| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inference-only MiniCPM-V 4.6 model (MiniCPMV4_6ForConditionalGeneration).""" |
| 4 | |
| 5 | from collections.abc import Iterable, Mapping |
| 6 | from typing import Any |
| 7 | |
| 8 | import numpy as np |
| 9 | import torch |
| 10 | from PIL import Image as PILImage |
| 11 | from torch import nn |
| 12 | from transformers import MiniCPMV4_6Config |
| 13 | |
| 14 | from vllm.config import VllmConfig |
| 15 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 16 | from vllm.model_executor.layers.activation import get_act_fn |
| 17 | from vllm.model_executor.layers.attention import MMEncoderAttention |
| 18 | from vllm.model_executor.layers.linear import ( |
| 19 | QKVParallelLinear, |
| 20 | RowParallelLinear, |
| 21 | ) |
| 22 | from vllm.model_executor.layers.mamba.mamba_utils import ( |
| 23 | MambaStateCopyFuncCalculator, |
| 24 | MambaStateDtypeCalculator, |
| 25 | MambaStateShapeCalculator, |
| 26 | ) |
| 27 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 28 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 29 | from vllm.multimodal.inputs import ( |
| 30 | MultiModalFeatureSpec, |
| 31 | MultiModalFieldConfig, |
| 32 | NestedTensors, |
| 33 | ) |
| 34 | from vllm.multimodal.parse import ImageProcessorItems, ImageSize, VideoProcessorItems |
| 35 | from vllm.multimodal.processing.processor import ( |
| 36 | PromptReplacement, |
| 37 | PromptUpdateDetails, |
| 38 | ResolvedPromptUpdate, |
| 39 | cached_encode, |
| 40 | ) |
| 41 | from vllm.sequence import IntermediateTensors |
| 42 | |
| 43 | from .idefics2_vision_model import Idefics2VisionTransformer |
| 44 | from .interfaces import ( |
| 45 | HasInnerState, |
| 46 | IsHybrid, |
| 47 | MultiModalEmbeddings, |
| 48 | SupportsLoRA, |
| 49 | SupportsMRoPE, |
| 50 | SupportsMultiModal, |
| 51 | SupportsPP, |
| 52 | _require_is_multimodal, |
| 53 | ) |
| 54 | from .minicpmv import ( |
| 55 | MiniCPMVDummyInputsBuilder, |
| 56 | MiniCPMVImageEmbeddingInputs, |
| 57 | MiniCPMVImageEmbeddingItems, |
| 58 | MiniCPMVImagePixelInputs, |
| 59 | MiniCPMVMultiModalProcessor, |
| 60 | MiniCPMVProcessingInfo, |
| 61 | MiniCPMVVideoEmbeddingItems, |
| 62 | _image_kwargs_from_video, |
| 63 | ) |
| 64 | from .module_mapping import MultiModelKeys |
| 65 | from .qwen3_5 import Qwen3_5ForCausalLM |
| 66 | from .utils import ( |
| 67 | AutoWeightsLoader, |
| 68 | WeightsMapper, |
| 69 | _merge_multimodal_embeddings, |
| 70 | flatten_bn, |
| 71 | maybe_prefix, |
| 72 | ) |
| 73 | from .vision import is_vit_use_data_parallel |
| 74 | |
| 75 | |
| 76 | def _minicpmv4_6_field_config(hf_inputs: Mapping[str, torch.Tensor]): |
| 77 | fields = dict( |
| 78 | pixel_values=MultiModalFieldConfig.batched("image"), |
| 79 | tgt_sizes=MultiModalFieldConfig.batched("image"), |
| 80 | image_embeds=MultiModalFieldConfig.batched("image"), |
| 81 | video_pixel_values=MultiModalFieldConfig.batched("video"), |
| 82 | video_image_sizes=MultiModalFieldConfig.batched("video", keep_on_cpu=True), |
| 83 | video_tgt_sizes=MultiModalFieldConfig.batched("video"), |
| 84 | video_embeds=MultiModalFieldConfig.batched("video"), |
| 85 | ) |
| 86 | if "use_vit_merger" in hf_inputs: |
| 87 | fields["use_vit_merger"] = MultiModalFieldConfig.batched( |
| 88 | "image", keep_on_cpu=True |
| 89 | ) |
| 90 | return fields |
| 91 | |
| 92 | |
| 93 | class MiniCPMV4_6MultiModalProcessor(MiniCPMVMultiModalProcessor): |
| 94 | def _resolve_downsample_mode( |
| 95 | self, |
| 96 | mm_kwargs: Mapping[str, object], |
| 97 | ) -> str: |
| 98 | ds = mm_kwargs.get("downsample_mode") |
| 99 | if ds is not None: |
| 100 | return str(ds) |
| 101 | info = self.info |
| 102 | assert isinstance(info, MiniCPMV4_6ProcessingInfo) |
| 103 | return info._get_downsample_mode() |
| 104 | |
| 105 | def get_image_prompt_texts( |
| 106 | self, |
| 107 | image_size, |
| 108 | image_idx: int = 0, |
| 109 | downsample_mode: str | None = None, |
| 110 | ) -> str: |
| 111 | info = self.info |
| 112 | assert isinstance(info, MiniCPMV4_6ProcessingInfo) |
| 113 | return info.get_slice_image_placeholder( |
| 114 | image_size, |
| 115 | image_idx=image_idx, |
| 116 | downsample_mode=downsample_mode, |
| 117 | ) |
| 118 | |
| 119 | def get_video_prompt_texts( |
| 120 | self, |
| 121 | image_size, |
| 122 | num_frames: int, |
| 123 | downsample_mode: str | None = None, |
| 124 | video_idx: int = 0, |
| 125 | ) -> str: |
| 126 | # Match transformers v5.7+ MiniCPMV4_6Processor video formatting: |
| 127 | # <image_id>{video_idx}</image_id>(<image>VIDEO*src</image> |
| 128 | # <slice>VIDEO*patch</slice>...)*num_frames |
| 129 | # Crucially the visual token inside each frame is ``<|video_pad|>`` |
| 130 | # (tokenizer.video_token), NOT ``<|image_pad|>`` — they share the same |
| 131 | # embedding-injection role but the language model is conditioned on |
| 132 | # which one is used. Using image_token for video silently produces |
| 133 | # garbage descriptions. |
| 134 | info = self.info |
| 135 | assert isinstance(info, MiniCPMV4_6ProcessingInfo) |
| 136 | grids, source_tokens, patch_tokens = info._compute_visual_tokens( |
| 137 | image_size, |
| 138 | max_slice_nums=info.get_video_max_slice_num(), |
| 139 | downsample_mode=downsample_mode, |
| 140 | ) |
| 141 | tokenizer = info.get_tokenizer() |
| 142 | video_token = getattr(tokenizer, "video_token", "<|video_pad|>") |
| 143 | image_start = getattr(tokenizer, "image_start_token", "<image>") |
| 144 | image_end = getattr(tokenizer, "image_end_token", "</image>") |
| 145 | slice_start = getattr(tokenizer, "slice_start_token", "<slice>") |
| 146 | slice_end = getattr(tokenizer, "slice_end_token", "</slice>") |
| 147 | id_start = getattr(tokenizer, "image_id_start_token", "<image_id>") |
| 148 | id_end = getattr(tokenizer, "image_id_end_token", "</image_id>") |
| 149 | |
| 150 | per_frame = image_start + video_token * source_tokens + image_end |
| 151 | if grids[0] > 0 and grids[1] > 0 and patch_tokens > 0: |
| 152 | slice_ph = slice_start + video_token * patch_tokens + slice_end |
| 153 | rows = [slice_ph * grids[1] for _ in range(grids[0])] |
| 154 | per_frame += "\n".join(rows) |
| 155 | |
| 156 | body = per_frame * num_frames |
| 157 | return f"{id_start}{video_idx}{id_end}" + body |
| 158 | |
| 159 | def process_images( |
| 160 | self, |
| 161 | mm_data: Mapping[str, object], |
| 162 | mm_kwargs: Mapping[str, object], |
| 163 | ) -> Mapping[str, NestedTensors]: |
| 164 | if (images := mm_data.get("images")) is None: |
| 165 | return {} |
| 166 | |
| 167 | mm_items = self.info.parse_mm_data({"image": images}, validate=False) |
| 168 | parsed_images = mm_items.get_items( |
| 169 | "image", (MiniCPMVImageEmbeddingItems, ImageProcessorItems) |
| 170 | ) |
| 171 | |
| 172 | if isinstance(parsed_images, MiniCPMVImageEmbeddingItems): |
| 173 | return {} |
| 174 | |
| 175 | # transformers v5.7+ MiniCPMV4_6ImageProcessor returns |
| 176 | # `pixel_values` (1, C, P, sum_W) where all slices are fused along W |
| 177 | # (NaViT-style), and `target_sizes` (n_slices, 2). vLLM expects each |
| 178 | # image entry to be a 4D tensor (n_slices, C, P, L_max_padded). |
| 179 | n_images = len(parsed_images) |
| 180 | image_processor = self.info.get_image_processor() |
| 181 | patch_size = image_processor.patch_size |
| 182 | per_image_pixel_values: list[torch.Tensor] = [] |
| 183 | per_image_tgt_sizes: list[torch.Tensor] = [] |
| 184 | for image in parsed_images: |
| 185 | ip_out = image_processor([image], **mm_kwargs) |
| 186 | pv = ip_out["pixel_values"] # (1, C, P, sum_W) |
| 187 | ts = ip_out["target_sizes"] # (n_slices, 2) |
| 188 | if pv.ndim == 4 and pv.shape[0] == 1: |
| 189 | pv = pv.squeeze(0) # (C, P, sum_W) |
| 190 | ts_long = ts.to(torch.long) |
| 191 | split_widths = (ts_long[:, 0] * ts_long[:, 1] * patch_size).tolist() |
| 192 | slices = torch.split(pv, split_widths, dim=-1) |
| 193 | n_slices = len(slices) |
| 194 | l_max = max(s.shape[-1] for s in slices) |
| 195 | out = torch.zeros( |
| 196 | n_slices, |
| 197 | pv.shape[0], |
| 198 | pv.shape[1], |
| 199 | l_max, |
| 200 | dtype=pv.dtype, |
| 201 | device=pv.device, |
| 202 | ) |
| 203 | for i, s in enumerate(slices): |
| 204 | out[i, :, :, : s.shape[-1]] = s |
| 205 | per_image_pixel_values.append(out) |
| 206 | per_image_tgt_sizes.append(ts_long) |
| 207 | |
| 208 | image_inputs: dict = { |
| 209 | "pixel_values": per_image_pixel_values, |
| 210 | "tgt_sizes": per_image_tgt_sizes, |
| 211 | } |
| 212 | |
| 213 | ds_mode = self._resolve_downsample_mode(mm_kwargs) |
| 214 | insert_layer_id = getattr( |
| 215 | self.info.get_hf_config(), |
| 216 | "insert_layer_id", |
| 217 | -1, |
| 218 | ) |
| 219 | merger_flag = ds_mode != "4x" and insert_layer_id >= 0 |
| 220 | image_inputs["use_vit_merger"] = [ |
| 221 | torch.tensor([merger_flag], dtype=torch.bool) for _ in range(n_images) |
| 222 | ] |
| 223 | return image_inputs |
| 224 | |
| 225 | def process_videos( |
| 226 | self, |
| 227 | mm_data: Mapping[str, object], |
| 228 | mm_kwargs: Mapping[str, object], |
| 229 | ) -> Mapping[str, NestedTensors]: |
| 230 | if (videos := mm_data.get("videos")) is None: |
| 231 | return {} |
| 232 | |
| 233 | mm_items = self.info.parse_mm_data({"video": videos}, validate=False) |
| 234 | parsed_videos = mm_items.get_items( |
| 235 | "video", (MiniCPMVVideoEmbeddingItems, VideoProcessorItems) |
| 236 | ) |
| 237 | |
| 238 | if isinstance(parsed_videos, MiniCPMVVideoEmbeddingItems): |
| 239 | return {} |
| 240 | |
| 241 | # Treat each video as a sequence of frames. The transformers v5.7+ |
| 242 | # `MiniCPMV4_6ImageProcessor` returns NaViT-style fused `pixel_values`; |
| 243 | # we run it per-frame, split the slices, then re-pack each video into |
| 244 | # a single 4D tensor (sum_slices, C, P, L_max_video). |
| 245 | image_processor = self.info.get_image_processor() |
| 246 | patch_size = image_processor.patch_size |
| 247 | video_max_slice = self.info.get_video_max_slice_num() |
| 248 | video_mm_kwargs = {**mm_kwargs, "max_slice_nums": video_max_slice} |
| 249 | |
| 250 | per_video_pixel_values: list[torch.Tensor] = [] |
| 251 | per_video_tgt_sizes: list[torch.Tensor] = [] |
| 252 | per_video_image_sizes: list[torch.Tensor] = [] |
| 253 | |
| 254 | for video in parsed_videos: |
| 255 | # video is iterable of frames (PIL Image or numpy array). |
| 256 | all_slices: list[torch.Tensor] = [] |
| 257 | ts_list: list[torch.Tensor] = [] |
| 258 | frame_sizes: list[torch.Tensor] = [] |
| 259 | for frame in video: |
| 260 | # Record per-frame (W, H) for video_image_sizes so that |
| 261 | # get_video_prompt_texts can consume a consistent frame size. |
| 262 | if isinstance(frame, PILImage.Image): |
| 263 | w, h = frame.size |
| 264 | elif isinstance(frame, np.ndarray): |
| 265 | if frame.ndim == 3 and frame.shape[-1] in (1, 3, 4): |
| 266 | # HWC (e.g. from np.array(PIL.Image)) |
| 267 | h, w = frame.shape[0], frame.shape[1] |
| 268 | else: |
| 269 | # CHW |
| 270 | _, h, w = frame.shape |
| 271 | elif isinstance(frame, torch.Tensor): |
| 272 | if frame.ndim == 3 and frame.shape[-1] in (1, 3, 4): |
| 273 | h, w = frame.shape[0], frame.shape[1] |
| 274 | else: |
| 275 | _, h, w = frame.shape |
| 276 | else: |
| 277 | raise TypeError(f"Unsupported frame type: {type(frame)}") |
| 278 | frame_sizes.append(torch.tensor([w, h], dtype=torch.long, device="cpu")) |
| 279 | |
| 280 | ip_out = image_processor([frame], **video_mm_kwargs) |
| 281 | pv = ip_out["pixel_values"] # (1, C, P, sum_W) |
| 282 | ts = ip_out["target_sizes"] # (n_slices, 2) |
| 283 | if pv.ndim == 4 and pv.shape[0] == 1: |
| 284 | pv = pv.squeeze(0) # (C, P, sum_W) |
| 285 | ts_long = ts.to(torch.long) |
| 286 | split_widths = (ts_long[:, 0] * ts_long[:, 1] * patch_size).tolist() |
| 287 | slices = torch.split(pv, split_widths, dim=-1) |
| 288 | all_slices.extend(slices) |
| 289 | ts_list.append(ts_long) |
| 290 | |
| 291 | if not all_slices: |
| 292 | continue |
| 293 | |
| 294 | l_max = max(s.shape[-1] for s in all_slices) |
| 295 | n_total = len(all_slices) |
| 296 | C, P = all_slices[0].shape[0], all_slices[0].shape[1] |
| 297 | out = torch.zeros( |
| 298 | n_total, |
| 299 | C, |
| 300 | P, |
| 301 | l_max, |
| 302 | dtype=all_slices[0].dtype, |
| 303 | device=all_slices[0].device, |
| 304 | ) |
| 305 | for i, s in enumerate(all_slices): |
| 306 | out[i, :, :, : s.shape[-1]] = s |
| 307 | |
| 308 | per_video_pixel_values.append(out) |
| 309 | per_video_tgt_sizes.append(torch.cat(ts_list, dim=0)) |
| 310 | per_video_image_sizes.append(torch.stack(frame_sizes)) |
| 311 | |
| 312 | if not per_video_pixel_values: |
| 313 | return {} |
| 314 | |
| 315 | return { |
| 316 | "video_pixel_values": per_video_pixel_values, |
| 317 | "video_tgt_sizes": per_video_tgt_sizes, |
| 318 | "video_image_sizes": per_video_image_sizes, |
| 319 | } |
| 320 | |
| 321 | def _get_prompt_updates( |
| 322 | self, |
| 323 | mm_items, |
| 324 | hf_processor_mm_kwargs: Mapping[str, object], |
| 325 | out_mm_kwargs, |
| 326 | ): |
| 327 | ds_mode = self._resolve_downsample_mode(hf_processor_mm_kwargs) |
| 328 | |
| 329 | placeholders = [ |
| 330 | ("image", self.info.image_pattern), |
| 331 | ("video", self.info.video_pattern), |
| 332 | ] |
| 333 | tokenizer = self.info.get_tokenizer() |
| 334 | additional_placeholders = [] |
| 335 | for modality, pattern in placeholders: |
| 336 | sub_pattern = tokenizer.decode( |
| 337 | cached_encode(tokenizer, pattern, add_special_tokens=False) |
| 338 | ) |
| 339 | if sub_pattern != pattern: |
| 340 | additional_placeholders.append((modality, sub_pattern)) |
| 341 | placeholders += additional_placeholders |
| 342 | |
| 343 | # The 4.6 chat_template emits `<|image_pad|>` / `<|video_pad|>` rather |
| 344 | # than `<unk>`, so use those tokens as the embedding selector. |
| 345 | image_embed_text = getattr(tokenizer, "image_token", "<|image_pad|>") |
| 346 | video_embed_text = getattr(tokenizer, "video_token", "<|video_pad|>") |
| 347 | vocab = tokenizer.get_vocab() |
| 348 | image_embed_ids = [vocab[image_embed_text]] |
| 349 | video_embed_ids = [vocab[video_embed_text]] |
| 350 | |
| 351 | def get_image_replacement(item_idx: int): |
| 352 | images = mm_items.get_items( |
| 353 | "image", |
| 354 | (MiniCPMVImageEmbeddingItems, ImageProcessorItems), |
| 355 | ) |
| 356 | image_size = images.get_image_size(item_idx) |
| 357 | return PromptUpdateDetails.select_token_ids( |
| 358 | cached_encode( |
| 359 | tokenizer, |
| 360 | self.get_image_prompt_texts( |
| 361 | image_size, |
| 362 | item_idx, |
| 363 | downsample_mode=ds_mode, |
| 364 | ), |
| 365 | add_special_tokens=False, |
| 366 | ), |
| 367 | image_embed_ids, |
| 368 | ) |
| 369 | |
| 370 | def get_video_replacement(item_idx: int): |
| 371 | # Prefer video_image_sizes from processed data so that the |
| 372 | # placeholder count is driven by the same frame sizes that the |
| 373 | # vision tower will actually consume. |
| 374 | video_mm_kwargs = out_mm_kwargs.get("video") |
| 375 | if video_mm_kwargs is not None and item_idx < len(video_mm_kwargs): |
| 376 | video_item = video_mm_kwargs[item_idx] |
| 377 | image_sizes_elem = video_item.get("video_image_sizes") |
| 378 | if image_sizes_elem is not None and image_sizes_elem.data is not None: |
| 379 | # image_sizes_elem.data: (num_frames, 2) – each row is [W, H] |
| 380 | image_sizes = image_sizes_elem.data |
| 381 | num_frames = image_sizes.shape[0] |
| 382 | frame_size = ImageSize( |
| 383 | width=int(image_sizes[0, 0].item()), |
| 384 | height=int(image_sizes[0, 1].item()), |
| 385 | ) |
| 386 | return PromptUpdateDetails.select_token_ids( |
| 387 | cached_encode( |
| 388 | tokenizer, |
| 389 | self.get_video_prompt_texts( |
| 390 | frame_size, |
| 391 | num_frames, |
| 392 | downsample_mode=ds_mode, |
| 393 | video_idx=item_idx, |
| 394 | ), |
| 395 | add_special_tokens=False, |
| 396 | ), |
| 397 | video_embed_ids, |
| 398 | ) |
| 399 | |
| 400 | videos = mm_items.get_items( |
| 401 | "video", |
| 402 | (MiniCPMVVideoEmbeddingItems, VideoProcessorItems), |
| 403 | ) |
| 404 | frame_size = videos.get_frame_size(item_idx) |
| 405 | num_frames = videos.get_num_frames(item_idx) |
| 406 | return PromptUpdateDetails.select_token_ids( |
| 407 | cached_encode( |
| 408 | tokenizer, |
| 409 | self.get_video_prompt_texts( |
| 410 | frame_size, |
| 411 | num_frames, |
| 412 | downsample_mode=ds_mode, |
| 413 | video_idx=item_idx, |
| 414 | ), |
| 415 | add_special_tokens=False, |
| 416 | ), |
| 417 | video_embed_ids, |
| 418 | ) |
| 419 | |
| 420 | get_replacement = { |
| 421 | "image": get_image_replacement, |
| 422 | "video": get_video_replacement, |
| 423 | } |
| 424 | |
| 425 | return [ |
| 426 | PromptReplacement( |
| 427 | modality=modality, |
| 428 | target=cached_encode(tokenizer, pattern, add_special_tokens=False), |
| 429 | replacement=get_replacement[modality], |
| 430 | ) |
| 431 | for modality, pattern in placeholders |
| 432 | ] |
| 433 | |
| 434 | def _recompute_cached_prompt_update( |
| 435 | self, cached_update: ResolvedPromptUpdate, new_item_idx: int |
| 436 | ) -> ResolvedPromptUpdate: |
| 437 | new_update = super()._recompute_cached_prompt_update( |
| 438 | cached_update, new_item_idx |
| 439 | ) |
| 440 | # MiniCPM-V 4.6 prefixes video placeholders with `<image_id>{idx}</image_id>` |
| 441 | # (the base class only rewrites the image modality). |
| 442 | if cached_update.modality == "video": |
| 443 | tokenizer = self.info.get_tokenizer() |
| 444 | id_start = getattr(tokenizer, "image_id_start_token", "<image_id>") |
| 445 | id_end = getattr(tokenizer, "image_id_end_token", "</image_id>") |
| 446 | video_token = getattr(tokenizer, "video_token", "<|video_pad|>") |
| 447 | |
| 448 | text = tokenizer.decode(cached_update.content.full) |
| 449 | prev_item_idx = cached_update.item_idx |
| 450 | |
| 451 | new_update = new_update.with_content( |
| 452 | PromptUpdateDetails.select_token_ids( |
| 453 | cached_encode( |
| 454 | tokenizer, |
| 455 | text.replace( |
| 456 | f"{id_start}{prev_item_idx}{id_end}", |
| 457 | f"{id_start}{new_item_idx}{id_end}", |
| 458 | 1, |
| 459 | ), |
| 460 | add_special_tokens=False, |
| 461 | ), |
| 462 | cached_encode(tokenizer, video_token, add_special_tokens=False), |
| 463 | ) |
| 464 | ) |
| 465 | return new_update |
| 466 | |
| 467 | def _get_mm_fields_config( |
| 468 | self, |
| 469 | hf_inputs, |
| 470 | hf_processor_mm_kwargs: Mapping[str, object], |
| 471 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 472 | return _minicpmv4_6_field_config(hf_inputs) |
| 473 | |
| 474 | |
| 475 | class MiniCPMV4_6ProcessingInfo(MiniCPMVProcessingInfo): |
| 476 | # transformers v5.7+ chat_template emits these as image/video placeholders. |
| 477 | image_pattern = "<|image_pad|>" |
| 478 | video_pattern = "<|video_pad|>" |
| 479 | |
| 480 | def get_hf_config(self): |
| 481 | return self.ctx.get_hf_config() |
| 482 | |
| 483 | def get_hf_processor(self, **kwargs: object): |
| 484 | # MiniCPM-V 4.6 keeps the native transformers MiniCPMV4_6Processor: |
| 485 | # this model has its own image/video handling and prompt-update logic |
| 486 | # below, so it does not need (and is incompatible with) the vendored |
| 487 | # MiniCPMVProcessor used by 2.x/4.0/4.5, whose __init__ assumes a |
| 488 | # legacy `image_processor.version` attribute that 4.6 no longer has. |
| 489 | hf_processor = self.ctx.get_hf_processor(**kwargs) |
| 490 | |
| 491 | # NumPy arrays are considered as Iterable but not Sequence in |
| 492 | # https://github.com/huggingface/transformers/blob/main/src/transformers/image_transforms.py#L428 |
| 493 | image_processor = getattr(hf_processor, "image_processor", None) |
| 494 | if image_processor is not None: |
| 495 | # transformers v5+ renamed `mean`/`std` -> `image_mean`/`image_std` |
| 496 | for attr in ("mean", "std", "image_mean", "image_std"): |
| 497 | val = getattr(image_processor, attr, None) |
| 498 | if isinstance(val, np.ndarray): |
| 499 | setattr(image_processor, attr, val.tolist()) |
| 500 | |
| 501 | return hf_processor |
| 502 | |
| 503 | def _get_expected_hidden_size(self) -> int: |
| 504 | config = self.get_hf_config() |
| 505 | if hasattr(config, "text_config") and config.text_config is not None: |
| 506 | return config.text_config.hidden_size |
| 507 | return config.hidden_size |
| 508 | |
| 509 | def get_model_version(self): |
| 510 | return (4, 6) |
| 511 | |
| 512 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 513 | return {"image": None, "video": None} |
| 514 | |
| 515 | def get_image_max_slice_num(self) -> int: |
| 516 | config = self.get_hf_config() |
| 517 | if hasattr(config, "slice_config") and config.slice_config is not None: |
| 518 | return getattr(config.slice_config, "max_slice_nums", 9) |
| 519 | return getattr(config, "max_slice_nums", 9) |
| 520 | |
| 521 | def get_video_max_slice_num(self) -> int: |
| 522 | # Override the base class default of 1: transformers v5.7+ |
| 523 | # `MiniCPMV4_6VideoProcessor` keeps the same max_slice_nums (default 9) |
| 524 | # as the image processor so that high-res frames get sliced. |
| 525 | try: |
| 526 | hf_processor = self.get_hf_processor() |
| 527 | video_processor = getattr(hf_processor, "video_processor", None) |
| 528 | if video_processor is not None: |
| 529 | return int(getattr(video_processor, "max_slice_nums", 9)) |
| 530 | except Exception: |
| 531 | pass |
| 532 | return self.get_image_max_slice_num() |
| 533 | |
| 534 | def _get_downsample_mode( |
| 535 | self, |
| 536 | downsample_mode: str | None = None, |
| 537 | ) -> str: |
| 538 | if downsample_mode is not None: |
| 539 | return downsample_mode |
| 540 | image_processor = self.get_image_processor() |
| 541 | return getattr(image_processor, "downsample_mode", "16x") |
| 542 | |
| 543 | def _compute_visual_tokens( |
| 544 | self, |
| 545 | image_size: ImageSize, |
| 546 | max_slice_nums: int | None = None, |
| 547 | downsample_mode: str | None = None, |
| 548 | ) -> tuple[list[int], int, int]: |
| 549 | """Compute grid, source_image_visual_tokens and patch_visual_tokens. |
| 550 | |
| 551 | Args: |
| 552 | image_size: Size of the source image. |
| 553 | max_slice_nums: Maximum number of slices, or None for the default. |
| 554 | downsample_mode: ``"16x"`` (default, full merge) or ``"4x"`` |
| 555 | (skip vit_merger, 4x more visual tokens). |
| 556 | |
| 557 | Returns: |
| 558 | (grids, source_image_visual_tokens, patch_visual_tokens) |
| 559 | grids is [0, 0] when no slicing occurs. |
| 560 | |
| 561 | """ |
| 562 | image_processor = self.get_image_processor() |
| 563 | if max_slice_nums is None: |
| 564 | max_slice_nums = image_processor.max_slice_nums |
| 565 | |
| 566 | patch_size = image_processor.patch_size |
| 567 | scale_res = image_processor.scale_resolution |
| 568 | downsample_mode = self._get_downsample_mode(downsample_mode) |
| 569 | token_divisor = 4 if downsample_mode == "4x" else 16 |
| 570 | |
| 571 | # vLLM ImageSize is (width, height); transformers expects (height, width) |
| 572 | hf_image_size = (image_size.height, image_size.width) |
| 573 | |
| 574 | # transformers v5.7+ requires `scale_resolution` arg |
| 575 | try: |
| 576 | grids = image_processor.get_sliced_grid( |
| 577 | hf_image_size, |
| 578 | max_slice_nums, |
| 579 | scale_res, |
| 580 | ) |
| 581 | except TypeError: |
| 582 | grids = image_processor.get_sliced_grid( |
| 583 | hf_image_size, |
| 584 | max_slice_nums, |
| 585 | ) |
| 586 | |
| 587 | if grids is None: |
| 588 | best_size = image_processor.find_best_resize( |
| 589 | hf_image_size, |
| 590 | scale_res, |
| 591 | patch_size, |
| 592 | allow_upscale=True, |
| 593 | ) |
| 594 | source_tokens = ( |
| 595 | best_size[0] * best_size[1] // (patch_size * patch_size * token_divisor) |
| 596 | ) |
| 597 | return [0, 0], source_tokens, 0 |
| 598 | |
| 599 | best_resize = image_processor.find_best_resize( |
| 600 | hf_image_size, |
| 601 | scale_res, |
| 602 | patch_size, |
| 603 | ) |
| 604 | source_tokens = ( |
| 605 | best_resize[0] * best_resize[1] // (patch_size * patch_size * token_divisor) |
| 606 | ) |
| 607 | refine_size = image_processor.get_refine_size( |
| 608 | hf_image_size, |
| 609 | grids, |
| 610 | scale_res, |
| 611 | patch_size, |
| 612 | allow_upscale=True, |
| 613 | ) |
| 614 | patch_w = refine_size[0] // grids[0] |
| 615 | patch_h = refine_size[1] // grids[1] |
| 616 | patch_tokens = patch_w * patch_h // (patch_size * patch_size * token_divisor) |
| 617 | return grids, source_tokens, patch_tokens |
| 618 | |
| 619 | def get_slice_image_placeholder( |
| 620 | self, |
| 621 | image_size, |
| 622 | image_idx: int = 0, |
| 623 | max_slice_nums: int | None = None, |
| 624 | use_image_id: bool = True, |
| 625 | downsample_mode: str | None = None, |
| 626 | ) -> str: |
| 627 | grids, source_tokens, patch_tokens = self._compute_visual_tokens( |
| 628 | image_size, |
| 629 | max_slice_nums, |
| 630 | downsample_mode=downsample_mode, |
| 631 | ) |
| 632 | image_processor = self.get_image_processor() |
| 633 | # transformers v5.7+ removed `get_slice_image_placeholder` from the |
| 634 | # image_processor and moved the logic into MiniCPMV4_6Processor. |
| 635 | # Replicate it here using tokenizer special tokens. |
| 636 | if hasattr(image_processor, "get_slice_image_placeholder"): |
| 637 | return image_processor.get_slice_image_placeholder( |
| 638 | grids, |
| 639 | image_idx=image_idx, |
| 640 | max_slice_nums=max_slice_nums, |
| 641 | use_image_id=use_image_id, |
| 642 | source_image_visual_tokens=source_tokens, |
| 643 | patch_visual_tokens=patch_tokens, |
| 644 | ) |
| 645 | tokenizer = self.get_tokenizer() |
| 646 | image_token = getattr(tokenizer, "image_token", "<|image_pad|>") |
| 647 | image_start = getattr(tokenizer, "image_start_token", "<image>") |
| 648 | image_end = getattr(tokenizer, "image_end_token", "</image>") |
| 649 | slice_start = getattr(tokenizer, "slice_start_token", "<slice>") |
| 650 | slice_end = getattr(tokenizer, "slice_end_token", "</slice>") |
| 651 | id_start = getattr(tokenizer, "image_id_start_token", "<image_id>") |
| 652 | id_end = getattr(tokenizer, "image_id_end_token", "</image_id>") |
| 653 | |
| 654 | placeholder = image_start + image_token * source_tokens + image_end |
| 655 | if use_image_id: |
| 656 | placeholder = f"{id_start}{image_idx}{id_end}" + placeholder |
| 657 | |
| 658 | num_rows, num_cols = grids[0], grids[1] |
| 659 | if num_cols > 0 and num_rows > 0 and patch_tokens > 0: |
| 660 | slice_ph = slice_start + image_token * patch_tokens + slice_end |
| 661 | slices = [slice_ph * num_cols for _ in range(num_rows)] |
| 662 | placeholder += "\n".join(slices) |
| 663 | return placeholder |
| 664 | |
| 665 | def get_num_image_tokens( |
| 666 | self, |
| 667 | image_size, |
| 668 | max_slice_nums: int | None = None, |
| 669 | downsample_mode: str | None = None, |
| 670 | ) -> int: |
| 671 | grids, source_tokens, patch_tokens = self._compute_visual_tokens( |
| 672 | image_size, |
| 673 | max_slice_nums, |
| 674 | downsample_mode=downsample_mode, |
| 675 | ) |
| 676 | return source_tokens + grids[0] * grids[1] * patch_tokens |
| 677 | |
| 678 | |
| 679 | class MiniCPMV4_6ViTWindowAttentionSelfAttn(nn.Module): |
| 680 | def __init__( |
| 681 | self, |
| 682 | config, |
| 683 | quant_config: QuantizationConfig | None = None, |
| 684 | prefix: str = "", |
| 685 | ): |
| 686 | super().__init__() |
| 687 | use_data_parallel = is_vit_use_data_parallel() |
| 688 | self.embed_dim = config.hidden_size |
| 689 | self.num_heads = config.num_attention_heads |
| 690 | self.head_dim = self.embed_dim // self.num_heads |
| 691 | self.scale = self.head_dim**-0.5 |
| 692 | |
| 693 | tp_size = 1 if use_data_parallel else get_tensor_model_parallel_world_size() |
| 694 | assert self.num_heads % tp_size == 0 |
| 695 | self.num_heads_per_partition = self.num_heads // tp_size |
| 696 | |
| 697 | self.qkv_proj = QKVParallelLinear( |
| 698 | self.embed_dim, |
| 699 | self.head_dim, |
| 700 | self.num_heads, |
| 701 | quant_config=quant_config, |
| 702 | prefix=f"{prefix}.qkv_proj", |
| 703 | disable_tp=use_data_parallel, |
| 704 | ) |
| 705 | self.out_proj = RowParallelLinear( |
| 706 | self.embed_dim, |
| 707 | self.embed_dim, |
| 708 | bias=True, |
| 709 | quant_config=quant_config, |
| 710 | prefix=f"{prefix}.out_proj", |
| 711 | disable_tp=use_data_parallel, |
| 712 | ) |
| 713 | self.attn = MMEncoderAttention( |
| 714 | self.num_heads_per_partition, |
| 715 | self.head_dim, |
| 716 | self.scale, |
| 717 | prefix=f"{prefix}.attn", |
| 718 | ) |
| 719 | |
| 720 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 721 | qkv, _ = self.qkv_proj(hidden_states) |
| 722 | q, k, v = qkv.chunk(3, dim=-1) |
| 723 | attn_out = self.attn(q, k, v) |
| 724 | out, _ = self.out_proj(attn_out) |
| 725 | return out |
| 726 | |
| 727 | |
| 728 | class MiniCPMV4_6ViTWindowAttentionMerger(nn.Module): |
| 729 | def __init__( |
| 730 | self, |
| 731 | config, |
| 732 | quant_config: QuantizationConfig | None = None, |
| 733 | prefix: str = "", |
| 734 | ): |
| 735 | super().__init__() |
| 736 | self.window_kernel_size = (2, 2) |
| 737 | self.embed_dim = config.hidden_size |
| 738 | |
| 739 | self.self_attn = MiniCPMV4_6ViTWindowAttentionSelfAttn( |
| 740 | config, |
| 741 | quant_config=quant_config, |
| 742 | prefix=f"{prefix}.self_attn", |
| 743 | ) |
| 744 | self.layer_norm1 = nn.LayerNorm( |
| 745 | self.embed_dim, |
| 746 | eps=config.layer_norm_eps, |
| 747 | ) |
| 748 | |
| 749 | hidden_4x = self.embed_dim * 4 |
| 750 | inter_4x = config.intermediate_size * 4 |
| 751 | |
| 752 | self.pre_norm = nn.LayerNorm(hidden_4x, eps=config.layer_norm_eps) |
| 753 | self.linear_1 = nn.Linear(hidden_4x, inter_4x, bias=True) |
| 754 | self.act = get_act_fn("gelu_pytorch_tanh") |
| 755 | self.linear_2 = nn.Linear(inter_4x, self.embed_dim, bias=True) |
| 756 | |
| 757 | def _apply_window_attention( |
| 758 | self, |
| 759 | valid_states: torch.Tensor, |
| 760 | H: int, |
| 761 | W: int, |
| 762 | ) -> torch.Tensor: |
| 763 | D = valid_states.shape[-1] |
| 764 | wh, ww = self.window_kernel_size |
| 765 | nh, nw = H // wh, W // ww |
| 766 | num_windows = nh * nw |
| 767 | |
| 768 | x = valid_states.view(H, W, D) |
| 769 | x = x.view(nh, wh, nw, ww, D).permute(0, 2, 1, 3, 4).contiguous() |
| 770 | x = x.view(num_windows, wh * ww, D) |
| 771 | |
| 772 | x = self.self_attn(x) |
| 773 | |
| 774 | x = x.view(nh, nw, wh, ww, D).permute(0, 2, 1, 3, 4).contiguous() |
| 775 | return x.view(H * W, D) |
| 776 | |
| 777 | def _apply_mlp_downsample( |
| 778 | self, |
| 779 | valid_states: torch.Tensor, |
| 780 | H: int, |
| 781 | W: int, |
| 782 | ) -> torch.Tensor: |
| 783 | D = valid_states.shape[-1] |
| 784 | wh, ww = self.window_kernel_size |
| 785 | nh, nw = H // wh, W // ww |
| 786 | |
| 787 | x = valid_states.view(H, W, D) |
| 788 | x = x.view(nh, wh, nw, ww, D).permute(0, 2, 1, 3, 4).contiguous() |
| 789 | |
| 790 | residual = x.reshape(nh * nw, wh * ww, D).mean(dim=1) |
| 791 | x = x.reshape(nh * nw, wh * ww * D) |
| 792 | |
| 793 | x = self.pre_norm(x) |
| 794 | x = self.linear_1(x) |
| 795 | x = self.act(x) |
| 796 | x = self.linear_2(x) |
| 797 | return x + residual |
| 798 | |
| 799 | def forward( |
| 800 | self, |
| 801 | hidden_states: torch.Tensor, |
| 802 | tgt_sizes: torch.Tensor, |
| 803 | attention_mask: torch.Tensor | None, |
| 804 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor | None]: |
| 805 | B, _L, D = hidden_states.shape |
| 806 | device = hidden_states.device |
| 807 | dtype = hidden_states.dtype |
| 808 | |
| 809 | all_merged = [] |
| 810 | new_tgt_sizes = torch.zeros_like(tgt_sizes) |
| 811 | |
| 812 | for b in range(B): |
| 813 | H, W = tgt_sizes[b].tolist() |
| 814 | hs = hidden_states[b, : H * W, :] |
| 815 | |
| 816 | residual = hs |
| 817 | hs = self.layer_norm1(hs) |
| 818 | hs = residual + self._apply_window_attention(hs, H, W) |
| 819 | |
| 820 | wh, ww = self.window_kernel_size |
| 821 | new_H, new_W = H // wh, W // ww |
| 822 | all_merged.append(self._apply_mlp_downsample(hs, H, W)) |
| 823 | new_tgt_sizes[b] = torch.tensor( |
| 824 | [new_H, new_W], |
| 825 | device=device, |
| 826 | dtype=tgt_sizes.dtype, |
| 827 | ) |
| 828 | |
| 829 | new_num_patches = new_tgt_sizes[:, 0] * new_tgt_sizes[:, 1] |
| 830 | new_max_patches = int(new_num_patches.max().item()) |
| 831 | new_hidden = torch.zeros( |
| 832 | B, |
| 833 | new_max_patches, |
| 834 | D, |
| 835 | device=device, |
| 836 | dtype=dtype, |
| 837 | ) |
| 838 | for b, merged in enumerate(all_merged): |
| 839 | new_hidden[b, : merged.shape[0], :] = merged |
| 840 | |
| 841 | # Build new attention mask after spatial downsampling |
| 842 | new_attention_mask: torch.Tensor | None = None |
| 843 | if attention_mask is not None: |
| 844 | mask = torch.zeros( |
| 845 | B, |
| 846 | new_max_patches, |
| 847 | dtype=torch.bool, |
| 848 | device=device, |
| 849 | ) |
| 850 | for b in range(B): |
| 851 | mask[b, : int(new_num_patches[b].item())] = True |
| 852 | min_val = torch.finfo(dtype).min |
| 853 | new_attention_mask = (~mask).to(dtype=dtype) * min_val |
| 854 | new_attention_mask = new_attention_mask[:, None, None, :] |
| 855 | |
| 856 | return new_hidden, new_tgt_sizes, new_attention_mask |
| 857 | |
| 858 | |
| 859 | class MiniCPMV4_6DownsampleMLP(nn.Module): |
| 860 | """Match HF (transformers v5.7+) parameter naming: pre_norm/linear_1/ |
| 861 | act/linear_2 (instead of pre_norm + Sequential(mlp.0/mlp.2)).""" |
| 862 | |
| 863 | def __init__( |
| 864 | self, |
| 865 | hidden_size: int, |
| 866 | llm_embed_dim: int, |
| 867 | merge_kernel_size: tuple[int, int] = (2, 2), |
| 868 | ): |
| 869 | super().__init__() |
| 870 | self.merge_kernel_size = merge_kernel_size |
| 871 | self.hidden_size = hidden_size * merge_kernel_size[0] * merge_kernel_size[1] |
| 872 | self.pre_norm = nn.LayerNorm(self.hidden_size, eps=1e-6) |
| 873 | self.linear_1 = nn.Linear(self.hidden_size, self.hidden_size, bias=True) |
| 874 | self.act = get_act_fn("gelu") |
| 875 | self.linear_2 = nn.Linear(self.hidden_size, llm_embed_dim, bias=True) |
| 876 | |
| 877 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 878 | x = self.pre_norm(x) |
| 879 | x = self.linear_1(x) |
| 880 | x = self.act(x) |
| 881 | x = self.linear_2(x) |
| 882 | return x |
| 883 | |
| 884 | |
| 885 | class MiniCPMV4_6Merger(nn.Module): |
| 886 | def __init__( |
| 887 | self, |
| 888 | hidden_size: int, |
| 889 | llm_embed_dim: int, |
| 890 | merge_kernel_size: tuple[int, int] = (2, 2), |
| 891 | times: int = 1, |
| 892 | ): |
| 893 | super().__init__() |
| 894 | self.merge_kernel_size = merge_kernel_size |
| 895 | self.times = times |
| 896 | self.mlp = nn.ModuleList( |
| 897 | [ |
| 898 | MiniCPMV4_6DownsampleMLP( |
| 899 | hidden_size, |
| 900 | llm_embed_dim if i == times - 1 else hidden_size, |
| 901 | merge_kernel_size, |
| 902 | ) |
| 903 | for i in range(times) |
| 904 | ] |
| 905 | ) |
| 906 | |
| 907 | def forward( |
| 908 | self, |
| 909 | hidden_states: torch.Tensor, |
| 910 | tgt_sizes: torch.Tensor, |
| 911 | ) -> list[torch.Tensor]: |
| 912 | """Args: |
| 913 | hidden_states: (B, max_patches, D) padded batch. |
| 914 | tgt_sizes: (B, 2) actual (H, W) per sample. |
| 915 | |
| 916 | """ |
| 917 | m1, m2 = self.merge_kernel_size |
| 918 | results = [] |
| 919 | |
| 920 | for b in range(len(tgt_sizes)): |
| 921 | h, w = tgt_sizes[b].tolist() |
| 922 | n_patches = h * w |
| 923 | hs = hidden_states[b, :n_patches, :] |
| 924 | |
| 925 | hs = hs.reshape(h // m1, m1, w // m2, m2, -1) |
| 926 | hs = hs.permute(0, 2, 1, 3, 4).reshape( |
| 927 | (h // m1) * (w // m2), |
| 928 | m1 * m2 * hs.shape[-1], |
| 929 | ) |
| 930 | hs = self.mlp[0](hs) |
| 931 | |
| 932 | if self.times > 1: |
| 933 | cur_h, cur_w = h // m1, w // m2 |
| 934 | for t in range(1, self.times): |
| 935 | cur_h, cur_w = cur_h // m1, cur_w // m2 |
| 936 | hs = hs.reshape(cur_h, m1, cur_w, m2, -1) |
| 937 | hs = hs.permute(0, 2, 1, 3, 4).reshape( |
| 938 | cur_h * cur_w, |
| 939 | m1 * m2 * hs.shape[-1], |
| 940 | ) |
| 941 | hs = self.mlp[t](hs) |
| 942 | |
| 943 | results.append(hs) |
| 944 | |
| 945 | return results |
| 946 | |
| 947 | |
| 948 | @MULTIMODAL_REGISTRY.register_processor( |
| 949 | MiniCPMV4_6MultiModalProcessor, |
| 950 | info=MiniCPMV4_6ProcessingInfo, |
| 951 | dummy_inputs=MiniCPMVDummyInputsBuilder, |
| 952 | ) |
| 953 | class MiniCPMV4_6ForConditionalGeneration( |
| 954 | nn.Module, |
| 955 | SupportsMultiModal, |
| 956 | SupportsLoRA, |
| 957 | SupportsPP, |
| 958 | HasInnerState, |
| 959 | IsHybrid, |
| 960 | SupportsMRoPE, |
| 961 | ): |
| 962 | supports_encoder_tp_data = True |
| 963 | |
| 964 | hf_to_vllm_mapper = WeightsMapper( |
| 965 | orig_to_new_stacked={ |
| 966 | ".q_proj": (".qkv_proj", "q"), |
| 967 | ".k_proj": (".qkv_proj", "k"), |
| 968 | ".v_proj": (".qkv_proj", "v"), |
| 969 | }, |
| 970 | orig_to_new_prefix={ |
| 971 | # transformers v5.7+ uses `vision_tower` and nests `vit_merger` |
| 972 | # inside it. Order matters: more specific prefix must come first. |
| 973 | "model.vision_tower.vit_merger.": "vit_merger.", |
| 974 | "model.vision_tower.": "vpm.", |
| 975 | "model.vpm.": "vpm.", |
| 976 | "model.vit_merger.": "vit_merger.", |
| 977 | "model.merger.": "merger.", |
| 978 | "model.language_model.": "language_model.model.", |
| 979 | "lm_head.": "language_model.lm_head.", |
| 980 | "mtp.": None, |
| 981 | }, |
| 982 | ) |
| 983 | |
| 984 | packed_modules_mapping = { |
| 985 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 986 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 987 | "in_proj_qkvz": ["in_proj_qkv", "in_proj_z"], |
| 988 | "in_proj_ba": ["in_proj_b", "in_proj_a"], |
| 989 | } |
| 990 | |
| 991 | @classmethod |
| 992 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 993 | # transformers v5.7+ chat_template uses these tokens. |
| 994 | if modality.startswith("image"): |
| 995 | return "<|image_pad|>" |
| 996 | if modality.startswith("video"): |
| 997 | return "<|video_pad|>" |
| 998 | raise ValueError("Only image or video modality is supported") |
| 999 | |
| 1000 | def get_mrope_input_positions( |
| 1001 | self, |
| 1002 | input_tokens: list[int], |
| 1003 | mm_features: list["MultiModalFeatureSpec"], |
| 1004 | ) -> tuple[torch.Tensor, int]: |
| 1005 | """MiniCPM-V uses embedding injection for vision, not spatial M-RoPE. |
| 1006 | |
| 1007 | All tokens (text and vision placeholders) get identical sequential |
| 1008 | positions duplicated across the 3 M-RoPE channels expected by the |
| 1009 | Qwen3.5 backbone. |
| 1010 | """ |
| 1011 | seq_len = len(input_tokens) |
| 1012 | positions = torch.arange(seq_len).unsqueeze(0).expand(3, -1) |
| 1013 | return positions, 0 |
| 1014 | |
| 1015 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1016 | super().__init__() |
| 1017 | config: MiniCPMV4_6Config = vllm_config.model_config.hf_config |
| 1018 | quant_config = vllm_config.quant_config |
| 1019 | multimodal_config = vllm_config.model_config.multimodal_config |
| 1020 | assert multimodal_config is not None |
| 1021 | |
| 1022 | self.config = config |
| 1023 | self.multimodal_config = multimodal_config |
| 1024 | self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data" |
| 1025 | |
| 1026 | # --- Vision tower --- |
| 1027 | with self._mark_tower_model(vllm_config, {"image"}): |
| 1028 | self.vpm = Idefics2VisionTransformer( |
| 1029 | config.vision_config, |
| 1030 | quant_config=quant_config, |
| 1031 | apply_encoder_attention_mask=True, |
| 1032 | prefix=maybe_prefix(prefix, "vpm"), |
| 1033 | ) |
| 1034 | if config.drop_vision_last_layer: |
| 1035 | self.vpm.encoder.layers = self.vpm.encoder.layers[:-1] |
| 1036 | |
| 1037 | self.vit_merger = MiniCPMV4_6ViTWindowAttentionMerger( |
| 1038 | config.vision_config, |
| 1039 | quant_config=quant_config, |
| 1040 | prefix=maybe_prefix(prefix, "vit_merger"), |
| 1041 | ) |
| 1042 | self.merger = MiniCPMV4_6Merger( |
| 1043 | hidden_size=config.vision_config.hidden_size, |
| 1044 | llm_embed_dim=config.text_config.hidden_size, |
| 1045 | ) |
| 1046 | |
| 1047 | # --- Language model --- |
| 1048 | # Temporarily swap top-level model_type so that Qwen3_5ForCausalLM |
| 1049 | # picks up the expected text config when introspecting the hf config. |
| 1050 | with self._mark_language_model(vllm_config): |
| 1051 | saved_model_type = config.model_type |
| 1052 | config.model_type = "qwen3_5_text" |
| 1053 | try: |
| 1054 | self.language_model = Qwen3_5ForCausalLM( |
| 1055 | vllm_config=vllm_config, |
| 1056 | prefix=maybe_prefix(prefix, "language_model"), |
| 1057 | ) |
| 1058 | finally: |
| 1059 | config.model_type = saved_model_type |
| 1060 | |
| 1061 | self.make_empty_intermediate_tensors = ( |
| 1062 | self.language_model.make_empty_intermediate_tensors |
| 1063 | ) |
| 1064 | |
| 1065 | # ----- Multimodal parsing ----- |
| 1066 | |
| 1067 | def _parse_and_validate_vision_input( |
| 1068 | self, |
| 1069 | **kwargs: object, |
| 1070 | ) -> MiniCPMVImagePixelInputs | MiniCPMVImageEmbeddingInputs | None: |
| 1071 | pixel_values = kwargs.pop("pixel_values", None) |
| 1072 | image_embeds = kwargs.pop("image_embeds", None) |
| 1073 | |
| 1074 | if pixel_values is None and image_embeds is None: |
| 1075 | return None |
| 1076 | |
| 1077 | if image_embeds is not None: |
| 1078 | return MiniCPMVImageEmbeddingInputs( |
| 1079 | type="image_embeds", |
| 1080 | image_embeds=image_embeds, |
| 1081 | ) |
| 1082 | |
| 1083 | assert isinstance(pixel_values, torch.Tensor | list) |
| 1084 | tgt_sizes = kwargs.pop("tgt_sizes") |
| 1085 | num_slices_flat = torch.tensor([len(ps) for ps in pixel_values]) |
| 1086 | pixel_values_flat = flatten_bn(pixel_values) |
| 1087 | tgt_sizes_flat = flatten_bn(tgt_sizes, concat=True) |
| 1088 | |
| 1089 | return MiniCPMVImagePixelInputs( |
| 1090 | type="pixel_values", |
| 1091 | pixel_values=pixel_values_flat, |
| 1092 | tgt_sizes=tgt_sizes_flat, |
| 1093 | num_slices=num_slices_flat, |
| 1094 | ) |
| 1095 | |
| 1096 | # ----- Vision forward ----- |
| 1097 | |
| 1098 | def get_vision_hidden_states( |
| 1099 | self, |
| 1100 | data: MiniCPMVImagePixelInputs, |
| 1101 | downsample_mode: str | None = None, |
| 1102 | ) -> list[torch.Tensor]: |
| 1103 | pixel_values = data["pixel_values"] |
| 1104 | tgt_sizes = data["tgt_sizes"] |
| 1105 | |
| 1106 | B = len(pixel_values) |
| 1107 | P = pixel_values[0].shape[-2] |
| 1108 | L = max(item.shape[-1] for item in pixel_values) |
| 1109 | device = pixel_values[0].device |
| 1110 | target_dtype = self.vpm.embeddings.patch_embedding.weight.dtype |
| 1111 | |
| 1112 | all_pixel_values = torch.zeros( |
| 1113 | B, |
| 1114 | 3, |
| 1115 | P, |
| 1116 | L, |
| 1117 | dtype=target_dtype, |
| 1118 | device=device, |
| 1119 | ) |
| 1120 | for i, pv in enumerate(pixel_values): |
| 1121 | all_pixel_values[i, ..., : pv.shape[-1]] = pv.to(target_dtype) |
| 1122 | |
| 1123 | num_patches = tgt_sizes.prod(-1) |
| 1124 | max_patches = int(num_patches.max().item()) |
| 1125 | patch_attn_mask = torch.zeros( |
| 1126 | B, |
| 1127 | max_patches, |
| 1128 | dtype=torch.bool, |
| 1129 | device=device, |
| 1130 | ) |
| 1131 | for i in range(B): |
| 1132 | patch_attn_mask[i, : num_patches[i]] = True |
| 1133 | |
| 1134 | hidden_states = self.vpm.embeddings( |
| 1135 | all_pixel_values, |
| 1136 | patch_attention_mask=patch_attn_mask.unsqueeze(1), |
| 1137 | tgt_sizes=tgt_sizes, |
| 1138 | ) |
| 1139 | |
| 1140 | if torch.any(~patch_attn_mask): |
| 1141 | mask_dtype = hidden_states.dtype |
| 1142 | min_val = torch.finfo(mask_dtype).min |
| 1143 | attention_mask = (~patch_attn_mask).to(dtype=mask_dtype) * min_val |
| 1144 | attention_mask = attention_mask[:, None, None, :] |
| 1145 | else: |
| 1146 | attention_mask = None |
| 1147 | |
| 1148 | # Encoder layers with mid-encoder merger injection |
| 1149 | insert_layer_id = getattr(self.config, "insert_layer_id", -1) |
| 1150 | if downsample_mode is None: |
| 1151 | downsample_mode = getattr(self.config, "downsample_mode", "16x") |
| 1152 | use_vit_merger = downsample_mode != "4x" and insert_layer_id >= 0 |
| 1153 | |
| 1154 | for layer in self.vpm.encoder.layers[: insert_layer_id + 1]: |
| 1155 | hidden_states = layer(hidden_states, attention_mask=attention_mask) |
| 1156 | |
| 1157 | if use_vit_merger: |
| 1158 | hidden_states, tgt_sizes, attention_mask = self.vit_merger( |
| 1159 | hidden_states, |
| 1160 | tgt_sizes, |
| 1161 | attention_mask, |
| 1162 | ) |
| 1163 | |
| 1164 | for layer in self.vpm.encoder.layers[insert_layer_id + 1 :]: |
| 1165 | hidden_states = layer(hidden_states, attention_mask=attention_mask) |
| 1166 | |
| 1167 | # 4. Post layernorm |
| 1168 | hidden_states = self.vpm.post_layernorm(hidden_states) |
| 1169 | |
| 1170 | # 5. MLP merger → list of per-slice tensors |
| 1171 | return self.merger(hidden_states, tgt_sizes) |
| 1172 | |
| 1173 | def _process_vision_input(self, image_input, use_vit_merger=None): |
| 1174 | if image_input["type"] == "image_embeds": |
| 1175 | return image_input["image_embeds"] |
| 1176 | |
| 1177 | downsample_mode = None |
| 1178 | if use_vit_merger is not None: |
| 1179 | downsample_mode = "16x" if use_vit_merger else "4x" |
| 1180 | image_features = self.get_vision_hidden_states( |
| 1181 | image_input, |
| 1182 | downsample_mode=downsample_mode, |
| 1183 | ) |
| 1184 | num_slices = image_input["num_slices"] |
| 1185 | results = [] |
| 1186 | idx = 0 |
| 1187 | for n in num_slices.tolist(): |
| 1188 | group = image_features[idx : idx + n] |
| 1189 | results.append(torch.cat(group, dim=0)) |
| 1190 | idx += n |
| 1191 | return results |
| 1192 | |
| 1193 | # ----- Multimodal embedding interface ----- |
| 1194 | |
| 1195 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 1196 | use_vit_merger_tensors = kwargs.pop("use_vit_merger", None) |
| 1197 | use_vit_merger = None |
| 1198 | if use_vit_merger_tensors is not None: |
| 1199 | if isinstance(use_vit_merger_tensors, torch.Tensor): |
| 1200 | use_vit_merger = bool(use_vit_merger_tensors.any().item()) |
| 1201 | elif isinstance(use_vit_merger_tensors, list | tuple): |
| 1202 | use_vit_merger = any( |
| 1203 | bool(t.any().item()) if isinstance(t, torch.Tensor) else bool(t) |
| 1204 | for t in use_vit_merger_tensors |
| 1205 | ) |
| 1206 | |
| 1207 | image_kwargs = { |
| 1208 | k: v |
| 1209 | for k, v in kwargs.items() |
| 1210 | if k in ("pixel_values", "image_embeds", "tgt_sizes") |
| 1211 | } |
| 1212 | video_kwargs = _image_kwargs_from_video(kwargs) |
| 1213 | |
| 1214 | multimodal_embeddings: tuple[torch.Tensor, ...] = () |
| 1215 | |
| 1216 | if ( |
| 1217 | image_kwargs.get("pixel_values") is not None |
| 1218 | or image_kwargs.get("image_embeds") is not None |
| 1219 | ): |
| 1220 | image_input = self._parse_and_validate_vision_input(**image_kwargs) |
| 1221 | if image_input is not None: |
| 1222 | multimodal_embeddings += tuple( |
| 1223 | self._process_vision_input( |
| 1224 | image_input, |
| 1225 | use_vit_merger=use_vit_merger, |
| 1226 | ) |
| 1227 | ) |
| 1228 | |
| 1229 | if ( |
| 1230 | video_kwargs.get("pixel_values") is not None |
| 1231 | or video_kwargs.get("image_embeds") is not None |
| 1232 | ): |
| 1233 | video_input = self._parse_and_validate_vision_input(**video_kwargs) |
| 1234 | if video_input is not None: |
| 1235 | multimodal_embeddings += tuple( |
| 1236 | self._process_vision_input( |
| 1237 | video_input, |
| 1238 | use_vit_merger=use_vit_merger, |
| 1239 | ) |
| 1240 | ) |
| 1241 | |
| 1242 | if not multimodal_embeddings: |
| 1243 | return [] |
| 1244 | return multimodal_embeddings |
| 1245 | |
| 1246 | def embed_input_ids( |
| 1247 | self, |
| 1248 | input_ids: torch.Tensor, |
| 1249 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 1250 | *, |
| 1251 | is_multimodal: torch.Tensor | None = None, |
| 1252 | ) -> torch.Tensor: |
| 1253 | inputs_embeds = self._embed_text_input_ids( |
| 1254 | input_ids, |
| 1255 | self.language_model.embed_input_ids, |
| 1256 | is_multimodal=is_multimodal, |
| 1257 | ) |
| 1258 | if multimodal_embeddings is None or len(multimodal_embeddings) == 0: |
| 1259 | return inputs_embeds |
| 1260 | |
| 1261 | is_multimodal = _require_is_multimodal(is_multimodal) |
| 1262 | return _merge_multimodal_embeddings( |
| 1263 | inputs_embeds=inputs_embeds, |
| 1264 | multimodal_embeddings=multimodal_embeddings, |
| 1265 | is_multimodal=is_multimodal, |
| 1266 | ) |
| 1267 | |
| 1268 | # ----- Forward / Logits ----- |
| 1269 | |
| 1270 | def forward( |
| 1271 | self, |
| 1272 | input_ids: torch.Tensor | None, |
| 1273 | positions: torch.Tensor, |
| 1274 | intermediate_tensors: IntermediateTensors | None = None, |
| 1275 | inputs_embeds: torch.Tensor | None = None, |
| 1276 | **kwargs: Any, |
| 1277 | ) -> torch.Tensor: |
| 1278 | if intermediate_tensors is not None: |
| 1279 | inputs_embeds = None |
| 1280 | |
| 1281 | return self.language_model.model( |
| 1282 | input_ids=input_ids, |
| 1283 | positions=positions, |
| 1284 | intermediate_tensors=intermediate_tensors, |
| 1285 | inputs_embeds=inputs_embeds, |
| 1286 | ) |
| 1287 | |
| 1288 | def compute_logits( |
| 1289 | self, |
| 1290 | hidden_states: torch.Tensor, |
| 1291 | ) -> torch.Tensor | None: |
| 1292 | return self.language_model.compute_logits(hidden_states) |
| 1293 | |
| 1294 | # ----- Weight loading ----- |
| 1295 | |
| 1296 | def load_weights( |
| 1297 | self, |
| 1298 | weights: Iterable[tuple[str, torch.Tensor]], |
| 1299 | ) -> set[str]: |
| 1300 | loader = AutoWeightsLoader(self) |
| 1301 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 1302 | |
| 1303 | def get_mm_mapping(self) -> MultiModelKeys: |
| 1304 | return MultiModelKeys.from_string_field( |
| 1305 | language_model="language_model", |
| 1306 | connector=["vit_merger", "merger"], |
| 1307 | tower_model="vpm", |
| 1308 | ) |
| 1309 | |
| 1310 | # ----- Mamba / Hybrid state helpers (same as Qwen3.5 VLM) ----- |
| 1311 | |
| 1312 | @classmethod |
| 1313 | def get_mamba_state_dtype_from_config(cls, vllm_config): |
| 1314 | return MambaStateDtypeCalculator.gated_delta_net_state_dtype( |
| 1315 | vllm_config.model_config.dtype, |
| 1316 | vllm_config.cache_config.mamba_cache_dtype, |
| 1317 | vllm_config.cache_config.mamba_ssm_cache_dtype, |
| 1318 | ) |
| 1319 | |
| 1320 | @classmethod |
| 1321 | def get_mamba_state_shape_from_config(cls, vllm_config): |
| 1322 | parallel_config = vllm_config.parallel_config |
| 1323 | hf_config = vllm_config.model_config.hf_text_config |
| 1324 | tp_size = parallel_config.tensor_parallel_size |
| 1325 | num_spec = ( |
| 1326 | vllm_config.speculative_config.num_speculative_tokens |
| 1327 | if vllm_config.speculative_config |
| 1328 | else 0 |
| 1329 | ) |
| 1330 | return MambaStateShapeCalculator.gated_delta_net_state_shape( |
| 1331 | tp_size, |
| 1332 | hf_config.linear_num_key_heads, |
| 1333 | hf_config.linear_num_value_heads, |
| 1334 | hf_config.linear_key_head_dim, |
| 1335 | hf_config.linear_value_head_dim, |
| 1336 | hf_config.linear_conv_kernel_dim, |
| 1337 | num_spec, |
| 1338 | ) |
| 1339 | |
| 1340 | @classmethod |
| 1341 | def get_mamba_state_copy_func(cls): |
| 1342 | return MambaStateCopyFuncCalculator.gated_delta_net_state_copy_func() |