| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import math |
| 4 | from collections.abc import Iterable |
| 5 | |
| 6 | import torch |
| 7 | import torch.nn as nn |
| 8 | from transformers import AutoModel, PreTrainedConfig |
| 9 | |
| 10 | from vllm.config import VllmConfig |
| 11 | from vllm.model_executor.layers.fusion.mm_input_norm import ( |
| 12 | FusedMMInputNorm, |
| 13 | IdentityInputNorm, |
| 14 | ) |
| 15 | from vllm.model_executor.layers.linear import ReplicatedLinear |
| 16 | from vllm.model_executor.layers.pooler import DispatchPooler |
| 17 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 18 | from vllm.model_executor.layers.quantization.auto_awq import AutoAWQConfig |
| 19 | from vllm.model_executor.models.internvl import ( |
| 20 | BaseInternVLDummyInputsBuilder, |
| 21 | BaseInternVLMultiModalProcessor, |
| 22 | BaseInternVLProcessingInfo, |
| 23 | InternVLImageEmbeddingInputs, |
| 24 | InternVLImageInputs, |
| 25 | InternVLImagePixelInputs, |
| 26 | ) |
| 27 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 28 | from vllm.model_executor.models.siglip import SiglipVisionModel |
| 29 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 30 | from vllm.sequence import IntermediateTensors |
| 31 | from vllm.transformers_utils.processor import cached_image_processor_from_config |
| 32 | from vllm.transformers_utils.processors.nemotron_vl import ( |
| 33 | SIGLIP_MEAN, |
| 34 | SIGLIP_STD, |
| 35 | LlamaNemotronNanoVLImageProcessor, |
| 36 | LlamaNemotronNanoVLProcessor, |
| 37 | LlamaNemotronVLEmbedImageProcessor, |
| 38 | LlamaNemotronVLEmbedProcessor, |
| 39 | ) |
| 40 | from vllm.transformers_utils.repo_utils import get_hf_file_to_dict |
| 41 | |
| 42 | from .interfaces import ( |
| 43 | MultiModalEmbeddings, |
| 44 | SupportsCrossEncoding, |
| 45 | SupportsLoRA, |
| 46 | SupportsMultiModal, |
| 47 | SupportsPP, |
| 48 | ) |
| 49 | from .interfaces_base import VllmModelForPooling |
| 50 | from .utils import ( |
| 51 | AutoWeightsLoader, |
| 52 | WeightsMapper, |
| 53 | init_vllm_registered_model, |
| 54 | maybe_prefix, |
| 55 | ) |
| 56 | |
| 57 | |
| 58 | class NemotronVLProcessingInfo(BaseInternVLProcessingInfo): |
| 59 | """Processing info for Nemotron VL models.""" |
| 60 | |
| 61 | def get_image_processor(self, **kwargs: object): |
| 62 | kwargs = self.ctx.get_merged_mm_kwargs(kwargs) |
| 63 | orig_processor = cached_image_processor_from_config( |
| 64 | self.ctx.model_config, **kwargs |
| 65 | ) |
| 66 | |
| 67 | return LlamaNemotronNanoVLImageProcessor( |
| 68 | image_size=orig_processor.image_size, |
| 69 | min_dynamic_patch=1, |
| 70 | max_dynamic_patch=orig_processor.max_num_tiles, |
| 71 | dynamic_image_size=True, |
| 72 | use_thumbnail=orig_processor.use_thumbnail, |
| 73 | ) |
| 74 | |
| 75 | def get_hf_processor(self, **kwargs: object) -> LlamaNemotronNanoVLProcessor: |
| 76 | config = self.get_hf_config() |
| 77 | vision_config = config.vision_config |
| 78 | |
| 79 | image_processor = self.get_image_processor(**kwargs) |
| 80 | image_size = image_processor.image_size |
| 81 | patch_size = vision_config.patch_size |
| 82 | downsample_ratio = config.downsample_ratio |
| 83 | image_seq_length = int((image_size // patch_size) ** 2 * (downsample_ratio**2)) |
| 84 | |
| 85 | return LlamaNemotronNanoVLProcessor( |
| 86 | tokenizer=self.get_tokenizer(), |
| 87 | image_processor=image_processor, |
| 88 | image_seq_length=image_seq_length, |
| 89 | ) |
| 90 | |
| 91 | |
| 92 | @MULTIMODAL_REGISTRY.register_processor( |
| 93 | BaseInternVLMultiModalProcessor[NemotronVLProcessingInfo], |
| 94 | info=NemotronVLProcessingInfo, |
| 95 | dummy_inputs=BaseInternVLDummyInputsBuilder[NemotronVLProcessingInfo], |
| 96 | ) |
| 97 | class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA): |
| 98 | # Ignore registered buffers, see |
| 99 | # https://huggingface.co/nvidia/C-RADIOv2-H/blob/main/input_conditioner.py#L28 |
| 100 | hf_to_vllm_mapper = WeightsMapper( |
| 101 | orig_to_new_substr={"norm_mean": None, "norm_std": None} |
| 102 | ) |
| 103 | |
| 104 | @classmethod |
| 105 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 106 | if modality.startswith("image"): |
| 107 | return "<image>" |
| 108 | |
| 109 | raise ValueError("Only image modality is supported") |
| 110 | |
| 111 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 112 | super().__init__() |
| 113 | |
| 114 | config = vllm_config.model_config.hf_config |
| 115 | quant_config = vllm_config.quant_config |
| 116 | multimodal_config = vllm_config.model_config.multimodal_config |
| 117 | |
| 118 | self.config = config |
| 119 | self.model_config = vllm_config.model_config |
| 120 | self.multimodal_config = multimodal_config |
| 121 | self._patch_quant_config(config, quant_config) |
| 122 | |
| 123 | image_size = config.force_image_size or config.vision_config.image_size |
| 124 | patch_size = config.vision_config.patch_size |
| 125 | self.patch_size = patch_size |
| 126 | self.num_image_token = int( |
| 127 | (image_size // patch_size) ** 2 * (config.downsample_ratio**2) |
| 128 | ) |
| 129 | self.downsample_ratio = config.downsample_ratio |
| 130 | self.ps_version = config.ps_version |
| 131 | |
| 132 | with self._mark_tower_model(vllm_config, "image"): |
| 133 | self.vision_model = self._init_vision_model( |
| 134 | config, |
| 135 | quant_config=quant_config, |
| 136 | prefix=maybe_prefix(prefix, "vision_model"), |
| 137 | ) |
| 138 | self.mlp1 = self._init_mlp1(config) |
| 139 | |
| 140 | with self._mark_language_model(vllm_config): |
| 141 | self.language_model = init_vllm_registered_model( |
| 142 | vllm_config=vllm_config, |
| 143 | hf_config=config.get_text_config(), |
| 144 | prefix=maybe_prefix(prefix, "language_model"), |
| 145 | ) |
| 146 | |
| 147 | self.img_context_token_id: int | None = None |
| 148 | |
| 149 | self.visual_token_mask = None |
| 150 | self.make_empty_intermediate_tensors = ( |
| 151 | self.language_model.make_empty_intermediate_tensors |
| 152 | ) |
| 153 | |
| 154 | def _patch_quant_config( |
| 155 | self, config: PreTrainedConfig, quant_config: QuantizationConfig | None |
| 156 | ): |
| 157 | # the awq models from OpenGVLab missing `modules_to_not_convert` |
| 158 | # patch the quant_config to add `modules_to_not_convert` back |
| 159 | if isinstance(quant_config, AutoAWQConfig): |
| 160 | text_config = config.get_text_config() |
| 161 | llm_quant_config = getattr(text_config, "quantization_config", None) |
| 162 | if (not quant_config.modules_to_not_convert) and ( |
| 163 | llm_quant_config is not None |
| 164 | ): |
| 165 | quant_config.modules_to_not_convert.append("vision_model") |
| 166 | |
| 167 | def _init_vision_model( |
| 168 | self, |
| 169 | config: PreTrainedConfig, |
| 170 | quant_config: QuantizationConfig | None, |
| 171 | *, |
| 172 | prefix: str, |
| 173 | ): |
| 174 | return AutoModel.from_config( |
| 175 | config.vision_config, |
| 176 | trust_remote_code=self.model_config.trust_remote_code, |
| 177 | ) |
| 178 | |
| 179 | def _init_mlp1( |
| 180 | self, |
| 181 | config: PreTrainedConfig, |
| 182 | vit_hidden_size: int | None = None, |
| 183 | vision_projection_hidden_size: int | None = None, |
| 184 | ) -> nn.Module: |
| 185 | if vit_hidden_size is None: |
| 186 | vit_hidden_size = config.vit_hidden_size |
| 187 | if vision_projection_hidden_size is None: |
| 188 | vision_projection_hidden_size = config.projector_hidden_size |
| 189 | llm_hidden_size = config.get_text_config().hidden_size |
| 190 | |
| 191 | return nn.Sequential( |
| 192 | nn.LayerNorm( |
| 193 | vit_hidden_size * int(1 / self.downsample_ratio) ** 2, bias=True |
| 194 | ), |
| 195 | nn.Linear( |
| 196 | vit_hidden_size * int(1 / self.downsample_ratio) ** 2, |
| 197 | vision_projection_hidden_size, |
| 198 | bias=True, |
| 199 | ), |
| 200 | nn.GELU(), |
| 201 | nn.Linear(vision_projection_hidden_size, llm_hidden_size), |
| 202 | ) |
| 203 | |
| 204 | def pixel_shuffle(self, x, scale_factor=0.5): |
| 205 | n, w, h, c = x.size() |
| 206 | # N, W, H, C --> N, W, H * scale, C // scale |
| 207 | x = x.view(n, w, int(h * scale_factor), int(c / scale_factor)) |
| 208 | # N, W, H * scale, C // scale --> N, H * scale, W, C // scale |
| 209 | x = x.permute(0, 2, 1, 3).contiguous() |
| 210 | x = x.view( |
| 211 | n, |
| 212 | int(h * scale_factor), |
| 213 | int(w * scale_factor), |
| 214 | int(c / (scale_factor * scale_factor)), |
| 215 | ) |
| 216 | if self.ps_version == "v1": |
| 217 | pass |
| 218 | else: |
| 219 | x = x.permute(0, 2, 1, 3).contiguous() |
| 220 | return x |
| 221 | |
| 222 | def _call_vision_model(self, pixel_values: torch.Tensor) -> torch.Tensor: |
| 223 | """Call vision model and return embeddings. |
| 224 | |
| 225 | Override this method in subclasses to handle different vision model |
| 226 | interfaces (e.g., SigLIP vs C-RADIO). |
| 227 | """ |
| 228 | vit_embeds = self.vision_model(x=pixel_values).features |
| 229 | return vit_embeds.to(dtype=torch.bfloat16) |
| 230 | |
| 231 | def extract_feature(self, pixel_values: torch.Tensor) -> torch.Tensor: |
| 232 | # https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1/blob/main/modeling.py#L177 |
| 233 | vit_embeds = self._call_vision_model(pixel_values) |
| 234 | |
| 235 | h = w = int(vit_embeds.shape[1] ** 0.5) |
| 236 | vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1) |
| 237 | vit_embeds = self.pixel_shuffle(vit_embeds, scale_factor=self.downsample_ratio) |
| 238 | vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1]) |
| 239 | vit_embeds = self.mlp1(vit_embeds) |
| 240 | return vit_embeds |
| 241 | |
| 242 | def _parse_and_validate_image_input( |
| 243 | self, **kwargs: object |
| 244 | ) -> InternVLImageInputs | None: |
| 245 | pixel_values_flat = kwargs.pop("pixel_values_flat", None) |
| 246 | image_num_patches = kwargs.pop("image_num_patches", None) |
| 247 | image_embeds = kwargs.pop("image_embeds", None) |
| 248 | |
| 249 | if pixel_values_flat is None and image_embeds is None: |
| 250 | return None |
| 251 | |
| 252 | if image_embeds is not None: |
| 253 | return InternVLImageEmbeddingInputs( |
| 254 | type="image_embeds", |
| 255 | data=image_embeds, |
| 256 | ) |
| 257 | |
| 258 | image_token_id = kwargs["image_token_id"] |
| 259 | if isinstance(image_token_id, torch.Tensor): |
| 260 | image_token_id = image_token_id.flatten().unique().item() |
| 261 | |
| 262 | assert isinstance(image_token_id, int) |
| 263 | self.img_context_token_id = image_token_id |
| 264 | |
| 265 | if pixel_values_flat is not None: |
| 266 | return InternVLImagePixelInputs( |
| 267 | type="pixel_values", |
| 268 | pixel_values_flat=pixel_values_flat, |
| 269 | num_patches=image_num_patches, |
| 270 | resolve_bindings={ |
| 271 | "h": self.config.force_image_size, |
| 272 | "w": self.config.force_image_size, |
| 273 | }, |
| 274 | ) |
| 275 | |
| 276 | raise AssertionError("This line should be unreachable.") |
| 277 | |
| 278 | def _process_image_input( |
| 279 | self, |
| 280 | image_input: InternVLImageInputs, |
| 281 | ) -> tuple[torch.Tensor, ...]: |
| 282 | if image_input["type"] == "image_embeds": |
| 283 | return image_input["data"] |
| 284 | |
| 285 | image_embeds = self.extract_feature(image_input["pixel_values_flat"]) |
| 286 | |
| 287 | num_patches = image_input["num_patches"] |
| 288 | hidden_size = self.config.get_text_config().hidden_size |
| 289 | |
| 290 | # Only one image in the current batch |
| 291 | if len(num_patches) == 1: |
| 292 | return (image_embeds.view(-1, hidden_size),) |
| 293 | |
| 294 | # NOTE: Image embeddings are split into separate tensors for each image |
| 295 | # by the size of each embedding. |
| 296 | feature_size = image_embeds.shape[1] |
| 297 | image_embeds = image_embeds.view(-1, hidden_size) |
| 298 | image_feature_sizes = [ |
| 299 | num_patches * feature_size for num_patches in num_patches |
| 300 | ] |
| 301 | return image_embeds.split(image_feature_sizes) |
| 302 | |
| 303 | def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict: |
| 304 | modalities = {} |
| 305 | |
| 306 | # Preserve the order of modalities if there are multiple of them |
| 307 | # from the order of kwargs. |
| 308 | for input_key in kwargs: |
| 309 | if ( |
| 310 | input_key in ("pixel_values_flat", "image_embeds") |
| 311 | and "images" not in modalities |
| 312 | ): |
| 313 | modalities["images"] = self._parse_and_validate_image_input(**kwargs) |
| 314 | |
| 315 | return modalities |
| 316 | |
| 317 | def _set_visual_token_mask(self, input_ids: torch.Tensor) -> None: |
| 318 | self.visual_token_mask = None |
| 319 | |
| 320 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 321 | modalities = self._parse_and_validate_multimodal_inputs(**kwargs) |
| 322 | if not modalities: |
| 323 | return [] |
| 324 | |
| 325 | # The result multimodal_embeddings is tuple of tensors, with each |
| 326 | # tensor corresponding to a multimodal data item (image). |
| 327 | multimodal_embeddings: tuple[torch.Tensor, ...] = () |
| 328 | |
| 329 | # NOTE: It is important to iterate over the keys in this dictionary |
| 330 | # to preserve the order of the modalities. |
| 331 | for modality in modalities: |
| 332 | if modality == "images": |
| 333 | image_input = modalities["images"] |
| 334 | image_embeddings = self._process_image_input(image_input) |
| 335 | multimodal_embeddings += tuple(image_embeddings) |
| 336 | |
| 337 | return multimodal_embeddings |
| 338 | |
| 339 | def embed_input_ids( |
| 340 | self, |
| 341 | input_ids: torch.Tensor, |
| 342 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 343 | *, |
| 344 | is_multimodal: torch.Tensor | None = None, |
| 345 | ) -> torch.Tensor: |
| 346 | if multimodal_embeddings is not None and len(multimodal_embeddings) > 0: |
| 347 | self._set_visual_token_mask(input_ids) |
| 348 | |
| 349 | # This is to satisfy the type checker for each overload |
| 350 | if multimodal_embeddings is None or is_multimodal is None: |
| 351 | return super().embed_input_ids(input_ids) |
| 352 | |
| 353 | return super().embed_input_ids( |
| 354 | input_ids, |
| 355 | multimodal_embeddings=multimodal_embeddings, |
| 356 | is_multimodal=is_multimodal, |
| 357 | ) |
| 358 | |
| 359 | def forward( |
| 360 | self, |
| 361 | input_ids: torch.Tensor | None, |
| 362 | positions: torch.Tensor, |
| 363 | intermediate_tensors: IntermediateTensors | None = None, |
| 364 | inputs_embeds: torch.Tensor | None = None, |
| 365 | **kwargs: object, |
| 366 | ) -> IntermediateTensors: |
| 367 | if intermediate_tensors is not None: |
| 368 | inputs_embeds = None |
| 369 | |
| 370 | forward_kwargs = { |
| 371 | "input_ids": input_ids, |
| 372 | "positions": positions, |
| 373 | "intermediate_tensors": intermediate_tensors, |
| 374 | "inputs_embeds": inputs_embeds, |
| 375 | } |
| 376 | |
| 377 | # Only required if the model is mono-architecture |
| 378 | if self.visual_token_mask is not None: |
| 379 | forward_kwargs.update({"visual_token_mask": self.visual_token_mask}) |
| 380 | self.visual_token_mask = None |
| 381 | |
| 382 | hidden_states = self.language_model.model(**forward_kwargs) |
| 383 | return hidden_states |
| 384 | |
| 385 | def compute_logits( |
| 386 | self, |
| 387 | hidden_states: torch.Tensor, |
| 388 | ) -> torch.Tensor | None: |
| 389 | return self.language_model.compute_logits(hidden_states) |
| 390 | |
| 391 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 392 | loader = AutoWeightsLoader(self) |
| 393 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 394 | |
| 395 | def get_mm_mapping(self) -> MultiModelKeys: |
| 396 | """Get the module prefix in multimodal models.""" |
| 397 | return MultiModelKeys.from_string_field( |
| 398 | language_model="language_model", |
| 399 | connector="mlp1", |
| 400 | tower_model="vision_model", |
| 401 | ) |
| 402 | |
| 403 | |
| 404 | # -------------------------------------------------------- |
| 405 | # LlamaNemotronVL Embedding Model (nvidia/llama-nemotron-embed-vl-1b-v2) |
| 406 | # Extends LlamaNemotronVLChatModel for embedding/pooling tasks: |
| 407 | # - SigLIP vision encoder (instead of C-RADIO) |
| 408 | # - Bidirectional (non-causal) LLaMA language model |
| 409 | # - Pooler output instead of generative logits |
| 410 | # -------------------------------------------------------- |
| 411 | |
| 412 | |
| 413 | class LlamaNemotronVLEmbedProcessingInfo(BaseInternVLProcessingInfo): |
| 414 | """Processing info for LlamaNemotronVL embedding model.""" |
| 415 | |
| 416 | def get_image_processor(self, **kwargs): |
| 417 | model_config = self.ctx.model_config |
| 418 | |
| 419 | config = self.get_hf_config() |
| 420 | processor_config = ( |
| 421 | get_hf_file_to_dict( |
| 422 | "processor_config.json", |
| 423 | model_config.model, |
| 424 | model_config.revision, |
| 425 | ) |
| 426 | or {} |
| 427 | ) |
| 428 | |
| 429 | min_dynamic_patch = processor_config.get( |
| 430 | "min_input_tiles", |
| 431 | getattr(config, "min_dynamic_patch", 1), |
| 432 | ) |
| 433 | max_dynamic_patch = processor_config.get( |
| 434 | "max_input_tiles", |
| 435 | getattr(config, "max_dynamic_patch", 1), |
| 436 | ) |
| 437 | dynamic_image_size = processor_config.get( |
| 438 | "dynamic_image_size", |
| 439 | getattr(config, "dynamic_image_size", True), |
| 440 | ) |
| 441 | |
| 442 | kwargs = self.ctx.get_merged_mm_kwargs(kwargs) |
| 443 | kwargs.setdefault("image_size", config.force_image_size) |
| 444 | kwargs.setdefault("min_dynamic_patch", min_dynamic_patch) |
| 445 | kwargs.setdefault("max_dynamic_patch", max_dynamic_patch) |
| 446 | kwargs.setdefault("dynamic_image_size", dynamic_image_size) |
| 447 | kwargs.setdefault("use_thumbnail", True) |
| 448 | |
| 449 | return LlamaNemotronVLEmbedImageProcessor(**kwargs) |
| 450 | |
| 451 | def get_hf_processor(self, **kwargs: object) -> LlamaNemotronVLEmbedProcessor: |
| 452 | config = self.get_hf_config() |
| 453 | vision_config = config.vision_config |
| 454 | |
| 455 | image_processor = self.get_image_processor(**kwargs) |
| 456 | image_size = image_processor.image_size |
| 457 | patch_size = vision_config.patch_size |
| 458 | downsample_ratio = config.downsample_ratio |
| 459 | image_seq_length = int((image_size // patch_size) ** 2 * (downsample_ratio**2)) |
| 460 | |
| 461 | return LlamaNemotronVLEmbedProcessor( |
| 462 | tokenizer=self.get_tokenizer(), |
| 463 | image_processor=image_processor, |
| 464 | image_seq_length=image_seq_length, |
| 465 | ) |
| 466 | |
| 467 | |
| 468 | @MULTIMODAL_REGISTRY.register_processor( |
| 469 | BaseInternVLMultiModalProcessor[LlamaNemotronVLEmbedProcessingInfo], |
| 470 | info=LlamaNemotronVLEmbedProcessingInfo, |
| 471 | dummy_inputs=BaseInternVLDummyInputsBuilder[LlamaNemotronVLEmbedProcessingInfo], |
| 472 | ) |
| 473 | class LlamaNemotronVLForEmbedding(LlamaNemotronVLChatModel, VllmModelForPooling): |
| 474 | """LlamaNemotronVL model for embeddings. |
| 475 | |
| 476 | Inherits from LlamaNemotronVLChatModel and specializes it for embedding tasks: |
| 477 | - Uses SigLIP vision encoder instead of C-RADIO |
| 478 | - Uses bidirectional LLaMA (via llm_config) instead of causal LLaMA |
| 479 | - Adds pooler for embedding output instead of generating logits |
| 480 | """ |
| 481 | |
| 482 | is_pooling_model = True |
| 483 | supports_mm_device_do_normalize = True |
| 484 | |
| 485 | # Weight mapping from checkpoint format to vLLM format |
| 486 | # Different from parent class due to different vision model structure |
| 487 | hf_to_vllm_mapper = WeightsMapper( |
| 488 | orig_to_new_prefix={ |
| 489 | # Language model mapping |
| 490 | "language_model.layers.": "language_model.model.layers.", |
| 491 | "language_model.embed_tokens.": "language_model.model.embed_tokens.", |
| 492 | "language_model.norm.": "language_model.model.norm.", |
| 493 | # Vision model mapping (SiglipVisionModel has nested vision_model) |
| 494 | "vision_model.encoder.": "vision_model.vision_model.encoder.", |
| 495 | "vision_model.embeddings.": "vision_model.vision_model.embeddings.", |
| 496 | "vision_model.post_layernorm.": "vision_model.vision_model.post_layernorm.", |
| 497 | } |
| 498 | ) |
| 499 | |
| 500 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 501 | super().__init__(vllm_config=vllm_config, prefix=prefix) |
| 502 | |
| 503 | config = vllm_config.model_config.hf_config |
| 504 | |
| 505 | # Override: get img_context_token_id from config (parent sets None) |
| 506 | self.img_context_token_id = getattr(config, "img_context_token_id", None) |
| 507 | |
| 508 | if self.model_config.get_multimodal_config().mm_device_do_normalize: |
| 509 | self.input_norm = FusedMMInputNorm( |
| 510 | image_mean=list(SIGLIP_MEAN), |
| 511 | image_std=list(SIGLIP_STD), |
| 512 | rescale_factor=1 / 255, |
| 513 | ) |
| 514 | else: |
| 515 | self.input_norm = IdentityInputNorm() |
| 516 | |
| 517 | # Initialize pooler for embedding output |
| 518 | pooler_config = vllm_config.model_config.pooler_config |
| 519 | assert pooler_config is not None |
| 520 | self.pooler = DispatchPooler.for_embedding(pooler_config) |
| 521 | |
| 522 | def _init_vision_model( |
| 523 | self, |
| 524 | config: PreTrainedConfig, |
| 525 | quant_config, |
| 526 | *, |
| 527 | prefix: str, |
| 528 | ) -> nn.Module: |
| 529 | """Override to use SigLIP instead of C-RADIO.""" |
| 530 | return SiglipVisionModel( |
| 531 | config.vision_config, |
| 532 | quant_config=quant_config, |
| 533 | prefix=prefix, |
| 534 | use_head=False, |
| 535 | ) |
| 536 | |
| 537 | def _init_mlp1( |
| 538 | self, |
| 539 | config: PreTrainedConfig, |
| 540 | vit_hidden_size: int | None = None, |
| 541 | vision_projection_hidden_size: int | None = None, |
| 542 | ) -> nn.Module: |
| 543 | """Override to use different MLP structure for embedding model.""" |
| 544 | if vit_hidden_size is None: |
| 545 | vit_hidden_size = config.vision_config.hidden_size |
| 546 | if vision_projection_hidden_size is None: |
| 547 | vision_projection_hidden_size = config.get_text_config().hidden_size |
| 548 | return super()._init_mlp1( |
| 549 | config, |
| 550 | vit_hidden_size=vit_hidden_size, |
| 551 | vision_projection_hidden_size=vision_projection_hidden_size, |
| 552 | ) |
| 553 | |
| 554 | def _call_vision_model(self, pixel_values: torch.Tensor) -> torch.Tensor: |
| 555 | """Override to handle SigLIP interface.""" |
| 556 | original_shape = pixel_values.shape |
| 557 | pixel_values = self.input_norm( |
| 558 | pixel_values.flatten(start_dim=1), self.vision_model.dtype |
| 559 | ).view(original_shape) |
| 560 | return self.vision_model(pixel_values) |
| 561 | |
| 562 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 563 | """Override to use different weight mapping for SigLIP.""" |
| 564 | loader = AutoWeightsLoader(self) |
| 565 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 566 | |
| 567 | |
| 568 | class LlamaNemotronVLForSequenceClassification( |
| 569 | LlamaNemotronVLForEmbedding, SupportsCrossEncoding |
| 570 | ): |
| 571 | """LlamaNemotronVL model variant for sequence classification / reranking.""" |
| 572 | |
| 573 | # Reranker checkpoint places base model weights under `model.*`, |
| 574 | # while `score.*` remains at the top level. |
| 575 | hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={"model.": ""}) | ( |
| 576 | LlamaNemotronVLForEmbedding.hf_to_vllm_mapper |
| 577 | ) |
| 578 | |
| 579 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 580 | super().__init__(vllm_config=vllm_config, prefix=prefix) |
| 581 | |
| 582 | text_config = vllm_config.model_config.hf_config.get_text_config() |
| 583 | model_config = vllm_config.model_config |
| 584 | quant_config = vllm_config.quant_config |
| 585 | |
| 586 | self.score = ReplicatedLinear( |
| 587 | model_config.get_hidden_size(), |
| 588 | text_config.num_labels, |
| 589 | bias=False, |
| 590 | params_dtype=model_config.head_dtype, |
| 591 | quant_config=quant_config, |
| 592 | return_bias=False, |
| 593 | prefix=maybe_prefix(prefix, "score"), |
| 594 | ) |
| 595 | |
| 596 | pooler_config = model_config.pooler_config |
| 597 | assert pooler_config is not None |
| 598 | self.pooler = DispatchPooler.for_seq_cls(pooler_config, classifier=self.score) |
| 599 | |
| 600 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 601 | loaded_weights = super().load_weights(weights) |
| 602 | |
| 603 | # reranker checkpoint omits the inner LM seq-cls head |
| 604 | # (`language_model.score.*`). It is unused by this outer model, but |
| 605 | # the default loader expects all parameters to be initialized. |
| 606 | for name, param in self.named_parameters(): |
| 607 | if not name.startswith("language_model.score.") or name in loaded_weights: |
| 608 | continue |
| 609 | |
| 610 | if name.endswith(".weight"): |
| 611 | torch.nn.init.kaiming_uniform_(param, a=math.sqrt(5)) |
| 612 | elif name.endswith(".bias"): |
| 613 | torch.nn.init.zeros_(param) |
| 614 | else: |
| 615 | torch.nn.init.normal_(param, mean=0.0, std=0.02) |
| 616 | |
| 617 | loaded_weights.add(name) |
| 618 | |
| 619 | return loaded_weights |