_executor.models.mistral3.Mistral3ImagePixelInputs] |
| 594 | |
| 595 | """ |
| 596 | if intermediate_tensors is not None: |
| 597 | inputs_embeds = None |
| 598 | |
| 599 | hidden_states = self.language_model.model( |
| 600 | input_ids, positions, intermediate_tensors, inputs_embeds=inputs_embeds |
| 601 | ) |
| 602 | |
| 603 | return hidden_states |
| 604 | |
| 605 | def compute_logits( |
| 606 | self, |
| 607 | hidden_states: torch.Tensor, |
| 608 | ) -> torch.Tensor | None: |
| 609 | return self.language_model.compute_logits(hidden_states) |
| 610 | |
| 611 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 612 | loader = AutoWeightsLoader(self) |
| 613 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 614 | |
| 615 | def get_mm_mapping(self) -> MultiModelKeys: |
| 616 | """Get the module prefix in multimodal models.""" |
| 617 | return MultiModelKeys.from_string_field( |
| 618 | language_model="language_model", |
| 619 | connector="multi_modal_projector", |
| 620 | tower_model="vision_tower", |
| 621 | ) |