| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved. |
| 4 | # |
| 5 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 6 | # you may not use this file except in compliance with the License. |
| 7 | # You may obtain a copy of the License at |
| 8 | # |
| 9 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 10 | # |
| 11 | # Unless required by applicable law or agreed to in writing, software |
| 12 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 13 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | # See the License for the specific language governing permissions and |
| 15 | # limitations under the License. |
| 16 | |
| 17 | import math |
| 18 | from collections.abc import Iterable, Iterator, Mapping, Sequence |
| 19 | from functools import partial |
| 20 | from typing import Annotated, Literal |
| 21 | |
| 22 | import numpy as np |
| 23 | import torch |
| 24 | import torch.nn as nn |
| 25 | from einops import rearrange |
| 26 | from transformers import BaseImageProcessor, BatchFeature, PreTrainedConfig |
| 27 | from transformers.activations import GELUActivation |
| 28 | from transformers.image_utils import ChannelDimension |
| 29 | from transformers.modeling_outputs import ( |
| 30 | BaseModelOutputWithPooling, |
| 31 | ) |
| 32 | from transformers.utils import torch_int |
| 33 | |
| 34 | from vllm.config import VllmConfig |
| 35 | from vllm.config.multimodal import MultiModalDummyOptions |
| 36 | from vllm.distributed import parallel_state |
| 37 | from vllm.distributed import utils as dist_utils |
| 38 | from vllm.inputs import MultiModalDataDict |
| 39 | from vllm.model_executor.layers.attention import ( |
| 40 | MMEncoderAttention, |
| 41 | ) |
| 42 | from vllm.model_executor.layers.conv import Conv2dLayer |
| 43 | from vllm.model_executor.layers.linear import ( |
| 44 | QKVParallelLinear, |
| 45 | RowParallelLinear, |
| 46 | ) |
| 47 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 48 | from vllm.model_executor.layers.rotary_embedding.common import ( |
| 49 | ApplyRotaryEmb, |
| 50 | ) |
| 51 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 52 | from vllm.multimodal.inputs import ( |
| 53 | MultiModalFeatureSpec, |
| 54 | MultiModalFieldConfig, |
| 55 | MultiModalKwargsItems, |
| 56 | ) |
| 57 | from vllm.multimodal.parse import ( |
| 58 | ImageProcessorItems, |
| 59 | ImageSize, |
| 60 | MultiModalDataItems, |
| 61 | ) |
| 62 | from vllm.multimodal.processing import ( |
| 63 | BaseDummyInputsBuilder, |
| 64 | BaseMultiModalProcessor, |
| 65 | BaseProcessingInfo, |
| 66 | PromptReplacement, |
| 67 | PromptUpdate, |
| 68 | ) |
| 69 | from vllm.sequence import IntermediateTensors |
| 70 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 71 | from vllm.utils.torch_utils import PIN_MEMORY, async_tensor_h2d |
| 72 | from vllm.v1.attention.backends.registry import AttentionBackendEnum |
| 73 | |
| 74 | from .ernie45 import Ernie4_5ForCausalLM |
| 75 | from .interfaces import MultiModalEmbeddings, SupportsMRoPE, SupportsMultiModal |
| 76 | from .siglip import SiglipMLP |
| 77 | from .utils import ( |
| 78 | AutoWeightsLoader, |
| 79 | PPMissingLayer, |
| 80 | WeightsMapper, |
| 81 | maybe_prefix, |
| 82 | ) |
| 83 | from .vision import get_vit_attn_backend |
| 84 | |
| 85 | |
| 86 | def smart_resize( |
| 87 | height: int, |
| 88 | width: int, |
| 89 | factor: int = 28, |
| 90 | min_pixels: int = 28 * 28 * 130, |
| 91 | max_pixels: int = 28 * 28 * 1280, |
| 92 | ): |
| 93 | """Rescales the image so that the following conditions are met: |
| 94 | |
| 95 | 1. Both dimensions (height and width) are divisible by 'factor'. |
| 96 | |
| 97 | 2. The total number of pixels is within the range ['min_pixels', 'max_pixels']. |
| 98 | |
| 99 | 3. The aspect ratio of the image is maintained as closely as possible. |
| 100 | |
| 101 | """ |
| 102 | if height < factor: |
| 103 | width = round((width * factor) / height) |
| 104 | height = factor |
| 105 | |
| 106 | if width < factor: |
| 107 | height = round((height * factor) / width) |
| 108 | width = factor |
| 109 | |
| 110 | if max(height, width) / min(height, width) > 200: |
| 111 | raise ValueError( |
| 112 | f"absolute aspect ratio must be smaller than 200, " |
| 113 | f"got {max(height, width) / min(height, width)}" |
| 114 | ) |
| 115 | h_bar = round(height / factor) * factor |
| 116 | w_bar = round(width / factor) * factor |
| 117 | if h_bar * w_bar > max_pixels: |
| 118 | beta = math.sqrt((height * width) / max_pixels) |
| 119 | h_bar = math.floor(height / beta / factor) * factor |
| 120 | w_bar = math.floor(width / beta / factor) * factor |
| 121 | elif h_bar * w_bar < min_pixels: |
| 122 | beta = math.sqrt(min_pixels / (height * width)) |
| 123 | h_bar = math.ceil(height * beta / factor) * factor |
| 124 | w_bar = math.ceil(width * beta / factor) * factor |
| 125 | return h_bar, w_bar |
| 126 | |
| 127 | |
| 128 | class PaddleOCRVLProcessingInfo(BaseProcessingInfo): |
| 129 | def get_hf_config(self): |
| 130 | return self.ctx.get_hf_config() |
| 131 | |
| 132 | def get_hf_processor(self, **kwargs: object): |
| 133 | return self.ctx.get_hf_processor(**kwargs) |
| 134 | |
| 135 | def get_image_processor(self, **kwargs: object): |
| 136 | return self.get_hf_processor(**kwargs).image_processor |
| 137 | |
| 138 | def get_supported_mm_limits(self): |
| 139 | return {"image": None} |
| 140 | |
| 141 | def get_num_image_tokens( |
| 142 | self, |
| 143 | *, |
| 144 | image_width: int, |
| 145 | image_height: int, |
| 146 | image_processor: BaseImageProcessor, |
| 147 | mm_kwargs: Mapping[str, object], |
| 148 | ) -> int: |
| 149 | hf_config = self.get_hf_config() |
| 150 | vision_config = hf_config.vision_config |
| 151 | patch_size = vision_config.patch_size |
| 152 | merge_size = vision_config.spatial_merge_size |
| 153 | |
| 154 | if self.ctx.model_config.trust_remote_code: |
| 155 | # Defined in HF Hub repo |
| 156 | min_pixels_key = "min_pixels" |
| 157 | max_pixels_key = "max_pixels" |
| 158 | else: |
| 159 | # Defined in Transformers library (requires v5.0 or above) |
| 160 | min_pixels_key = "shortest_edge" |
| 161 | max_pixels_key = "longest_edge" |
| 162 | |
| 163 | mm_kwargs = self.ctx.get_merged_mm_kwargs(mm_kwargs) |
| 164 | size = image_processor.size |
| 165 | if override_size := mm_kwargs.get("size"): |
| 166 | size = size | override_size |
| 167 | if (override_min_pixels := mm_kwargs.get("min_pixels")) is not None: |
| 168 | size = size | {min_pixels_key: override_min_pixels} |
| 169 | if (override_max_pixels := mm_kwargs.get("max_pixels")) is not None: |
| 170 | size = size | {max_pixels_key: override_max_pixels} |
| 171 | |
| 172 | resized_height, resized_width = smart_resize( |
| 173 | height=image_height, |
| 174 | width=image_width, |
| 175 | factor=patch_size * merge_size, |
| 176 | min_pixels=size[min_pixels_key], |
| 177 | max_pixels=size[max_pixels_key], |
| 178 | ) |
| 179 | preprocessed_size = ImageSize(width=resized_width, height=resized_height) |
| 180 | |
| 181 | grid_t = 1 |
| 182 | grid_h = preprocessed_size.height // patch_size |
| 183 | grid_w = preprocessed_size.width // patch_size |
| 184 | |
| 185 | num_patches = grid_t * grid_h * grid_w |
| 186 | num_image_tokens = num_patches // (merge_size**2) |
| 187 | |
| 188 | return num_image_tokens |
| 189 | |
| 190 | def get_image_size_with_most_features(self) -> ImageSize: |
| 191 | hf_config = self.get_hf_config() |
| 192 | image_processor = self.get_image_processor() |
| 193 | |
| 194 | # See `smart_resize` for the calculation of the image size. |
| 195 | merge_size = hf_config.vision_config.spatial_merge_size |
| 196 | patch_size = hf_config.vision_config.patch_size |
| 197 | factor = merge_size * patch_size |
| 198 | if self.ctx.model_config.trust_remote_code: |
| 199 | # Defined in HF Hub repo |
| 200 | max_pixels = image_processor.max_pixels |
| 201 | else: |
| 202 | # Defined in Transformers library (requires v5.0 or above) |
| 203 | max_pixels = image_processor.size.longest_edge |
| 204 | max_num_tokens = max_pixels // (factor**2) |
| 205 | # Find factors of max_num_tokens close to its square root |
| 206 | # to create a dummy image with a reasonable aspect ratio. |
| 207 | h_patches = int(math.sqrt(max_num_tokens)) |
| 208 | max_num_tokens -= max_num_tokens % h_patches |
| 209 | w_patches = max_num_tokens // h_patches |
| 210 | return ImageSize(height=h_patches * factor, width=w_patches * factor) |
| 211 | |
| 212 | |
| 213 | class PaddleOCRVLDummyInputsBuilder(BaseDummyInputsBuilder[PaddleOCRVLProcessingInfo]): |
| 214 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 215 | num_images = mm_counts.get("image", 0) |
| 216 | |
| 217 | processor = self.info.get_hf_processor() |
| 218 | image_token = processor.image_token |
| 219 | |
| 220 | return image_token * num_images |
| 221 | |
| 222 | def get_dummy_mm_data( |
| 223 | self, |
| 224 | seq_len: int, |
| 225 | mm_counts: Mapping[str, int], |
| 226 | mm_options: MultiModalDummyOptions, |
| 227 | ) -> MultiModalDataDict: |
| 228 | max_image_size = self.info.get_image_size_with_most_features() |
| 229 | |
| 230 | return { |
| 231 | "image": self._get_dummy_images( |
| 232 | width=max_image_size.width, |
| 233 | height=max_image_size.height, |
| 234 | num_images=mm_counts.get("image", 0), |
| 235 | overrides=mm_options.get("image"), |
| 236 | ) |
| 237 | } |
| 238 | |
| 239 | |
| 240 | class PaddleOCRVLMultiModalProcessor( |
| 241 | BaseMultiModalProcessor[PaddleOCRVLProcessingInfo] |
| 242 | ): |
| 243 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 244 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 245 | |
| 246 | def _call_hf_processor( |
| 247 | self, |
| 248 | hf_data: Mapping[str, object], |
| 249 | hf_kwargs: Mapping[str, object], |
| 250 | ) -> BatchFeature: |
| 251 | final_mm_kwargs = dict(hf_kwargs or {}) |
| 252 | final_mm_kwargs.setdefault("images_kwargs", {}) |
| 253 | # vLLM use PIL.Image, always set channel_last |
| 254 | final_mm_kwargs["input_data_format"] = ChannelDimension.LAST |
| 255 | return self.info.ctx.call_hf_processor( |
| 256 | self.info.get_hf_processor(**final_mm_kwargs), |
| 257 | hf_data, |
| 258 | hf_kwargs, |
| 259 | ) |
| 260 | |
| 261 | def _postprocess_hf_mm_data( |
| 262 | self, |
| 263 | hf_data: Mapping[str, object], |
| 264 | hf_kwargs: Mapping[str, object], |
| 265 | processed_data: BatchFeature, |
| 266 | ) -> BatchFeature: |
| 267 | if not hf_data: |
| 268 | return processed_data |
| 269 | |
| 270 | num_patches_per_image = processed_data["image_grid_thw"].prod(-1) |
| 271 | processed_data["pixel_values"] = processed_data["pixel_values"].split( |
| 272 | num_patches_per_image.tolist() |
| 273 | ) |
| 274 | return processed_data |
| 275 | |
| 276 | def _get_mm_fields_config( |
| 277 | self, |
| 278 | hf_inputs: BatchFeature, |
| 279 | hf_processor_mm_kwargs: Mapping[str, object], |
| 280 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 281 | return dict( |
| 282 | pixel_values=MultiModalFieldConfig.batched("image"), |
| 283 | image_grid_thw=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 284 | ) |
| 285 | |
| 286 | def _get_prompt_updates( |
| 287 | self, |
| 288 | mm_items: MultiModalDataItems, |
| 289 | hf_processor_mm_kwargs: Mapping[str, object], |
| 290 | out_mm_kwargs: MultiModalKwargsItems, |
| 291 | ) -> Sequence[PromptUpdate]: |
| 292 | image_processor = self.info.get_image_processor(**hf_processor_mm_kwargs) |
| 293 | hf_config = self.info.get_hf_config() |
| 294 | image_token_id = hf_config.image_token_id |
| 295 | |
| 296 | def get_replacement(item_idx: int, image_processor): |
| 297 | images = mm_items.get_items("image", ImageProcessorItems) |
| 298 | |
| 299 | image_size = images.get_image_size(item_idx) |
| 300 | num_image_tokens = self.info.get_num_image_tokens( |
| 301 | image_width=image_size.width, |
| 302 | image_height=image_size.height, |
| 303 | image_processor=image_processor, |
| 304 | mm_kwargs=hf_processor_mm_kwargs, |
| 305 | ) |
| 306 | |
| 307 | return [image_token_id] * num_image_tokens |
| 308 | |
| 309 | return [ |
| 310 | PromptReplacement( |
| 311 | modality="image", |
| 312 | target=[image_token_id], |
| 313 | replacement=partial(get_replacement, image_processor=image_processor), |
| 314 | ), |
| 315 | ] |
| 316 | |
| 317 | |
| 318 | class Projector(nn.Module): |
| 319 | def __init__( |
| 320 | self, |
| 321 | text_config: PreTrainedConfig, |
| 322 | vision_config: PreTrainedConfig, |
| 323 | prefix: str = "", |
| 324 | ): |
| 325 | super().__init__() |
| 326 | self.text_config = text_config |
| 327 | self.vision_config = vision_config |
| 328 | self.merge_kernel_size = (2, 2) |
| 329 | |
| 330 | self.hidden_size = ( |
| 331 | self.vision_config.hidden_size |
| 332 | * self.merge_kernel_size[0] |
| 333 | * self.merge_kernel_size[1] |
| 334 | ) |
| 335 | |
| 336 | self.pre_norm = torch.nn.LayerNorm(self.vision_config.hidden_size, eps=1e-05) |
| 337 | self.linear_1 = nn.Linear(self.hidden_size, self.hidden_size, bias=True) |
| 338 | self.act = GELUActivation() |
| 339 | self.linear_2 = nn.Linear( |
| 340 | self.hidden_size, self.text_config.hidden_size, bias=True |
| 341 | ) |
| 342 | |
| 343 | def forward( |
| 344 | self, |
| 345 | image_features: torch.Tensor, |
| 346 | image_grid_thw: torch.Tensor, |
| 347 | ) -> torch.Tensor: |
| 348 | m1, m2 = self.merge_kernel_size |
| 349 | if isinstance(image_features, (list, tuple)): |
| 350 | processed_features = list() |
| 351 | for image_feature, image_grid in zip(image_features, image_grid_thw): |
| 352 | image_feature = self.pre_norm(image_feature) |
| 353 | t, h, w = image_grid |
| 354 | |
| 355 | image_feature = rearrange( |
| 356 | image_feature, |
| 357 | "(t h p1 w p2) d -> (t h w) (p1 p2 d)", |
| 358 | t=t, |
| 359 | h=h // m1, |
| 360 | p1=m1, |
| 361 | w=w // m2, |
| 362 | p2=m2, |
| 363 | ) |
| 364 | hidden_states = self.linear_1(image_feature) |
| 365 | hidden_states = self.act(hidden_states) |
| 366 | hidden_states = self.linear_2(hidden_states) |
| 367 | processed_features.append(hidden_states) |
| 368 | |
| 369 | return processed_features |
| 370 | |
| 371 | dims = image_features.shape[:-1] |
| 372 | dim = image_features.shape[-1] |
| 373 | image_features = image_features.view(np.prod(dims), dim) |
| 374 | hidden_states = self.pre_norm(image_features).view(-1, self.hidden_size) |
| 375 | hidden_states = self.linear_1(hidden_states) |
| 376 | hidden_states = self.act(hidden_states) |
| 377 | hidden_states = self.linear_2(hidden_states) |
| 378 | |
| 379 | return hidden_states.view(*dims, -1) |
| 380 | |
| 381 | |
| 382 | class PaddleOCRImagePixelInputs(TensorSchema): |
| 383 | type: Literal["pixel_values"] |
| 384 | pixel_values: Annotated[ |
| 385 | torch.Tensor, |
| 386 | TensorShape("bn", "p", 3, "patch_size", "patch_size", dynamic_dims={"p"}), |
| 387 | ] |
| 388 | image_grid_thw: Annotated[ |
| 389 | torch.Tensor, |
| 390 | TensorShape("bn", 3), |
| 391 | ] |
| 392 | |
| 393 | |
| 394 | class SiglipVisionEmbeddings(nn.Module): |
| 395 | def __init__(self, config: PreTrainedConfig): |
| 396 | super().__init__() |
| 397 | self.config = config |
| 398 | self.embed_dim = config.hidden_size |
| 399 | self.image_size = config.image_size |
| 400 | self.patch_size = config.patch_size |
| 401 | |
| 402 | self.patch_embedding = Conv2dLayer( |
| 403 | in_channels=config.num_channels, |
| 404 | out_channels=self.embed_dim, |
| 405 | kernel_size=self.patch_size, |
| 406 | stride=self.patch_size, |
| 407 | padding="valid", |
| 408 | ) |
| 409 | |
| 410 | self.num_patches = (self.image_size // self.patch_size) ** 2 |
| 411 | self.num_positions = self.num_patches |
| 412 | self.cache_position_embedding: dict[tuple[int, int], torch.Tensor] = {} |
| 413 | self.cache_position_count: dict[tuple[int, int], int] = {} |
| 414 | self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim) |
| 415 | |
| 416 | self.register_buffer( |
| 417 | "position_ids", |
| 418 | torch.arange(self.num_positions).expand((1, -1)), |
| 419 | persistent=False, |
| 420 | ) |
| 421 | |
| 422 | def interpolate_pos_encoding( |
| 423 | self, |
| 424 | embeddings: torch.Tensor, |
| 425 | height: int, |
| 426 | width: int, |
| 427 | is_after_patchify: bool = False, |
| 428 | ) -> torch.Tensor: |
| 429 | num_positions = self.position_embedding.weight.shape[0] |
| 430 | |
| 431 | patch_pos_embed = self.position_embedding.weight.unsqueeze(0) |
| 432 | |
| 433 | dim = embeddings.shape[-1] |
| 434 | |
| 435 | if is_after_patchify: |
| 436 | new_height = height |
| 437 | new_width = width |
| 438 | else: |
| 439 | new_height = height // self.patch_size |
| 440 | new_width = width // self.patch_size |
| 441 | |
| 442 | sqrt_num_positions = torch_int(num_positions**0.5) |
| 443 | patch_pos_embed = patch_pos_embed.reshape( |
| 444 | 1, sqrt_num_positions, sqrt_num_positions, dim |
| 445 | ) |
| 446 | patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2) |
| 447 | |
| 448 | patch_pos_embed = nn.functional.interpolate( |
| 449 | patch_pos_embed, |
| 450 | size=(new_height, new_width), |
| 451 | mode="bilinear", |
| 452 | align_corners=False, |
| 453 | ) |
| 454 | |
| 455 | patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim) |
| 456 | return patch_pos_embed |
| 457 | |
| 458 | def fetch_position_embedding_lfu_cache( |
| 459 | self, embeddings: torch.Tensor, h: int, w: int, max_cache: int = 20 |
| 460 | ): |
| 461 | grid = (h, w) |
| 462 | if grid in self.cache_position_embedding: |
| 463 | self.cache_position_count[grid] += 1 |
| 464 | return self.cache_position_embedding[grid] |
| 465 | |
| 466 | if len(self.cache_position_embedding) >= max_cache: |
| 467 | min_hit_grid = min( |
| 468 | self.cache_position_count, |
| 469 | key=self.cache_position_count.__getitem__, |
| 470 | ) |
| 471 | self.cache_position_count.pop(min_hit_grid) |
| 472 | self.cache_position_embedding.pop(min_hit_grid) |
| 473 | |
| 474 | position_embedding = self.interpolate_pos_encoding(embeddings, h, w, True) |
| 475 | self.cache_position_count[grid] = 1 |
| 476 | self.cache_position_embedding[grid] = position_embedding |
| 477 | return position_embedding |
| 478 | |
| 479 | def forward( |
| 480 | self, |
| 481 | pixel_values: torch.FloatTensor, |
| 482 | position_ids: torch.Tensor | None = None, |
| 483 | image_grid_thw: Sequence[tuple[int, int, int]] | None = None, |
| 484 | interpolate_pos_encoding=False, |
| 485 | ) -> torch.Tensor: |
| 486 | if pixel_values.dim() == 4: |
| 487 | pixel_values = pixel_values.unsqueeze(0) |
| 488 | if pixel_values.dim() == 5: |
| 489 | if position_ids is None: |
| 490 | raise ValueError( |
| 491 | "position_ids cannot be None when pixel_values.dim() is 5." |
| 492 | ) |
| 493 | ( |
| 494 | batch_size, |
| 495 | sequence_len, |
| 496 | channel, |
| 497 | height, |
| 498 | width, |
| 499 | ) = pixel_values.shape |
| 500 | target_dtype = self.patch_embedding.weight.dtype |
| 501 | pixel_values = rearrange(pixel_values, "b l c h w -> (b l) c h w") |
| 502 | patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype)) |
| 503 | embeddings = patch_embeds.flatten(-2).squeeze(-1) |
| 504 | |
| 505 | start = 0 |
| 506 | tmp_embeddings = list() |
| 507 | assert image_grid_thw is not None |
| 508 | for image_grid in image_grid_thw: |
| 509 | t, h, w = image_grid |
| 510 | end = start + t * h * w |
| 511 | image_embeddings = embeddings[start:end, :] |
| 512 | position_embedding = ( |
| 513 | self.interpolate_pos_encoding(image_embeddings, h, w, True) |
| 514 | .squeeze(0) |
| 515 | .repeat(t, 1) |
| 516 | ) |
| 517 | image_embeddings = image_embeddings + position_embedding |
| 518 | tmp_embeddings.append(image_embeddings) |
| 519 | start = end |
| 520 | embeddings = torch.concat(tmp_embeddings, dim=0).unsqueeze(0) |
| 521 | |
| 522 | return embeddings |
| 523 | else: |
| 524 | raise ValueError( |
| 525 | "Unsupported pixel_values dimension:" |
| 526 | f" {pixel_values.dim()}. Expected 4 or 5." |
| 527 | ) |
| 528 | |
| 529 | |
| 530 | def all_gather_interleave(local_tensor: torch.Tensor, hidden_size: int, tp_size: int): |
| 531 | """All-gather the input tensor interleavely across model parallel group.""" |
| 532 | import torch.distributed as dist |
| 533 | |
| 534 | gathered_tensors = [torch.zeros_like(local_tensor) for _ in range(tp_size)] |
| 535 | dist.all_gather( |
| 536 | gathered_tensors, local_tensor, group=parallel_state.get_tp_group().device_group |
| 537 | ) |
| 538 | |
| 539 | gathered_tensors_split = [ |
| 540 | torch.split(tensor, hidden_size // tp_size, -1) for tensor in gathered_tensors |
| 541 | ] |
| 542 | ordered_tensors = [ |
| 543 | tensor for pair in zip(*gathered_tensors_split) for tensor in pair |
| 544 | ] |
| 545 | result_tensor = torch.cat(ordered_tensors, dim=-1) |
| 546 | return result_tensor |
| 547 | |
| 548 | |
| 549 | class SiglipAttention(nn.Module): |
| 550 | """SigLIP vision attention adapted from Qwen2.5-VisionAttention.""" |
| 551 | |
| 552 | def __init__( |
| 553 | self, |
| 554 | *, |
| 555 | embed_dim: int, |
| 556 | num_heads: int, |
| 557 | projection_size: int, |
| 558 | quant_config: QuantizationConfig | None = None, |
| 559 | prefix: str = "", |
| 560 | ) -> None: |
| 561 | super().__init__() |
| 562 | |
| 563 | self.tp_size = parallel_state.get_tensor_model_parallel_world_size() |
| 564 | self.tp_rank = parallel_state.get_tensor_model_parallel_rank() |
| 565 | self.hidden_size_per_attention_head = dist_utils.divide( |
| 566 | projection_size, num_heads |
| 567 | ) |
| 568 | self.num_attention_heads_per_partition = dist_utils.divide( |
| 569 | num_heads, self.tp_size |
| 570 | ) |
| 571 | |
| 572 | self.qkv_proj = QKVParallelLinear( |
| 573 | hidden_size=embed_dim, |
| 574 | head_size=self.hidden_size_per_attention_head, |
| 575 | total_num_heads=num_heads, |
| 576 | total_num_kv_heads=num_heads, |
| 577 | bias=True, |
| 578 | quant_config=quant_config, |
| 579 | prefix=f"{prefix}.qkv_proj", |
| 580 | ) |
| 581 | self.out_proj = RowParallelLinear( |
| 582 | input_size=projection_size, |
| 583 | output_size=embed_dim, |
| 584 | quant_config=quant_config, |
| 585 | prefix=f"{prefix}.out_proj", |
| 586 | ) |
| 587 | self.attn = MMEncoderAttention( |
| 588 | num_heads=self.num_attention_heads_per_partition, |
| 589 | head_size=self.hidden_size_per_attention_head, |
| 590 | scale=self.hidden_size_per_attention_head**-0.5, |
| 591 | prefix=f"{prefix}.attn", |
| 592 | ) |
| 593 | self.apply_rotary_emb = ApplyRotaryEmb( |
| 594 | enforce_enable=True, |
| 595 | enable_fp32_compute=True, |
| 596 | ) |
| 597 | |
| 598 | def split_qkv(self, qkv: torch.Tensor) -> tuple[torch.Tensor, ...]: |
| 599 | seq_len, bs, _ = qkv.shape |
| 600 | if self.tp_size > 1: |
| 601 | qkv = all_gather_interleave(qkv, self.qkv_proj.hidden_size, self.tp_size) |
| 602 | |
| 603 | q, k, v = qkv.chunk(3, dim=2) |
| 604 | |
| 605 | if self.tp_size > 1: |
| 606 | splitter = partial( |
| 607 | dist_utils.split_tensor_along_last_dim, num_partitions=self.tp_size |
| 608 | ) |
| 609 | q = splitter(q)[self.tp_rank] |
| 610 | k = splitter(k)[self.tp_rank] |
| 611 | v = splitter(v)[self.tp_rank] |
| 612 | |
| 613 | new_shape = ( |
| 614 | seq_len, |
| 615 | bs, |
| 616 | self.num_attention_heads_per_partition, |
| 617 | self.hidden_size_per_attention_head, |
| 618 | ) |
| 619 | q, k, v = (x.view(*new_shape) for x in (q, k, v)) |
| 620 | return q, k, v |
| 621 | |
| 622 | def forward( |
| 623 | self, |
| 624 | hidden_states: torch.Tensor, |
| 625 | *, |
| 626 | cu_seqlens: torch.Tensor, |
| 627 | rotary_pos_emb: torch.Tensor | None, |
| 628 | max_seqlen: torch.Tensor | None, |
| 629 | ) -> torch.Tensor: |
| 630 | batch_size, _, _ = hidden_states.shape |
| 631 | |
| 632 | x = rearrange(hidden_states, "b s d -> s b d") |
| 633 | x, _ = self.qkv_proj(x) |
| 634 | q, k, v = self.split_qkv(x) |
| 635 | q, k, v = (rearrange(t, "s b h d -> b s h d") for t in (q, k, v)) |
| 636 | |
| 637 | if rotary_pos_emb is not None: |
| 638 | qk_concat = torch.cat([q, k], dim=0) |
| 639 | qk_rotated = self.apply_rotary_emb( |
| 640 | qk_concat, |
| 641 | rotary_pos_emb.cos(), |
| 642 | rotary_pos_emb.sin(), |
| 643 | ) |
| 644 | q, k = torch.chunk(qk_rotated, 2, dim=0) |
| 645 | |
| 646 | context_layer = self.attn( |
| 647 | query=q, |
| 648 | key=k, |
| 649 | value=v, |
| 650 | cu_seqlens=cu_seqlens, |
| 651 | max_seqlen=max_seqlen, |
| 652 | ) |
| 653 | context_layer = rearrange(context_layer, "b s h d -> b s (h d)") |
| 654 | |
| 655 | output, _ = self.out_proj(context_layer) |
| 656 | return output |
| 657 | |
| 658 | |
| 659 | class SigLIPRotaryEmbedding(nn.Module): |
| 660 | def __init__(self, dim: int, theta: float = 10000.0) -> None: |
| 661 | super().__init__() |
| 662 | self.dim = dim |
| 663 | self.theta = theta |
| 664 | self.rope_init() |
| 665 | |
| 666 | def rope_init(self): |
| 667 | inv_freq = 1.0 / ( |
| 668 | self.theta ** (torch.arange(0, self.dim, 2, dtype=torch.float) / self.dim) |
| 669 | ) |
| 670 | self.register_buffer("inv_freq", inv_freq, persistent=False) |
| 671 | |
| 672 | def forward(self, seqlen: int) -> torch.Tensor: |
| 673 | seq = torch.arange( |
| 674 | seqlen, |
| 675 | device=self.inv_freq.device, |
| 676 | dtype=self.inv_freq.dtype, |
| 677 | ) |
| 678 | freqs = torch.outer(seq, self.inv_freq) |
| 679 | return freqs |
| 680 | |
| 681 | |
| 682 | class SiglipEncoderLayer(nn.Module): |
| 683 | def __init__( |
| 684 | self, |
| 685 | config: PreTrainedConfig, |
| 686 | quant_config: QuantizationConfig | None = None, |
| 687 | prefix: str = "", |
| 688 | ): |
| 689 | super().__init__() |
| 690 | self.embed_dim = config.hidden_size |
| 691 | self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps) |
| 692 | self.self_attn = SiglipAttention( |
| 693 | embed_dim=config.hidden_size, |
| 694 | num_heads=config.num_attention_heads, |
| 695 | projection_size=config.hidden_size, |
| 696 | quant_config=quant_config, |
| 697 | prefix=f"{prefix}.self_attn", |
| 698 | ) |
| 699 | self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps) |
| 700 | self.mlp = SiglipMLP( |
| 701 | config, |
| 702 | quant_config=quant_config, |
| 703 | prefix=f"{prefix}.mlp", |
| 704 | ) |
| 705 | |
| 706 | def forward( |
| 707 | self, |
| 708 | hidden_states: torch.Tensor, |
| 709 | *, |
| 710 | cu_seqlens: torch.Tensor, |
| 711 | rotary_pos_emb: torch.Tensor | None, |
| 712 | max_seqlen: torch.Tensor | None, |
| 713 | ) -> torch.Tensor: |
| 714 | residual = hidden_states |
| 715 | |
| 716 | hidden_states = self.layer_norm1(hidden_states) |
| 717 | hidden_states = self.self_attn( |
| 718 | hidden_states=hidden_states, |
| 719 | cu_seqlens=cu_seqlens, |
| 720 | rotary_pos_emb=rotary_pos_emb, |
| 721 | max_seqlen=max_seqlen, |
| 722 | ) |
| 723 | |
| 724 | hidden_states = residual + hidden_states |
| 725 | |
| 726 | residual = hidden_states |
| 727 | hidden_states = self.layer_norm2(hidden_states) |
| 728 | hidden_states = self.mlp(hidden_states) |
| 729 | |
| 730 | hidden_states = residual + hidden_states |
| 731 | |
| 732 | return hidden_states |
| 733 | |
| 734 | |
| 735 | class SiglipEncoder(nn.Module): |
| 736 | def __init__( |
| 737 | self, |
| 738 | config: PreTrainedConfig, |
| 739 | quant_config: QuantizationConfig | None = None, |
| 740 | prefix: str = "", |
| 741 | ): |
| 742 | super().__init__() |
| 743 | self.config = config |
| 744 | embed_dim = config.hidden_size |
| 745 | num_heads = config.num_attention_heads |
| 746 | head_dim = embed_dim // num_heads |
| 747 | |
| 748 | self.attn_backend = get_vit_attn_backend( |
| 749 | head_size=head_dim, |
| 750 | dtype=torch.get_default_dtype(), |
| 751 | ) |
| 752 | |
| 753 | self.layers = nn.ModuleList( |
| 754 | [ |
| 755 | SiglipEncoderLayer( |
| 756 | config, |
| 757 | quant_config=quant_config, |
| 758 | prefix=f"{prefix}.layers.{layer_idx}", |
| 759 | ) |
| 760 | for layer_idx in range(config.num_hidden_layers) |
| 761 | ] |
| 762 | ) |
| 763 | self.rotary_pos_emb = SigLIPRotaryEmbedding(head_dim // 2) |
| 764 | |
| 765 | @staticmethod |
| 766 | def flatten_list(image_grid_thw): |
| 767 | tmp_image_grid_thw = list() |
| 768 | for image_grid in image_grid_thw: |
| 769 | if isinstance(image_grid, list): |
| 770 | tmp_image_grid_thw.extend(image_grid) |
| 771 | else: |
| 772 | tmp_image_grid_thw.append(image_grid) |
| 773 | return tmp_image_grid_thw |
| 774 | |
| 775 | def forward( |
| 776 | self, |
| 777 | inputs_embeds, |
| 778 | cu_seqlens: torch.Tensor | None = None, |
| 779 | image_grid_thw: list[tuple[int, int, int] | list[tuple[int, int, int]]] |
| 780 | | None = None, |
| 781 | height_position_ids: torch.Tensor | None = None, |
| 782 | width_position_ids: torch.Tensor | None = None, |
| 783 | ) -> torch.Tensor: |
| 784 | device = inputs_embeds.device |
| 785 | hidden_states = inputs_embeds |
| 786 | |
| 787 | flatten_image_grid_thw = self.flatten_list(image_grid_thw) |
| 788 | |
| 789 | if width_position_ids is None or height_position_ids is None: |
| 790 | split_hids = list() |
| 791 | split_wids = list() |
| 792 | for t, h, w in flatten_image_grid_thw: |
| 793 | image_pids = torch.arange(t * h * w, device=device) % (h * w) |
| 794 | sample_hids = image_pids // w |
| 795 | sample_wids = image_pids % w |
| 796 | split_hids.append(sample_hids) |
| 797 | split_wids.append(sample_wids) |
| 798 | width_position_ids = torch.concat(split_wids, dim=0) |
| 799 | height_position_ids = torch.concat(split_hids, dim=0) |
| 800 | |
| 801 | pids = torch.stack( |
| 802 | [height_position_ids, width_position_ids], |
| 803 | dim=-1, |
| 804 | ) |
| 805 | # The ids are built from the grids above, so `h`/`w` bound them |
| 806 | # and the table size is known on the host. |
| 807 | max_grid_size = max(max(h, w) for _, h, w in flatten_image_grid_thw) |
| 808 | rope_emb_max_grid = self.rotary_pos_emb(max_grid_size) |
| 809 | rotary_pos_emb = rope_emb_max_grid[pids].flatten(1) |
| 810 | |
| 811 | if cu_seqlens is None: |
| 812 | raise ValueError("cu_seqlens cannot be None for SiglipEncoder.") |
| 813 | if not isinstance(cu_seqlens, torch.Tensor): |
| 814 | cu_seqlens = async_tensor_h2d(cu_seqlens, dtype=torch.int32, device=device) |
| 815 | else: |
| 816 | cu_seqlens = cu_seqlens.to(device=device, non_blocking=True) |
| 817 | |
| 818 | max_seqlen = None |
| 819 | if self.attn_backend in { |
| 820 | AttentionBackendEnum.FLASH_ATTN, |
| 821 | AttentionBackendEnum.ROCM_AITER_FA, |
| 822 | AttentionBackendEnum.TRITON_ATTN, |
| 823 | }: |
| 824 | max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max() |
| 825 | |
| 826 | hidden_states = inputs_embeds |
| 827 | for encoder_layer in self.layers: |
| 828 | hidden_states = encoder_layer( |
| 829 | hidden_states, |
| 830 | cu_seqlens=cu_seqlens, |
| 831 | rotary_pos_emb=rotary_pos_emb, |
| 832 | max_seqlen=max_seqlen, |
| 833 | ) |
| 834 | return hidden_states |
| 835 | |
| 836 | |
| 837 | class SiglipVisionTransformer(nn.Module): |
| 838 | def __init__( |
| 839 | self, |
| 840 | config: PreTrainedConfig, |
| 841 | quant_config: QuantizationConfig | None = None, |
| 842 | prefix: str = "", |
| 843 | ): |
| 844 | super().__init__() |
| 845 | self.config = config |
| 846 | embed_dim = config.hidden_size |
| 847 | |
| 848 | self.embeddings = SiglipVisionEmbeddings(config) |
| 849 | self.encoder = SiglipEncoder( |
| 850 | config, |
| 851 | quant_config=quant_config, |
| 852 | prefix=f"{prefix}.encoder", |
| 853 | ) |
| 854 | self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) |
| 855 | |
| 856 | def forward( |
| 857 | self, |
| 858 | pixel_values: torch.Tensor, |
| 859 | interpolate_pos_encoding: bool | None = False, |
| 860 | position_ids: torch.Tensor | None = None, |
| 861 | height_position_ids: torch.Tensor | None = None, |
| 862 | width_position_ids: torch.Tensor | None = None, |
| 863 | cu_seqlens: torch.Tensor | None = None, |
| 864 | image_grid_thw: Sequence[tuple[int, int, int]] | None = None, |
| 865 | ) -> torch.Tensor: |
| 866 | hidden_states = self.embeddings( |
| 867 | pixel_values, |
| 868 | interpolate_pos_encoding=interpolate_pos_encoding, |
| 869 | position_ids=position_ids, |
| 870 | image_grid_thw=image_grid_thw, |
| 871 | ) |
| 872 | |
| 873 | last_hidden_state = self.encoder( |
| 874 | inputs_embeds=hidden_states, |
| 875 | cu_seqlens=cu_seqlens, |
| 876 | image_grid_thw=image_grid_thw, |
| 877 | height_position_ids=height_position_ids, |
| 878 | width_position_ids=width_position_ids, |
| 879 | ) |
| 880 | |
| 881 | last_hidden_state = self.post_layernorm(last_hidden_state) |
| 882 | return last_hidden_state |
| 883 | |
| 884 | |
| 885 | class SiglipVisionModel(nn.Module): |
| 886 | hf_to_vllm_mapper = WeightsMapper( |
| 887 | orig_to_new_stacked={ |
| 888 | ".q_proj": (".qkv_proj", "q"), |
| 889 | ".k_proj": (".qkv_proj", "k"), |
| 890 | ".v_proj": (".qkv_proj", "v"), |
| 891 | }, |
| 892 | # The SigLIP attention pooling head and packing pos embedding are |
| 893 | # present in the checkpoint but absent from this vision tower. |
| 894 | orig_to_new_substr={ |
| 895 | "head.attention": None, |
| 896 | "head.layernorm": None, |
| 897 | "head.mlp": None, |
| 898 | "head.probe": None, |
| 899 | "packing_position_embedding": None, |
| 900 | }, |
| 901 | ) |
| 902 | |
| 903 | def __init__( |
| 904 | self, |
| 905 | config, |
| 906 | quant_config: QuantizationConfig | None = None, |
| 907 | prefix: str = "", |
| 908 | ): |
| 909 | super().__init__() |
| 910 | |
| 911 | self.vision_model = SiglipVisionTransformer( |
| 912 | config, |
| 913 | quant_config=quant_config, |
| 914 | prefix=maybe_prefix(prefix, "vision_model"), |
| 915 | ) |
| 916 | self.quant_config = quant_config |
| 917 | |
| 918 | @property |
| 919 | def dtype(self) -> torch.dtype: |
| 920 | return self.vision_model.embeddings.patch_embedding.weight.dtype |
| 921 | |
| 922 | @property |
| 923 | def device(self) -> torch.device: |
| 924 | return self.vision_model.embeddings.patch_embedding.weight.device |
| 925 | |
| 926 | def get_input_embeddings(self) -> nn.Module: |
| 927 | return self.vision_model.embeddings.patch_embedding |
| 928 | |
| 929 | def forward( |
| 930 | self, |
| 931 | pixel_values, |
| 932 | interpolate_pos_encoding: bool = False, |
| 933 | position_ids: torch.Tensor | None = None, |
| 934 | image_grid_thw: Sequence[tuple[int, int, int]] | None = None, |
| 935 | cu_seqlens: torch.Tensor | None = None, |
| 936 | ) -> BaseModelOutputWithPooling: |
| 937 | return self.vision_model( |
| 938 | pixel_values=pixel_values, |
| 939 | interpolate_pos_encoding=interpolate_pos_encoding, |
| 940 | position_ids=position_ids, |
| 941 | image_grid_thw=image_grid_thw, |
| 942 | cu_seqlens=cu_seqlens, |
| 943 | ) |
| 944 | |
| 945 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 946 | loader = AutoWeightsLoader(self) |
| 947 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 948 | |
| 949 | |
| 950 | @MULTIMODAL_REGISTRY.register_processor( |
| 951 | PaddleOCRVLMultiModalProcessor, |
| 952 | info=PaddleOCRVLProcessingInfo, |
| 953 | dummy_inputs=PaddleOCRVLDummyInputsBuilder, |
| 954 | ) |
| 955 | class PaddleOCRVLForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsMRoPE): |
| 956 | hf_to_vllm_mapper = WeightsMapper( |
| 957 | orig_to_new_prefix={ |
| 958 | "model.": "language_model.model.", |
| 959 | "lm_head.": "language_model.lm_head.", |
| 960 | } |
| 961 | ) |
| 962 | |
| 963 | @classmethod |
| 964 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 965 | if modality.startswith("image"): |
| 966 | return "<|IMAGE_START|><|IMAGE_PLACEHOLDER|><|IMAGE_END|>" |
| 967 | |
| 968 | raise ValueError("Only image modality is supported") |
| 969 | |
| 970 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 971 | super().__init__() |
| 972 | config = vllm_config.model_config.hf_config |
| 973 | if hasattr(config, "text_config"): |
| 974 | text_config = config.text_config.to_dict() |
| 975 | unsafe_keys = ["model_type", "architectures", "tie_word_embeddings"] |
| 976 | for key in unsafe_keys: |
| 977 | text_config.pop(key, None) |
| 978 | config.update(text_config) |
| 979 | |
| 980 | quant_config = vllm_config.quant_config |
| 981 | |
| 982 | self.config = config |
| 983 | |
| 984 | with self._mark_tower_model(vllm_config, "image"): |
| 985 | self.visual = SiglipVisionModel( |
| 986 | config=config.vision_config, |
| 987 | quant_config=quant_config, |
| 988 | prefix=maybe_prefix(prefix, "visual"), |
| 989 | ) |
| 990 | self.mlp_AR = Projector(config, config.vision_config) |
| 991 | |
| 992 | with self._mark_language_model(vllm_config): |
| 993 | self.language_model = Ernie4_5ForCausalLM( |
| 994 | vllm_config=vllm_config, |
| 995 | prefix=maybe_prefix(prefix, "language_model"), |
| 996 | ) |
| 997 | |
| 998 | for layer in self.language_model.model.layers: |
| 999 | if not isinstance(layer, PPMissingLayer): |
| 1000 | layer.self_attn.rotary_emb.is_neox_style = True |
| 1001 | |
| 1002 | self.make_empty_intermediate_tensors = ( |
| 1003 | self.language_model.make_empty_intermediate_tensors |
| 1004 | ) |
| 1005 | |
| 1006 | def compute_logits( |
| 1007 | self, |
| 1008 | hidden_states: torch.Tensor, |
| 1009 | ) -> torch.Tensor | None: |
| 1010 | return self.language_model.compute_logits(hidden_states) |
| 1011 | |
| 1012 | def iter_mm_grid_thw( |
| 1013 | self, mm_features: list[MultiModalFeatureSpec] |
| 1014 | ) -> Iterator[tuple[int, int, int, int, float]]: |
| 1015 | """Iterate over multimodal features and yield grid information. |
| 1016 | |
| 1017 | Args: |
| 1018 | mm_features: List of multimodal feature specifications |
| 1019 | |
| 1020 | Yields: |
| 1021 | Tuple of (offset, grid_t, grid_h, grid_w, t_factor) for each frame/image |
| 1022 | |
| 1023 | """ |
| 1024 | spatial_merge_size = self.config.vision_config.spatial_merge_size |
| 1025 | tokens_per_second = getattr(self.config.vision_config, "tokens_per_second", 1.0) |
| 1026 | for mm_feature in sorted(mm_features, key=lambda f: f.mm_position.offset): |
| 1027 | offset = mm_feature.mm_position.offset |
| 1028 | feature_data = mm_feature.data |
| 1029 | assert feature_data is not None |
| 1030 | if mm_feature.modality == "image": |
| 1031 | grid_data = feature_data["image_grid_thw"].data |
| 1032 | assert isinstance(grid_data, torch.Tensor) |
| 1033 | t, h, w = grid_data.tolist() |
| 1034 | assert t == 1, f"Image must have 1 frame, got {t}" |
| 1035 | yield offset, 1, h // spatial_merge_size, w // spatial_merge_size, 1.0 |
| 1036 | elif mm_feature.modality == "video": |
| 1037 | grid_data = feature_data["video_grid_thw"].data |
| 1038 | assert isinstance(grid_data, torch.Tensor) |
| 1039 | t, h, w = grid_data.tolist() |
| 1040 | second_per_grid_ts = 1.0 |
| 1041 | second_per_grid_item = feature_data.get("second_per_grid_ts") |
| 1042 | if second_per_grid_item is not None: |
| 1043 | second_per_grid_data = second_per_grid_item.data |
| 1044 | assert isinstance(second_per_grid_data, torch.Tensor) |
| 1045 | second_per_grid_ts = second_per_grid_data.item() |
| 1046 | t_factor = second_per_grid_ts * tokens_per_second |
| 1047 | yield ( |
| 1048 | offset, |
| 1049 | t, |
| 1050 | h // spatial_merge_size, |
| 1051 | w // spatial_merge_size, |
| 1052 | t_factor, |
| 1053 | ) |
| 1054 | else: |
| 1055 | raise ValueError(f"Unsupported modality: {mm_feature.modality}") |
| 1056 | |
| 1057 | def get_mrope_input_positions( |
| 1058 | self, |
| 1059 | input_tokens: list[int], |
| 1060 | mm_features: list[MultiModalFeatureSpec], |
| 1061 | ) -> tuple[torch.Tensor, int]: |
| 1062 | llm_pos_ids_list: list = [] |
| 1063 | st = 0 |
| 1064 | |
| 1065 | for ( |
| 1066 | offset, |
| 1067 | llm_grid_t, |
| 1068 | llm_grid_h, |
| 1069 | llm_grid_w, |
| 1070 | t_factor, |
| 1071 | ) in self.iter_mm_grid_thw(mm_features): |
| 1072 | text_len = offset - st |
| 1073 | st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0 |
| 1074 | llm_pos_ids_list.append( |
| 1075 | np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx |
| 1076 | ) |
| 1077 | |
| 1078 | grid_indices = np.indices((llm_grid_t, llm_grid_h, llm_grid_w)) |
| 1079 | if t_factor != 1.0: |
| 1080 | grid_indices[0] = (grid_indices[0] * t_factor).astype(np.int64) |
| 1081 | |
| 1082 | llm_pos_ids_list.append(grid_indices.reshape(3, -1) + text_len + st_idx) |
| 1083 | st = offset + llm_grid_t * llm_grid_h * llm_grid_w |
| 1084 | |
| 1085 | if st < len(input_tokens): |
| 1086 | st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0 |
| 1087 | text_len = len(input_tokens) - st |
| 1088 | llm_pos_ids_list.append( |
| 1089 | np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx |
| 1090 | ) |
| 1091 | |
| 1092 | llm_positions = np.concatenate(llm_pos_ids_list, axis=1).reshape(3, -1) |
| 1093 | mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item() |
| 1094 | |
| 1095 | return torch.from_numpy(llm_positions), mrope_position_delta |
| 1096 | |
| 1097 | def _parse_and_validate_image_input( |
| 1098 | self, **kwargs: object |
| 1099 | ) -> PaddleOCRImagePixelInputs | None: |
| 1100 | pixel_values = kwargs.pop("pixel_values", None) |
| 1101 | image_grid_thw = kwargs.pop("image_grid_thw", None) |
| 1102 | |
| 1103 | if pixel_values is None: |
| 1104 | return None |
| 1105 | |
| 1106 | return PaddleOCRImagePixelInputs( |
| 1107 | type="pixel_values", |
| 1108 | pixel_values=pixel_values, |
| 1109 | image_grid_thw=image_grid_thw, |
| 1110 | ) |
| 1111 | |
| 1112 | def forward( |
| 1113 | self, |
| 1114 | input_ids: torch.Tensor | None, |
| 1115 | positions: torch.Tensor, |
| 1116 | intermediate_tensors: IntermediateTensors | None = None, |
| 1117 | inputs_embeds: torch.Tensor | None = None, |
| 1118 | **kwargs, |
| 1119 | ): |
| 1120 | if intermediate_tensors is not None: |
| 1121 | inputs_embeds = None |
| 1122 | |
| 1123 | return self.language_model( |
| 1124 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 1125 | ) |
| 1126 | |
| 1127 | def encode_image( |
| 1128 | self, pixel_values: torch.Tensor, image_grid_thw: torch.Tensor |
| 1129 | ) -> torch.Tensor: |
| 1130 | pixel_values = pixel_values.type(self.visual.dtype) |
| 1131 | siglip_position_ids = list() |
| 1132 | image_grid_hws = list() |
| 1133 | cu_seqlens = [0] |
| 1134 | |
| 1135 | thw_tuple = tuple(image_grid_thw.tolist()) |
| 1136 | numel = np.prod(thw_tuple) |
| 1137 | image_grid_hws.append(thw_tuple) |
| 1138 | image_position_ids = torch.arange(numel) % np.prod(thw_tuple[1:]) |
| 1139 | siglip_position_ids.append(image_position_ids) |
| 1140 | cu_seqlens.append(cu_seqlens[-1] + numel) |
| 1141 | |
| 1142 | # Both are built on the host; concat straight into a pinned buffer |
| 1143 | # so the H2D copy stays non-blocking. |
| 1144 | siglip_position_ids = torch.concat( |
| 1145 | siglip_position_ids, |
| 1146 | dim=0, |
| 1147 | out=torch.empty( |
| 1148 | sum(t.numel() for t in siglip_position_ids), |
| 1149 | dtype=torch.int64, |
| 1150 | pin_memory=PIN_MEMORY, |
| 1151 | ), |
| 1152 | ).to(pixel_values.device, non_blocking=True) |
| 1153 | cu_seqlens = async_tensor_h2d( |
| 1154 | cu_seqlens, dtype=torch.int32, device=pixel_values.device |
| 1155 | ) |
| 1156 | |
| 1157 | vision_outputs = self.visual( |
| 1158 | pixel_values=pixel_values, |
| 1159 | image_grid_thw=image_grid_hws, |
| 1160 | position_ids=siglip_position_ids, |
| 1161 | interpolate_pos_encoding=True, |
| 1162 | cu_seqlens=cu_seqlens, |
| 1163 | ) |
| 1164 | return vision_outputs |
| 1165 | |
| 1166 | def _process_image_input( |
| 1167 | self, image_input: PaddleOCRImagePixelInputs |
| 1168 | ) -> MultiModalEmbeddings: |
| 1169 | pixel_values = image_input.pixel_values |
| 1170 | image_grid_thw = image_input.image_grid_thw |
| 1171 | vision_outputs = tuple( |
| 1172 | self.encode_image(pixel, grid).squeeze(0) |
| 1173 | for pixel, grid in zip(pixel_values, image_grid_thw) |
| 1174 | ) |
| 1175 | image_embeds = self.mlp_AR(vision_outputs, image_grid_thw) |
| 1176 | return image_embeds |
| 1177 | |
| 1178 | def embed_multimodal(self, **kwargs) -> MultiModalEmbeddings: |
| 1179 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 1180 | if image_input is None: |
| 1181 | return () |
| 1182 | |
| 1183 | multimodal_embeddings: tuple[torch.Tensor, ...] = () |
| 1184 | image_embeds = self._process_image_input(image_input) |
| 1185 | multimodal_embeddings += tuple(image_embeds) |
| 1186 | |
| 1187 | return multimodal_embeddings |
| 1188 | |
| 1189 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1190 | loader = AutoWeightsLoader(self) |
| 1191 | autoloaded_weights = loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 1192 | return autoloaded_weights |