| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import math |
| 5 | from collections.abc import Iterable, Mapping, Sequence |
| 6 | from dataclasses import dataclass |
| 7 | from functools import partial |
| 8 | from itertools import islice |
| 9 | from typing import Annotated |
| 10 | |
| 11 | import numpy as np |
| 12 | import torch |
| 13 | import torch.nn as nn |
| 14 | import torch.nn.functional as F |
| 15 | from einops import rearrange |
| 16 | from transformers import ( |
| 17 | BaseImageProcessor, |
| 18 | BatchFeature, |
| 19 | PreTrainedConfig, |
| 20 | ) |
| 21 | |
| 22 | from vllm.compilation.decorators import support_torch_compile |
| 23 | from vllm.config import CacheConfig, VllmConfig |
| 24 | from vllm.config.multimodal import MultiModalDummyOptions |
| 25 | from vllm.distributed import ( |
| 26 | get_pp_group, |
| 27 | get_tensor_model_parallel_rank, |
| 28 | get_tensor_model_parallel_world_size, |
| 29 | split_tensor_along_last_dim, |
| 30 | tensor_model_parallel_all_gather, |
| 31 | ) |
| 32 | from vllm.inputs import MultiModalDataDict |
| 33 | from vllm.model_executor.layers.activation import MulAndSilu, QuickGELU, SiluAndMul |
| 34 | from vllm.model_executor.layers.attention import Attention, MMEncoderAttention |
| 35 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 36 | from vllm.model_executor.layers.linear import ( |
| 37 | ColumnParallelLinear, |
| 38 | MergedColumnParallelLinear, |
| 39 | QKVParallelLinear, |
| 40 | RowParallelLinear, |
| 41 | ) |
| 42 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 43 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 44 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 45 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 46 | ParallelLMHead, |
| 47 | VocabParallelEmbedding, |
| 48 | ) |
| 49 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 50 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 51 | from vllm.multimodal.inputs import ( |
| 52 | MultiModalFieldConfig, |
| 53 | MultiModalKwargsItems, |
| 54 | ) |
| 55 | from vllm.multimodal.parse import ImageProcessorItems, ImageSize, MultiModalDataItems |
| 56 | from vllm.multimodal.processing import ( |
| 57 | BaseDummyInputsBuilder, |
| 58 | BaseMultiModalProcessor, |
| 59 | BaseProcessingInfo, |
| 60 | PromptIndexTargets, |
| 61 | PromptInsertion, |
| 62 | PromptUpdate, |
| 63 | PromptUpdateDetails, |
| 64 | cached_encode, |
| 65 | ) |
| 66 | from vllm.sequence import IntermediateTensors |
| 67 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 68 | |
| 69 | from .interfaces import ( |
| 70 | MultiModalEmbeddings, |
| 71 | SupportsLoRA, |
| 72 | SupportsMultiModal, |
| 73 | SupportsPP, |
| 74 | SupportsQuant, |
| 75 | ) |
| 76 | from .utils import ( |
| 77 | AutoWeightsLoader, |
| 78 | WeightsMapper, |
| 79 | make_empty_intermediate_tensors_factory, |
| 80 | make_layers, |
| 81 | maybe_prefix, |
| 82 | ) |
| 83 | |
| 84 | # TODO: hard-coded for now. Consider making it configurable. |
| 85 | VIT_LAYERS = [-2, -9] |
| 86 | NUM_PREFIX_TOKENS = 1 |
| 87 | ADDITIONAL_VOCAB_SIZE = 128 |
| 88 | IMAGE_PATCH_TOKEN = "<im_patch>" |
| 89 | IM_COL_TOKEN = "<im_col>" |
| 90 | IM_START_TOKEN = "<im_start>" |
| 91 | IM_END_TOKEN = "<im_end>" |
| 92 | POOLING_SIZE = 2 |
| 93 | |
| 94 | |
| 95 | class MolmoImageInputs(TensorSchema): |
| 96 | """Dimensions: |
| 97 | - bn: Batch size * number of images |
| 98 | - bnc: Batch size * number of images * number of crops (dynamic) |
| 99 | - np: Number of patches |
| 100 | - tp: Token sequence positions |
| 101 | - pd: Patch dimension |
| 102 | """ |
| 103 | |
| 104 | images: Annotated[torch.Tensor, TensorShape("bnc", "np", "pd")] |
| 105 | |
| 106 | image_masks: Annotated[torch.Tensor | None, TensorShape("bnc", "np")] |
| 107 | |
| 108 | image_input_idx: Annotated[torch.Tensor, TensorShape("bnc", "tp")] |
| 109 | """An index tensor that maps image features to their corresponding patch tokens.""" |
| 110 | |
| 111 | num_crops: Annotated[torch.Tensor, TensorShape("bn")] |
| 112 | |
| 113 | |
| 114 | @dataclass |
| 115 | class VisionBackboneConfig: |
| 116 | image_default_input_size: tuple[int, int] = (336, 336) |
| 117 | image_patch_size: int = 14 |
| 118 | image_pos_patch_size: int = 14 |
| 119 | image_emb_dim: int = 1024 |
| 120 | image_num_heads: int = 16 |
| 121 | image_num_key_value_heads: int = 16 |
| 122 | image_num_layers: int = 23 |
| 123 | image_mlp_dim: int = 4096 |
| 124 | image_mlp_activations: str = "quick_gelu" |
| 125 | image_num_pos: int = 577 |
| 126 | image_norm_eps: float = 1e-5 |
| 127 | |
| 128 | def __post_init__(self): |
| 129 | self.image_default_input_size = tuple(self.image_default_input_size) # type: ignore[assignment] |
| 130 | |
| 131 | @property |
| 132 | def image_num_patch(self): |
| 133 | h, w = self.image_default_input_size |
| 134 | return h // self.image_patch_size, w // self.image_patch_size |
| 135 | |
| 136 | |
| 137 | class ViTMLP(nn.Module): |
| 138 | """MLP used in Vision Transformer.""" |
| 139 | |
| 140 | def __init__( |
| 141 | self, |
| 142 | config: VisionBackboneConfig, |
| 143 | quant_config: QuantizationConfig | None = None, |
| 144 | prefix: str = "", |
| 145 | ): |
| 146 | super().__init__() |
| 147 | self.w1 = ColumnParallelLinear( |
| 148 | config.image_emb_dim, |
| 149 | config.image_mlp_dim, |
| 150 | bias=True, |
| 151 | quant_config=quant_config, |
| 152 | prefix=f"{prefix}.w1", |
| 153 | ) |
| 154 | # Activation function. |
| 155 | assert config.image_mlp_activations == "quick_gelu" |
| 156 | self.act = QuickGELU() |
| 157 | self.w2 = RowParallelLinear( |
| 158 | config.image_mlp_dim, |
| 159 | config.image_emb_dim, |
| 160 | bias=True, |
| 161 | quant_config=quant_config, |
| 162 | prefix=f"{prefix}.w2", |
| 163 | ) |
| 164 | |
| 165 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 166 | x, _ = self.w1(x) |
| 167 | x = self.act(x) |
| 168 | x, _ = self.w2(x) |
| 169 | return x |
| 170 | |
| 171 | |
| 172 | class MultiHeadDotProductAttention(nn.Module): |
| 173 | """Multi-head attention used in Vision Transformer.""" |
| 174 | |
| 175 | def __init__( |
| 176 | self, |
| 177 | config: VisionBackboneConfig, |
| 178 | use_bias: bool = True, |
| 179 | nlayers: int = 1, |
| 180 | quant_config: QuantizationConfig | None = None, |
| 181 | prefix: str = "", |
| 182 | ): |
| 183 | super().__init__() |
| 184 | |
| 185 | self.hidden_size = config.image_emb_dim |
| 186 | self.total_num_heads = config.image_num_heads |
| 187 | tp_size = get_tensor_model_parallel_world_size() |
| 188 | |
| 189 | assert self.hidden_size % self.total_num_heads == 0 |
| 190 | assert self.total_num_heads % tp_size == 0 |
| 191 | |
| 192 | self.num_heads = self.total_num_heads // tp_size |
| 193 | self.head_dim = self.hidden_size // self.total_num_heads |
| 194 | |
| 195 | self.total_num_kv_heads = config.image_num_key_value_heads |
| 196 | if self.total_num_kv_heads >= tp_size: |
| 197 | assert self.total_num_kv_heads % tp_size == 0 |
| 198 | else: |
| 199 | assert tp_size % self.total_num_kv_heads == 0 |
| 200 | |
| 201 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 202 | |
| 203 | self.wq = ColumnParallelLinear( |
| 204 | nlayers * self.hidden_size, |
| 205 | self.total_num_heads * self.head_dim, |
| 206 | bias=use_bias, |
| 207 | quant_config=quant_config, |
| 208 | prefix=f"{prefix}.wq", |
| 209 | ) |
| 210 | self.wk = ColumnParallelLinear( |
| 211 | nlayers * self.hidden_size, |
| 212 | self.total_num_kv_heads * self.head_dim, |
| 213 | bias=use_bias, |
| 214 | quant_config=quant_config, |
| 215 | prefix=f"{prefix}.wk", |
| 216 | ) |
| 217 | self.wv = ColumnParallelLinear( |
| 218 | nlayers * self.hidden_size, |
| 219 | self.total_num_kv_heads * self.head_dim, |
| 220 | bias=use_bias, |
| 221 | quant_config=quant_config, |
| 222 | prefix=f"{prefix}.wv", |
| 223 | ) |
| 224 | self.wo = RowParallelLinear( |
| 225 | self.total_num_heads * self.head_dim, |
| 226 | self.hidden_size, |
| 227 | bias=use_bias, |
| 228 | quant_config=quant_config, |
| 229 | prefix=f"{prefix}.wo", |
| 230 | ) |
| 231 | |
| 232 | self.scale = self.head_dim**-0.5 |
| 233 | self.attn = MMEncoderAttention( |
| 234 | self.num_heads, |
| 235 | self.head_dim, |
| 236 | self.scale, |
| 237 | num_kv_heads=self.num_kv_heads, |
| 238 | prefix=f"{prefix}.attn", |
| 239 | ) |
| 240 | |
| 241 | def forward( |
| 242 | self, inputs_q: torch.Tensor, inputs_kv: torch.Tensor | None = None |
| 243 | ) -> torch.Tensor: |
| 244 | if inputs_kv is not None: |
| 245 | inputs_k = inputs_kv |
| 246 | inputs_v = inputs_kv |
| 247 | else: |
| 248 | inputs_k = inputs_q |
| 249 | inputs_v = inputs_q |
| 250 | |
| 251 | xq, _ = self.wq(inputs_q) |
| 252 | xk, _ = self.wk(inputs_k) |
| 253 | xv, _ = self.wv(inputs_v) |
| 254 | |
| 255 | output = self.attn(xq, xk, xv) |
| 256 | output, _ = self.wo(output) |
| 257 | |
| 258 | return output |
| 259 | |
| 260 | |
| 261 | class ResidualAttentionBlock(nn.Module): |
| 262 | """Residual attention block used in Vision Transformer.""" |
| 263 | |
| 264 | def __init__( |
| 265 | self, |
| 266 | config: VisionBackboneConfig, |
| 267 | quant_config: QuantizationConfig | None = None, |
| 268 | prefix: str = "", |
| 269 | ): |
| 270 | super().__init__() |
| 271 | self.attention = MultiHeadDotProductAttention( |
| 272 | config, quant_config=quant_config, prefix=f"{prefix}.attention" |
| 273 | ) |
| 274 | self.feed_forward = ViTMLP( |
| 275 | config, quant_config, prefix=f"{prefix}.feed_forward" |
| 276 | ) |
| 277 | self.attention_norm = nn.LayerNorm( |
| 278 | config.image_emb_dim, |
| 279 | eps=config.image_norm_eps, |
| 280 | ) |
| 281 | self.ffn_norm = nn.LayerNorm( |
| 282 | config.image_emb_dim, |
| 283 | eps=config.image_norm_eps, |
| 284 | ) |
| 285 | |
| 286 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 287 | x = x + self.attention(self.attention_norm(x)) |
| 288 | x = x + self.feed_forward(self.ffn_norm(x)) |
| 289 | return x |
| 290 | |
| 291 | |
| 292 | class BlockCollection(nn.Module): |
| 293 | """Collection of residual attention blocks used in Vision Transformer.""" |
| 294 | |
| 295 | def __init__( |
| 296 | self, |
| 297 | config: VisionBackboneConfig, |
| 298 | quant_config: QuantizationConfig | None = None, |
| 299 | prefix: str = "", |
| 300 | ): |
| 301 | super().__init__() |
| 302 | self.resblocks = nn.ModuleList( |
| 303 | [ |
| 304 | ResidualAttentionBlock( |
| 305 | config, quant_config, prefix=f"{prefix}.resblocks.{i}" |
| 306 | ) |
| 307 | for i in range(config.image_num_layers) |
| 308 | ] |
| 309 | ) |
| 310 | |
| 311 | def forward(self, x: torch.Tensor) -> list[torch.Tensor]: |
| 312 | hidden_states = [] |
| 313 | for r in self.resblocks: |
| 314 | x = r(x) |
| 315 | hidden_states.append(x) |
| 316 | return hidden_states |
| 317 | |
| 318 | |
| 319 | def _expand_token(token: torch.Tensor, batch_size: int) -> torch.Tensor: |
| 320 | return token.view(1, 1, -1).expand(batch_size, -1, -1) |
| 321 | |
| 322 | |
| 323 | class VisionTransformer(nn.Module): |
| 324 | """Vision Transformer used in Vision Backbone.""" |
| 325 | |
| 326 | def __init__( |
| 327 | self, |
| 328 | config: VisionBackboneConfig, |
| 329 | quant_config: QuantizationConfig | None = None, |
| 330 | prefix: str = "", |
| 331 | ): |
| 332 | super().__init__() |
| 333 | scale = config.image_emb_dim**-0.5 |
| 334 | self.patch_num: tuple[int, int] = config.image_num_patch |
| 335 | self.class_embedding = nn.Parameter(torch.randn(config.image_emb_dim) * scale) |
| 336 | self.num_prefix_tokens: int = NUM_PREFIX_TOKENS |
| 337 | self.positional_embedding = nn.Parameter( |
| 338 | torch.randn(config.image_num_pos, config.image_emb_dim) * scale |
| 339 | ) |
| 340 | image_patch_size = config.image_patch_size |
| 341 | self.patch_embedding = nn.Linear( |
| 342 | image_patch_size * image_patch_size * 3, |
| 343 | config.image_emb_dim, |
| 344 | bias=False, |
| 345 | ) |
| 346 | self.pre_ln = nn.LayerNorm(config.image_emb_dim, eps=config.image_norm_eps) |
| 347 | self.transformer = BlockCollection( |
| 348 | config, quant_config, prefix=f"{prefix}.transformer" |
| 349 | ) |
| 350 | |
| 351 | def add_pos_emb(self, x: torch.Tensor, patch_num: tuple[int, int]) -> torch.Tensor: |
| 352 | cls_emb = self.positional_embedding[0:1] |
| 353 | pos_emb = self.positional_embedding[1:] |
| 354 | |
| 355 | pos_emb = pos_emb.reshape( |
| 356 | ( |
| 357 | int(math.sqrt(pos_emb.shape[0])), |
| 358 | int(math.sqrt(pos_emb.shape[0])), |
| 359 | pos_emb.shape[1], |
| 360 | ) |
| 361 | ) |
| 362 | |
| 363 | (patch_num_0, patch_num_1) = patch_num |
| 364 | |
| 365 | if pos_emb.shape[0] != patch_num_0 or pos_emb.shape[1] != patch_num_1: |
| 366 | # from https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py |
| 367 | pos_emb = pos_emb.unsqueeze(0).permute(0, 3, 1, 2) |
| 368 | pos_emb = F.interpolate( |
| 369 | pos_emb, |
| 370 | size=(patch_num_0, patch_num_1), |
| 371 | mode="bicubic", |
| 372 | align_corners=False, |
| 373 | antialias=True, |
| 374 | ) |
| 375 | pos_emb = pos_emb.permute(0, 2, 3, 1).squeeze(0) |
| 376 | |
| 377 | pos_emb = pos_emb.reshape(-1, pos_emb.shape[-1]) |
| 378 | x = x + torch.cat([cls_emb[None, :, :], pos_emb[None, :, :]], dim=1).to(x.dtype) |
| 379 | return x |
| 380 | |
| 381 | def forward( |
| 382 | self, x: torch.Tensor, patch_num: tuple[int, int] | None = None |
| 383 | ) -> list[torch.Tensor]: |
| 384 | """: param x: (batch_size, num_patch, n_pixels).""" |
| 385 | if patch_num is None: |
| 386 | patch_num = self.patch_num |
| 387 | B, N, D = x.shape |
| 388 | |
| 389 | x = self.patch_embedding(x) |
| 390 | |
| 391 | # class embeddings and positional embeddings |
| 392 | x = torch.cat( |
| 393 | [_expand_token(self.class_embedding, x.shape[0]).to(x.dtype), x], dim=1 |
| 394 | ) |
| 395 | x = self.add_pos_emb(x, patch_num) |
| 396 | |
| 397 | x = self.pre_ln(x) |
| 398 | |
| 399 | hidden_states = self.transformer(x) |
| 400 | return hidden_states |
| 401 | |
| 402 | |
| 403 | class MolmoAttention(nn.Module): |
| 404 | """Molmo's LLM attention.""" |
| 405 | |
| 406 | def __init__( |
| 407 | self, |
| 408 | config: PreTrainedConfig, |
| 409 | cache_config: CacheConfig | None = None, |
| 410 | quant_config: QuantizationConfig | None = None, |
| 411 | prefix: str = "", |
| 412 | ) -> None: |
| 413 | super().__init__() |
| 414 | self.hidden_size = config.hidden_size |
| 415 | self.tp_size = get_tensor_model_parallel_world_size() |
| 416 | self.total_num_heads = config.num_attention_heads |
| 417 | |
| 418 | assert self.hidden_size % self.total_num_heads == 0 |
| 419 | assert self.total_num_heads % self.tp_size == 0 |
| 420 | |
| 421 | self.num_heads = self.total_num_heads // self.tp_size |
| 422 | self.total_num_kv_heads = config.num_key_value_heads or self.total_num_heads |
| 423 | if self.total_num_kv_heads >= self.tp_size: |
| 424 | assert self.total_num_kv_heads % self.tp_size == 0 |
| 425 | else: |
| 426 | assert self.tp_size % self.total_num_kv_heads == 0 |
| 427 | |
| 428 | self.num_kv_heads = max(1, self.total_num_kv_heads // self.tp_size) |
| 429 | self.head_dim = self.hidden_size // self.total_num_heads |
| 430 | self.q_size = self.num_heads * self.head_dim |
| 431 | self.kv_size = self.num_kv_heads * self.head_dim |
| 432 | self.max_position_embeddings = config.max_position_embeddings |
| 433 | |
| 434 | # Attention input projection. Projects x -> (q, k, v) |
| 435 | self.qkv_proj = QKVParallelLinear( |
| 436 | self.hidden_size, |
| 437 | self.head_dim, |
| 438 | self.total_num_heads, |
| 439 | self.total_num_kv_heads, |
| 440 | bias=config.qkv_bias, |
| 441 | quant_config=quant_config, |
| 442 | prefix=f"{prefix}.qkv_proj", |
| 443 | ) |
| 444 | |
| 445 | self.tp_rank: int | None = None |
| 446 | self.k_norm: nn.Module | None = None |
| 447 | self.q_norm: nn.Module | None = None |
| 448 | if config.attention_layer_norm: |
| 449 | self.tp_rank = get_tensor_model_parallel_rank() |
| 450 | self.k_norm = RMSNorm( |
| 451 | self.total_num_kv_heads * self.head_dim, eps=config.layer_norm_eps |
| 452 | ) |
| 453 | self.q_norm = RMSNorm(config.hidden_size, eps=config.layer_norm_eps) |
| 454 | |
| 455 | # Rotary embeddings. |
| 456 | self.rotary_emb = get_rope( |
| 457 | self.head_dim, |
| 458 | max_position=self.max_position_embeddings, |
| 459 | rope_parameters=config.rope_parameters, |
| 460 | ) |
| 461 | self.scaling = self.head_dim**-0.5 |
| 462 | self.attn = Attention( |
| 463 | self.num_heads, |
| 464 | self.head_dim, |
| 465 | self.scaling, |
| 466 | num_kv_heads=self.num_kv_heads, |
| 467 | cache_config=cache_config, |
| 468 | quant_config=quant_config, |
| 469 | prefix=f"{prefix}.attn", |
| 470 | ) |
| 471 | |
| 472 | # Attention output projection. |
| 473 | self.o_proj = RowParallelLinear( |
| 474 | self.total_num_heads * self.head_dim, |
| 475 | self.hidden_size, |
| 476 | bias=False, |
| 477 | quant_config=quant_config, |
| 478 | prefix=f"{prefix}.o_proj", |
| 479 | ) |
| 480 | |
| 481 | def _apply_qk_norm( |
| 482 | self, q: torch.Tensor, k: torch.Tensor |
| 483 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 484 | assert self.q_norm is not None |
| 485 | assert self.k_norm is not None |
| 486 | if self.tp_size > 1: |
| 487 | q = tensor_model_parallel_all_gather(q.contiguous()) |
| 488 | k = tensor_model_parallel_all_gather(k.contiguous()) |
| 489 | q = self.q_norm(q) |
| 490 | k = self.k_norm(k) |
| 491 | if self.tp_size > 1: |
| 492 | assert self.tp_rank is not None |
| 493 | splitter = partial(split_tensor_along_last_dim, num_partitions=self.tp_size) |
| 494 | q = splitter(q)[self.tp_rank] |
| 495 | k = splitter(k)[self.tp_rank] |
| 496 | return q, k |
| 497 | |
| 498 | def forward( |
| 499 | self, |
| 500 | positions: torch.Tensor, |
| 501 | hidden_states: torch.Tensor, |
| 502 | ) -> torch.Tensor: |
| 503 | qkv, _ = self.qkv_proj(hidden_states) |
| 504 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 505 | if self.q_norm is not None and self.k_norm is not None: |
| 506 | q, k = self._apply_qk_norm(q, k) |
| 507 | q, k = self.rotary_emb(positions, q, k) |
| 508 | attn_output = self.attn(q, k, v) |
| 509 | output, _ = self.o_proj(attn_output) |
| 510 | return output |
| 511 | |
| 512 | |
| 513 | class LanguageModelMLP(nn.Module): |
| 514 | """Molmo's LLM mlp.""" |
| 515 | |
| 516 | def __init__( |
| 517 | self, |
| 518 | config: PreTrainedConfig, |
| 519 | input_dim: int | None = None, |
| 520 | quant_config: QuantizationConfig | None = None, |
| 521 | prefix: str = "", |
| 522 | ) -> None: |
| 523 | super().__init__() |
| 524 | self.hidden_size = config.hidden_size |
| 525 | self.intermediate_size = config.intermediate_size // 2 |
| 526 | |
| 527 | self.gate_up_proj = MergedColumnParallelLinear( |
| 528 | input_dim or self.hidden_size, |
| 529 | [self.intermediate_size] * 2, |
| 530 | bias=False, |
| 531 | quant_config=quant_config, |
| 532 | prefix=f"{prefix}.gate_up_proj", |
| 533 | ) |
| 534 | # Activation function. |
| 535 | self.act_fn = MulAndSilu() |
| 536 | # Feed-forward output projection. |
| 537 | self.down_proj = RowParallelLinear( |
| 538 | self.intermediate_size, |
| 539 | self.hidden_size, |
| 540 | bias=False, |
| 541 | quant_config=quant_config, |
| 542 | prefix=f"{prefix}.down_proj", |
| 543 | ) |
| 544 | |
| 545 | def forward( |
| 546 | self, |
| 547 | x: torch.Tensor, |
| 548 | ) -> torch.Tensor: |
| 549 | gate_up, _ = self.gate_up_proj(x) |
| 550 | x = self.act_fn(gate_up) |
| 551 | x, _ = self.down_proj(x) |
| 552 | return x |
| 553 | |
| 554 | |
| 555 | class ImageProjectorMLP(nn.Module): |
| 556 | """Molmo's image_projector mlp.""" |
| 557 | |
| 558 | def __init__( |
| 559 | self, |
| 560 | config: PreTrainedConfig, |
| 561 | input_dim: int | None = None, |
| 562 | quant_config: QuantizationConfig | None = None, |
| 563 | prefix: str = "", |
| 564 | ) -> None: |
| 565 | super().__init__() |
| 566 | self.hidden_size = config.hidden_size |
| 567 | self.intermediate_size = config.intermediate_size // 2 |
| 568 | |
| 569 | self.merged_linear = MergedColumnParallelLinear( |
| 570 | input_dim or self.hidden_size, |
| 571 | [self.intermediate_size] * 2, |
| 572 | bias=False, |
| 573 | quant_config=quant_config, |
| 574 | prefix=f"{prefix}.merged_linear", |
| 575 | ) |
| 576 | # Activation function. |
| 577 | self.act_fn = SiluAndMul() |
| 578 | |
| 579 | # Feed-forward output projection. |
| 580 | self.down_proj = RowParallelLinear( |
| 581 | self.intermediate_size, |
| 582 | self.hidden_size, |
| 583 | bias=False, |
| 584 | quant_config=quant_config, |
| 585 | prefix=f"{prefix}.down_proj", |
| 586 | ) |
| 587 | |
| 588 | def forward( |
| 589 | self, |
| 590 | x: torch.Tensor, |
| 591 | ) -> torch.Tensor: |
| 592 | gate_up, _ = self.merged_linear(x) |
| 593 | x = self.act_fn(gate_up) |
| 594 | x, _ = self.down_proj(x) |
| 595 | return x |
| 596 | |
| 597 | |
| 598 | class MolmoDecoderLayer(nn.Module): |
| 599 | def __init__( |
| 600 | self, |
| 601 | config: PreTrainedConfig, |
| 602 | cache_config: CacheConfig | None = None, |
| 603 | quant_config: QuantizationConfig | None = None, |
| 604 | prefix: str = "", |
| 605 | ) -> None: |
| 606 | super().__init__() |
| 607 | # Attention block. |
| 608 | self.self_attn = MolmoAttention( |
| 609 | config, cache_config, quant_config, prefix=f"{prefix}.self_attn" |
| 610 | ) |
| 611 | |
| 612 | # MLP block. |
| 613 | self.mlp = LanguageModelMLP( |
| 614 | config, quant_config=quant_config, prefix=f"{prefix}.mlp" |
| 615 | ) |
| 616 | |
| 617 | # LayerNorm |
| 618 | assert config.layer_norm_type == "rms" |
| 619 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.layer_norm_eps) |
| 620 | self.post_attention_layernorm = RMSNorm( |
| 621 | config.hidden_size, eps=config.layer_norm_eps |
| 622 | ) |
| 623 | |
| 624 | def forward( |
| 625 | self, |
| 626 | positions: torch.Tensor, |
| 627 | hidden_states: torch.Tensor, |
| 628 | residual: torch.Tensor | None, |
| 629 | ) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]: |
| 630 | # Self Attention |
| 631 | if residual is None: |
| 632 | residual = hidden_states |
| 633 | hidden_states = self.input_layernorm(hidden_states) |
| 634 | else: |
| 635 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 636 | hidden_states = self.self_attn( |
| 637 | positions=positions, |
| 638 | hidden_states=hidden_states, |
| 639 | ) |
| 640 | |
| 641 | hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| 642 | hidden_states = self.mlp(hidden_states) |
| 643 | return hidden_states, residual |
| 644 | |
| 645 | |
| 646 | class MolmoDecoderNormAfterLayer(MolmoDecoderLayer): |
| 647 | def forward( |
| 648 | self, |
| 649 | positions: torch.Tensor, |
| 650 | hidden_states: torch.Tensor, |
| 651 | residual: torch.Tensor | None, |
| 652 | ) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]: |
| 653 | # Self Attention |
| 654 | residual = hidden_states |
| 655 | hidden_states = self.self_attn( |
| 656 | positions=positions, |
| 657 | hidden_states=hidden_states, |
| 658 | ) |
| 659 | |
| 660 | hidden_states = self.input_layernorm(hidden_states) |
| 661 | hidden_states = hidden_states + residual |
| 662 | residual = hidden_states |
| 663 | |
| 664 | hidden_states = self.mlp(hidden_states) |
| 665 | hidden_states = self.post_attention_layernorm(hidden_states) |
| 666 | hidden_states = hidden_states + residual |
| 667 | residual = None |
| 668 | return hidden_states, residual |
| 669 | |
| 670 | |
| 671 | class MolmoVisionBackbone(nn.Module, SupportsQuant): |
| 672 | packed_modules_mapping = {"merged_linear": ["gate_proj", "up_proj"]} |
| 673 | |
| 674 | hf_to_vllm_mapper = WeightsMapper( |
| 675 | orig_to_new_stacked={ |
| 676 | # image_projector gate_up merge |
| 677 | "gate_proj": ("merged_linear", 0), |
| 678 | "up_proj": ("merged_linear", 1), |
| 679 | } |
| 680 | ) |
| 681 | |
| 682 | def __init__( |
| 683 | self, |
| 684 | config: PreTrainedConfig, |
| 685 | vision_config: VisionBackboneConfig, |
| 686 | quant_config: QuantizationConfig | None = None, |
| 687 | prefix: str = "", |
| 688 | ) -> None: |
| 689 | super().__init__() |
| 690 | self.vit_layers = VIT_LAYERS |
| 691 | self.image_num_patch = vision_config.image_num_patch |
| 692 | self.llm_patches_per_crop = ( |
| 693 | (self.image_num_patch[0] + 1) // POOLING_SIZE, |
| 694 | (self.image_num_patch[1] + 1) // POOLING_SIZE, |
| 695 | ) |
| 696 | self.image_vit = VisionTransformer( |
| 697 | vision_config, quant_config=quant_config, prefix=f"{prefix}.image_vit" |
| 698 | ) |
| 699 | self.num_prefix_tokens = self.image_vit.num_prefix_tokens |
| 700 | assert self.num_prefix_tokens in {0, 1}, ( |
| 701 | "Only 0 or 1 prefix tokens are supported" |
| 702 | ) |
| 703 | self.image_pooling_2d = MultiHeadDotProductAttention( |
| 704 | vision_config, |
| 705 | nlayers=len(self.vit_layers), |
| 706 | quant_config=quant_config, |
| 707 | prefix=f"{prefix}.image_pooling_2d", |
| 708 | ) |
| 709 | self.image_projector = ImageProjectorMLP( |
| 710 | config, |
| 711 | input_dim=vision_config.image_emb_dim, |
| 712 | quant_config=quant_config, |
| 713 | prefix=f"{prefix}.image_projector", |
| 714 | ) |
| 715 | |
| 716 | image_dim = vision_config.image_emb_dim * len(self.vit_layers) |
| 717 | self.pad_embed = nn.Parameter(torch.zeros((2, image_dim))) |
| 718 | |
| 719 | @property |
| 720 | def dtype(self) -> torch.dtype: |
| 721 | return self.image_vit.patch_embedding.weight.dtype |
| 722 | |
| 723 | @property |
| 724 | def device(self) -> torch.device: |
| 725 | return self.image_vit.patch_embedding.weight.device |
| 726 | |
| 727 | def encode_image(self, images: torch.Tensor) -> torch.Tensor: |
| 728 | """: param images: (batch_size, num_crops, num_patch, n_pixels).""" |
| 729 | B, T, N, D = images.shape |
| 730 | |
| 731 | mask = ~torch.all(images.view(B * T, N, D) == -1, dim=(1, 2), keepdim=True) |
| 732 | |
| 733 | images = images.view(B * T, N, D) |
| 734 | image_features = self.image_vit(images) |
| 735 | |
| 736 | if self.vit_layers is not None: |
| 737 | features = [] |
| 738 | for layer in self.vit_layers: |
| 739 | features.append(image_features[layer]) |
| 740 | image_features = torch.cat(features, dim=-1) |
| 741 | else: |
| 742 | image_features = image_features[-1] |
| 743 | |
| 744 | if self.num_prefix_tokens > 0: |
| 745 | image_features = image_features[:, 1:] |
| 746 | |
| 747 | image_features = image_features * mask |
| 748 | image_features = image_features.view(B, T, N, -1) |
| 749 | |
| 750 | return image_features |
| 751 | |
| 752 | def forward( |
| 753 | self, |
| 754 | images: torch.Tensor, |
| 755 | image_masks: torch.Tensor, |
| 756 | ) -> torch.Tensor: |
| 757 | # image_features: (batch_size, num_crops(=num_image), num_patch, nximage_emb_dim) # noqa: E501 |
| 758 | batch_size, num_image = images.shape[:2] |
| 759 | images = images.to(device=self.device, dtype=self.dtype) |
| 760 | image_features = self.encode_image(images) |
| 761 | |
| 762 | og_dtype = image_features.dtype |
| 763 | assert image_masks is not None |
| 764 | pad_embed = self.pad_embed[:, None, None, None, :] |
| 765 | all_pad = image_masks == 0 |
| 766 | partial_pad = torch.logical_and(image_masks < 1, torch.logical_not(all_pad)).to( |
| 767 | dtype=torch.float32 |
| 768 | ) |
| 769 | all_pad = all_pad.to(dtype=torch.float32) |
| 770 | image_features = image_features + pad_embed[0] * torch.unsqueeze(all_pad, -1) |
| 771 | image_features = image_features + pad_embed[1] * torch.unsqueeze( |
| 772 | partial_pad, -1 |
| 773 | ) |
| 774 | |
| 775 | image_features = image_features.to(og_dtype) |
| 776 | |
| 777 | image_features = image_features.reshape( |
| 778 | (batch_size, num_image) + self.image_num_patch + (-1,), |
| 779 | ) |
| 780 | |
| 781 | if missing_w := self.image_num_patch[0] % POOLING_SIZE: |
| 782 | # Padding for image pooling (see below) |
| 783 | image_features = F.pad( |
| 784 | image_features, |
| 785 | (0, 0, 0, missing_w, 0, missing_w, 0, 0, 0, 0), |
| 786 | ) |
| 787 | |
| 788 | # image pooling |
| 789 | image_features = rearrange( |
| 790 | image_features, |
| 791 | "b n (h dh) (w dw) c -> (b n h w) (dh dw) c", |
| 792 | dh=POOLING_SIZE, |
| 793 | dw=POOLING_SIZE, |
| 794 | ) |
| 795 | |
| 796 | query = image_features.mean(-2, keepdim=True) |
| 797 | image_features = self.image_pooling_2d(query, image_features) |
| 798 | |
| 799 | h, w = self.llm_patches_per_crop |
| 800 | image_features = image_features.view(batch_size, num_image, h * w, -1) |
| 801 | |
| 802 | image_features = self.image_projector(image_features) |
| 803 | |
| 804 | # image_features: (batch_size, num_image, num_patch, d_model) |
| 805 | return image_features |
| 806 | |
| 807 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 808 | loader = AutoWeightsLoader(self) |
| 809 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 810 | |
| 811 | |
| 812 | @support_torch_compile |
| 813 | class MolmoModel(nn.Module, SupportsQuant): |
| 814 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 815 | super().__init__() |
| 816 | |
| 817 | config = vllm_config.model_config.hf_config |
| 818 | cache_config = vllm_config.cache_config |
| 819 | quant_config = vllm_config.quant_config |
| 820 | |
| 821 | self.config = config |
| 822 | |
| 823 | self.embedding_size = config.embedding_size or config.vocab_size |
| 824 | self.embedding_size += ADDITIONAL_VOCAB_SIZE |
| 825 | self.embed_tokens = VocabParallelEmbedding( |
| 826 | self.embedding_size, |
| 827 | config.hidden_size, |
| 828 | quant_config=quant_config, |
| 829 | ) |
| 830 | |
| 831 | decoder_layer = ( |
| 832 | MolmoDecoderNormAfterLayer if config.norm_after else MolmoDecoderLayer |
| 833 | ) |
| 834 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 835 | config.num_hidden_layers, |
| 836 | lambda prefix: decoder_layer( |
| 837 | config, cache_config, quant_config, prefix=prefix |
| 838 | ), |
| 839 | prefix=f"{prefix}.layers", |
| 840 | ) |
| 841 | |
| 842 | assert config.layer_norm_type == "rms" |
| 843 | self.norm = RMSNorm(config.hidden_size, config.layer_norm_eps) |
| 844 | |
| 845 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 846 | ["hidden_states", "residual"], config.hidden_size |
| 847 | ) |
| 848 | |
| 849 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 850 | return self.embed_tokens(input_ids) |
| 851 | |
| 852 | def forward( |
| 853 | self, |
| 854 | input_ids: torch.Tensor | None, |
| 855 | positions: torch.Tensor, |
| 856 | intermediate_tensors: IntermediateTensors | None = None, |
| 857 | inputs_embeds: torch.Tensor | None = None, |
| 858 | ) -> torch.Tensor: |
| 859 | if get_pp_group().is_first_rank: |
| 860 | if inputs_embeds is not None: |
| 861 | hidden_states = inputs_embeds |
| 862 | else: |
| 863 | hidden_states = self.embed_tokens(input_ids) |
| 864 | residual = None |
| 865 | else: |
| 866 | assert intermediate_tensors is not None |
| 867 | hidden_states = intermediate_tensors["hidden_states"] |
| 868 | residual = intermediate_tensors["residual"] |
| 869 | |
| 870 | # Apply blocks one-by-one. |
| 871 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 872 | hidden_states, residual = layer( |
| 873 | positions, |
| 874 | hidden_states, |
| 875 | residual, |
| 876 | ) |
| 877 | if not get_pp_group().is_last_rank: |
| 878 | return IntermediateTensors( |
| 879 | {"hidden_states": hidden_states, "residual": residual} |
| 880 | ) |
| 881 | if residual is not None: |
| 882 | hidden_states, _ = self.norm(hidden_states, residual) |
| 883 | else: |
| 884 | hidden_states = self.norm(hidden_states) |
| 885 | return hidden_states |
| 886 | |
| 887 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 888 | return AutoWeightsLoader(self).load_weights(weights) |
| 889 | |
| 890 | |
| 891 | def _lowest_multiple(x: int, k: int) -> int: |
| 892 | return (x // k) * k |
| 893 | |
| 894 | |
| 895 | def get_num_patches( |
| 896 | num_tiles: int, |
| 897 | *, |
| 898 | crop_patches: int, |
| 899 | left_margin: int, |
| 900 | right_margin: int, |
| 901 | pooling_size: int, |
| 902 | ) -> int: |
| 903 | if num_tiles == 1: |
| 904 | return _lowest_multiple(crop_patches + pooling_size - 1, pooling_size) |
| 905 | |
| 906 | crop_window_patches = crop_patches - (left_margin + right_margin) |
| 907 | |
| 908 | left_num = _lowest_multiple( |
| 909 | crop_window_patches + left_margin + pooling_size - 1, |
| 910 | pooling_size, |
| 911 | ) |
| 912 | middle_num = _lowest_multiple( |
| 913 | crop_window_patches + pooling_size - 1, |
| 914 | pooling_size, |
| 915 | ) |
| 916 | right_num = _lowest_multiple( |
| 917 | crop_window_patches + right_margin + pooling_size - 1, |
| 918 | pooling_size, |
| 919 | ) |
| 920 | |
| 921 | return left_num + (num_tiles - 2) * middle_num + right_num |
| 922 | |
| 923 | |
| 924 | def get_patches_grid_size( |
| 925 | *, |
| 926 | tiling_h: int, |
| 927 | tiling_w: int, |
| 928 | crop_patches: int, |
| 929 | left_margin: int, |
| 930 | right_margin: int, |
| 931 | pooling_size: int, |
| 932 | ) -> tuple[int, int]: |
| 933 | nrows = get_num_patches( |
| 934 | tiling_h, |
| 935 | crop_patches=crop_patches, |
| 936 | left_margin=left_margin, |
| 937 | right_margin=right_margin, |
| 938 | pooling_size=pooling_size, |
| 939 | ) |
| 940 | ncols = get_num_patches( |
| 941 | tiling_w, |
| 942 | crop_patches=crop_patches, |
| 943 | left_margin=left_margin, |
| 944 | right_margin=right_margin, |
| 945 | pooling_size=pooling_size, |
| 946 | ) |
| 947 | |
| 948 | return nrows, ncols |
| 949 | |
| 950 | |
| 951 | def get_candidate_tilings(max_num: int) -> list[tuple[int, int]]: |
| 952 | tilings = [ |
| 953 | (i, j) |
| 954 | for i in range(1, max_num + 1) |
| 955 | for j in range(1, max_num + 1) |
| 956 | if i * j <= max_num |
| 957 | ] |
| 958 | return sorted(tilings, key=lambda x: x[0] * x[1]) |
| 959 | |
| 960 | |
| 961 | def select_tiling( |
| 962 | *, |
| 963 | height: int, |
| 964 | width: int, |
| 965 | patch_size: int, |
| 966 | max_num_patches: int, |
| 967 | ): |
| 968 | tilings = get_candidate_tilings(max_num_patches) |
| 969 | candidate_tilings = np.array(tilings, dtype=np.int32) |
| 970 | candidate_resolutions = candidate_tilings * patch_size |
| 971 | |
| 972 | original_size = np.array([height, width], dtype=np.float32) |
| 973 | required_scale_d = candidate_resolutions.astype(np.float32) / original_size |
| 974 | required_scale = required_scale_d.min(axis=-1, keepdims=True) |
| 975 | |
| 976 | if (required_scale < 1).all(): |
| 977 | ix = required_scale.argmax() |
| 978 | else: |
| 979 | ix = np.where(required_scale < 1.0, 10e9, required_scale).argmin() |
| 980 | |
| 981 | return candidate_tilings[ix] |
| 982 | |
| 983 | |
| 984 | def _as_2tuple(x: int | tuple[int, int]) -> tuple[int, int]: |
| 985 | if isinstance(x, int): |
| 986 | return x, x |
| 987 | |
| 988 | return x |
| 989 | |
| 990 | |
| 991 | class MolmoProcessingInfo(BaseProcessingInfo): |
| 992 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 993 | return {"image": None} |
| 994 | |
| 995 | def select_tiling( |
| 996 | self, |
| 997 | *, |
| 998 | image_width: int, |
| 999 | image_height: int, |
| 1000 | image_processor: BaseImageProcessor, |
| 1001 | ) -> tuple[int, int]: |
| 1002 | max_crops = image_processor.max_crops |
| 1003 | left_margin, right_margin = image_processor.overlap_margins |
| 1004 | base_image_input_size = _as_2tuple(image_processor.base_image_input_size) |
| 1005 | base_image_input_d = image_processor.image_patch_size |
| 1006 | |
| 1007 | total_margin_pixels = base_image_input_d * (right_margin + left_margin) |
| 1008 | crop_patches = base_image_input_size[0] // base_image_input_d |
| 1009 | crop_window_patches = crop_patches - (right_margin + left_margin) |
| 1010 | crop_window_size = crop_window_patches * base_image_input_d |
| 1011 | tiling_h, tiling_w = select_tiling( |
| 1012 | height=image_height - total_margin_pixels, |
| 1013 | width=image_width - total_margin_pixels, |
| 1014 | patch_size=crop_window_size, |
| 1015 | max_num_patches=max_crops, |
| 1016 | ) |
| 1017 | |
| 1018 | return tiling_w, tiling_h |
| 1019 | |
| 1020 | def get_patches_grid_size( |
| 1021 | self, |
| 1022 | *, |
| 1023 | image_width: int, |
| 1024 | image_height: int, |
| 1025 | image_processor: BaseImageProcessor, |
| 1026 | ) -> tuple[int, int]: |
| 1027 | left_margin, right_margin = image_processor.overlap_margins |
| 1028 | base_image_input_size = _as_2tuple(image_processor.base_image_input_size) |
| 1029 | base_image_input_d = image_processor.image_patch_size |
| 1030 | pooling_size = POOLING_SIZE |
| 1031 | |
| 1032 | crop_patches = base_image_input_size[0] // base_image_input_d |
| 1033 | tiling_w, tiling_h = self.select_tiling( |
| 1034 | image_height=image_height, |
| 1035 | image_width=image_width, |
| 1036 | image_processor=image_processor, |
| 1037 | ) |
| 1038 | |
| 1039 | nrows, ncols = get_patches_grid_size( |
| 1040 | tiling_h=tiling_h, |
| 1041 | tiling_w=tiling_w, |
| 1042 | crop_patches=crop_patches, |
| 1043 | left_margin=left_margin, |
| 1044 | right_margin=right_margin, |
| 1045 | pooling_size=pooling_size, |
| 1046 | ) |
| 1047 | |
| 1048 | return ncols, nrows |
| 1049 | |
| 1050 | def get_num_image_tokens( |
| 1051 | self, |
| 1052 | *, |
| 1053 | image_width: int, |
| 1054 | image_height: int, |
| 1055 | image_processor: BaseImageProcessor, |
| 1056 | ) -> int: |
| 1057 | ncols, nrows = self.get_patches_grid_size( |
| 1058 | image_width=image_width, |
| 1059 | image_height=image_height, |
| 1060 | image_processor=image_processor, |
| 1061 | ) |
| 1062 | pooling_size = POOLING_SIZE |
| 1063 | |
| 1064 | image_token_length_w = image_processor.image_token_length_w |
| 1065 | image_token_length_h = image_processor.image_token_length_h |
| 1066 | |
| 1067 | # Calculate total tokens: 2 for start/end + (w+1)*h for column separators |
| 1068 | extra = 2 + (image_token_length_w + 1) * image_token_length_h |
| 1069 | joint = 2 + ((ncols + 1) // pooling_size + 1) * ((nrows + 1) // pooling_size) |
| 1070 | |
| 1071 | return extra + joint |
| 1072 | |
| 1073 | def get_image_size_with_most_features(self) -> ImageSize: |
| 1074 | processor = self.get_hf_processor() |
| 1075 | image_processor = processor.image_processor |
| 1076 | |
| 1077 | tilings = get_candidate_tilings(image_processor.max_crops) |
| 1078 | base_h, base_w = _as_2tuple(image_processor.base_image_input_size) |
| 1079 | |
| 1080 | largest_feature_size, largest_feature_pinpoint = 0, None |
| 1081 | for wr, hr in tilings: |
| 1082 | width, height = base_w * wr, base_h * hr |
| 1083 | |
| 1084 | feat_size = self.get_num_image_tokens( |
| 1085 | image_width=width, |
| 1086 | image_height=height, |
| 1087 | image_processor=image_processor, |
| 1088 | ) |
| 1089 | if feat_size > largest_feature_size: |
| 1090 | largest_feature_size = feat_size |
| 1091 | largest_feature_pinpoint = ImageSize(width=width, height=height) |
| 1092 | |
| 1093 | if largest_feature_size == 0 or largest_feature_pinpoint is None: |
| 1094 | raise ValueError("Cannot have a largest feature size of 0!") |
| 1095 | |
| 1096 | return largest_feature_pinpoint |
| 1097 | |
| 1098 | |
| 1099 | class MolmoDummyInputsBuilder(BaseDummyInputsBuilder[MolmoProcessingInfo]): |
| 1100 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 1101 | return "" |
| 1102 | |
| 1103 | def get_dummy_mm_data( |
| 1104 | self, |
| 1105 | seq_len: int, |
| 1106 | mm_counts: Mapping[str, int], |
| 1107 | mm_options: MultiModalDummyOptions, |
| 1108 | ) -> MultiModalDataDict: |
| 1109 | target_width, target_height = self.info.get_image_size_with_most_features() |
| 1110 | |
| 1111 | return { |
| 1112 | "image": self._get_dummy_images( |
| 1113 | width=target_width, |
| 1114 | height=target_height, |
| 1115 | num_images=mm_counts.get("image", 0), |
| 1116 | overrides=mm_options.get("image"), |
| 1117 | ) |
| 1118 | } |
| 1119 | |
| 1120 | |
| 1121 | class MolmoMultiModalProcessor(BaseMultiModalProcessor[MolmoProcessingInfo]): |
| 1122 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 1123 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 1124 | |
| 1125 | def _postprocess_prompt(self, prompt: list[int]) -> list[int]: |
| 1126 | processor = self.info.get_hf_processor() |
| 1127 | |
| 1128 | # The chat template is already applied to the prompt tokens |
| 1129 | # Use message_format="none" to avoid applying it again |
| 1130 | # Prepend an empty space if `always_start_with_space` is True |
| 1131 | tokens = processor.get_tokens_input( |
| 1132 | self.info.get_tokenizer().decode(prompt), |
| 1133 | message_format="none", |
| 1134 | always_start_with_space=True, |
| 1135 | ) |
| 1136 | |
| 1137 | # Prepend a BOS token id to the tokens |
| 1138 | return self.info.ctx.call_hf_processor( |
| 1139 | processor.process, |
| 1140 | dict(tokens=tokens), |
| 1141 | )["input_ids"].tolist() |
| 1142 | |
| 1143 | def _call_hf_processor( |
| 1144 | self, |
| 1145 | hf_data: Mapping[str, object], |
| 1146 | hf_kwargs: Mapping[str, object], |
| 1147 | ) -> BatchFeature: |
| 1148 | hf_processor = self.info.get_hf_processor(**hf_kwargs) |
| 1149 | return self.info.ctx.call_hf_processor( |
| 1150 | hf_processor.process, |
| 1151 | hf_data, |
| 1152 | hf_kwargs, |
| 1153 | ) |
| 1154 | |
| 1155 | def _postprocess_hf_mm_data( |
| 1156 | self, |
| 1157 | hf_data: Mapping[str, object], |
| 1158 | hf_kwargs: Mapping[str, object], |
| 1159 | processed_data: BatchFeature, |
| 1160 | ) -> BatchFeature: |
| 1161 | if not hf_data: |
| 1162 | return processed_data |
| 1163 | |
| 1164 | hf_processor = self.info.get_hf_processor(**hf_kwargs) |
| 1165 | tokenizer = hf_processor.tokenizer |
| 1166 | image_patch_id = tokenizer.vocab[IMAGE_PATCH_TOKEN] |
| 1167 | |
| 1168 | image_processor = hf_processor.image_processor |
| 1169 | |
| 1170 | processed_data.pop("input_ids") |
| 1171 | |
| 1172 | if (images := hf_data.get("images")) is not None: |
| 1173 | mm_items = self.info.parse_mm_data({"image": images}, validate=False) |
| 1174 | parsed_images = mm_items.get_items("image", ImageProcessorItems) |
| 1175 | image_sizes = [ |
| 1176 | parsed_images.get_image_size(i) for i in range(len(parsed_images)) |
| 1177 | ] |
| 1178 | |
| 1179 | feat_is_patch = processed_data["image_input_idx"] >= 0 |
| 1180 | |
| 1181 | tilings = [ |
| 1182 | self.info.select_tiling( |
| 1183 | image_width=image_size.width, |
| 1184 | image_height=image_size.height, |
| 1185 | image_processor=image_processor, |
| 1186 | ) |
| 1187 | for image_size in image_sizes |
| 1188 | ] |
| 1189 | # For each image: tiling_h * tiling_w + extra |
| 1190 | num_crops = torch.tensor(tilings).prod(-1) + 1 |
| 1191 | assert num_crops.sum() == len(feat_is_patch) |
| 1192 | |
| 1193 | processed_data["num_crops"] = num_crops |
| 1194 | processed_data["img_patch_id"] = image_patch_id |
| 1195 | |
| 1196 | return processed_data |
| 1197 | |
| 1198 | def _get_mm_fields_config( |
| 1199 | self, |
| 1200 | hf_inputs: BatchFeature, |
| 1201 | hf_processor_mm_kwargs: Mapping[str, object], |
| 1202 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 1203 | num_crops = hf_inputs.get("num_crops", torch.empty(0)) |
| 1204 | num_images = len(num_crops) |
| 1205 | |
| 1206 | return dict( |
| 1207 | images=MultiModalFieldConfig.flat_from_sizes("image", num_crops), |
| 1208 | image_masks=MultiModalFieldConfig.flat_from_sizes("image", num_crops), |
| 1209 | image_input_idx=MultiModalFieldConfig.flat_from_sizes("image", num_crops), |
| 1210 | num_crops=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 1211 | img_patch_id=MultiModalFieldConfig.shared( |
| 1212 | "image", num_images, keep_on_cpu=True |
| 1213 | ), |
| 1214 | ) |
| 1215 | |
| 1216 | def _get_prompt_updates( |
| 1217 | self, |
| 1218 | mm_items: MultiModalDataItems, |
| 1219 | hf_processor_mm_kwargs: Mapping[str, object], |
| 1220 | out_mm_kwargs: MultiModalKwargsItems, |
| 1221 | ) -> Sequence[PromptUpdate]: |
| 1222 | tokenizer = self.info.get_tokenizer() |
| 1223 | vocab = tokenizer.get_vocab() |
| 1224 | img_patch_id = vocab[IMAGE_PATCH_TOKEN] |
| 1225 | img_col_id = vocab[IM_COL_TOKEN] |
| 1226 | img_start_id = vocab[IM_START_TOKEN] |
| 1227 | img_end_id = vocab[IM_END_TOKEN] |
| 1228 | |
| 1229 | processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 1230 | image_processor = processor.image_processor |
| 1231 | image_token_length_w = image_processor.image_token_length_w |
| 1232 | image_token_length_h = image_processor.image_token_length_h |
| 1233 | pooling_size = POOLING_SIZE |
| 1234 | |
| 1235 | extra_row = [img_patch_id] * image_token_length_w + [img_col_id] |
| 1236 | extra_joint = [img_start_id] + extra_row * image_token_length_h + [img_end_id] |
| 1237 | |
| 1238 | def get_insertion_molmo(item_idx: int): |
| 1239 | images = mm_items.get_items("image", ImageProcessorItems) |
| 1240 | image_size = images.get_image_size(item_idx) |
| 1241 | |
| 1242 | ncols, nrows = self.info.get_patches_grid_size( |
| 1243 | image_width=image_size.width, |
| 1244 | image_height=image_size.height, |
| 1245 | image_processor=image_processor, |
| 1246 | ) |
| 1247 | |
| 1248 | joint_row = [img_patch_id] * ((ncols + 1) // pooling_size) + [img_col_id] |
| 1249 | joint = ( |
| 1250 | [img_start_id] |
| 1251 | + joint_row * ((nrows + 1) // pooling_size) |
| 1252 | + [img_end_id] |
| 1253 | ) |
| 1254 | |
| 1255 | return PromptUpdateDetails.select_token_id( |
| 1256 | extra_joint + joint, |
| 1257 | embed_token_id=img_patch_id, |
| 1258 | ) |
| 1259 | |
| 1260 | return [ |
| 1261 | PromptInsertion( |
| 1262 | modality="image", |
| 1263 | target=PromptIndexTargets.prefix( |
| 1264 | cached_encode(tokenizer, "<|endoftext|>", add_special_tokens=False), |
| 1265 | ), |
| 1266 | insertion=get_insertion_molmo, |
| 1267 | ) |
| 1268 | ] |
| 1269 | |
| 1270 | |
| 1271 | @MULTIMODAL_REGISTRY.register_processor( |
| 1272 | MolmoMultiModalProcessor, |
| 1273 | info=MolmoProcessingInfo, |
| 1274 | dummy_inputs=MolmoDummyInputsBuilder, |
| 1275 | ) |
| 1276 | class MolmoForCausalLM( |
| 1277 | nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA, SupportsQuant |
| 1278 | ): |
| 1279 | hf_to_vllm_mapper = WeightsMapper( |
| 1280 | orig_to_new_substr={ |
| 1281 | # vision backbone mapping |
| 1282 | "image_projector.w1.": "image_projector.gate_proj.", |
| 1283 | "image_projector.w3.": "image_projector.up_proj.", |
| 1284 | "image_projector.w2.": "image_projector.down_proj.", |
| 1285 | # language backbone mapping |
| 1286 | "att_proj": "self_attn.qkv_proj", |
| 1287 | "attn_out": "self_attn.o_proj", |
| 1288 | "q_norm": "self_attn.q_norm", |
| 1289 | "k_norm": "self_attn.k_norm", |
| 1290 | "ff_proj": "mlp.gate_up_proj", |
| 1291 | "ff_out": "mlp.down_proj", |
| 1292 | "attn_norm": "input_layernorm", |
| 1293 | "ff_norm": "post_attention_layernorm", |
| 1294 | }, |
| 1295 | orig_to_new_prefix={ |
| 1296 | # vision backbone mapping |
| 1297 | "model.vision_backbone.": "vision_backbone.", |
| 1298 | # language backbone mapping |
| 1299 | "model.transformer.blocks.": "model.layers.", |
| 1300 | "model.transformer.ln_f.": "model.norm.", |
| 1301 | # lm_head is renamed to model.transformer.mlp.down_proj firstly, |
| 1302 | # we need to run a second renaming for it |
| 1303 | "model.transformer.mlp.down_proj.": "lm_head.", |
| 1304 | }, |
| 1305 | ) |
| 1306 | |
| 1307 | packed_modules_mapping = { |
| 1308 | "qkv_proj": ["qkv_proj"], |
| 1309 | "gate_up_proj": ["gate_up_proj"], # language model |
| 1310 | "merged_linear": ["gate_proj", "up_proj"], # image_projector |
| 1311 | } |
| 1312 | |
| 1313 | @classmethod |
| 1314 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 1315 | if modality.startswith("image"): |
| 1316 | return None |
| 1317 | |
| 1318 | raise ValueError("Only image modality is supported") |
| 1319 | |
| 1320 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1321 | super().__init__() |
| 1322 | config = vllm_config.model_config.hf_config |
| 1323 | quant_config = vllm_config.quant_config |
| 1324 | multimodal_config = vllm_config.model_config.multimodal_config |
| 1325 | |
| 1326 | self.config = config |
| 1327 | self.multimodal_config = multimodal_config |
| 1328 | |
| 1329 | vision_config = VisionBackboneConfig() |
| 1330 | |
| 1331 | with self._mark_tower_model(vllm_config, "image"): |
| 1332 | self.vision_backbone = MolmoVisionBackbone( |
| 1333 | config, |
| 1334 | vision_config, |
| 1335 | quant_config, |
| 1336 | prefix=maybe_prefix(prefix, "vision_backbone"), |
| 1337 | ) |
| 1338 | |
| 1339 | with self._mark_language_model(vllm_config): |
| 1340 | self.model = MolmoModel( |
| 1341 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 1342 | ) |
| 1343 | |
| 1344 | self.img_patch_id: int | None = None |
| 1345 | |
| 1346 | self.lm_head = ParallelLMHead( |
| 1347 | config.embedding_size or config.vocab_size, |
| 1348 | config.hidden_size, |
| 1349 | quant_config=quant_config, |
| 1350 | prefix=maybe_prefix(prefix, "lm_head"), |
| 1351 | ) |
| 1352 | if self.config.weight_tying: |
| 1353 | self.lm_head = self.lm_head.tie_weights(self.model.transformer.wte) |
| 1354 | |
| 1355 | self.logits_processor = LogitsProcessor( |
| 1356 | config.embedding_size or config.vocab_size |
| 1357 | ) |
| 1358 | |
| 1359 | self.make_empty_intermediate_tensors = ( |
| 1360 | self.model.make_empty_intermediate_tensors |
| 1361 | ) |
| 1362 | |
| 1363 | def _parse_and_validate_image_input( |
| 1364 | self, |
| 1365 | **kwargs: object, |
| 1366 | ) -> MolmoImageInputs | None: |
| 1367 | images = kwargs.pop("images", None) |
| 1368 | image_masks = kwargs.pop("image_masks", None) |
| 1369 | image_input_idx = kwargs.pop("image_input_idx", None) |
| 1370 | num_crops = kwargs.pop("num_crops", None) |
| 1371 | |
| 1372 | if images is None: |
| 1373 | return None |
| 1374 | |
| 1375 | img_patch_id = kwargs.pop("img_patch_id", None) |
| 1376 | if isinstance(img_patch_id, torch.Tensor): |
| 1377 | img_patch_id = img_patch_id.item() |
| 1378 | |
| 1379 | assert isinstance(img_patch_id, int) |
| 1380 | self.img_patch_id = img_patch_id |
| 1381 | |
| 1382 | return MolmoImageInputs( |
| 1383 | images=images, |
| 1384 | image_masks=image_masks, |
| 1385 | image_input_idx=image_input_idx, |
| 1386 | num_crops=num_crops, |
| 1387 | ) |
| 1388 | |
| 1389 | def _process_image_input( |
| 1390 | self, |
| 1391 | image_input: MolmoImageInputs, |
| 1392 | ) -> list[torch.Tensor]: |
| 1393 | images = image_input["images"] |
| 1394 | image_masks = image_input["image_masks"] |
| 1395 | image_input_idx = image_input["image_input_idx"] |
| 1396 | num_crops = image_input["num_crops"] |
| 1397 | |
| 1398 | # Call the vision backbone on the whole batch at once |
| 1399 | image_features = self.vision_backbone( |
| 1400 | images=images.unsqueeze(0), |
| 1401 | image_masks=None if image_masks is None else image_masks.unsqueeze(0), |
| 1402 | ).squeeze(0) |
| 1403 | |
| 1404 | # Only the features corresponding to patch tokens are relevant |
| 1405 | # Re-order the features using the image_input_idx tensor |
| 1406 | results = [] |
| 1407 | num_crops_list = num_crops.tolist() |
| 1408 | for feats, img_idx in zip( |
| 1409 | image_features.split(num_crops_list), |
| 1410 | image_input_idx.split(num_crops_list), |
| 1411 | ): |
| 1412 | is_valid = img_idx >= 0 |
| 1413 | valid_img_idx = img_idx[is_valid] |
| 1414 | order = torch.argsort(valid_img_idx) |
| 1415 | results.append(feats[is_valid][order]) |
| 1416 | return results |
| 1417 | |
| 1418 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 1419 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 1420 | if image_input is None: |
| 1421 | return [] |
| 1422 | |
| 1423 | return self._process_image_input(image_input) |
| 1424 | |
| 1425 | def forward( |
| 1426 | self, |
| 1427 | input_ids: torch.LongTensor, |
| 1428 | positions: torch.LongTensor, |
| 1429 | intermediate_tensors: IntermediateTensors | None = None, |
| 1430 | inputs_embeds: torch.Tensor | None = None, |
| 1431 | **kwargs: object, |
| 1432 | ) -> torch.Tensor: |
| 1433 | if intermediate_tensors is not None: |
| 1434 | inputs_embeds = None |
| 1435 | |
| 1436 | hidden_states = self.model( |
| 1437 | input_ids, positions, intermediate_tensors, inputs_embeds=inputs_embeds |
| 1438 | ) |
| 1439 | |
| 1440 | return hidden_states |
| 1441 | |
| 1442 | def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 1443 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 1444 | return logits |
| 1445 | |
| 1446 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): |
| 1447 | loader = AutoWeightsLoader(self) |
| 1448 | weights = _get_weights_with_merged_embedding(weights) |
| 1449 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 1450 | |
| 1451 | def get_mm_mapping(self) -> MultiModelKeys: |
| 1452 | """Get the module prefix in multimodal models.""" |
| 1453 | return MultiModelKeys.from_string_field( |
| 1454 | language_model="model", |
| 1455 | connector="vision_backbone.image_projector", |
| 1456 | tower_model="vision_backbone", |
| 1457 | ) |
| 1458 | |
| 1459 | |
| 1460 | def _get_weights_with_merged_embedding( |
| 1461 | weights: Iterable[tuple[str, torch.Tensor]], |
| 1462 | ) -> Iterable[tuple[str, torch.Tensor]]: |
| 1463 | embedding_weights = {} |
| 1464 | for name, weight in weights: |
| 1465 | if "wte.embedding" in name: |
| 1466 | embedding_weights["embedding"] = weight |
| 1467 | elif "wte.new_embedding" in name: |
| 1468 | embedding_weights["new_embedding"] = weight |
| 1469 | else: |
| 1470 | yield (name, weight) |
| 1471 | # this is compatible with most of quantization, |
| 1472 | # because they won't quantize embed_tokens |
| 1473 | embedding_weights = torch.cat( |
| 1474 | [embedding_weights["embedding"], embedding_weights["new_embedding"]], |
| 1475 | dim=0, |
| 1476 | ) |
| 1477 | yield ("model.embed_tokens.weight", embedding_weights) |