| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Adapted from vllm/model_executor/models/aya_vision.py |
| 4 | """Command-A-Vision (Cohere2Vision) multimodal model implementation for vLLM.""" |
| 5 | |
| 6 | from collections.abc import Iterable, Mapping, Sequence |
| 7 | from typing import Annotated, Literal |
| 8 | |
| 9 | import torch |
| 10 | from torch import nn |
| 11 | from transformers import BatchFeature, PreTrainedConfig |
| 12 | from transformers.models.cohere2_vision import Cohere2VisionConfig |
| 13 | from transformers.models.cohere2_vision.image_processing_cohere2_vision import ( # noqa: E501 |
| 14 | Cohere2VisionImageProcessor, |
| 15 | ) |
| 16 | from transformers.models.cohere2_vision.processing_cohere2_vision import ( |
| 17 | Cohere2VisionProcessor, |
| 18 | ) |
| 19 | |
| 20 | from vllm.config import VllmConfig |
| 21 | from vllm.config.multimodal import MultiModalDummyOptions |
| 22 | from vllm.inputs import MultiModalDataDict |
| 23 | from vllm.model_executor.layers.activation import MulAndSilu |
| 24 | from vllm.model_executor.layers.linear import ( |
| 25 | MergedColumnParallelLinear, |
| 26 | RowParallelLinear, |
| 27 | ) |
| 28 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 29 | from vllm.model_executor.layers.quantization.auto_awq import AutoAWQConfig |
| 30 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 31 | from vllm.multimodal.inputs import ( |
| 32 | MultiModalFieldConfig, |
| 33 | MultiModalKwargsItems, |
| 34 | ) |
| 35 | from vllm.multimodal.parse import ImageProcessorItems, ImageSize, MultiModalDataItems |
| 36 | from vllm.multimodal.processing import ( |
| 37 | BaseDummyInputsBuilder, |
| 38 | BaseMultiModalProcessor, |
| 39 | BaseProcessingInfo, |
| 40 | PromptReplacement, |
| 41 | PromptUpdate, |
| 42 | PromptUpdateDetails, |
| 43 | cached_encode, |
| 44 | ) |
| 45 | from vllm.sequence import IntermediateTensors |
| 46 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 47 | |
| 48 | from .interfaces import ( |
| 49 | MultiModalEmbeddings, |
| 50 | SupportsMultiModal, |
| 51 | SupportsPP, |
| 52 | SupportsQuant, |
| 53 | ) |
| 54 | from .siglip import SiglipVisionModel |
| 55 | from .utils import ( |
| 56 | AutoWeightsLoader, |
| 57 | WeightsMapper, |
| 58 | init_vllm_registered_model, |
| 59 | maybe_prefix, |
| 60 | ) |
| 61 | |
| 62 | |
| 63 | class Cohere2VisionImagePixelInputs(TensorSchema): |
| 64 | """Dimensions: |
| 65 | - np: The total number of patches over each image over each prompt in |
| 66 | the batch |
| 67 | - c: Number of channels |
| 68 | - h: Height of each image patch |
| 69 | - w: Width of each image patch |
| 70 | - bn: Batch size * number of images |
| 71 | """ |
| 72 | |
| 73 | type: Literal["pixel_values"] |
| 74 | |
| 75 | pixel_values: Annotated[ |
| 76 | torch.Tensor, |
| 77 | TensorShape("np", 3, "h", "w"), |
| 78 | ] |
| 79 | |
| 80 | num_patches: Annotated[ |
| 81 | torch.Tensor, |
| 82 | TensorShape("bn"), |
| 83 | ] |
| 84 | |
| 85 | |
| 86 | class Cohere2VisionMultiModalProjector(nn.Module): |
| 87 | """Multimodal projector that maps vision features to text embedding space. |
| 88 | |
| 89 | Uses pixel shuffle downsampling followed by SwiGLU activation. |
| 90 | """ |
| 91 | |
| 92 | def __init__(self, config: Cohere2VisionConfig, prefix: str = ""): |
| 93 | super().__init__() |
| 94 | self.downsample_factor = config.downsample_factor |
| 95 | |
| 96 | # Input dimension after pixel shuffle downsampling |
| 97 | input_dim = config.vision_config.hidden_size * (config.downsample_factor**2) |
| 98 | # MergedColumnParallelLinear expects the intermediate size to be a list |
| 99 | # of sizes, so that it will load the weights as two separate linear |
| 100 | # layers before applying any parallelism. |
| 101 | # We need to divide the alignment intermediate size by 2 because |
| 102 | # the weights are merged weights of two linear layers for SwiGLU. |
| 103 | self.intermediate_size = config.alignment_intermediate_size // 2 |
| 104 | |
| 105 | self.linear_1 = MergedColumnParallelLinear( |
| 106 | input_dim, |
| 107 | [self.intermediate_size] * 2, |
| 108 | bias=True, |
| 109 | return_bias=False, |
| 110 | prefix=f"{prefix}.linear_1", |
| 111 | ) |
| 112 | self.act = MulAndSilu() |
| 113 | self.linear_2 = RowParallelLinear( |
| 114 | self.intermediate_size, |
| 115 | config.text_config.hidden_size, |
| 116 | bias=True, |
| 117 | return_bias=False, |
| 118 | prefix=f"{prefix}.linear_2", |
| 119 | ) |
| 120 | |
| 121 | def forward(self, image_features): |
| 122 | image_features = self.pixel_shuffle(image_features) |
| 123 | hidden_states = self.linear_1(image_features) |
| 124 | hidden_states = self.act(hidden_states) |
| 125 | hidden_states = self.linear_2(hidden_states) |
| 126 | return hidden_states |
| 127 | |
| 128 | def pixel_shuffle(self, image_features: torch.Tensor) -> torch.Tensor: |
| 129 | """Apply pixel shuffle downsampling to reduce spatial dimensions. |
| 130 | |
| 131 | Args: |
| 132 | image_features: Input tensor of shape [B, S, D] where S = H*W |
| 133 | |
| 134 | Returns: |
| 135 | Downsampled tensor with increased channel dimension |
| 136 | |
| 137 | """ |
| 138 | height = width = int(image_features.shape[1] ** 0.5) |
| 139 | x = image_features.reshape(image_features.shape[0], width, height, -1) |
| 140 | n, h, w, c = x.size() |
| 141 | scale_factor = 1.0 / self.downsample_factor |
| 142 | nh = int(h * scale_factor) |
| 143 | nw = int(w * scale_factor) |
| 144 | x = x.reshape(n, nh, self.downsample_factor, nw, self.downsample_factor, c) |
| 145 | x = x.permute(0, 1, 3, 2, 4, 5).contiguous() |
| 146 | x = x.reshape(n, nh, nw, -1) |
| 147 | return x |
| 148 | |
| 149 | |
| 150 | class Cohere2VisionProcessingInfo(BaseProcessingInfo): |
| 151 | def get_hf_config(self) -> Cohere2VisionConfig: |
| 152 | return self.ctx.get_hf_config(Cohere2VisionConfig) |
| 153 | |
| 154 | def get_hf_processor(self, **kwargs: object) -> Cohere2VisionProcessor: |
| 155 | return self.ctx.get_hf_processor(Cohere2VisionProcessor, **kwargs) |
| 156 | |
| 157 | def get_image_processor(self, **kwargs: object): |
| 158 | return self.get_hf_processor(**kwargs).image_processor |
| 159 | |
| 160 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 161 | return {"image": None} |
| 162 | |
| 163 | def get_image_size_with_most_features(self) -> ImageSize: |
| 164 | image_processor = self.get_image_processor() |
| 165 | height = image_processor.size["height"] |
| 166 | width = image_processor.size["width"] |
| 167 | max_patches = image_processor.max_patches |
| 168 | return ImageSize(height=height * max_patches, width=width) |
| 169 | |
| 170 | def get_num_patches( |
| 171 | self, |
| 172 | *, |
| 173 | image_width: int, |
| 174 | image_height: int, |
| 175 | processor: Cohere2VisionProcessor, |
| 176 | mm_kwargs: Mapping[str, object], |
| 177 | ) -> int: |
| 178 | """Calculate the number of image patches for a given image. |
| 179 | Uses the HF processor to determine the actual number of patches. |
| 180 | """ |
| 181 | image_processor: Cohere2VisionImageProcessor = processor.image_processor |
| 182 | |
| 183 | return image_processor.get_number_of_image_patches( |
| 184 | image_height, |
| 185 | image_width, |
| 186 | self.ctx.get_merged_mm_kwargs(mm_kwargs), |
| 187 | ) |
| 188 | |
| 189 | |
| 190 | class Cohere2VisionDummyInputsBuilder( |
| 191 | BaseDummyInputsBuilder[Cohere2VisionProcessingInfo] |
| 192 | ): |
| 193 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 194 | num_images = mm_counts.get("image", 0) |
| 195 | |
| 196 | processor = self.info.get_hf_processor() |
| 197 | image_token = processor.image_token |
| 198 | |
| 199 | return image_token * num_images |
| 200 | |
| 201 | def get_dummy_mm_data( |
| 202 | self, |
| 203 | seq_len: int, |
| 204 | mm_counts: Mapping[str, int], |
| 205 | mm_options: MultiModalDummyOptions, |
| 206 | ) -> MultiModalDataDict: |
| 207 | image_size = self.info.get_image_size_with_most_features() |
| 208 | |
| 209 | return { |
| 210 | "image": self._get_dummy_images( |
| 211 | width=image_size.width, |
| 212 | height=image_size.height, |
| 213 | num_images=mm_counts.get("image", 0), |
| 214 | overrides=mm_options.get("image"), |
| 215 | ) |
| 216 | } |
| 217 | |
| 218 | |
| 219 | class Cohere2VisionMultiModalProcessor( |
| 220 | BaseMultiModalProcessor[Cohere2VisionProcessingInfo] |
| 221 | ): |
| 222 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 223 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 224 | |
| 225 | def _postprocess_hf_mm_data( |
| 226 | self, |
| 227 | mm_data: Mapping[str, object], |
| 228 | hf_processor_mm_kwargs: Mapping[str, object], |
| 229 | processed_data: BatchFeature, |
| 230 | ) -> BatchFeature: |
| 231 | if not mm_data: |
| 232 | return processed_data |
| 233 | |
| 234 | if ( |
| 235 | "num_patches" not in processed_data |
| 236 | and (images := mm_data.get("images")) is not None |
| 237 | ): |
| 238 | hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 239 | |
| 240 | # Fallback calculation if HF processor didn't provide num_patches |
| 241 | mm_items = self.info.parse_mm_data({"image": images}, validate=False) |
| 242 | parsed_images = mm_items.get_items("image", ImageProcessorItems) |
| 243 | |
| 244 | num_patches = [ |
| 245 | self.info.get_num_patches( |
| 246 | image_width=parsed_images.get_image_size(i).width, |
| 247 | image_height=parsed_images.get_image_size(i).height, |
| 248 | processor=hf_processor, |
| 249 | mm_kwargs=hf_processor_mm_kwargs, |
| 250 | ) |
| 251 | for i in range(len(parsed_images)) |
| 252 | ] |
| 253 | processed_data["num_patches"] = torch.tensor(num_patches) |
| 254 | |
| 255 | return processed_data |
| 256 | |
| 257 | def _get_mm_fields_config( |
| 258 | self, |
| 259 | hf_inputs: BatchFeature, |
| 260 | hf_processor_mm_kwargs: Mapping[str, object], |
| 261 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 262 | num_patches = hf_inputs.get("num_patches", torch.empty(0)) |
| 263 | return dict( |
| 264 | pixel_values=MultiModalFieldConfig.flat_from_sizes("image", num_patches), |
| 265 | num_patches=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 266 | image_embeds=MultiModalFieldConfig.batched("image"), |
| 267 | ) |
| 268 | |
| 269 | def _get_prompt_updates( |
| 270 | self, |
| 271 | mm_items: MultiModalDataItems, |
| 272 | hf_processor_mm_kwargs: Mapping[str, object], |
| 273 | out_mm_kwargs: MultiModalKwargsItems, |
| 274 | ) -> Sequence[PromptUpdate]: |
| 275 | hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 276 | image_token = hf_processor.image_token |
| 277 | image_token_id = hf_processor.image_token_id |
| 278 | img_tokens_per_tile = int(hf_processor.patch_size**2) |
| 279 | img_line_break_token = hf_processor.img_line_break_token |
| 280 | boi_token = hf_processor.boi_token |
| 281 | eoi_token = hf_processor.eoi_token |
| 282 | tokenizer = self.info.get_tokenizer() |
| 283 | |
| 284 | def get_replacement(item_idx: int): |
| 285 | images = mm_items.get_items("image", ImageProcessorItems) |
| 286 | image_size: ImageSize = images.get_image_size(item_idx) |
| 287 | |
| 288 | num_patches = self.info.get_num_patches( |
| 289 | image_width=image_size.width, |
| 290 | image_height=image_size.height, |
| 291 | processor=hf_processor, |
| 292 | mm_kwargs=hf_processor_mm_kwargs, |
| 293 | ) |
| 294 | patch_tokens = image_token * img_tokens_per_tile + img_line_break_token |
| 295 | repl = f"{boi_token}{patch_tokens * num_patches}{eoi_token}" |
| 296 | |
| 297 | repl_ids = cached_encode(tokenizer, repl, add_special_tokens=False) |
| 298 | return PromptUpdateDetails.select_token_id(repl_ids, image_token_id) |
| 299 | |
| 300 | return [ |
| 301 | PromptReplacement( |
| 302 | modality="image", |
| 303 | target=[image_token_id], |
| 304 | replacement=get_replacement, |
| 305 | ) |
| 306 | ] |
| 307 | |
| 308 | |
| 309 | @MULTIMODAL_REGISTRY.register_processor( |
| 310 | Cohere2VisionMultiModalProcessor, |
| 311 | info=Cohere2VisionProcessingInfo, |
| 312 | dummy_inputs=Cohere2VisionDummyInputsBuilder, |
| 313 | ) |
| 314 | class Cohere2VisionForConditionalGeneration( |
| 315 | nn.Module, SupportsMultiModal, SupportsPP, SupportsQuant |
| 316 | ): |
| 317 | hf_to_vllm_mapper = WeightsMapper( |
| 318 | orig_to_new_prefix={ |
| 319 | "model.vision_tower.": "vision_tower.", |
| 320 | "model.multi_modal_projector.": "multi_modal_projector.", |
| 321 | "model.language_model.": "language_model.model.", |
| 322 | } |
| 323 | ) |
| 324 | |
| 325 | packed_modules_mapping = { |
| 326 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 327 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 328 | } |
| 329 | |
| 330 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 331 | super().__init__() |
| 332 | config: Cohere2VisionConfig = vllm_config.model_config.hf_config |
| 333 | quant_config = vllm_config.quant_config |
| 334 | multimodal_config = vllm_config.model_config.multimodal_config |
| 335 | self.config = config |
| 336 | self.quant_config = quant_config |
| 337 | self.multimodal_config = multimodal_config |
| 338 | self._patch_quant_config(config, quant_config) |
| 339 | |
| 340 | with self._mark_tower_model(vllm_config, "image"): |
| 341 | self.vision_tower = SiglipVisionModel( |
| 342 | config.vision_config, |
| 343 | quant_config, |
| 344 | prefix=maybe_prefix(prefix, "vision_tower"), |
| 345 | ) |
| 346 | self.multi_modal_projector = Cohere2VisionMultiModalProjector( |
| 347 | config, prefix=maybe_prefix(prefix, "multi_modal_projector") |
| 348 | ) |
| 349 | |
| 350 | with self._mark_language_model(vllm_config): |
| 351 | self.language_model = init_vllm_registered_model( |
| 352 | vllm_config=vllm_config, |
| 353 | hf_config=config.text_config, |
| 354 | prefix=maybe_prefix(prefix, "language_model"), |
| 355 | architectures=config.text_config.architectures, |
| 356 | ) |
| 357 | |
| 358 | self.make_empty_intermediate_tensors = ( |
| 359 | self.language_model.make_empty_intermediate_tensors |
| 360 | ) |
| 361 | |
| 362 | @property |
| 363 | def dtype(self): |
| 364 | return next(self.parameters()).dtype |
| 365 | |
| 366 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 367 | loader = AutoWeightsLoader(self) |
| 368 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 369 | |
| 370 | def _process_image_input( |
| 371 | self, image_input: Cohere2VisionImagePixelInputs, **kwargs: object |
| 372 | ) -> list[torch.Tensor]: |
| 373 | """Process image pixels through vision tower and projector. |
| 374 | |
| 375 | Args: |
| 376 | image_input: Validated image input containing pixel values and |
| 377 | patch counts |
| 378 | **kwargs: Unused; accepted for interface compatibility. |
| 379 | |
| 380 | Returns: |
| 381 | List of flattened image embeddings, one per image |
| 382 | |
| 383 | """ |
| 384 | pixel_values = image_input["pixel_values"] |
| 385 | num_patches = image_input["num_patches"] |
| 386 | |
| 387 | # Extract visual features |
| 388 | image_features = self.vision_tower(pixel_values) |
| 389 | |
| 390 | # Project to text embedding space |
| 391 | image_embeds = self.multi_modal_projector(image_features) |
| 392 | |
| 393 | # Split and flatten embeddings per image |
| 394 | return [e.flatten(0, 2) for e in image_embeds.split(num_patches.tolist())] |
| 395 | |
| 396 | def _parse_and_validate_image_input( |
| 397 | self, **kwargs: object |
| 398 | ) -> Cohere2VisionImagePixelInputs | None: |
| 399 | pixel_values = kwargs.pop("pixel_values", None) |
| 400 | num_patches = kwargs.pop("num_patches", None) |
| 401 | image_embeds = kwargs.pop("image_embeds", None) |
| 402 | assert image_embeds is None, "Cohere2Vision does not support image_embeds." |
| 403 | |
| 404 | if pixel_values is None: |
| 405 | return None |
| 406 | |
| 407 | return Cohere2VisionImagePixelInputs( |
| 408 | type="pixel_values", |
| 409 | pixel_values=pixel_values, |
| 410 | num_patches=num_patches, |
| 411 | resolve_bindings={ |
| 412 | "h": self.config.vision_config.image_size, |
| 413 | "w": self.config.vision_config.image_size, |
| 414 | }, |
| 415 | ) |
| 416 | |
| 417 | def _patch_quant_config( |
| 418 | self, config: PreTrainedConfig, quant_config: QuantizationConfig | None |
| 419 | ): |
| 420 | # the awq models from OpenGVLab missing `modules_to_not_convert` |
| 421 | # patch the quant_config to add `modules_to_not_convert` back |
| 422 | if isinstance(quant_config, AutoAWQConfig): |
| 423 | text_config = config.text_config |
| 424 | llm_quant_config = getattr(text_config, "quantization_config", None) |
| 425 | if (not quant_config.modules_to_not_convert) and ( |
| 426 | llm_quant_config is not None |
| 427 | ): |
| 428 | quant_config.modules_to_not_convert.append("vision_tower") |
| 429 | |
| 430 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 431 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 432 | if image_input is None: |
| 433 | return [] |
| 434 | |
| 435 | return self._process_image_input(image_input, **kwargs) |
| 436 | |
| 437 | def forward( |
| 438 | self, |
| 439 | input_ids: torch.Tensor | None, |
| 440 | positions: torch.Tensor, |
| 441 | intermediate_tensors: IntermediateTensors | None = None, |
| 442 | inputs_embeds: torch.Tensor | None = None, |
| 443 | **kwargs: object, |
| 444 | ) -> torch.Tensor | IntermediateTensors: |
| 445 | if intermediate_tensors is not None: |
| 446 | inputs_embeds = None |
| 447 | |
| 448 | hidden_states = self.language_model.model( |
| 449 | input_ids=input_ids, |
| 450 | positions=positions, |
| 451 | intermediate_tensors=intermediate_tensors, |
| 452 | inputs_embeds=inputs_embeds, |
| 453 | ) |
| 454 | return hidden_states |
| 455 | |
| 456 | def compute_logits( |
| 457 | self, |
| 458 | hidden_states: torch.Tensor, |
| 459 | ) -> torch.Tensor | None: |
| 460 | return self.language_model.compute_logits(hidden_states) |