| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import itertools |
| 5 | import math |
| 6 | import typing |
| 7 | from collections.abc import Hashable, Iterable, Mapping, Sequence |
| 8 | from typing import Annotated, Any, Literal |
| 9 | |
| 10 | import torch |
| 11 | import torch.nn as nn |
| 12 | from transformers import BatchFeature |
| 13 | from transformers.activations import ACT2FN |
| 14 | from transformers.models.lfm2_vl import Lfm2VlProcessor |
| 15 | from transformers.models.lfm2_vl.configuration_lfm2_vl import Lfm2VlConfig |
| 16 | from transformers.models.lfm2_vl.image_processing_lfm2_vl import ( |
| 17 | Lfm2VlImageProcessor, |
| 18 | find_closest_aspect_ratio, |
| 19 | round_by_factor, |
| 20 | ) |
| 21 | from typing_extensions import Buffer |
| 22 | |
| 23 | from vllm.config import VllmConfig |
| 24 | from vllm.config.multimodal import MultiModalDummyOptions |
| 25 | from vllm.forward_context import set_forward_context |
| 26 | from vllm.inputs import MultiModalDataDict |
| 27 | from vllm.model_executor.layers.linear import ReplicatedLinear |
| 28 | from vllm.model_executor.layers.mamba.mamba_utils import ( |
| 29 | MambaStateCopyFunc, |
| 30 | MambaStateCopyFuncCalculator, |
| 31 | MambaStateDtypeCalculator, |
| 32 | MambaStateShapeCalculator, |
| 33 | ) |
| 34 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 35 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 36 | from vllm.multimodal.inputs import ( |
| 37 | MultiModalFieldConfig, |
| 38 | MultiModalKwargsItem, |
| 39 | MultiModalKwargsItems, |
| 40 | ) |
| 41 | from vllm.multimodal.parse import ImageProcessorItems, ImageSize, MultiModalDataItems |
| 42 | from vllm.multimodal.processing import ( |
| 43 | BaseDummyInputsBuilder, |
| 44 | BaseMultiModalProcessor, |
| 45 | BaseProcessingInfo, |
| 46 | PromptReplacement, |
| 47 | PromptUpdateDetails, |
| 48 | cached_encode, |
| 49 | ) |
| 50 | from vllm.renderers import TokenizeParams |
| 51 | from vllm.sequence import IntermediateTensors |
| 52 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 53 | |
| 54 | from .interfaces import ( |
| 55 | IsHybrid, |
| 56 | MultiModalEmbeddings, |
| 57 | SupportsEncoderCudaGraph, |
| 58 | SupportsLoRA, |
| 59 | SupportsMultiModal, |
| 60 | SupportsPP, |
| 61 | ) |
| 62 | from .lfm2_siglip2 import Siglip2Model |
| 63 | from .utils import ( |
| 64 | AutoWeightsLoader, |
| 65 | WeightsMapper, |
| 66 | init_vllm_registered_model, |
| 67 | maybe_prefix, |
| 68 | ) |
| 69 | from .vision import is_vit_use_data_parallel |
| 70 | |
| 71 | |
| 72 | def _pad_cumulative_seqlens_buffer( |
| 73 | dst: torch.Tensor, |
| 74 | src: torch.Tensor, |
| 75 | ) -> None: |
| 76 | n = src.shape[0] |
| 77 | dst.zero_() |
| 78 | dst[:n].copy_(src) |
| 79 | if n < dst.shape[0]: |
| 80 | dst[n:] = src[-1] |
| 81 | |
| 82 | |
| 83 | class Lfm2VLImagePixelInputs(TensorSchema): |
| 84 | """Dimensions: |
| 85 | - b: Number of images in the prompt |
| 86 | - bn: Batch size * number of images |
| 87 | - d: Number of dimensions |
| 88 | - fd: Number of features per dimension |
| 89 | """ |
| 90 | |
| 91 | type: Literal["pixel_values"] = "pixel_values" |
| 92 | pixel_values: Annotated[torch.Tensor, TensorShape("bn", "d", "fd")] |
| 93 | spatial_shapes: Annotated[torch.Tensor, TensorShape("bn", 2)] |
| 94 | num_patches: Annotated[torch.Tensor, TensorShape("b")] |
| 95 | |
| 96 | |
| 97 | LFM2VLImageInputs = Lfm2VLImagePixelInputs |
| 98 | |
| 99 | |
| 100 | class Lfm2VLProcessingInfo(BaseProcessingInfo): |
| 101 | def get_hf_config(self): |
| 102 | return self.ctx.get_hf_config(Lfm2VlConfig) |
| 103 | |
| 104 | def get_hf_processor(self, **kwargs): |
| 105 | return self.ctx.get_hf_processor(Lfm2VlProcessor, **kwargs) |
| 106 | |
| 107 | def get_image_processor(self, **kwargs: object) -> Lfm2VlImageProcessor: |
| 108 | return self.get_hf_processor(**kwargs).image_processor |
| 109 | |
| 110 | def get_default_tok_params(self) -> TokenizeParams: |
| 111 | return super().get_default_tok_params().with_kwargs(add_special_tokens=False) |
| 112 | |
| 113 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 114 | return {"image": None} |
| 115 | |
| 116 | def get_image_size_with_most_features(self) -> ImageSize: |
| 117 | processor = self.get_image_processor() |
| 118 | max_image_tokens = processor.max_image_tokens |
| 119 | encoder_patch_size = processor.encoder_patch_size |
| 120 | downsample_factor = processor.downsample_factor |
| 121 | max_pixels = max_image_tokens * (encoder_patch_size**2) * (downsample_factor**2) |
| 122 | side = int(math.sqrt(max_pixels)) |
| 123 | return ImageSize(width=side, height=side) |
| 124 | |
| 125 | def _is_image_too_large( |
| 126 | self, |
| 127 | height: int, |
| 128 | width: int, |
| 129 | max_image_tokens: int, |
| 130 | encoder_patch_size: int, |
| 131 | downsample_factor: int, |
| 132 | max_pixels_tolerance: float, |
| 133 | ) -> bool: |
| 134 | """Check if the image is too large to be processed as one tile.""" |
| 135 | total_factor = encoder_patch_size * downsample_factor |
| 136 | |
| 137 | h_bar = max(encoder_patch_size, round_by_factor(height, total_factor)) |
| 138 | w_bar = max(encoder_patch_size, round_by_factor(width, total_factor)) |
| 139 | return ( |
| 140 | h_bar * w_bar |
| 141 | > max_image_tokens |
| 142 | * encoder_patch_size**2 |
| 143 | * downsample_factor**2 |
| 144 | * max_pixels_tolerance |
| 145 | ) |
| 146 | |
| 147 | def smart_resize( |
| 148 | self, |
| 149 | height: int, |
| 150 | width: int, |
| 151 | downsample_factor: int, |
| 152 | min_image_tokens: int, |
| 153 | max_image_tokens: int, |
| 154 | encoder_patch_size: int, |
| 155 | ) -> tuple[int, int]: |
| 156 | total_factor = encoder_patch_size * downsample_factor |
| 157 | smart_resize_min_pixels = ( |
| 158 | min_image_tokens * encoder_patch_size**2 * downsample_factor**2 |
| 159 | ) |
| 160 | smart_resize_max_pixels = ( |
| 161 | max_image_tokens * encoder_patch_size**2 * downsample_factor**2 |
| 162 | ) |
| 163 | |
| 164 | h_bar = max(total_factor, round_by_factor(height, total_factor)) |
| 165 | w_bar = max(total_factor, round_by_factor(width, total_factor)) |
| 166 | |
| 167 | if h_bar * w_bar > smart_resize_max_pixels: |
| 168 | beta = math.sqrt((height * width) / smart_resize_max_pixels) |
| 169 | h_bar = max( |
| 170 | total_factor, math.floor(height / beta / total_factor) * total_factor |
| 171 | ) |
| 172 | w_bar = max( |
| 173 | total_factor, math.floor(width / beta / total_factor) * total_factor |
| 174 | ) |
| 175 | elif h_bar * w_bar < smart_resize_min_pixels: |
| 176 | beta = math.sqrt(smart_resize_min_pixels / (height * width)) |
| 177 | h_bar = math.ceil(height * beta / total_factor) * total_factor |
| 178 | w_bar = math.ceil(width * beta / total_factor) * total_factor |
| 179 | |
| 180 | return w_bar, h_bar |
| 181 | |
| 182 | def _target_ratios(self, min_tiles: int, max_tiles: int) -> list[tuple[int, int]]: |
| 183 | ratios = [ |
| 184 | (w, h) |
| 185 | for n in range(min_tiles, max_tiles + 1) |
| 186 | for w in range(1, n + 1) |
| 187 | for h in range(1, n + 1) |
| 188 | if min_tiles <= w * h <= max_tiles |
| 189 | ] |
| 190 | return sorted(set(ratios), key=lambda x: x[0] * x[1]) |
| 191 | |
| 192 | def _get_grid_layout( |
| 193 | self, |
| 194 | height: int, |
| 195 | width: int, |
| 196 | min_tiles: int, |
| 197 | max_tiles: int, |
| 198 | tile_size: int, |
| 199 | ) -> tuple[int, int, int]: |
| 200 | aspect_ratio = width / height |
| 201 | target_ratios = self._target_ratios(min_tiles, max_tiles) |
| 202 | # find best matching grid configuration |
| 203 | grid_width, grid_height = find_closest_aspect_ratio( |
| 204 | aspect_ratio, target_ratios, width, height, tile_size |
| 205 | ) |
| 206 | total_patches = grid_width * grid_height |
| 207 | return grid_width, grid_height, total_patches |
| 208 | |
| 209 | def _get_image_feature_grid_size( |
| 210 | self, |
| 211 | image_width: int, |
| 212 | image_height: int, |
| 213 | processor: Lfm2VlProcessor, |
| 214 | mm_kwargs: Mapping[str, object], |
| 215 | ) -> tuple[int, int, int]: |
| 216 | image_processor: Lfm2VlImageProcessor = processor.image_processor |
| 217 | |
| 218 | mm_kwargs = self.ctx.get_merged_mm_kwargs(mm_kwargs) |
| 219 | downsample_factor = mm_kwargs.get( |
| 220 | "downsample_factor", image_processor.downsample_factor |
| 221 | ) |
| 222 | encoder_patch_size = mm_kwargs.get( |
| 223 | "encoder_patch_size", image_processor.encoder_patch_size |
| 224 | ) |
| 225 | max_pixels_tolerance = mm_kwargs.get( |
| 226 | "max_pixels_tolerance", image_processor.max_pixels_tolerance |
| 227 | ) |
| 228 | min_tiles = mm_kwargs.get("min_tiles", image_processor.min_tiles) |
| 229 | max_tiles = mm_kwargs.get("max_tiles", image_processor.max_tiles) |
| 230 | max_image_tokens = mm_kwargs.get( |
| 231 | "max_image_tokens", image_processor.max_image_tokens |
| 232 | ) |
| 233 | tile_size = mm_kwargs.get("tile_size", image_processor.tile_size) |
| 234 | assert isinstance(downsample_factor, int) |
| 235 | assert isinstance(encoder_patch_size, int) |
| 236 | assert isinstance(max_pixels_tolerance, int | float) |
| 237 | assert isinstance(min_tiles, int) |
| 238 | assert isinstance(max_tiles, int) |
| 239 | assert isinstance(max_image_tokens, int) |
| 240 | assert isinstance(tile_size, int) |
| 241 | |
| 242 | do_image_splitting = not min_tiles == max_tiles == 1 |
| 243 | is_image_large = self._is_image_too_large( |
| 244 | height=image_height, |
| 245 | width=image_width, |
| 246 | max_image_tokens=max_image_tokens, |
| 247 | encoder_patch_size=encoder_patch_size, |
| 248 | downsample_factor=downsample_factor, |
| 249 | max_pixels_tolerance=max_pixels_tolerance, |
| 250 | ) |
| 251 | |
| 252 | # Big image will be cropped into patches and small images are just resized |
| 253 | if is_image_large and do_image_splitting: |
| 254 | grid_width, grid_height, total_patches = self._get_grid_layout( |
| 255 | image_height, |
| 256 | image_width, |
| 257 | min_tiles=min_tiles, |
| 258 | max_tiles=max_tiles, |
| 259 | tile_size=tile_size, |
| 260 | ) |
| 261 | else: |
| 262 | grid_width = grid_height = total_patches = 1 |
| 263 | |
| 264 | if grid_width * grid_height != 1: # Thumbnail |
| 265 | total_patches += 1 |
| 266 | |
| 267 | return grid_width, grid_height, total_patches |
| 268 | |
| 269 | def get_num_patches( |
| 270 | self, |
| 271 | *, |
| 272 | image_width: int, |
| 273 | image_height: int, |
| 274 | processor: Lfm2VlProcessor, |
| 275 | mm_kwargs: Mapping[str, object], |
| 276 | ) -> int: |
| 277 | _, _, total_patches = self._get_image_feature_grid_size( |
| 278 | image_width=image_width, |
| 279 | image_height=image_height, |
| 280 | processor=processor, |
| 281 | mm_kwargs=mm_kwargs, |
| 282 | ) |
| 283 | return total_patches |
| 284 | |
| 285 | def get_image_repl( |
| 286 | self, |
| 287 | image_width: int, |
| 288 | image_height: int, |
| 289 | spatial_shapes: torch.Tensor, |
| 290 | processor: Lfm2VlProcessor, |
| 291 | mm_kwargs: Mapping[str, object], |
| 292 | ) -> str: |
| 293 | grid_placeholder = "<|img_row_{n_h}_col_{n_w}|>" |
| 294 | image_token = processor.image_token |
| 295 | image_start_token = processor.image_start_token |
| 296 | image_end_token = processor.image_end_token |
| 297 | image_thumbnail_token = processor.image_thumbnail_token |
| 298 | |
| 299 | num_thumbnail_tokens, num_tokens_per_tile = self.get_num_image_tokens( |
| 300 | spatial_shapes=spatial_shapes, |
| 301 | processor=processor, |
| 302 | mm_kwargs=mm_kwargs, |
| 303 | ) |
| 304 | tile_img_placeholder = grid_placeholder + (image_token * num_tokens_per_tile) |
| 305 | |
| 306 | grid_w, grid_h, _ = self._get_image_feature_grid_size( |
| 307 | image_width=image_width, |
| 308 | image_height=image_height, |
| 309 | processor=processor, |
| 310 | mm_kwargs=mm_kwargs, |
| 311 | ) |
| 312 | |
| 313 | if grid_w > 1 or grid_h > 1: |
| 314 | tiles_placeholder: list[str] = [ |
| 315 | tile_img_placeholder.format(n_h=i + 1, n_w=j + 1) |
| 316 | for i in range(grid_h) |
| 317 | for j in range(grid_w) |
| 318 | ] |
| 319 | |
| 320 | if num_thumbnail_tokens > 0: |
| 321 | tiles_placeholder.append( |
| 322 | image_thumbnail_token + (image_token * num_thumbnail_tokens) |
| 323 | ) |
| 324 | else: |
| 325 | tiles_placeholder = [image_token * num_thumbnail_tokens] |
| 326 | |
| 327 | placeholder = "".join( |
| 328 | itertools.chain([image_start_token], tiles_placeholder, [image_end_token]) |
| 329 | ) |
| 330 | return placeholder |
| 331 | |
| 332 | def get_num_image_tokens( |
| 333 | self, |
| 334 | *, |
| 335 | spatial_shapes: torch.Tensor, |
| 336 | processor: Lfm2VlProcessor, |
| 337 | mm_kwargs: Mapping[str, object], |
| 338 | ) -> tuple[int, int]: |
| 339 | image_processor: Lfm2VlImageProcessor = processor.image_processor |
| 340 | |
| 341 | mm_kwargs = self.ctx.get_merged_mm_kwargs(mm_kwargs) |
| 342 | downsample_factor = mm_kwargs.get( |
| 343 | "downsample_factor", image_processor.downsample_factor |
| 344 | ) |
| 345 | encoder_patch_size = mm_kwargs.get( |
| 346 | "encoder_patch_size", image_processor.encoder_patch_size |
| 347 | ) |
| 348 | tile_size = mm_kwargs.get("tile_size", image_processor.tile_size) |
| 349 | assert isinstance(downsample_factor, int) |
| 350 | assert isinstance(encoder_patch_size, int) |
| 351 | assert isinstance(tile_size, int) |
| 352 | |
| 353 | thumbnail_height_patches = int(spatial_shapes[-1][0].item()) |
| 354 | thumbnail_width_patches = int(spatial_shapes[-1][1].item()) |
| 355 | # HF computes thumbnail tokens as |
| 356 | # ceil(h_patches / downsample_factor) * ceil(w_patches / downsample_factor). |
| 357 | # We assert divisibility here so any processor/model drift is surfaced |
| 358 | # immediately instead of being hidden by floor division. |
| 359 | assert thumbnail_height_patches % downsample_factor == 0, ( |
| 360 | "LFM2-VL thumbnail height patch grid must be divisible by " |
| 361 | f"downsample_factor, got height_patches={thumbnail_height_patches}, " |
| 362 | f"downsample_factor={downsample_factor}" |
| 363 | ) |
| 364 | assert thumbnail_width_patches % downsample_factor == 0, ( |
| 365 | "LFM2-VL thumbnail width patch grid must be divisible by " |
| 366 | f"downsample_factor, got width_patches={thumbnail_width_patches}, " |
| 367 | f"downsample_factor={downsample_factor}" |
| 368 | ) |
| 369 | num_thumbnail_tokens = math.ceil( |
| 370 | thumbnail_height_patches / downsample_factor |
| 371 | ) * math.ceil(thumbnail_width_patches / downsample_factor) |
| 372 | num_patches_tile = tile_size // encoder_patch_size |
| 373 | dwn_num_patches_tile = math.ceil(num_patches_tile / downsample_factor) |
| 374 | num_tiles_tokens = dwn_num_patches_tile * dwn_num_patches_tile |
| 375 | |
| 376 | return num_thumbnail_tokens, num_tiles_tokens |
| 377 | |
| 378 | |
| 379 | class Lfm2VLDummyInputsBuilder(BaseDummyInputsBuilder[Lfm2VLProcessingInfo]): |
| 380 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 381 | num_images = mm_counts.get("image", 0) |
| 382 | processor = self.info.get_hf_processor() |
| 383 | image_token = processor.image_token |
| 384 | return image_token * num_images |
| 385 | |
| 386 | def get_dummy_mm_data( |
| 387 | self, |
| 388 | seq_len: int, |
| 389 | mm_counts: Mapping[str, int], |
| 390 | mm_options: MultiModalDummyOptions, |
| 391 | ) -> MultiModalDataDict: |
| 392 | target_width, target_height = self.info.get_image_size_with_most_features() |
| 393 | |
| 394 | return { |
| 395 | "image": self._get_dummy_images( |
| 396 | width=target_width, |
| 397 | height=target_height, |
| 398 | num_images=mm_counts.get("image", 0), |
| 399 | overrides=mm_options.get("image"), |
| 400 | ), |
| 401 | } |
| 402 | |
| 403 | |
| 404 | class Lfm2VLMultiModalProcessor(BaseMultiModalProcessor[Lfm2VLProcessingInfo]): |
| 405 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 406 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 407 | |
| 408 | def _postprocess_hf_mm_data( |
| 409 | self, |
| 410 | mm_data: Mapping[str, object], |
| 411 | hf_processor_mm_kwargs: Mapping[str, object], |
| 412 | processed_data: BatchFeature, |
| 413 | ) -> BatchFeature: |
| 414 | if not mm_data: |
| 415 | return processed_data |
| 416 | |
| 417 | images = mm_data.get("images", []) |
| 418 | mm_items = self.info.parse_mm_data({"image": images}, validate=False) |
| 419 | parsed_images = mm_items.get_items("image", ImageProcessorItems) |
| 420 | image_sizes = [ |
| 421 | parsed_images.get_image_size(i) for i in range(len(parsed_images)) |
| 422 | ] |
| 423 | hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 424 | |
| 425 | num_patches = [ |
| 426 | self.info.get_num_patches( |
| 427 | image_width=size.width, |
| 428 | image_height=size.height, |
| 429 | processor=hf_processor, |
| 430 | mm_kwargs=hf_processor_mm_kwargs, |
| 431 | ) |
| 432 | for size in image_sizes |
| 433 | ] |
| 434 | processed_data["num_patches"] = torch.tensor(num_patches) |
| 435 | |
| 436 | return processed_data |
| 437 | |
| 438 | def _get_mm_fields_config( |
| 439 | self, |
| 440 | hf_inputs: BatchFeature, |
| 441 | hf_processor_mm_kwargs: Mapping[str, object], |
| 442 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 443 | num_patches = hf_inputs.get("num_patches", torch.empty(0)) |
| 444 | |
| 445 | return dict[str, MultiModalFieldConfig]( |
| 446 | pixel_values=MultiModalFieldConfig.flat_from_sizes("image", num_patches), |
| 447 | spatial_shapes=MultiModalFieldConfig.flat_from_sizes( |
| 448 | "image", num_patches, keep_on_cpu=True |
| 449 | ), |
| 450 | num_patches=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 451 | ) |
| 452 | |
| 453 | def _get_prompt_updates( |
| 454 | self, |
| 455 | mm_items: MultiModalDataItems, |
| 456 | hf_processor_mm_kwargs: Mapping[str, object], |
| 457 | out_mm_kwargs: MultiModalKwargsItems, |
| 458 | ) -> Sequence[PromptReplacement]: |
| 459 | hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 460 | image_token_id = hf_processor.image_token_id |
| 461 | tokenizer = self.info.get_tokenizer() |
| 462 | |
| 463 | def get_image_replacement_lfm2vl(item_idx: int): |
| 464 | images = mm_items.get_items("image", ImageProcessorItems) |
| 465 | image_size = images.get_image_size(item_idx) |
| 466 | out_item = out_mm_kwargs["image"][item_idx] |
| 467 | spatial_shapes = out_item["spatial_shapes"].data |
| 468 | assert isinstance(spatial_shapes, torch.Tensor) |
| 469 | image_repl = self.info.get_image_repl( |
| 470 | image_width=image_size.width, |
| 471 | image_height=image_size.height, |
| 472 | spatial_shapes=spatial_shapes, |
| 473 | processor=hf_processor, |
| 474 | mm_kwargs=hf_processor_mm_kwargs, |
| 475 | ) |
| 476 | image_repl_ids = cached_encode( |
| 477 | tokenizer, image_repl, add_special_tokens=False |
| 478 | ) |
| 479 | return PromptUpdateDetails.select_token_id( |
| 480 | image_repl_ids, |
| 481 | image_token_id, |
| 482 | ) |
| 483 | |
| 484 | return [ |
| 485 | PromptReplacement( |
| 486 | modality="image", |
| 487 | target=[image_token_id], |
| 488 | replacement=get_image_replacement_lfm2vl, |
| 489 | ) |
| 490 | ] |
| 491 | |
| 492 | |
| 493 | class Lfm2VLMultiModalProjector(nn.Module): |
| 494 | def __init__( |
| 495 | self, |
| 496 | config: Lfm2VlConfig, |
| 497 | prefix: str = "", |
| 498 | ): |
| 499 | super().__init__() |
| 500 | self.use_data_parallel = is_vit_use_data_parallel() |
| 501 | |
| 502 | in_channels = config.vision_config.hidden_size * (config.downsample_factor**2) |
| 503 | self.factor = config.downsample_factor |
| 504 | self.projector_use_layernorm = config.projector_use_layernorm |
| 505 | if self.projector_use_layernorm: |
| 506 | self.layer_norm = nn.LayerNorm(in_channels) |
| 507 | self.linear_1 = ReplicatedLinear( |
| 508 | in_channels, |
| 509 | config.projector_hidden_size, |
| 510 | bias=config.projector_bias, |
| 511 | prefix=maybe_prefix(prefix, "linear_1"), |
| 512 | return_bias=False, |
| 513 | ) |
| 514 | self.act = ACT2FN[config.projector_hidden_act] |
| 515 | self.linear_2 = ReplicatedLinear( |
| 516 | config.projector_hidden_size, |
| 517 | config.text_config.hidden_size, |
| 518 | bias=config.projector_bias, |
| 519 | prefix=maybe_prefix(prefix, "linear_2"), |
| 520 | return_bias=False, |
| 521 | ) |
| 522 | |
| 523 | def forward( |
| 524 | self, |
| 525 | vision_features_packed: torch.Tensor, |
| 526 | spatial_shapes: torch.Tensor, |
| 527 | ) -> torch.Tensor: |
| 528 | """Project packed vision features without materializing padded tensors. |
| 529 | |
| 530 | Args: |
| 531 | vision_features_packed: (total_tokens, hidden_size) packed in tile order. |
| 532 | spatial_shapes: (num_tiles, 2) on CPU (height, width) per tile. |
| 533 | |
| 534 | Returns: |
| 535 | projected_packed: (total_projected_tokens, text_hidden_size) |
| 536 | |
| 537 | """ |
| 538 | assert spatial_shapes.device.type == "cpu", ( |
| 539 | "Expected `spatial_shapes` on CPU to avoid device-to-host sync in " |
| 540 | "variable-length packing." |
| 541 | ) |
| 542 | factor = self.factor |
| 543 | device = vision_features_packed.device |
| 544 | hidden_size = vision_features_packed.shape[-1] |
| 545 | |
| 546 | spatial_shapes_list: list[list[int]] = spatial_shapes.tolist() |
| 547 | lengths_list = [h * w for h, w in spatial_shapes_list] |
| 548 | |
| 549 | gather_idx_parts: list[torch.Tensor] = [] |
| 550 | offset = 0 |
| 551 | |
| 552 | dh = torch.arange(factor, dtype=torch.int64) |
| 553 | dw = torch.arange(factor, dtype=torch.int64) |
| 554 | dh_grid, dw_grid = torch.meshgrid(dh, dw, indexing="ij") |
| 555 | dh_flat = dh_grid.reshape(-1) |
| 556 | dw_flat = dw_grid.reshape(-1) |
| 557 | |
| 558 | for (height, width), length in zip(spatial_shapes_list, lengths_list): |
| 559 | if length <= 0: |
| 560 | continue |
| 561 | if height % factor != 0 or width % factor != 0: |
| 562 | raise ValueError( |
| 563 | "spatial_shapes must be divisible by downsample_factor: " |
| 564 | f"got ({height}, {width}) with factor={factor}." |
| 565 | ) |
| 566 | height_out = height // factor |
| 567 | width_out = width // factor |
| 568 | |
| 569 | rows_out = torch.arange(height_out, dtype=torch.int64) |
| 570 | cols_out = torch.arange(width_out, dtype=torch.int64) |
| 571 | rr, cc = torch.meshgrid(rows_out, cols_out, indexing="ij") |
| 572 | rr = rr.reshape(-1) |
| 573 | cc = cc.reshape(-1) |
| 574 | |
| 575 | token_idx = (rr[:, None] * factor + dh_flat[None, :]) * width + ( |
| 576 | cc[:, None] * factor + dw_flat[None, :] |
| 577 | ) |
| 578 | gather_idx_parts.append(token_idx.reshape(-1) + offset) |
| 579 | offset += length |
| 580 | |
| 581 | if gather_idx_parts: |
| 582 | gather_idx = torch.cat(gather_idx_parts).to(device=device) |
| 583 | return self.forward_with_gather_idx(vision_features_packed, gather_idx) |
| 584 | else: |
| 585 | unshuffled = vision_features_packed.new_empty( |
| 586 | (0, factor * factor * hidden_size) |
| 587 | ) |
| 588 | |
| 589 | return self.forward_from_unshuffled(unshuffled) |
| 590 | |
| 591 | def forward_with_gather_idx( |
| 592 | self, |
| 593 | vision_features_packed: torch.Tensor, |
| 594 | gather_idx: torch.Tensor, |
| 595 | ) -> torch.Tensor: |
| 596 | hidden_size = vision_features_packed.shape[-1] |
| 597 | factor = self.factor |
| 598 | gathered = vision_features_packed.index_select(0, gather_idx) |
| 599 | unshuffled = gathered.reshape(-1, factor * factor * hidden_size) |
| 600 | return self.forward_from_unshuffled(unshuffled) |
| 601 | |
| 602 | def forward_from_unshuffled(self, unshuffled: torch.Tensor) -> torch.Tensor: |
| 603 | if self.projector_use_layernorm: |
| 604 | unshuffled = self.layer_norm(unshuffled) |
| 605 | hidden_states = self.linear_1(unshuffled) |
| 606 | hidden_states = self.act(hidden_states) |
| 607 | projected_packed = self.linear_2(hidden_states) |
| 608 | return projected_packed |
| 609 | |
| 610 | |
| 611 | @MULTIMODAL_REGISTRY.register_processor( |
| 612 | Lfm2VLMultiModalProcessor, |
| 613 | info=Lfm2VLProcessingInfo, |
| 614 | dummy_inputs=Lfm2VLDummyInputsBuilder, |
| 615 | ) |
| 616 | class Lfm2VLForConditionalGeneration( |
| 617 | nn.Module, |
| 618 | SupportsMultiModal, |
| 619 | SupportsEncoderCudaGraph, |
| 620 | SupportsLoRA, |
| 621 | SupportsPP, |
| 622 | IsHybrid, |
| 623 | ): |
| 624 | supports_tower_connector_lora = True |
| 625 | merge_by_field_config = True |
| 626 | |
| 627 | hf_to_vllm_mapper = WeightsMapper( |
| 628 | orig_to_new_prefix={ |
| 629 | "lm_head.": "language_model.lm_head.", |
| 630 | "model.language_model.": "language_model.model.", |
| 631 | "model.vision_tower.": "vision_tower.", |
| 632 | "model.multi_modal_projector.": "multi_modal_projector.", |
| 633 | } |
| 634 | ) |
| 635 | |
| 636 | @classmethod |
| 637 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 638 | if modality.startswith("image"): |
| 639 | return "<image>" |
| 640 | |
| 641 | raise ValueError("Only image modality is supported") |
| 642 | |
| 643 | @classmethod |
| 644 | def get_mamba_state_dtype_from_config( |
| 645 | cls, |
| 646 | vllm_config: "VllmConfig", |
| 647 | ) -> tuple[torch.dtype, ...]: |
| 648 | return MambaStateDtypeCalculator.short_conv_state_dtype( |
| 649 | vllm_config.model_config.dtype, |
| 650 | vllm_config.cache_config.mamba_cache_dtype, |
| 651 | ) |
| 652 | |
| 653 | @classmethod |
| 654 | def get_mamba_state_shape_from_config( |
| 655 | cls, |
| 656 | vllm_config: "VllmConfig", |
| 657 | ) -> tuple[tuple[int, int]]: |
| 658 | """Calculate shapes for LFM2's convolutional cache. |
| 659 | |
| 660 | Args: |
| 661 | vllm_config: vLLM config |
| 662 | |
| 663 | Returns: |
| 664 | Tuple containing: |
| 665 | - conv_state_shape: Shape for convolutional state cache |
| 666 | |
| 667 | """ |
| 668 | parallel_config = vllm_config.parallel_config |
| 669 | hf_language_config = vllm_config.model_config.hf_config.text_config |
| 670 | |
| 671 | return MambaStateShapeCalculator.short_conv_state_shape( |
| 672 | tp_world_size=parallel_config.tensor_parallel_size, |
| 673 | intermediate_size=hf_language_config.hidden_size, |
| 674 | conv_kernel=hf_language_config.conv_L_cache, |
| 675 | ) |
| 676 | |
| 677 | @classmethod |
| 678 | def get_mamba_state_copy_func(cls) -> tuple[MambaStateCopyFunc]: |
| 679 | return MambaStateCopyFuncCalculator.short_conv_state_copy_func() |
| 680 | |
| 681 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "model"): |
| 682 | super().__init__() |
| 683 | config: Lfm2VlConfig = vllm_config.model_config.hf_config |
| 684 | multimodal_config = vllm_config.model_config.multimodal_config |
| 685 | assert multimodal_config is not None |
| 686 | vision_config = config.vision_config |
| 687 | quant_config = vllm_config.quant_config |
| 688 | |
| 689 | self.config = config |
| 690 | self.vllm_config = vllm_config |
| 691 | self.model_config = vllm_config.model_config |
| 692 | self.multimodal_config = multimodal_config |
| 693 | self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data" |
| 694 | |
| 695 | with self._mark_tower_model(vllm_config, "image"): |
| 696 | if vision_config.model_type == "siglip2_vision_model": |
| 697 | self.vision_tower = Siglip2Model( |
| 698 | config=vision_config, |
| 699 | quant_config=quant_config, |
| 700 | prefix=maybe_prefix(prefix, "vision_tower"), |
| 701 | ) |
| 702 | else: |
| 703 | raise ValueError( |
| 704 | f"Unsupported visual tokenizer type: {vision_config.model_type}" |
| 705 | ) |
| 706 | |
| 707 | self.multi_modal_projector = Lfm2VLMultiModalProjector( |
| 708 | config=config, |
| 709 | prefix=maybe_prefix(prefix, "multi_modal_projector"), |
| 710 | ) |
| 711 | |
| 712 | with self._mark_language_model(vllm_config): |
| 713 | self.language_model = init_vllm_registered_model( |
| 714 | vllm_config=vllm_config, |
| 715 | hf_config=config.text_config, |
| 716 | prefix=maybe_prefix(prefix, "language"), |
| 717 | architectures=config.text_config.architectures, |
| 718 | ) |
| 719 | |
| 720 | self.make_empty_intermediate_tensors = ( |
| 721 | self.language_model.make_empty_intermediate_tensors |
| 722 | ) |
| 723 | |
| 724 | def _parse_and_validate_image_input( |
| 725 | self, **kwargs: object |
| 726 | ) -> LFM2VLImageInputs | None: |
| 727 | pixel_values = kwargs.pop("pixel_values", None) |
| 728 | spatial_shapes = kwargs.pop("spatial_shapes", None) |
| 729 | num_patches = kwargs.pop("num_patches", None) |
| 730 | if pixel_values is None: |
| 731 | return None |
| 732 | |
| 733 | return LFM2VLImageInputs( |
| 734 | type="pixel_values", |
| 735 | pixel_values=pixel_values, |
| 736 | spatial_shapes=spatial_shapes, |
| 737 | num_patches=num_patches, |
| 738 | ) |
| 739 | |
| 740 | def image_pixels_to_features( |
| 741 | self, |
| 742 | pixel_values: torch.FloatTensor, |
| 743 | spatial_shapes: torch.Tensor, |
| 744 | ) -> list[torch.Tensor]: |
| 745 | assert spatial_shapes.device.type == "cpu", ( |
| 746 | "Expected `spatial_shapes` on CPU to avoid device-to-host sync in " |
| 747 | "variable-length packing." |
| 748 | ) |
| 749 | |
| 750 | pixel_values = pixel_values.to( |
| 751 | dtype=self.vision_tower.vision_model.embeddings.patch_embedding.weight.dtype |
| 752 | ) # fp16 compatibility |
| 753 | |
| 754 | # LFM2-VL's HF processor pads patch sequences with trailing zeros. |
| 755 | # Pack patch tokens upfront so the vision tower runs entirely unpadded. |
| 756 | spatial_shapes_list: list[list[int]] = spatial_shapes.tolist() |
| 757 | lengths_list = [h * w for h, w in spatial_shapes_list] |
| 758 | total_tokens = int(sum(lengths_list)) |
| 759 | lengths_cpu = (spatial_shapes[:, 0] * spatial_shapes[:, 1]).to( |
| 760 | dtype=torch.int32 |
| 761 | ) |
| 762 | max_seqlen = ( |
| 763 | lengths_cpu.max().reshape(1) |
| 764 | if lengths_cpu.numel() |
| 765 | else torch.tensor([0], dtype=torch.int32) |
| 766 | ) |
| 767 | |
| 768 | if total_tokens == 0: |
| 769 | return [] |
| 770 | |
| 771 | packed_pixel_values = pixel_values.new_empty( |
| 772 | (total_tokens, pixel_values.shape[-1]) |
| 773 | ) |
| 774 | offset = 0 |
| 775 | for i, length in enumerate(lengths_list): |
| 776 | if length <= 0: |
| 777 | continue |
| 778 | packed_pixel_values[offset : offset + length].copy_( |
| 779 | pixel_values[i, :length] |
| 780 | ) |
| 781 | offset += length |
| 782 | packed_pixel_values = packed_pixel_values.unsqueeze(0) |
| 783 | |
| 784 | lengths = torch.tensor( |
| 785 | lengths_list, dtype=torch.int32, device=pixel_values.device |
| 786 | ) |
| 787 | cu_seqlens = torch.zeros( |
| 788 | lengths.shape[0] + 1, |
| 789 | dtype=torch.int32, |
| 790 | device=pixel_values.device, |
| 791 | ) |
| 792 | cu_seqlens[1:] = torch.cumsum(lengths, dim=0) |
| 793 | |
| 794 | with set_forward_context(None, self.vllm_config): |
| 795 | vision_outputs = self.vision_tower( |
| 796 | pixel_values_packed=packed_pixel_values, |
| 797 | spatial_shapes=spatial_shapes, |
| 798 | cu_seqlens=cu_seqlens, |
| 799 | max_seqlen=max_seqlen, |
| 800 | ) |
| 801 | image_outputs_packed = getattr( |
| 802 | vision_outputs, "last_hidden_state", vision_outputs |
| 803 | ) |
| 804 | vision_features_packed = image_outputs_packed[0] |
| 805 | |
| 806 | projected_packed = self.multi_modal_projector( |
| 807 | vision_features_packed=vision_features_packed, |
| 808 | spatial_shapes=spatial_shapes, |
| 809 | ) |
| 810 | projected_lengths_list = self._get_lfm2vl_tile_output_lengths( |
| 811 | spatial_shapes_list |
| 812 | ) |
| 813 | |
| 814 | image_features: list[torch.Tensor] = [] |
| 815 | offset = 0 |
| 816 | for out_len in projected_lengths_list: |
| 817 | image_features.append(projected_packed[offset : offset + out_len]) |
| 818 | offset += out_len |
| 819 | |
| 820 | return image_features |
| 821 | |
| 822 | def _process_image_input( |
| 823 | self, |
| 824 | image_input: LFM2VLImageInputs, |
| 825 | ) -> torch.Tensor | list[torch.Tensor]: |
| 826 | pixel_values = image_input["pixel_values"] |
| 827 | spatial_shapes = image_input["spatial_shapes"] |
| 828 | num_patches = image_input["num_patches"] |
| 829 | |
| 830 | image_features = self.image_pixels_to_features( |
| 831 | pixel_values, |
| 832 | spatial_shapes=spatial_shapes, |
| 833 | ) |
| 834 | |
| 835 | # Group patches by image - num_patches is on CPU (keep_on_cpu=True) |
| 836 | # so .tolist() is instant with no DtoH sync |
| 837 | num_patches_list = num_patches.tolist() |
| 838 | batched_features: list[torch.Tensor] = [] |
| 839 | patch_idx = 0 |
| 840 | for count in num_patches_list: |
| 841 | # Slice the list of patch tensors for this image |
| 842 | image_patches = image_features[patch_idx : patch_idx + count] |
| 843 | # Concatenate patches for this image |
| 844 | batched_features.append(torch.cat(image_patches, dim=0)) |
| 845 | patch_idx += count |
| 846 | |
| 847 | return batched_features |
| 848 | |
| 849 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 850 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 851 | if image_input is None: |
| 852 | return [] |
| 853 | |
| 854 | return self._process_image_input(image_input) |
| 855 | |
| 856 | def get_encoder_cudagraph_config(self): |
| 857 | from vllm.v1.worker.encoder_cudagraph_defs import ( |
| 858 | EncoderCudaGraphConfig, |
| 859 | ) |
| 860 | |
| 861 | return EncoderCudaGraphConfig( |
| 862 | modalities=["image"], |
| 863 | buffer_keys=[ |
| 864 | "pixel_values_packed", |
| 865 | "pos_embeds", |
| 866 | "cu_seqlens", |
| 867 | "max_seqlen", |
| 868 | "gather_idx", |
| 869 | ], |
| 870 | out_hidden_size=self.config.text_config.hidden_size, |
| 871 | padding_logics={ |
| 872 | "cu_seqlens": _pad_cumulative_seqlens_buffer, |
| 873 | }, |
| 874 | ) |
| 875 | |
| 876 | def get_max_frames_per_video(self) -> int: |
| 877 | return 0 |
| 878 | |
| 879 | def get_encoder_cudagraph_budget_range( |
| 880 | self, |
| 881 | vllm_config: VllmConfig, |
| 882 | ) -> tuple[int, int]: |
| 883 | min_budget = self._get_lfm2vl_min_image_tokens() |
| 884 | max_budget = min( |
| 885 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 886 | self.model_config.max_model_len, |
| 887 | ) |
| 888 | return min_budget, max_budget |
| 889 | |
| 890 | def _get_spatial_shapes_list( |
| 891 | self, |
| 892 | spatial_shapes: torch.Tensor, |
| 893 | ) -> list[list[int]]: |
| 894 | assert spatial_shapes.device.type == "cpu", ( |
| 895 | "Expected `spatial_shapes` on CPU to avoid device-to-host sync in " |
| 896 | "variable-length packing." |
| 897 | ) |
| 898 | return spatial_shapes.tolist() |
| 899 | |
| 900 | @staticmethod |
| 901 | def _get_lfm2vl_tile_input_lengths( |
| 902 | spatial_shapes_list: list[list[int]], |
| 903 | ) -> list[int]: |
| 904 | return [height * width for height, width in spatial_shapes_list] |
| 905 | |
| 906 | def _get_lfm2vl_tile_output_lengths( |
| 907 | self, |
| 908 | spatial_shapes_list: list[list[int]], |
| 909 | ) -> list[int]: |
| 910 | factor = self.multi_modal_projector.factor |
| 911 | output_lengths: list[int] = [] |
| 912 | for height, width in spatial_shapes_list: |
| 913 | if height % factor != 0 or width % factor != 0: |
| 914 | raise ValueError( |
| 915 | "spatial_shapes must be divisible by downsample_factor: " |
| 916 | f"got ({height}, {width}) with factor={factor}." |
| 917 | ) |
| 918 | output_lengths.append((height // factor) * (width // factor)) |
| 919 | return output_lengths |
| 920 | |
| 921 | def _get_lfm2vl_mm_processor_kwargs(self) -> Mapping[str, object]: |
| 922 | return self.multimodal_config.mm_processor_kwargs or {} |
| 923 | |
| 924 | def _get_lfm2vl_min_image_tokens(self) -> int: |
| 925 | value = self._get_lfm2vl_mm_processor_kwargs().get( |
| 926 | "min_image_tokens", |
| 927 | getattr(self.config, "min_image_tokens", None) or 64, |
| 928 | ) |
| 929 | if not isinstance( |
| 930 | value, |
| 931 | (str, Buffer, typing.SupportsInt, typing.SupportsIndex), |
| 932 | ): |
| 933 | raise TypeError( |
| 934 | "int() argument must be a string, a bytes-like object " |
| 935 | f"or a real number, not '{type(value).__name__}'" |
| 936 | ) |
| 937 | return max(1, int(value)) |
| 938 | |
| 939 | def _get_lfm2vl_item_tile_slices( |
| 940 | self, |
| 941 | num_patches: torch.Tensor, |
| 942 | ) -> list[tuple[int, int]]: |
| 943 | num_patches_list = [int(x) for x in num_patches.tolist()] |
| 944 | starts = [0] |
| 945 | for count in num_patches_list: |
| 946 | starts.append(starts[-1] + count) |
| 947 | return list(zip(starts[:-1], starts[1:])) |
| 948 | |
| 949 | def get_encoder_cudagraph_item_specs( |
| 950 | self, |
| 951 | mm_kwargs: dict[str, Any], |
| 952 | ): |
| 953 | from vllm.v1.worker.encoder_cudagraph_defs import EncoderItemSpec |
| 954 | |
| 955 | spatial_shapes = mm_kwargs["spatial_shapes"] |
| 956 | num_patches = mm_kwargs["num_patches"] |
| 957 | spatial_shapes_list = self._get_spatial_shapes_list(spatial_shapes) |
| 958 | input_lengths = self._get_lfm2vl_tile_input_lengths(spatial_shapes_list) |
| 959 | output_lengths = self._get_lfm2vl_tile_output_lengths(spatial_shapes_list) |
| 960 | |
| 961 | return [ |
| 962 | EncoderItemSpec( |
| 963 | input_size=sum(input_lengths[start:end]), |
| 964 | output_tokens=sum(output_lengths[start:end]), |
| 965 | ) |
| 966 | for start, end in self._get_lfm2vl_item_tile_slices(num_patches) |
| 967 | ] |
| 968 | |
| 969 | def select_encoder_cudagraph_items( |
| 970 | self, |
| 971 | mm_kwargs: dict[str, Any], |
| 972 | indices: list[int], |
| 973 | ) -> dict[str, Any]: |
| 974 | pixel_values = mm_kwargs["pixel_values"] |
| 975 | spatial_shapes = mm_kwargs["spatial_shapes"] |
| 976 | num_patches = mm_kwargs["num_patches"] |
| 977 | |
| 978 | tile_slices = self._get_lfm2vl_item_tile_slices(num_patches) |
| 979 | |
| 980 | if len(indices) == 0: |
| 981 | return { |
| 982 | "pixel_values": pixel_values[:0], |
| 983 | "spatial_shapes": spatial_shapes[:0], |
| 984 | "num_patches": num_patches[:0], |
| 985 | } |
| 986 | |
| 987 | tile_indices: list[int] = [] |
| 988 | for image_idx in indices: |
| 989 | start, end = tile_slices[image_idx] |
| 990 | tile_indices.extend(range(start, end)) |
| 991 | |
| 992 | return { |
| 993 | "pixel_values": pixel_values[tile_indices], |
| 994 | "spatial_shapes": spatial_shapes[tile_indices], |
| 995 | "num_patches": num_patches[indices], |
| 996 | } |
| 997 | |
| 998 | def _pack_lfm2vl_pixel_values( |
| 999 | self, |
| 1000 | pixel_values: torch.Tensor, |
| 1001 | spatial_shapes_list: list[list[int]], |
| 1002 | ) -> torch.Tensor: |
| 1003 | input_lengths = self._get_lfm2vl_tile_input_lengths(spatial_shapes_list) |
| 1004 | total_tokens = sum(input_lengths) |
| 1005 | packed = pixel_values.new_empty((total_tokens, pixel_values.shape[-1])) |
| 1006 | |
| 1007 | offset = 0 |
| 1008 | for i, length in enumerate(input_lengths): |
| 1009 | if length <= 0: |
| 1010 | continue |
| 1011 | packed[offset : offset + length].copy_(pixel_values[i, :length]) |
| 1012 | offset += length |
| 1013 | return packed |
| 1014 | |
| 1015 | def _get_lfm2vl_pos_embeds( |
| 1016 | self, |
| 1017 | spatial_shapes: torch.Tensor, |
| 1018 | spatial_shapes_list: list[list[int]], |
| 1019 | ) -> torch.Tensor: |
| 1020 | embeddings = self.vision_tower.vision_model.embeddings |
| 1021 | positional_embeddings = embeddings.position_embedding.weight.reshape( |
| 1022 | embeddings.position_embedding_size, |
| 1023 | embeddings.position_embedding_size, |
| 1024 | -1, |
| 1025 | ) |
| 1026 | lengths_list = self._get_lfm2vl_tile_input_lengths(spatial_shapes_list) |
| 1027 | return embeddings.resize_positional_embeddings_packed( |
| 1028 | positional_embeddings, |
| 1029 | spatial_shapes, |
| 1030 | lengths_list=lengths_list, |
| 1031 | ) |
| 1032 | |
| 1033 | def _get_lfm2vl_cu_seqlens( |
| 1034 | self, |
| 1035 | spatial_shapes_list: list[list[int]], |
| 1036 | device: torch.device, |
| 1037 | ) -> torch.Tensor: |
| 1038 | lengths = torch.tensor( |
| 1039 | self._get_lfm2vl_tile_input_lengths(spatial_shapes_list), |
| 1040 | dtype=torch.int32, |
| 1041 | device=device, |
| 1042 | ) |
| 1043 | cu_seqlens = torch.zeros( |
| 1044 | lengths.shape[0] + 1, |
| 1045 | dtype=torch.int32, |
| 1046 | device=device, |
| 1047 | ) |
| 1048 | if lengths.numel() > 0: |
| 1049 | cu_seqlens[1:] = torch.cumsum(lengths, dim=0) |
| 1050 | return cu_seqlens |
| 1051 | |
| 1052 | def _get_lfm2vl_max_seqlen( |
| 1053 | self, |
| 1054 | spatial_shapes_list: list[list[int]], |
| 1055 | ) -> torch.Tensor: |
| 1056 | input_lengths = self._get_lfm2vl_tile_input_lengths(spatial_shapes_list) |
| 1057 | max_seqlen = max(input_lengths) if input_lengths else 0 |
| 1058 | return torch.tensor(max_seqlen, dtype=torch.int32) |
| 1059 | |
| 1060 | def _get_lfm2vl_projector_gather_idx( |
| 1061 | self, |
| 1062 | spatial_shapes_list: list[list[int]], |
| 1063 | device: torch.device, |
| 1064 | ) -> torch.Tensor: |
| 1065 | factor = self.multi_modal_projector.factor |
| 1066 | dh = torch.arange(factor, dtype=torch.int64) |
| 1067 | dw = torch.arange(factor, dtype=torch.int64) |
| 1068 | dh_grid, dw_grid = torch.meshgrid(dh, dw, indexing="ij") |
| 1069 | dh_flat = dh_grid.reshape(-1) |
| 1070 | dw_flat = dw_grid.reshape(-1) |
| 1071 | |
| 1072 | gather_idx_parts: list[torch.Tensor] = [] |
| 1073 | offset = 0 |
| 1074 | for height, width in spatial_shapes_list: |
| 1075 | length = height * width |
| 1076 | if length <= 0: |
| 1077 | continue |
| 1078 | if height % factor != 0 or width % factor != 0: |
| 1079 | raise ValueError( |
| 1080 | "spatial_shapes must be divisible by downsample_factor: " |
| 1081 | f"got ({height}, {width}) with factor={factor}." |
| 1082 | ) |
| 1083 | |
| 1084 | rows_out = torch.arange(height // factor, dtype=torch.int64) |
| 1085 | cols_out = torch.arange(width // factor, dtype=torch.int64) |
| 1086 | rr, cc = torch.meshgrid(rows_out, cols_out, indexing="ij") |
| 1087 | rr = rr.reshape(-1) |
| 1088 | cc = cc.reshape(-1) |
| 1089 | token_idx = (rr[:, None] * factor + dh_flat[None, :]) * width + ( |
| 1090 | cc[:, None] * factor + dw_flat[None, :] |
| 1091 | ) |
| 1092 | gather_idx_parts.append(token_idx.reshape(-1) + offset) |
| 1093 | offset += length |
| 1094 | |
| 1095 | if not gather_idx_parts: |
| 1096 | return torch.empty(0, dtype=torch.int64, device=device) |
| 1097 | return torch.cat(gather_idx_parts).to(device=device) |
| 1098 | |
| 1099 | def _prepare_lfm2vl_cudagraph_values( |
| 1100 | self, |
| 1101 | pixel_values: torch.Tensor, |
| 1102 | spatial_shapes: torch.Tensor, |
| 1103 | ) -> dict[str, torch.Tensor]: |
| 1104 | spatial_shapes_list = self._get_spatial_shapes_list(spatial_shapes) |
| 1105 | pixel_values_packed = self._pack_lfm2vl_pixel_values( |
| 1106 | pixel_values, |
| 1107 | spatial_shapes_list, |
| 1108 | ) |
| 1109 | pos_embeds = self._get_lfm2vl_pos_embeds(spatial_shapes, spatial_shapes_list) |
| 1110 | device = pixel_values.device |
| 1111 | |
| 1112 | return { |
| 1113 | "pixel_values_packed": pixel_values_packed, |
| 1114 | "pos_embeds": pos_embeds, |
| 1115 | "cu_seqlens": self._get_lfm2vl_cu_seqlens(spatial_shapes_list, device), |
| 1116 | "max_seqlen": self._get_lfm2vl_max_seqlen(spatial_shapes_list), |
| 1117 | "gather_idx": self._get_lfm2vl_projector_gather_idx( |
| 1118 | spatial_shapes_list, |
| 1119 | device, |
| 1120 | ), |
| 1121 | } |
| 1122 | |
| 1123 | def _get_lfm2vl_capture_spatial_shapes( |
| 1124 | self, |
| 1125 | token_budget: int, |
| 1126 | ) -> torch.Tensor: |
| 1127 | factor = self.multi_modal_projector.factor |
| 1128 | min_image_tokens = self._get_lfm2vl_min_image_tokens() |
| 1129 | remaining = token_budget |
| 1130 | shapes: list[list[int]] = [] |
| 1131 | |
| 1132 | while remaining > 0: |
| 1133 | out_tokens = min(remaining, min_image_tokens) |
| 1134 | shapes.append([factor, out_tokens * factor]) |
| 1135 | remaining -= out_tokens |
| 1136 | |
| 1137 | return torch.tensor(shapes, dtype=torch.int64) |
| 1138 | |
| 1139 | def prepare_encoder_cudagraph_capture_inputs( |
| 1140 | self, |
| 1141 | token_budget: int, |
| 1142 | max_batch_size: int, |
| 1143 | max_frames_per_batch: int, |
| 1144 | device: torch.device, |
| 1145 | dtype: torch.dtype, |
| 1146 | path: str = "default", |
| 1147 | axis_keys: tuple[Hashable, ...] | None = None, |
| 1148 | ): |
| 1149 | from vllm.v1.worker.encoder_cudagraph_defs import ( |
| 1150 | EncoderCudaGraphCaptureInputs, |
| 1151 | ) |
| 1152 | |
| 1153 | spatial_shapes = self._get_lfm2vl_capture_spatial_shapes(token_budget) |
| 1154 | spatial_shapes_list = self._get_spatial_shapes_list(spatial_shapes) |
| 1155 | input_lengths = self._get_lfm2vl_tile_input_lengths(spatial_shapes_list) |
| 1156 | total_input_tokens = sum(input_lengths) |
| 1157 | |
| 1158 | patch_dim = ( |
| 1159 | self.vision_tower.vision_model.embeddings.patch_embedding.weight.shape[1] |
| 1160 | ) |
| 1161 | dummy_pixel_values = torch.randn( |
| 1162 | total_input_tokens, |
| 1163 | patch_dim, |
| 1164 | device=device, |
| 1165 | dtype=dtype, |
| 1166 | ) |
| 1167 | pos_embeds = self._get_lfm2vl_pos_embeds( |
| 1168 | spatial_shapes, |
| 1169 | spatial_shapes_list, |
| 1170 | ).to(device=device, dtype=dtype) |
| 1171 | |
| 1172 | # max_seqlen.item() is baked into the captured ViT attention graph, so |
| 1173 | # capture with a budget-level upper bound that covers any replay item. |
| 1174 | max_tile_input_tokens = token_budget * self.multi_modal_projector.factor**2 |
| 1175 | values = { |
| 1176 | "pixel_values_packed": dummy_pixel_values, |
| 1177 | "pos_embeds": pos_embeds, |
| 1178 | "cu_seqlens": self._get_lfm2vl_cu_seqlens(spatial_shapes_list, device), |
| 1179 | "max_seqlen": torch.tensor(max_tile_input_tokens, dtype=torch.int32), |
| 1180 | "gather_idx": self._get_lfm2vl_projector_gather_idx( |
| 1181 | spatial_shapes_list, |
| 1182 | device, |
| 1183 | ), |
| 1184 | } |
| 1185 | |
| 1186 | return EncoderCudaGraphCaptureInputs(values=values) |
| 1187 | |
| 1188 | def prepare_encoder_cudagraph_replay_buffers( |
| 1189 | self, |
| 1190 | mm_kwargs: dict[str, Any], |
| 1191 | max_batch_size: int, |
| 1192 | max_frames_per_batch: int, |
| 1193 | path: str = "default", |
| 1194 | ): |
| 1195 | from vllm.v1.worker.encoder_cudagraph_defs import ( |
| 1196 | EncoderCudaGraphReplayBuffers, |
| 1197 | ) |
| 1198 | |
| 1199 | values = self._prepare_lfm2vl_cudagraph_values( |
| 1200 | mm_kwargs["pixel_values"], |
| 1201 | mm_kwargs["spatial_shapes"], |
| 1202 | ) |
| 1203 | return EncoderCudaGraphReplayBuffers(values=values) |
| 1204 | |
| 1205 | def encoder_cudagraph_forward( |
| 1206 | self, |
| 1207 | values: dict[str, torch.Tensor], |
| 1208 | path: str = "default", |
| 1209 | ) -> torch.Tensor: |
| 1210 | embeddings = self.vision_tower.vision_model.embeddings |
| 1211 | pixel_values = values["pixel_values_packed"].to( |
| 1212 | dtype=embeddings.patch_embedding.weight.dtype |
| 1213 | ) |
| 1214 | patch_embeds = embeddings.patch_embedding(pixel_values) |
| 1215 | hidden_states = (patch_embeds + values["pos_embeds"]).unsqueeze(0) |
| 1216 | |
| 1217 | with set_forward_context(None, self.vllm_config): |
| 1218 | encoder_outputs = self.vision_tower.vision_model.encoder( |
| 1219 | inputs_embeds=hidden_states, |
| 1220 | cu_seqlens=values["cu_seqlens"], |
| 1221 | max_seqlen=values["max_seqlen"], |
| 1222 | ) |
| 1223 | |
| 1224 | post_layernorm = self.vision_tower.vision_model.post_layernorm |
| 1225 | if post_layernorm is not None: |
| 1226 | encoder_outputs = post_layernorm(encoder_outputs) |
| 1227 | |
| 1228 | return self.multi_modal_projector.forward_with_gather_idx( |
| 1229 | vision_features_packed=encoder_outputs[0], |
| 1230 | gather_idx=values["gather_idx"], |
| 1231 | ) |
| 1232 | |
| 1233 | def encoder_eager_forward( |
| 1234 | self, |
| 1235 | mm_kwargs: dict[str, Any], |
| 1236 | path: str = "default", |
| 1237 | ) -> torch.Tensor: |
| 1238 | image_input = LFM2VLImageInputs( |
| 1239 | type="pixel_values", |
| 1240 | pixel_values=mm_kwargs["pixel_values"], |
| 1241 | spatial_shapes=mm_kwargs["spatial_shapes"], |
| 1242 | num_patches=mm_kwargs["num_patches"], |
| 1243 | ) |
| 1244 | return torch.cat(self._process_image_input(image_input), dim=0) |
| 1245 | |
| 1246 | def forward( |
| 1247 | self, |
| 1248 | input_ids: torch.Tensor | None, |
| 1249 | positions: torch.Tensor, |
| 1250 | intermediate_tensors: IntermediateTensors | None = None, |
| 1251 | inputs_embeds: torch.Tensor | None = None, |
| 1252 | **kwargs: object, |
| 1253 | ) -> torch.Tensor | IntermediateTensors: |
| 1254 | if intermediate_tensors is not None: |
| 1255 | inputs_embeds = None |
| 1256 | |
| 1257 | hidden_states = self.language_model( |
| 1258 | input_ids=input_ids, |
| 1259 | positions=positions, |
| 1260 | intermediate_tensors=intermediate_tensors, |
| 1261 | inputs_embeds=inputs_embeds, |
| 1262 | ) |
| 1263 | return hidden_states |
| 1264 | |
| 1265 | def compute_logits( |
| 1266 | self, |
| 1267 | hidden_states: torch.Tensor, |
| 1268 | ) -> torch.Tensor | None: |
| 1269 | return self.language_model.compute_logits(hidden_states) |
| 1270 | |
| 1271 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1272 | loader = AutoWeightsLoader(self) |
| 1273 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 1274 | |
| 1275 | def get_mm_mapping(self) -> MultiModelKeys: |
| 1276 | """Get the module prefix in multimodal models.""" |
| 1277 | return MultiModelKeys.from_string_field( |
| 1278 | language_model="language_model", |
| 1279 | connector="multi_modal_projector", |
| 1280 | tower_model="vision_tower", |
| 1281 | ) |
| 1282 | |
| 1283 | def get_mm_lora_token_counts( |
| 1284 | self, |
| 1285 | *, |
| 1286 | modality: str, |
| 1287 | mm_kwargs: MultiModalKwargsItem | None, |
| 1288 | num_mm_embeds: int, |
| 1289 | ) -> tuple[int, int | None]: |
| 1290 | del modality, mm_kwargs |
| 1291 | downsample_factor = self.config.downsample_factor |
| 1292 | return num_mm_embeds * downsample_factor**2, num_mm_embeds |