| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # |
| 4 | # Adapted from https://github.com/amalad/vllm/blob/nemotron_parse/vllm/model_executor/models/nemotron_parse.py |
| 5 | # that's based on https://huggingface.co/nvidia/NVIDIA-Nemotron-Parse-v1.1/blob/main/hf_nemotron_parse_modeling.py |
| 6 | # |
| 7 | # Bart classes based on old vLLM codebase: |
| 8 | # https://github.com/vllm-project/vllm/blob/v0.10.2/vllm/model_executor/models/bart.py |
| 9 | |
| 10 | import math |
| 11 | from collections.abc import Iterable, Mapping, Sequence |
| 12 | from typing import Annotated, Literal |
| 13 | |
| 14 | import torch |
| 15 | import torch.nn as nn |
| 16 | from einops import rearrange |
| 17 | from transformers import ( |
| 18 | BartConfig, |
| 19 | BatchFeature, |
| 20 | PreTrainedConfig, |
| 21 | ) |
| 22 | |
| 23 | from vllm.config import CacheConfig, VllmConfig |
| 24 | from vllm.config.lora import LoRAConfig |
| 25 | from vllm.config.multimodal import MultiModalDummyOptions |
| 26 | from vllm.inputs import MultiModalDataDict |
| 27 | from vllm.logger import init_logger |
| 28 | from vllm.model_executor.layers.activation import get_act_fn |
| 29 | from vllm.model_executor.layers.linear import ColumnParallelLinear, RowParallelLinear |
| 30 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 31 | from vllm.model_executor.layers.quantization.base_config import QuantizationConfig |
| 32 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 33 | ParallelLMHead, |
| 34 | VocabParallelEmbedding, |
| 35 | ) |
| 36 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader |
| 37 | from vllm.model_executor.models.interfaces import ( |
| 38 | MultiModalEmbeddings, |
| 39 | SupportsMultiModal, |
| 40 | ) |
| 41 | from vllm.model_executor.models.radio import RadioModel |
| 42 | from vllm.model_executor.models.whisper import WhisperAttention, WhisperCrossAttention |
| 43 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 44 | from vllm.multimodal.inputs import ( |
| 45 | MultiModalFieldConfig, |
| 46 | MultiModalKwargsItems, |
| 47 | ) |
| 48 | from vllm.multimodal.parse import MultiModalDataItems |
| 49 | from vllm.multimodal.processing import ( |
| 50 | BaseDummyInputsBuilder, |
| 51 | BaseProcessingInfo, |
| 52 | EncDecMultiModalProcessor, |
| 53 | PromptReplacement, |
| 54 | PromptUpdate, |
| 55 | ) |
| 56 | from vllm.renderers import TokenizeParams |
| 57 | from vllm.transformers_utils.configs.radio import RadioConfig |
| 58 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 59 | from vllm.v1.attention.backend import AttentionType |
| 60 | |
| 61 | logger = init_logger(__name__) |
| 62 | |
| 63 | |
| 64 | class BartScaledWordEmbedding(VocabParallelEmbedding): |
| 65 | """This module overrides VocabParallelEmbedding's |
| 66 | forward by multiplying with embeddings scale. |
| 67 | """ |
| 68 | |
| 69 | def __init__( |
| 70 | self, num_embeddings: int, embedding_dim: int, embed_scale: float = 1.0 |
| 71 | ): |
| 72 | super().__init__(num_embeddings, embedding_dim) |
| 73 | self.embed_scale = embed_scale |
| 74 | |
| 75 | def forward(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 76 | return super().forward(input_ids) * self.embed_scale |
| 77 | |
| 78 | |
| 79 | class BartDecoderLayer(nn.Module): |
| 80 | def __init__( |
| 81 | self, |
| 82 | config: BartConfig, |
| 83 | cache_config: CacheConfig | None = None, |
| 84 | quant_config: QuantizationConfig | None = None, |
| 85 | prefix: str = "", |
| 86 | ): |
| 87 | super().__init__() |
| 88 | self.embed_dim = config.d_model |
| 89 | |
| 90 | self.self_attn = WhisperAttention( |
| 91 | embed_dim=self.embed_dim, |
| 92 | num_heads=config.decoder_attention_heads, |
| 93 | attn_type=AttentionType.DECODER, |
| 94 | cache_config=cache_config, |
| 95 | quant_config=quant_config, |
| 96 | prefix=f"{prefix}.self_attn", |
| 97 | ) |
| 98 | self.activation_fn = get_act_fn(config.activation_function) |
| 99 | |
| 100 | self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim) |
| 101 | """ |
| 102 | afeldman-nm: personally I would call this "cross-attention", |
| 103 | however I left the name as "encoder_attn" to maintain consistency |
| 104 | with the name of the pretrained weights. |
| 105 | """ |
| 106 | self.encoder_attn = WhisperCrossAttention( |
| 107 | self.embed_dim, |
| 108 | config.decoder_attention_heads, |
| 109 | cache_config=cache_config, |
| 110 | quant_config=quant_config, |
| 111 | prefix=f"{prefix}.encoder_attn", |
| 112 | ) |
| 113 | self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim) |
| 114 | |
| 115 | ffn_hidden_size = self.embed_dim |
| 116 | ffn_intermediate_size = config.encoder_ffn_dim |
| 117 | ffn_has_bias = True |
| 118 | self.fc1 = ColumnParallelLinear( |
| 119 | ffn_hidden_size, |
| 120 | ffn_intermediate_size, |
| 121 | bias=ffn_has_bias, |
| 122 | quant_config=quant_config, |
| 123 | prefix=f"{prefix}.fc1", |
| 124 | ) |
| 125 | self.fc2 = RowParallelLinear( |
| 126 | ffn_intermediate_size, |
| 127 | ffn_hidden_size, |
| 128 | bias=ffn_has_bias, |
| 129 | quant_config=quant_config, |
| 130 | prefix=f"{prefix}.fc2", |
| 131 | ) |
| 132 | |
| 133 | self.final_layer_norm = nn.LayerNorm(self.embed_dim) |
| 134 | |
| 135 | def forward( |
| 136 | self, |
| 137 | decoder_hidden_states: torch.Tensor, |
| 138 | encoder_hidden_states: torch.Tensor | None = None, |
| 139 | ) -> torch.Tensor: |
| 140 | r"""Args: |
| 141 | decoder_hidden_states: torch.Tensor of *decoder* input embeddings. |
| 142 | encoder_hidden_states: torch.Tensor of *encoder* input embeddings. |
| 143 | |
| 144 | Returns: |
| 145 | Decoder layer output torch.Tensor |
| 146 | |
| 147 | """ |
| 148 | residual = decoder_hidden_states |
| 149 | |
| 150 | # Self Attention |
| 151 | hidden_states = self.self_attn(hidden_states=decoder_hidden_states) |
| 152 | |
| 153 | hidden_states = residual + hidden_states |
| 154 | hidden_states = self.self_attn_layer_norm(hidden_states) |
| 155 | |
| 156 | # Cross-Attention Block |
| 157 | |
| 158 | residual = hidden_states |
| 159 | |
| 160 | hidden_states = self.encoder_attn( |
| 161 | hidden_states=hidden_states, |
| 162 | encoder_hidden_states=encoder_hidden_states, |
| 163 | ) |
| 164 | |
| 165 | hidden_states = residual + hidden_states |
| 166 | hidden_states = self.encoder_attn_layer_norm(hidden_states) |
| 167 | |
| 168 | # Fully Connected |
| 169 | residual = hidden_states |
| 170 | fc1_out, _ = self.fc1(hidden_states) |
| 171 | hidden_states = self.activation_fn(fc1_out) |
| 172 | |
| 173 | hidden_states, _ = self.fc2(hidden_states) |
| 174 | |
| 175 | hidden_states = residual + hidden_states |
| 176 | hidden_states = self.final_layer_norm(hidden_states) |
| 177 | |
| 178 | return hidden_states |
| 179 | |
| 180 | |
| 181 | class MBartDecoderLayer(BartDecoderLayer): |
| 182 | def forward( |
| 183 | self, |
| 184 | decoder_hidden_states: torch.Tensor, |
| 185 | encoder_hidden_states: torch.Tensor | None = None, |
| 186 | ) -> torch.Tensor: |
| 187 | residual = decoder_hidden_states |
| 188 | hidden_states = self.self_attn_layer_norm(decoder_hidden_states) |
| 189 | |
| 190 | # Self Attention |
| 191 | hidden_states = self.self_attn(hidden_states=hidden_states) |
| 192 | |
| 193 | hidden_states = residual + hidden_states |
| 194 | |
| 195 | # Cross-Attention Block |
| 196 | |
| 197 | residual = hidden_states |
| 198 | hidden_states = self.encoder_attn_layer_norm(hidden_states) |
| 199 | |
| 200 | hidden_states = self.encoder_attn( |
| 201 | hidden_states=hidden_states, |
| 202 | encoder_hidden_states=encoder_hidden_states, |
| 203 | ) |
| 204 | |
| 205 | hidden_states = residual + hidden_states |
| 206 | |
| 207 | # Fully Connected |
| 208 | residual = hidden_states |
| 209 | hidden_states = self.final_layer_norm(hidden_states) |
| 210 | fc1_out, _ = self.fc1(hidden_states) |
| 211 | hidden_states = self.activation_fn(fc1_out) |
| 212 | |
| 213 | hidden_states, _ = self.fc2(hidden_states) |
| 214 | |
| 215 | hidden_states = residual + hidden_states |
| 216 | |
| 217 | return hidden_states |
| 218 | |
| 219 | |
| 220 | class MBartDecoderNoPos(nn.Module): |
| 221 | """Transformer decoder consisting of *config.decoder_layers* layers. |
| 222 | Each layer is a [`BartDecoderLayer`] |
| 223 | |
| 224 | Args: |
| 225 | config: BartConfig |
| 226 | embed_tokens (nn.Embedding): output embedding |
| 227 | |
| 228 | """ |
| 229 | |
| 230 | def __init__( |
| 231 | self, |
| 232 | config: BartConfig, |
| 233 | cache_config: CacheConfig | None = None, |
| 234 | quant_config: QuantizationConfig | None = None, |
| 235 | lora_config: LoRAConfig | None = None, |
| 236 | embed_tokens: nn.Embedding | None = None, |
| 237 | prefix: str = "", |
| 238 | ): |
| 239 | super().__init__() |
| 240 | self.cache_config = cache_config |
| 241 | self.quant_config = quant_config |
| 242 | self.lora_config = lora_config |
| 243 | embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0 |
| 244 | |
| 245 | self.embed_tokens = BartScaledWordEmbedding( |
| 246 | config.vocab_size, config.d_model, embed_scale=embed_scale |
| 247 | ) |
| 248 | |
| 249 | if embed_tokens is not None: |
| 250 | self.embed_tokens.weight = embed_tokens.weight |
| 251 | |
| 252 | self.layers = nn.ModuleList( |
| 253 | [ |
| 254 | MBartDecoderLayer( |
| 255 | config, |
| 256 | cache_config, |
| 257 | quant_config, |
| 258 | prefix=f"{prefix}.layers.{layer_idx}", |
| 259 | ) |
| 260 | for layer_idx in range(config.decoder_layers) |
| 261 | ] |
| 262 | ) |
| 263 | |
| 264 | self.layernorm_embedding = nn.LayerNorm(config.d_model) |
| 265 | self.layer_norm = nn.LayerNorm(config.d_model) |
| 266 | |
| 267 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 268 | return self.embed_tokens(input_ids) |
| 269 | |
| 270 | def forward( |
| 271 | self, |
| 272 | decoder_input_ids: torch.Tensor | None, |
| 273 | *, |
| 274 | encoder_hidden_states: torch.Tensor | None, |
| 275 | inputs_embeds: torch.Tensor | None = None, |
| 276 | **kwargs, |
| 277 | ) -> torch.Tensor: |
| 278 | r"""Args: |
| 279 | decoder_input_ids: Indices of *decoder* input sequence tokens in the |
| 280 | vocabulary. Padding will be ignored by default should you provide it. |
| 281 | encoder_hidden_states: Tensor of encoder output embeddings |
| 282 | Returns: |
| 283 | Decoder output torch.Tensor |
| 284 | |
| 285 | """ |
| 286 | if inputs_embeds is None: |
| 287 | inputs_embeds = self.embed_tokens(decoder_input_ids) |
| 288 | |
| 289 | hidden_states = self.layernorm_embedding(inputs_embeds) |
| 290 | |
| 291 | # decoder layers |
| 292 | |
| 293 | for decoder_layer in self.layers: |
| 294 | hidden_states = decoder_layer( |
| 295 | decoder_hidden_states=hidden_states, |
| 296 | encoder_hidden_states=encoder_hidden_states, |
| 297 | ) |
| 298 | |
| 299 | hidden_states = self.layer_norm(hidden_states) |
| 300 | return hidden_states |
| 301 | |
| 302 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 303 | stacked_params_mapping = [ |
| 304 | # (param_name, shard_name, shard_id) |
| 305 | (".self_attn.qkv_proj", ".self_attn.q_proj", "q"), |
| 306 | (".self_attn.qkv_proj", ".self_attn.k_proj", "k"), |
| 307 | (".self_attn.qkv_proj", ".self_attn.v_proj", "v"), |
| 308 | # MergedColumnParallelLinear uses integer indices (0, 1) |
| 309 | (".encoder_attn.kv_proj", ".encoder_attn.k_proj", 0), |
| 310 | (".encoder_attn.kv_proj", ".encoder_attn.v_proj", 1), |
| 311 | ] |
| 312 | params_dict = dict(self.named_parameters()) |
| 313 | loaded_params: set[str] = set() |
| 314 | for name, loaded_weight in weights: |
| 315 | if name.startswith("embed_positions"): |
| 316 | continue |
| 317 | |
| 318 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 319 | if weight_name not in name: |
| 320 | continue |
| 321 | name = name.replace(weight_name, param_name) |
| 322 | # Skip loading extra bias for GPTQ models. |
| 323 | if name.endswith(".bias") and name not in params_dict: |
| 324 | continue |
| 325 | |
| 326 | param = params_dict[name] |
| 327 | weight_loader = param.weight_loader |
| 328 | weight_loader(param, loaded_weight, shard_id) |
| 329 | break |
| 330 | else: |
| 331 | # Skip loading extra bias for GPTQ models. |
| 332 | if name.endswith(".bias") and name not in params_dict: |
| 333 | continue |
| 334 | |
| 335 | param = params_dict[name] |
| 336 | weight_loader = getattr(param, "weight_loader", default_weight_loader) |
| 337 | weight_loader(param, loaded_weight) |
| 338 | loaded_params.add(name) |
| 339 | return loaded_params |
| 340 | |
| 341 | |
| 342 | class NemotronParsePixelInputs(TensorSchema): |
| 343 | """Dimensions: |
| 344 | - b: Batch size |
| 345 | - c: Number of channels (3) |
| 346 | - h: Height |
| 347 | - w: Width |
| 348 | """ |
| 349 | |
| 350 | type: Literal["pixel_values"] |
| 351 | data: Annotated[torch.Tensor, TensorShape("b", 3, "h", "w")] |
| 352 | |
| 353 | |
| 354 | class NemotronParseProcessingInfo(BaseProcessingInfo): |
| 355 | def get_default_tok_params(self) -> TokenizeParams: |
| 356 | return super().get_default_tok_params().with_kwargs(add_special_tokens=False) |
| 357 | |
| 358 | @property |
| 359 | def skip_prompt_length_check(self) -> bool: |
| 360 | return True # Because the encoder prompt is padded |
| 361 | |
| 362 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 363 | return {"image": 1} |
| 364 | |
| 365 | def get_num_image_tokens(self) -> int: |
| 366 | config = self.get_hf_config() |
| 367 | final_size = config.image_size |
| 368 | patch_size = config.encoder.patch_size |
| 369 | |
| 370 | return (final_size[0] // patch_size) * ((final_size[1] // patch_size) // 4) + 1 |
| 371 | |
| 372 | def get_mm_max_tokens_per_item( |
| 373 | self, |
| 374 | seq_len: int, |
| 375 | mm_counts: Mapping[str, int], |
| 376 | ) -> Mapping[str, int] | None: |
| 377 | image_tokens = self.get_num_image_tokens() |
| 378 | return {"image": image_tokens} |
| 379 | |
| 380 | |
| 381 | class NemotronParseDummyInputsBuilder( |
| 382 | BaseDummyInputsBuilder[NemotronParseProcessingInfo] |
| 383 | ): |
| 384 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 385 | return "" |
| 386 | |
| 387 | def get_dummy_mm_data( |
| 388 | self, |
| 389 | seq_len: int, |
| 390 | mm_counts: Mapping[str, int], |
| 391 | mm_options: MultiModalDummyOptions, |
| 392 | ) -> MultiModalDataDict: |
| 393 | num_images = mm_counts.get("image", 0) |
| 394 | |
| 395 | target_width, target_height = self.info.get_hf_config().image_size |
| 396 | |
| 397 | return { |
| 398 | "image": self._get_dummy_images( |
| 399 | width=target_width, height=target_height, num_images=num_images |
| 400 | ) |
| 401 | } |
| 402 | |
| 403 | |
| 404 | class NemotronParseMultiModalProcessor( |
| 405 | EncDecMultiModalProcessor[NemotronParseProcessingInfo] |
| 406 | ): |
| 407 | def create_encoder_prompt( |
| 408 | self, |
| 409 | prompt: list[int], |
| 410 | mm_items: MultiModalDataItems, |
| 411 | ) -> list[int]: |
| 412 | return [0] |
| 413 | |
| 414 | def _get_mm_fields_config( |
| 415 | self, |
| 416 | hf_inputs: BatchFeature, |
| 417 | hf_processor_mm_kwargs: Mapping[str, object], |
| 418 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 419 | return dict(pixel_values=MultiModalFieldConfig.batched("image")) |
| 420 | |
| 421 | def _get_prompt_updates( |
| 422 | self, |
| 423 | mm_items: MultiModalDataItems, |
| 424 | hf_processor_mm_kwargs: Mapping[str, object], |
| 425 | out_mm_kwargs: MultiModalKwargsItems, |
| 426 | ) -> Sequence[PromptUpdate]: |
| 427 | num_image_tokens = self.info.get_num_image_tokens() |
| 428 | |
| 429 | return [ |
| 430 | PromptReplacement( |
| 431 | modality="image", |
| 432 | target=[0], |
| 433 | replacement=[0] * num_image_tokens, |
| 434 | ) |
| 435 | ] |
| 436 | |
| 437 | |
| 438 | class RadioWithNeck(nn.Module): |
| 439 | """Vision encoder using RADIO model with custom neck.""" |
| 440 | |
| 441 | def __init__( |
| 442 | self, |
| 443 | config: PreTrainedConfig, |
| 444 | quant_config: QuantizationConfig | None = None, |
| 445 | prefix: str = "", |
| 446 | ): |
| 447 | super().__init__() |
| 448 | self.config = config.encoder |
| 449 | |
| 450 | self.model_encoder = self.get_vit_model_from_radio_config( |
| 451 | config, quant_config=quant_config |
| 452 | ) |
| 453 | |
| 454 | # Neck components |
| 455 | last_hidden_state = 1024 |
| 456 | self.conv1 = nn.Conv1d(1280, last_hidden_state, 1) |
| 457 | self.layer_norm1 = nn.LayerNorm( |
| 458 | last_hidden_state, eps=1e-06, elementwise_affine=True |
| 459 | ) |
| 460 | self.conv2 = nn.Conv2d( |
| 461 | last_hidden_state, |
| 462 | last_hidden_state, |
| 463 | kernel_size=(1, 4), |
| 464 | stride=(1, 4), |
| 465 | padding=0, |
| 466 | bias=False, |
| 467 | ) |
| 468 | self.layer_norm2 = nn.LayerNorm( |
| 469 | last_hidden_state, eps=1e-06, elementwise_affine=True |
| 470 | ) |
| 471 | self.sum_proj = ColumnParallelLinear( |
| 472 | 3840, |
| 473 | last_hidden_state, |
| 474 | quant_config=quant_config, |
| 475 | prefix=f"{prefix}.sum_proj", |
| 476 | ) |
| 477 | self.layer_norm3 = nn.LayerNorm( |
| 478 | last_hidden_state, eps=1e-06, elementwise_affine=True |
| 479 | ) |
| 480 | |
| 481 | def get_vit_model_from_radio_config( |
| 482 | self, |
| 483 | hf_config: PreTrainedConfig, |
| 484 | quant_config: QuantizationConfig | None = None, |
| 485 | ) -> RadioModel: |
| 486 | hf_config_vision = hf_config.encoder |
| 487 | model_name = hf_config_vision.args.get("model") |
| 488 | if model_name is None: |
| 489 | raise ValueError(f"Unsupported vit model type: {model_name}") |
| 490 | |
| 491 | radio_config = RadioConfig( |
| 492 | model_name=model_name, |
| 493 | image_size=hf_config.image_size, |
| 494 | **hf_config_vision.args, |
| 495 | ) |
| 496 | |
| 497 | return RadioModel(config=radio_config, quant_config=quant_config) |
| 498 | |
| 499 | def forward(self, pixel_values: torch.Tensor, **kwargs) -> torch.Tensor: |
| 500 | summary, feature = self.model_encoder(pixel_values) |
| 501 | |
| 502 | output = self.conv1(feature.permute(0, 2, 1)).permute(0, 2, 1) |
| 503 | output = self.layer_norm1(output) |
| 504 | |
| 505 | patch_size = self.config.patch_size |
| 506 | output = rearrange( |
| 507 | output, |
| 508 | "b (h w) d -> b d h w", |
| 509 | h=pixel_values.shape[-2] // patch_size, |
| 510 | w=pixel_values.shape[-1] // patch_size, |
| 511 | ) |
| 512 | |
| 513 | output = self.conv2(output) |
| 514 | output = rearrange(output, "b d h w -> b (h w) d") |
| 515 | output = self.layer_norm2(output) |
| 516 | summary = self.layer_norm3(self.sum_proj(summary)[0]) |
| 517 | output = torch.cat((output, summary.unsqueeze(1)), dim=1) |
| 518 | |
| 519 | return output |
| 520 | |
| 521 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): |
| 522 | model_encoder_weights = [] |
| 523 | adaptor_dict = { |
| 524 | name: param |
| 525 | for name, param in dict(self.named_parameters()).items() |
| 526 | if not name.startswith("model_encoder") |
| 527 | } |
| 528 | for name, w in weights: |
| 529 | if name.startswith("model_encoder"): |
| 530 | model_encoder_weights.append((".".join(name.split(".")[1:]), w)) |
| 531 | else: |
| 532 | param = adaptor_dict[name] |
| 533 | with torch.no_grad(): |
| 534 | default_weight_loader(param, w) |
| 535 | |
| 536 | self.model_encoder.load_weights(model_encoder_weights) |
| 537 | |
| 538 | |
| 539 | @MULTIMODAL_REGISTRY.register_processor( |
| 540 | NemotronParseMultiModalProcessor, |
| 541 | info=NemotronParseProcessingInfo, |
| 542 | dummy_inputs=NemotronParseDummyInputsBuilder, |
| 543 | ) |
| 544 | class NemotronParseForConditionalGeneration(nn.Module, SupportsMultiModal): |
| 545 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 546 | super().__init__() |
| 547 | config = vllm_config.model_config.hf_config |
| 548 | |
| 549 | self.config = config |
| 550 | self.vision_config = config.encoder |
| 551 | cache_config = vllm_config.cache_config |
| 552 | quant_config = vllm_config.quant_config |
| 553 | |
| 554 | with self._mark_tower_model(vllm_config, "image"): |
| 555 | self.encoder = RadioWithNeck( |
| 556 | config=config, quant_config=quant_config, prefix=f"{prefix}.encoder" |
| 557 | ) |
| 558 | |
| 559 | with self._mark_language_model(vllm_config): |
| 560 | self.decoder = MBartDecoderNoPos( |
| 561 | config.decoder, |
| 562 | cache_config=cache_config, |
| 563 | quant_config=quant_config, |
| 564 | prefix=f"{prefix}.decoder", |
| 565 | ) |
| 566 | |
| 567 | self.vocab_size = config.decoder.vocab_size |
| 568 | self.lm_head = ParallelLMHead( |
| 569 | config.decoder.vocab_size, config.decoder.d_model, quant_config=quant_config |
| 570 | ) |
| 571 | # Some checkpoints (e.g. compact exports) tie the output head to the |
| 572 | # decoder's input embeddings instead of materializing a separate |
| 573 | # lm_head.weight tensor. |
| 574 | tie_word_embeddings = bool( |
| 575 | getattr(config, "tie_word_embeddings", False) |
| 576 | or getattr(config.decoder, "tie_word_embeddings", False) |
| 577 | ) |
| 578 | if tie_word_embeddings: |
| 579 | self.lm_head = self.lm_head.tie_weights(self.decoder.embed_tokens) |
| 580 | self.logits_processor = LogitsProcessor( |
| 581 | self.vocab_size, config.decoder.vocab_size |
| 582 | ) |
| 583 | |
| 584 | @classmethod |
| 585 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 586 | if modality.startswith("image"): |
| 587 | return None |
| 588 | |
| 589 | raise ValueError("Only image modality is supported") |
| 590 | |
| 591 | def _parse_and_validate_image_input( |
| 592 | self, **kwargs: object |
| 593 | ) -> NemotronParsePixelInputs | None: |
| 594 | pixel_values = kwargs.pop("pixel_values", None) |
| 595 | image_embeds = kwargs.pop("image_embeds", None) |
| 596 | |
| 597 | if pixel_values is None and image_embeds is None: |
| 598 | return None |
| 599 | |
| 600 | if pixel_values is not None and image_embeds is not None: |
| 601 | raise ValueError("Both pixel values and image embeds are provided.") |
| 602 | |
| 603 | if pixel_values is not None: |
| 604 | h, w = self.config.image_size |
| 605 | return NemotronParsePixelInputs( |
| 606 | type="pixel_values", |
| 607 | data=pixel_values, |
| 608 | resolve_bindings={ |
| 609 | "h": h, |
| 610 | "w": w, |
| 611 | }, |
| 612 | ) |
| 613 | |
| 614 | if image_embeds is not None: |
| 615 | raise NotImplementedError |
| 616 | |
| 617 | raise AssertionError("This line should be unreachable.") |
| 618 | |
| 619 | def _process_image_input( |
| 620 | self, image_input: NemotronParsePixelInputs |
| 621 | ) -> torch.Tensor: |
| 622 | assert image_input["type"] == "pixel_values" |
| 623 | pixel_values = image_input["data"] |
| 624 | dtype = next(self.encoder.parameters()).dtype |
| 625 | pixel_values = pixel_values.to(dtype) |
| 626 | return self.encoder(pixel_values) |
| 627 | |
| 628 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings | None: |
| 629 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 630 | if image_input is None: |
| 631 | return None |
| 632 | vision_embeddings = self._process_image_input(image_input) |
| 633 | return vision_embeddings |
| 634 | |
| 635 | def forward( |
| 636 | self, |
| 637 | input_ids: torch.Tensor | None, |
| 638 | positions: torch.Tensor, |
| 639 | encoder_outputs: list[torch.Tensor] | None = None, |
| 640 | **kwargs, |
| 641 | ) -> torch.Tensor: |
| 642 | r"""Args: |
| 643 | input_ids: torch.Tensor of *decoder* input token ids. |
| 644 | positions: torch.Tensor of *decoder* position indices. |
| 645 | encoder_outputs: List of encoder output tensors (vision embeddings). |
| 646 | During profiling, this may be None or empty. |
| 647 | |
| 648 | Returns: |
| 649 | Output torch.Tensor |
| 650 | |
| 651 | """ |
| 652 | inputs_embeds = None |
| 653 | if encoder_outputs: |
| 654 | inputs_embeds = torch.cat(encoder_outputs, dim=0) |
| 655 | hidden_states = self.decoder( |
| 656 | decoder_input_ids=input_ids, encoder_hidden_states=inputs_embeds |
| 657 | ) |
| 658 | return hidden_states |
| 659 | |
| 660 | def compute_logits( |
| 661 | self, |
| 662 | hidden_states: torch.Tensor, |
| 663 | ) -> torch.Tensor | None: |
| 664 | return self.logits_processor(self.lm_head, hidden_states) |
| 665 | |
| 666 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): |
| 667 | lm_head_dict = dict(self.lm_head.named_parameters()) |
| 668 | |
| 669 | def is_encoder(name: str) -> bool: |
| 670 | return name.startswith("encoder") |
| 671 | |
| 672 | def is_decoder(name: str) -> bool: |
| 673 | return name.startswith("decoder") |
| 674 | |
| 675 | def is_lm_head(name: str): |
| 676 | return name.startswith("lm_head") |
| 677 | |
| 678 | # Separate weights by component |
| 679 | encoder_weights = [] |
| 680 | decoder_weights = [] |
| 681 | |
| 682 | for name, w in weights: |
| 683 | if is_encoder(name): |
| 684 | encoder_weights.append((".".join(name.split(".")[1:]), w)) |
| 685 | elif is_decoder(name): |
| 686 | decoder_weights.append((".".join(name.split(".")[1:]), w)) |
| 687 | elif is_lm_head(name): |
| 688 | trimmed_name = ".".join(name.split(".")[1:]) |
| 689 | param = lm_head_dict[trimmed_name] |
| 690 | with torch.no_grad(): |
| 691 | default_weight_loader(param, w) |
| 692 | else: |
| 693 | logger.info("Found unexpected weight: %s", name) |
| 694 | |
| 695 | # Load encoder weights |
| 696 | self.encoder.load_weights(encoder_weights) |
| 697 | # Load decoder weights |
| 698 | self.decoder.load_weights(decoder_weights) |