| 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, Mapping, Sequence |
| 5 | from dataclasses import dataclass, fields |
| 6 | from functools import partial |
| 7 | from itertools import islice |
| 8 | from typing import Annotated, Any |
| 9 | |
| 10 | import numpy as np |
| 11 | import torch |
| 12 | import torch.nn as nn |
| 13 | import torch.nn.functional as F |
| 14 | from PIL.Image import Image |
| 15 | from transformers import ( |
| 16 | BaseImageProcessor, |
| 17 | BaseVideoProcessor, |
| 18 | BatchFeature, |
| 19 | PreTrainedConfig, |
| 20 | ProcessorMixin, |
| 21 | ) |
| 22 | from transformers.image_utils import ImageInput |
| 23 | from transformers.video_utils import VideoMetadata |
| 24 | from typing_extensions import TypedDict |
| 25 | |
| 26 | from vllm.compilation.decorators import support_torch_compile |
| 27 | from vllm.config import CacheConfig, VllmConfig |
| 28 | from vllm.config.multimodal import MultiModalDummyOptions, VideoDummyOptions |
| 29 | from vllm.distributed import ( |
| 30 | get_pp_group, |
| 31 | get_tensor_model_parallel_rank, |
| 32 | get_tensor_model_parallel_world_size, |
| 33 | split_tensor_along_last_dim, |
| 34 | tensor_model_parallel_all_gather, |
| 35 | ) |
| 36 | from vllm.inputs import MultiModalDataDict |
| 37 | from vllm.logger import init_logger |
| 38 | from vllm.model_executor.layers.activation import MulAndSilu, SiluAndMul, get_act_fn |
| 39 | from vllm.model_executor.layers.attention import Attention, MMEncoderAttention |
| 40 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 41 | from vllm.model_executor.layers.linear import ( |
| 42 | ColumnParallelLinear, |
| 43 | MergedColumnParallelLinear, |
| 44 | QKVParallelLinear, |
| 45 | RowParallelLinear, |
| 46 | ) |
| 47 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 48 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 49 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 50 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 51 | ParallelLMHead, |
| 52 | VocabParallelEmbedding, |
| 53 | ) |
| 54 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 55 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 56 | from vllm.multimodal.image import normalize_image |
| 57 | from vllm.multimodal.inputs import ( |
| 58 | MultiModalFieldConfig, |
| 59 | MultiModalKwargsItems, |
| 60 | VideoItem, |
| 61 | ) |
| 62 | from vllm.multimodal.parse import ( |
| 63 | ImageProcessorItems, |
| 64 | ImageSize, |
| 65 | MultiModalDataItems, |
| 66 | MultiModalDataParser, |
| 67 | ) |
| 68 | from vllm.multimodal.processing import ( |
| 69 | BaseMultiModalProcessor, |
| 70 | BaseProcessingInfo, |
| 71 | PromptReplacement, |
| 72 | PromptUpdate, |
| 73 | PromptUpdateDetails, |
| 74 | ) |
| 75 | from vllm.multimodal.processing.dummy_inputs import BaseDummyInputsBuilder |
| 76 | from vllm.sequence import IntermediateTensors |
| 77 | from vllm.utils.math_utils import round_down |
| 78 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 79 | |
| 80 | from .interfaces import ( |
| 81 | MultiModalEmbeddings, |
| 82 | SupportsLoRA, |
| 83 | SupportsMultiModal, |
| 84 | SupportsPP, |
| 85 | SupportsQuant, |
| 86 | ) |
| 87 | from .utils import ( |
| 88 | AutoWeightsLoader, |
| 89 | WeightsMapper, |
| 90 | _merge_multimodal_embeddings, |
| 91 | extract_layer_index, |
| 92 | make_empty_intermediate_tensors_factory, |
| 93 | make_layers, |
| 94 | maybe_prefix, |
| 95 | ) |
| 96 | |
| 97 | logger = init_logger(__name__) |
| 98 | |
| 99 | |
| 100 | # Special tokens. These should be present in any tokenizer we use |
| 101 | # because the preprocessor relies on them. |
| 102 | IMAGE_PROMPT = "<|image|>" |
| 103 | VIDEO_PROMPT = "<|video|>" |
| 104 | _MAX_VIDEO_FPS = 8 |
| 105 | |
| 106 | |
| 107 | class Molmo2ImageInputs(TensorSchema): |
| 108 | """Dimensions: |
| 109 | - nc: The total number of crops (dynamic) |
| 110 | - np: The total number of patches per crop |
| 111 | - cps: Number of channels * patch_size * patch_size |
| 112 | - npp: Number of pooled patches (dynamic) |
| 113 | - pp: pooling_size * pooling_size |
| 114 | - ni: Number of images |
| 115 | - nt: Number of image tokens (dynamic) |
| 116 | """ |
| 117 | |
| 118 | pixel_values: Annotated[torch.Tensor, TensorShape("nc", "np", "cps")] |
| 119 | |
| 120 | token_pooling: Annotated[torch.Tensor, TensorShape("npp", "pp")] |
| 121 | """ |
| 122 | An index tensor that maps image features to their corresponding |
| 123 | patch tokens before pooling. |
| 124 | """ |
| 125 | |
| 126 | num_pooled_patches: Annotated[torch.Tensor, TensorShape("ni")] |
| 127 | |
| 128 | image_tokens: Annotated[torch.BoolTensor, TensorShape("nt")] |
| 129 | |
| 130 | num_image_tokens: Annotated[torch.Tensor, TensorShape("ni")] |
| 131 | |
| 132 | |
| 133 | class Molmo2VideoInputs(TensorSchema): |
| 134 | """Dimensions: |
| 135 | - nc: The total number of frames (dynamic) |
| 136 | - np: The total number of patches per frame |
| 137 | - cps: Number of channels * patch_size * patch_size |
| 138 | - npp: Number of pooled patches (dynamic) |
| 139 | - pp: pooling_size * pooling_size |
| 140 | - nv: Number of videos |
| 141 | - nt: Number of video tokens (dynamic) |
| 142 | """ |
| 143 | |
| 144 | pixel_values_videos: Annotated[torch.Tensor, TensorShape("nc", "np", "cps")] |
| 145 | |
| 146 | token_pooling: Annotated[torch.Tensor, TensorShape("npp", "pp")] |
| 147 | """ |
| 148 | An index tensor that maps image features to their corresponding |
| 149 | patch tokens before pooling. |
| 150 | """ |
| 151 | |
| 152 | num_pooled_patches: Annotated[torch.Tensor, TensorShape("nv")] |
| 153 | |
| 154 | video_tokens: Annotated[torch.BoolTensor, TensorShape("nt")] |
| 155 | |
| 156 | num_video_tokens: Annotated[torch.Tensor, TensorShape("nv")] |
| 157 | |
| 158 | |
| 159 | class Molmo2MultiModalInputs(TypedDict, total=False): |
| 160 | images: Molmo2ImageInputs | None |
| 161 | videos: Molmo2VideoInputs | None |
| 162 | |
| 163 | |
| 164 | @dataclass |
| 165 | class VitConfig: |
| 166 | """Config for a vision transformer.""" |
| 167 | |
| 168 | hidden_size: int = 1152 |
| 169 | intermediate_size: int = 4304 |
| 170 | num_hidden_layers: int = 27 |
| 171 | num_attention_heads: int = 16 |
| 172 | num_key_value_heads: int = 16 |
| 173 | head_dim: int = 72 |
| 174 | hidden_act: str = "gelu_pytorch_tanh" |
| 175 | layer_norm_eps: float = 1e-6 |
| 176 | image_default_input_size: tuple[int, int] = (378, 378) |
| 177 | image_patch_size: int = 14 |
| 178 | image_num_pos: int = 577 |
| 179 | |
| 180 | def __post_init__(self): |
| 181 | self.image_default_input_size = tuple(self.image_default_input_size) # type: ignore[assignment] |
| 182 | |
| 183 | @property |
| 184 | def image_num_patch(self): |
| 185 | h, w = self.image_default_input_size |
| 186 | return h // self.image_patch_size, w // self.image_patch_size |
| 187 | |
| 188 | |
| 189 | @dataclass |
| 190 | class AdapterConfig: |
| 191 | """Config for a vit-llm adapter.""" |
| 192 | |
| 193 | vit_layers: tuple[int, int] = (-3, -9) |
| 194 | pooling_attention_mask: bool = False |
| 195 | hidden_size: int = 1152 |
| 196 | num_attention_heads: int = 16 |
| 197 | num_key_value_heads: int = 16 |
| 198 | head_dim: int = 72 |
| 199 | hidden_act: str = "silu" |
| 200 | intermediate_size: int = 18944 |
| 201 | text_hidden_size: int = 3584 |
| 202 | |
| 203 | |
| 204 | @dataclass |
| 205 | class TextConfig: |
| 206 | """Configuration for a text model transformer.""" |
| 207 | |
| 208 | hidden_size: int = 3584 |
| 209 | """ |
| 210 | The hidden size of the model. |
| 211 | """ |
| 212 | |
| 213 | num_attention_heads: int = 28 |
| 214 | """ |
| 215 | The number of self-attention heads. |
| 216 | """ |
| 217 | |
| 218 | num_key_value_heads: int = 4 |
| 219 | """ |
| 220 | The number of heads to use for keys and values. |
| 221 | """ |
| 222 | |
| 223 | head_dim: int = 128 |
| 224 | """ |
| 225 | The head dimensionality for the attention mechanism. |
| 226 | """ |
| 227 | |
| 228 | vocab_size: int = 152064 |
| 229 | """Vocabulary size of the model.""" |
| 230 | |
| 231 | additional_vocab_size: int = 128 |
| 232 | """Number of additional tokens to have the input embeddings for""" |
| 233 | |
| 234 | qkv_bias: bool = True |
| 235 | """ |
| 236 | Do QKV projection a bias |
| 237 | """ |
| 238 | |
| 239 | num_hidden_layers: int = 48 |
| 240 | """ |
| 241 | The number of layers/blocks. |
| 242 | """ |
| 243 | |
| 244 | intermediate_size: int = 18944 |
| 245 | """ |
| 246 | The hidden size for the MLP. |
| 247 | """ |
| 248 | |
| 249 | hidden_act: str = "silu" |
| 250 | """ |
| 251 | The activation function to use within the MLP layers. |
| 252 | """ |
| 253 | |
| 254 | max_position_embeddings: int = 4096 |
| 255 | """ |
| 256 | Max positional embeddings to use in RoPE cache |
| 257 | """ |
| 258 | |
| 259 | rope_theta: float = 1000000.0 |
| 260 | """ |
| 261 | RoPE theta parameter. |
| 262 | """ |
| 263 | |
| 264 | use_qk_norm: bool = False |
| 265 | """ |
| 266 | Apply layer norm to the keys and queries within the attention mechanism. |
| 267 | This can help stabilize training. |
| 268 | """ |
| 269 | |
| 270 | qk_norm_type: str = "olmo" |
| 271 | """ |
| 272 | The type of layer norm to use for the keys and queries. |
| 273 | Can be "olmo" or "qwen3". |
| 274 | """ |
| 275 | |
| 276 | layer_norm_eps: float = 1e-6 |
| 277 | """ |
| 278 | epsilon for layer norms |
| 279 | """ |
| 280 | |
| 281 | norm_after: bool = False |
| 282 | """Apply layer norm before and after the attention and MLP blocks.""" |
| 283 | |
| 284 | rope_scaling_layers: tuple[int, ...] | None = None |
| 285 | """ |
| 286 | RoPE scaling layers. |
| 287 | """ |
| 288 | |
| 289 | |
| 290 | class ViTMLP(nn.Module): |
| 291 | """MLP used in Vision Transformer.""" |
| 292 | |
| 293 | def __init__( |
| 294 | self, |
| 295 | dim: int, |
| 296 | hidden_dim: int, |
| 297 | hidden_act: str, |
| 298 | quant_config: QuantizationConfig | None = None, |
| 299 | prefix: str = "", |
| 300 | ) -> None: |
| 301 | super().__init__() |
| 302 | self.w1 = ColumnParallelLinear( |
| 303 | dim, |
| 304 | hidden_dim, |
| 305 | bias=True, |
| 306 | quant_config=quant_config, |
| 307 | prefix=f"{prefix}.w1", |
| 308 | ) |
| 309 | # Activation function. |
| 310 | self.act = get_act_fn(hidden_act) |
| 311 | self.w2 = RowParallelLinear( |
| 312 | hidden_dim, |
| 313 | dim, |
| 314 | bias=True, |
| 315 | quant_config=quant_config, |
| 316 | prefix=f"{prefix}.w2", |
| 317 | ) |
| 318 | |
| 319 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 320 | x, _ = self.w1(x) |
| 321 | x = self.act(x) |
| 322 | x, _ = self.w2(x) |
| 323 | return x |
| 324 | |
| 325 | |
| 326 | class ViTMultiHeadDotProductAttention(nn.Module): |
| 327 | """Multi-head attention used in Vision Transformer.""" |
| 328 | |
| 329 | def __init__( |
| 330 | self, |
| 331 | hidden_size: int, |
| 332 | num_heads: int, |
| 333 | num_key_value_heads: int, |
| 334 | head_dim: int, |
| 335 | use_bias: bool = True, |
| 336 | quant_config: QuantizationConfig | None = None, |
| 337 | prefix: str = "", |
| 338 | ) -> None: |
| 339 | super().__init__() |
| 340 | |
| 341 | self.hidden_size = hidden_size |
| 342 | self.total_num_heads = num_heads |
| 343 | tp_size = get_tensor_model_parallel_world_size() |
| 344 | |
| 345 | assert self.hidden_size % self.total_num_heads == 0 |
| 346 | assert self.total_num_heads % tp_size == 0 |
| 347 | |
| 348 | self.num_heads = self.total_num_heads // tp_size |
| 349 | self.head_dim = head_dim |
| 350 | |
| 351 | assert self.head_dim == self.hidden_size // self.total_num_heads |
| 352 | |
| 353 | self.total_num_kv_heads = num_key_value_heads |
| 354 | if self.total_num_kv_heads >= tp_size: |
| 355 | assert self.total_num_kv_heads % tp_size == 0 |
| 356 | else: |
| 357 | assert tp_size % self.total_num_kv_heads == 0 |
| 358 | |
| 359 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 360 | |
| 361 | self.q_size = self.num_heads * self.head_dim |
| 362 | self.kv_size = self.num_kv_heads * self.head_dim |
| 363 | |
| 364 | self.merged_qkv = QKVParallelLinear( |
| 365 | self.hidden_size, |
| 366 | self.head_dim, |
| 367 | self.total_num_heads, |
| 368 | self.total_num_kv_heads, |
| 369 | bias=use_bias, |
| 370 | quant_config=quant_config, |
| 371 | prefix=f"{prefix}.merged_qkv", |
| 372 | ) |
| 373 | self.wo = RowParallelLinear( |
| 374 | self.total_num_heads * self.head_dim, |
| 375 | self.hidden_size, |
| 376 | bias=use_bias, |
| 377 | quant_config=quant_config, |
| 378 | prefix=f"{prefix}.wo", |
| 379 | ) |
| 380 | self.scale = self.head_dim**-0.5 |
| 381 | self.attn = MMEncoderAttention( |
| 382 | self.num_heads, |
| 383 | self.head_dim, |
| 384 | self.scale, |
| 385 | num_kv_heads=self.num_kv_heads, |
| 386 | prefix=f"{prefix}.attn", |
| 387 | ) |
| 388 | |
| 389 | def forward(self, inputs: torch.Tensor) -> torch.Tensor: |
| 390 | qkv, _ = self.merged_qkv(inputs) |
| 391 | xq, xk, xv = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 392 | |
| 393 | output = self.attn(xq, xk, xv) |
| 394 | |
| 395 | output, _ = self.wo(output) |
| 396 | |
| 397 | return output |
| 398 | |
| 399 | |
| 400 | class Molmo2VisionBlock(nn.Module): |
| 401 | """Residual attention block used in Vision Transformer.""" |
| 402 | |
| 403 | def __init__( |
| 404 | self, |
| 405 | config: VitConfig, |
| 406 | quant_config: QuantizationConfig | None = None, |
| 407 | prefix: str = "", |
| 408 | ) -> None: |
| 409 | super().__init__() |
| 410 | self.attention = ViTMultiHeadDotProductAttention( |
| 411 | hidden_size=config.hidden_size, |
| 412 | num_heads=config.num_attention_heads, |
| 413 | num_key_value_heads=config.num_key_value_heads, |
| 414 | head_dim=config.head_dim, |
| 415 | quant_config=quant_config, |
| 416 | prefix=f"{prefix}.attention", |
| 417 | ) |
| 418 | self.feed_forward = ViTMLP( |
| 419 | dim=config.hidden_size, |
| 420 | hidden_dim=config.intermediate_size, |
| 421 | hidden_act=config.hidden_act, |
| 422 | quant_config=quant_config, |
| 423 | prefix=f"{prefix}.feed_forward", |
| 424 | ) |
| 425 | self.attention_norm = nn.LayerNorm( |
| 426 | config.hidden_size, |
| 427 | eps=config.layer_norm_eps, |
| 428 | ) |
| 429 | self.ffn_norm = nn.LayerNorm( |
| 430 | config.hidden_size, |
| 431 | eps=config.layer_norm_eps, |
| 432 | ) |
| 433 | |
| 434 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 435 | x = x + self.attention(self.attention_norm(x)) |
| 436 | x = x + self.feed_forward(self.ffn_norm(x)) |
| 437 | return x |
| 438 | |
| 439 | |
| 440 | class Molmo2VisionBlockCollection(nn.Module): |
| 441 | """Collection of residual attention blocks used in Vision Transformer.""" |
| 442 | |
| 443 | def __init__( |
| 444 | self, |
| 445 | config: VitConfig, |
| 446 | quant_config: QuantizationConfig | None = None, |
| 447 | prefix: str = "", |
| 448 | ) -> None: |
| 449 | super().__init__() |
| 450 | self.resblocks = nn.ModuleList( |
| 451 | [ |
| 452 | Molmo2VisionBlock( |
| 453 | config, |
| 454 | quant_config, |
| 455 | prefix=f"{prefix}.resblocks.{layer_idx}", |
| 456 | ) |
| 457 | for layer_idx in range(config.num_hidden_layers) |
| 458 | ] |
| 459 | ) |
| 460 | |
| 461 | def forward(self, x: torch.Tensor) -> list[torch.Tensor]: |
| 462 | hidden_states = [] |
| 463 | for r in self.resblocks: |
| 464 | x = r(x) |
| 465 | hidden_states.append(x) |
| 466 | return hidden_states |
| 467 | |
| 468 | |
| 469 | class Molmo2VisionTransformer(nn.Module): |
| 470 | """Vision Transformer used in Vision Backbone.""" |
| 471 | |
| 472 | def __init__( |
| 473 | self, |
| 474 | config: VitConfig, |
| 475 | quant_config: QuantizationConfig | None = None, |
| 476 | prefix: str = "", |
| 477 | ) -> None: |
| 478 | super().__init__() |
| 479 | scale = config.hidden_size**-0.5 |
| 480 | self.num_prefix_tokens: int = 0 # no class embeddings |
| 481 | self.patch_num: tuple[int, int] = config.image_num_patch |
| 482 | self.positional_embedding = nn.Parameter( |
| 483 | torch.randn(config.image_num_pos, config.hidden_size) * scale, |
| 484 | ) |
| 485 | image_patch_size = config.image_patch_size |
| 486 | self.patch_embedding = nn.Linear( |
| 487 | image_patch_size * image_patch_size * 3, |
| 488 | config.hidden_size, |
| 489 | bias=True, |
| 490 | ) |
| 491 | self.transformer = Molmo2VisionBlockCollection( |
| 492 | config, |
| 493 | quant_config, |
| 494 | prefix=maybe_prefix(prefix, "transformer"), |
| 495 | ) |
| 496 | |
| 497 | def add_pos_emb(self, x: torch.Tensor, patch_num: tuple[int, int]) -> torch.Tensor: |
| 498 | pos_emb = self.positional_embedding |
| 499 | |
| 500 | pos_emb = pos_emb.reshape( |
| 501 | ( |
| 502 | int(math.sqrt(pos_emb.shape[0])), |
| 503 | int(math.sqrt(pos_emb.shape[0])), |
| 504 | pos_emb.shape[1], |
| 505 | ) |
| 506 | ) |
| 507 | |
| 508 | (patch_num_0, patch_num_1) = patch_num |
| 509 | |
| 510 | if pos_emb.shape[0] != patch_num_0 or pos_emb.shape[1] != patch_num_1: |
| 511 | # from https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py |
| 512 | pos_emb = pos_emb.unsqueeze(0).permute(0, 3, 1, 2) |
| 513 | pos_emb = F.interpolate( |
| 514 | pos_emb, |
| 515 | size=(patch_num_0, patch_num_1), |
| 516 | mode="bicubic", |
| 517 | align_corners=False, |
| 518 | antialias=True, |
| 519 | ) |
| 520 | pos_emb = pos_emb.permute(0, 2, 3, 1).squeeze(0) |
| 521 | |
| 522 | pos_emb = pos_emb.reshape(-1, pos_emb.shape[-1]) |
| 523 | x = x + pos_emb[None, :, :].to(x.dtype) |
| 524 | return x |
| 525 | |
| 526 | def forward( |
| 527 | self, |
| 528 | x: torch.Tensor, |
| 529 | patch_num: tuple[int, int] | None = None, |
| 530 | ) -> list[torch.Tensor]: |
| 531 | """: param x: (batch_size, num_patch, n_pixels).""" |
| 532 | if patch_num is None: |
| 533 | patch_num = self.patch_num |
| 534 | |
| 535 | x = self.patch_embedding(x) |
| 536 | |
| 537 | x = self.add_pos_emb(x, patch_num) |
| 538 | |
| 539 | hidden_states = self.transformer(x) |
| 540 | return hidden_states |
| 541 | |
| 542 | |
| 543 | class ImagePoolingAttention(nn.Module): |
| 544 | """Multi-head attention used for image pooling.""" |
| 545 | |
| 546 | def __init__( |
| 547 | self, |
| 548 | input_dim: int, |
| 549 | hidden_size: int, |
| 550 | num_heads: int, |
| 551 | num_key_value_heads: int, |
| 552 | head_dim: int, |
| 553 | use_bias: bool = True, |
| 554 | use_pytorch_sdpa: bool = False, |
| 555 | quant_config: QuantizationConfig | None = None, |
| 556 | prefix: str = "", |
| 557 | ) -> None: |
| 558 | super().__init__() |
| 559 | |
| 560 | self.input_dim = input_dim |
| 561 | self.hidden_size = hidden_size |
| 562 | self.total_num_heads = num_heads |
| 563 | tp_size = get_tensor_model_parallel_world_size() |
| 564 | |
| 565 | assert self.hidden_size % self.total_num_heads == 0 |
| 566 | assert self.total_num_heads % tp_size == 0 |
| 567 | |
| 568 | self.num_heads = self.total_num_heads // tp_size |
| 569 | self.head_dim = head_dim |
| 570 | |
| 571 | assert self.head_dim == self.hidden_size // self.total_num_heads |
| 572 | |
| 573 | self.total_num_kv_heads = num_key_value_heads |
| 574 | if self.total_num_kv_heads >= tp_size: |
| 575 | assert self.total_num_kv_heads % tp_size == 0 |
| 576 | else: |
| 577 | assert tp_size % self.total_num_kv_heads == 0 |
| 578 | |
| 579 | self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) |
| 580 | |
| 581 | self.kv_size = self.num_kv_heads * self.head_dim |
| 582 | |
| 583 | self.q_proj = ColumnParallelLinear( |
| 584 | self.input_dim, |
| 585 | self.total_num_heads * self.head_dim, |
| 586 | bias=use_bias, |
| 587 | quant_config=quant_config, |
| 588 | prefix=f"{prefix}.q_proj", |
| 589 | ) |
| 590 | self.merged_kv = MergedColumnParallelLinear( |
| 591 | self.input_dim, |
| 592 | [self.total_num_kv_heads * self.head_dim] * 2, |
| 593 | bias=use_bias, |
| 594 | quant_config=quant_config, |
| 595 | prefix=f"{prefix}.merged_kv", |
| 596 | ) |
| 597 | self.o_proj = RowParallelLinear( |
| 598 | self.total_num_heads * self.head_dim, |
| 599 | self.hidden_size, |
| 600 | bias=use_bias, |
| 601 | quant_config=quant_config, |
| 602 | prefix=f"{prefix}.o_proj", |
| 603 | ) |
| 604 | self.scale = self.head_dim**-0.5 |
| 605 | self.use_pytorch_sdpa = use_pytorch_sdpa |
| 606 | if use_pytorch_sdpa: |
| 607 | self.attn = None |
| 608 | else: |
| 609 | self.attn = MMEncoderAttention( |
| 610 | self.num_heads, |
| 611 | self.head_dim, |
| 612 | self.scale, |
| 613 | num_kv_heads=self.num_kv_heads, |
| 614 | prefix=f"{prefix}.attn", |
| 615 | ) |
| 616 | |
| 617 | def forward_sdpa( |
| 618 | self, |
| 619 | query: torch.Tensor, |
| 620 | key: torch.Tensor, |
| 621 | value: torch.Tensor, |
| 622 | attn_mask: torch.Tensor | None = None, |
| 623 | ) -> torch.Tensor: |
| 624 | bsz, q_len, _ = query.size() |
| 625 | kv_len = key.size(1) |
| 626 | |
| 627 | query = query.view(bsz, q_len, self.num_heads, self.head_dim) |
| 628 | key = key.view(bsz, kv_len, self.num_kv_heads, self.head_dim) |
| 629 | value = value.view(bsz, kv_len, self.num_kv_heads, self.head_dim) |
| 630 | |
| 631 | query, key, value = (x.transpose(1, 2) for x in (query, key, value)) |
| 632 | |
| 633 | out = F.scaled_dot_product_attention( |
| 634 | query, |
| 635 | key, |
| 636 | value, |
| 637 | attn_mask=attn_mask, |
| 638 | is_causal=False, |
| 639 | enable_gqa=self.num_heads > self.num_kv_heads, |
| 640 | ).transpose(1, 2) |
| 641 | |
| 642 | return out.reshape(bsz, q_len, -1) |
| 643 | |
| 644 | def forward( |
| 645 | self, |
| 646 | inputs_q: torch.Tensor, |
| 647 | inputs_kv: torch.Tensor, |
| 648 | attn_mask: torch.Tensor | None = None, |
| 649 | ) -> torch.Tensor: |
| 650 | xq, _ = self.q_proj(inputs_q) |
| 651 | kv, _ = self.merged_kv(inputs_kv) |
| 652 | xk, xv = kv.split([self.kv_size, self.kv_size], dim=-1) |
| 653 | |
| 654 | if self.use_pytorch_sdpa: |
| 655 | output = self.forward_sdpa(xq, xk, xv, attn_mask) |
| 656 | else: |
| 657 | assert self.attn is not None |
| 658 | output = self.attn(xq, xk, xv) |
| 659 | |
| 660 | output, _ = self.o_proj(output) |
| 661 | |
| 662 | return output |
| 663 | |
| 664 | |
| 665 | class ImageProjectorMLP(nn.Module): |
| 666 | """MLP used for the image projector.""" |
| 667 | |
| 668 | def __init__( |
| 669 | self, |
| 670 | input_dim: int, |
| 671 | hidden_dim: int, |
| 672 | output_dim: int, |
| 673 | hidden_act: str, |
| 674 | quant_config: QuantizationConfig | None = None, |
| 675 | prefix: str = "", |
| 676 | ) -> None: |
| 677 | super().__init__() |
| 678 | |
| 679 | self.merged_linear = MergedColumnParallelLinear( |
| 680 | input_dim, |
| 681 | [hidden_dim] * 2, |
| 682 | bias=False, |
| 683 | quant_config=quant_config, |
| 684 | prefix=f"{prefix}.merged_linear", |
| 685 | ) |
| 686 | # Activation function. |
| 687 | assert hidden_act == "silu" |
| 688 | self.act_fn = SiluAndMul() |
| 689 | |
| 690 | # Feed-forward output projection. |
| 691 | self.down_proj = RowParallelLinear( |
| 692 | hidden_dim, |
| 693 | output_dim, |
| 694 | bias=False, |
| 695 | quant_config=quant_config, |
| 696 | prefix=f"{prefix}.down_proj", |
| 697 | ) |
| 698 | |
| 699 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 700 | x, _ = self.merged_linear(x) |
| 701 | x = self.act_fn(x) |
| 702 | x, _ = self.down_proj(x) |
| 703 | return x |
| 704 | |
| 705 | |
| 706 | class Molmo2VisionBackbone(nn.Module, SupportsQuant): |
| 707 | packed_modules_mapping = { |
| 708 | "merged_qkv": ["wq", "wk", "wv"], # vision backbone |
| 709 | "merged_kv": ["k_proj", "v_proj"], # image_pooling_2d |
| 710 | "merged_linear": ["gate_proj", "up_proj"], |
| 711 | } |
| 712 | |
| 713 | # Runs after the top-level mapper, so image_pooling_2d/image_projector |
| 714 | # source names are already renamed to q/k/v_proj and gate/up_proj. |
| 715 | hf_to_vllm_mapper = WeightsMapper( |
| 716 | orig_to_new_stacked={ |
| 717 | "wq": ("merged_qkv", "q"), |
| 718 | "wk": ("merged_qkv", "k"), |
| 719 | "wv": ("merged_qkv", "v"), |
| 720 | "k_proj": ("merged_kv", 0), |
| 721 | "v_proj": ("merged_kv", 1), |
| 722 | "gate_proj": ("merged_linear", 0), |
| 723 | "up_proj": ("merged_linear", 1), |
| 724 | }, |
| 725 | ) |
| 726 | |
| 727 | def __init__( |
| 728 | self, |
| 729 | vit_config: VitConfig, |
| 730 | adapter_config: AdapterConfig, |
| 731 | quant_config: QuantizationConfig | None = None, |
| 732 | prefix: str = "", |
| 733 | ) -> None: |
| 734 | super().__init__() |
| 735 | self.vit_config = vit_config |
| 736 | self.adapter_config = adapter_config |
| 737 | |
| 738 | self.vit_layers = [] |
| 739 | for layer in adapter_config.vit_layers: |
| 740 | if layer >= 0: |
| 741 | self.vit_layers.append(layer) |
| 742 | else: |
| 743 | self.vit_layers.append(layer + vit_config.num_hidden_layers) |
| 744 | |
| 745 | last_layer_needed = max(self.vit_layers) + 1 |
| 746 | if last_layer_needed < vit_config.num_hidden_layers: |
| 747 | vit_config.num_hidden_layers = last_layer_needed |
| 748 | self.image_vit = Molmo2VisionTransformer( |
| 749 | vit_config, |
| 750 | quant_config, |
| 751 | prefix=f"{prefix}.image_vit", |
| 752 | ) |
| 753 | |
| 754 | self.num_prefix_tokens: int = self.image_vit.num_prefix_tokens |
| 755 | |
| 756 | pool_dim = vit_config.hidden_size * len(adapter_config.vit_layers) |
| 757 | self.image_pooling_2d = ImagePoolingAttention( |
| 758 | input_dim=pool_dim, |
| 759 | hidden_size=adapter_config.hidden_size, |
| 760 | num_heads=adapter_config.num_attention_heads, |
| 761 | num_key_value_heads=adapter_config.num_key_value_heads, |
| 762 | head_dim=adapter_config.head_dim, |
| 763 | use_pytorch_sdpa=adapter_config.pooling_attention_mask, |
| 764 | quant_config=quant_config, |
| 765 | prefix=f"{prefix}.image_pooling_2d", |
| 766 | ) |
| 767 | self.image_projector = ImageProjectorMLP( |
| 768 | input_dim=adapter_config.hidden_size, |
| 769 | hidden_dim=adapter_config.intermediate_size, |
| 770 | output_dim=adapter_config.text_hidden_size, |
| 771 | hidden_act=adapter_config.hidden_act, |
| 772 | quant_config=quant_config, |
| 773 | prefix=f"{prefix}.image_projector", |
| 774 | ) |
| 775 | |
| 776 | @property |
| 777 | def dtype(self) -> torch.dtype: |
| 778 | return self.image_vit.patch_embedding.weight.dtype |
| 779 | |
| 780 | @property |
| 781 | def device(self) -> torch.device: |
| 782 | return self.image_vit.patch_embedding.weight.device |
| 783 | |
| 784 | def encode_image(self, images: torch.Tensor) -> torch.Tensor: |
| 785 | """: param images: (batch_size, num_crops, num_patch, n_pixels).""" |
| 786 | B, T, N, D = images.shape |
| 787 | images = images.view(B * T, N, D) |
| 788 | image_features = self.image_vit(images) |
| 789 | |
| 790 | features = [] |
| 791 | for layer in self.vit_layers: |
| 792 | features.append(image_features[layer]) |
| 793 | image_features = torch.cat(features, dim=-1) |
| 794 | |
| 795 | if self.num_prefix_tokens > 0: |
| 796 | image_features = image_features[:, 1:] |
| 797 | image_features = image_features.view(B, T, N, -1) |
| 798 | return image_features |
| 799 | |
| 800 | def forward( |
| 801 | self, |
| 802 | images: torch.Tensor, |
| 803 | token_pooling: torch.Tensor, |
| 804 | ) -> torch.Tensor: |
| 805 | # image_features shape: |
| 806 | # (batch_size, num_crops(=num_image), num_patch, nximage_emb_dim) |
| 807 | batch_size, num_image = images.shape[:2] |
| 808 | images = images.to(device=self.device, dtype=self.dtype) |
| 809 | image_features = self.encode_image(images) |
| 810 | |
| 811 | dim = image_features.shape[-1] |
| 812 | valid = token_pooling >= 0 |
| 813 | valid_token = torch.any(valid, -1) |
| 814 | |
| 815 | # Use `token_pooling` to arange the features for image pooling |
| 816 | batch_idx = torch.arange( |
| 817 | token_pooling.shape[0], |
| 818 | dtype=torch.long, |
| 819 | device=token_pooling.device, |
| 820 | ) |
| 821 | batch_idx = torch.tile( |
| 822 | batch_idx.view(batch_size, 1, 1), |
| 823 | [1, token_pooling.shape[1], token_pooling.shape[2]], |
| 824 | ) |
| 825 | |
| 826 | # Now [batch, num_features, num_pooled_patches, dim] |
| 827 | to_pool = image_features.reshape(batch_size, -1, dim)[ |
| 828 | batch_idx, torch.clip(token_pooling, 0) |
| 829 | ] |
| 830 | to_pool = to_pool * valid.to(self.dtype)[:, :, :, None] |
| 831 | to_pool = to_pool.reshape([-1, token_pooling.shape[-1], dim]) |
| 832 | if self.adapter_config.pooling_attention_mask: |
| 833 | attn_mask = valid.reshape([-1, 1, 1, valid.shape[-1]]) |
| 834 | denom = valid.view(-1, to_pool.shape[-2]).float().sum(-1) |
| 835 | denom = torch.where(denom == 0, 1, denom) |
| 836 | query = to_pool.sum(-2, keepdim=True) / denom[:, None, None].to( |
| 837 | to_pool.dtype |
| 838 | ) |
| 839 | else: |
| 840 | attn_mask = None |
| 841 | query = to_pool.mean(-2, keepdim=True) |
| 842 | |
| 843 | pooled_features = self.image_pooling_2d(query, to_pool, attn_mask=attn_mask) |
| 844 | pooled_features = pooled_features.reshape( |
| 845 | [batch_size, -1, pooled_features.shape[-1]] |
| 846 | ) |
| 847 | |
| 848 | # MLP layer to map the feature. |
| 849 | pooled_features = self.image_projector(pooled_features) |
| 850 | return pooled_features.view(-1, pooled_features.shape[-1])[ |
| 851 | valid_token.flatten() |
| 852 | ] |
| 853 | |
| 854 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 855 | loader = AutoWeightsLoader(self) |
| 856 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 857 | |
| 858 | |
| 859 | class Molmo2Attention(nn.Module): |
| 860 | """Molmo2's LLM Attention.""" |
| 861 | |
| 862 | def __init__( |
| 863 | self, |
| 864 | config: TextConfig, |
| 865 | rope_parameters: dict[str, Any], |
| 866 | cache_config: CacheConfig | None = None, |
| 867 | quant_config: QuantizationConfig | None = None, |
| 868 | prefix: str = "", |
| 869 | ) -> None: |
| 870 | super().__init__() |
| 871 | self.hidden_size = config.hidden_size |
| 872 | self.tp_size = get_tensor_model_parallel_world_size() |
| 873 | self.total_num_heads = config.num_attention_heads |
| 874 | |
| 875 | assert self.hidden_size % self.total_num_heads == 0 |
| 876 | assert self.total_num_heads % self.tp_size == 0 |
| 877 | |
| 878 | self.num_heads = self.total_num_heads // self.tp_size |
| 879 | self.total_num_kv_heads = config.num_key_value_heads |
| 880 | if self.total_num_kv_heads >= self.tp_size: |
| 881 | assert self.total_num_kv_heads % self.tp_size == 0 |
| 882 | else: |
| 883 | assert self.tp_size % self.total_num_kv_heads == 0 |
| 884 | self.num_kv_heads = max(1, self.total_num_kv_heads // self.tp_size) |
| 885 | self.head_dim = config.head_dim |
| 886 | |
| 887 | self.q_size = self.num_heads * self.head_dim |
| 888 | self.kv_size = self.num_kv_heads * self.head_dim |
| 889 | self.max_position_embeddings = config.max_position_embeddings |
| 890 | self.rope_theta = config.rope_theta |
| 891 | |
| 892 | # Attention input projection. Projects x -> (q, k, v) |
| 893 | self.qkv_proj = QKVParallelLinear( |
| 894 | self.hidden_size, |
| 895 | self.head_dim, |
| 896 | self.total_num_heads, |
| 897 | self.total_num_kv_heads, |
| 898 | bias=config.qkv_bias, |
| 899 | quant_config=quant_config, |
| 900 | ) |
| 901 | |
| 902 | self.tp_rank: int | None = None |
| 903 | self.k_norm: nn.Module | None = None |
| 904 | self.q_norm: nn.Module | None = None |
| 905 | self.qk_norm_type: str | None = None |
| 906 | if config.use_qk_norm: |
| 907 | k_norm_size = ( |
| 908 | self.head_dim |
| 909 | if config.qk_norm_type == "qwen3" |
| 910 | else self.total_num_kv_heads * self.head_dim |
| 911 | ) |
| 912 | self.tp_rank = get_tensor_model_parallel_rank() |
| 913 | self.k_norm = RMSNorm(k_norm_size, eps=config.layer_norm_eps) |
| 914 | q_norm_size = ( |
| 915 | self.head_dim |
| 916 | if config.qk_norm_type == "qwen3" |
| 917 | else self.total_num_heads * self.head_dim |
| 918 | ) |
| 919 | self.q_norm = RMSNorm(q_norm_size, eps=config.layer_norm_eps) |
| 920 | self.qk_norm_type = config.qk_norm_type |
| 921 | # Rotary embeddings. Rope scaling is only applied on full attention layers. |
| 922 | layer_idx = extract_layer_index(prefix) |
| 923 | if ( |
| 924 | config.rope_scaling_layers is not None |
| 925 | and layer_idx not in config.rope_scaling_layers |
| 926 | ): |
| 927 | rope_theta = rope_parameters["rope_theta"] |
| 928 | rope_parameters = {"rope_type": "default", "rope_theta": rope_theta} |
| 929 | self.rotary_emb = get_rope( |
| 930 | self.head_dim, |
| 931 | max_position=self.max_position_embeddings, |
| 932 | rope_parameters=rope_parameters, |
| 933 | ) |
| 934 | self.scaling = self.head_dim**-0.5 |
| 935 | self.attn = Attention( |
| 936 | self.num_heads, |
| 937 | self.head_dim, |
| 938 | self.scaling, |
| 939 | num_kv_heads=self.num_kv_heads, |
| 940 | cache_config=cache_config, |
| 941 | quant_config=quant_config, |
| 942 | prefix=f"{prefix}.attn", |
| 943 | ) |
| 944 | |
| 945 | # Attention output projection. |
| 946 | self.o_proj = RowParallelLinear( |
| 947 | self.total_num_heads * self.head_dim, |
| 948 | self.hidden_size, |
| 949 | bias=False, |
| 950 | quant_config=quant_config, |
| 951 | ) |
| 952 | |
| 953 | def _apply_qk_norm( |
| 954 | self, |
| 955 | q: torch.Tensor, |
| 956 | k: torch.Tensor, |
| 957 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 958 | assert self.q_norm is not None |
| 959 | assert self.k_norm is not None |
| 960 | if self.tp_size > 1: |
| 961 | q = tensor_model_parallel_all_gather(q.contiguous()) |
| 962 | k = tensor_model_parallel_all_gather(k.contiguous()) |
| 963 | q = self.q_norm(q) |
| 964 | k = self.k_norm(k) |
| 965 | if self.tp_size > 1: |
| 966 | assert self.tp_rank is not None |
| 967 | splitter = partial(split_tensor_along_last_dim, num_partitions=self.tp_size) |
| 968 | q = splitter(q)[self.tp_rank] |
| 969 | k = splitter(k)[self.tp_rank] |
| 970 | return q, k |
| 971 | |
| 972 | def forward( |
| 973 | self, |
| 974 | positions: torch.Tensor, |
| 975 | hidden_states: torch.Tensor, |
| 976 | **kwargs: object, |
| 977 | ) -> torch.Tensor: |
| 978 | qkv, _ = self.qkv_proj(hidden_states) |
| 979 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 980 | if ( |
| 981 | self.q_norm is not None |
| 982 | and self.k_norm is not None |
| 983 | and self.qk_norm_type == "olmo" |
| 984 | ): |
| 985 | q, k = self._apply_qk_norm(q, k) |
| 986 | elif self.q_norm is not None and self.k_norm is not None: |
| 987 | q_by_head = q.view( |
| 988 | *q.shape[:-1], |
| 989 | q.shape[-1] // self.head_dim, |
| 990 | self.head_dim, |
| 991 | ) |
| 992 | q_by_head = self.q_norm(q_by_head) |
| 993 | q = q_by_head.view(q.shape) |
| 994 | k_by_head = k.view( |
| 995 | *k.shape[:-1], |
| 996 | k.shape[-1] // self.head_dim, |
| 997 | self.head_dim, |
| 998 | ) |
| 999 | k_by_head = self.k_norm(k_by_head) |
| 1000 | k = k_by_head.view(k.shape) |
| 1001 | q, k = self.rotary_emb(positions, q, k) |
| 1002 | attn_output = self.attn(q, k, v) |
| 1003 | |
| 1004 | output, _ = self.o_proj(attn_output) |
| 1005 | return output |
| 1006 | |
| 1007 | |
| 1008 | class LanguageModelMLP(nn.Module): |
| 1009 | """Molmo2's LLM mlp.""" |
| 1010 | |
| 1011 | def __init__( |
| 1012 | self, |
| 1013 | input_dim: int, |
| 1014 | intermediate_size: int, |
| 1015 | hidden_act: str, |
| 1016 | quant_config: QuantizationConfig | None = None, |
| 1017 | ) -> None: |
| 1018 | super().__init__() |
| 1019 | |
| 1020 | self.up_gate_proj = MergedColumnParallelLinear( |
| 1021 | input_dim, |
| 1022 | [intermediate_size] * 2, |
| 1023 | bias=False, |
| 1024 | quant_config=quant_config, |
| 1025 | ) |
| 1026 | # Activation function. |
| 1027 | assert hidden_act == "silu" |
| 1028 | self.act_fn = MulAndSilu() |
| 1029 | # Feed-forward output projection. |
| 1030 | self.down_proj = RowParallelLinear( |
| 1031 | intermediate_size, |
| 1032 | input_dim, |
| 1033 | bias=False, |
| 1034 | quant_config=quant_config, |
| 1035 | ) |
| 1036 | |
| 1037 | def forward( |
| 1038 | self, |
| 1039 | x: torch.Tensor, |
| 1040 | ) -> torch.Tensor: |
| 1041 | up_gate, _ = self.up_gate_proj(x) |
| 1042 | x = self.act_fn(up_gate) |
| 1043 | x, _ = self.down_proj(x) |
| 1044 | return x |
| 1045 | |
| 1046 | |
| 1047 | class Molmo2DecoderLayer(nn.Module): |
| 1048 | def __init__( |
| 1049 | self, |
| 1050 | config: TextConfig, |
| 1051 | rope_parameters: dict[str, Any], |
| 1052 | cache_config: CacheConfig | None = None, |
| 1053 | quant_config: QuantizationConfig | None = None, |
| 1054 | prefix: str = "", |
| 1055 | ) -> None: |
| 1056 | super().__init__() |
| 1057 | # Attention block. |
| 1058 | self.self_attn = Molmo2Attention( |
| 1059 | config, |
| 1060 | rope_parameters, |
| 1061 | cache_config, |
| 1062 | quant_config, |
| 1063 | prefix=f"{prefix}.self_attn", |
| 1064 | ) |
| 1065 | |
| 1066 | # MLP block. |
| 1067 | self.mlp = LanguageModelMLP( |
| 1068 | config.hidden_size, |
| 1069 | config.intermediate_size, |
| 1070 | config.hidden_act, |
| 1071 | quant_config, |
| 1072 | ) |
| 1073 | |
| 1074 | # LayerNorm |
| 1075 | self.input_layernorm = RMSNorm(config.hidden_size, eps=config.layer_norm_eps) |
| 1076 | self.post_attention_layernorm = RMSNorm( |
| 1077 | config.hidden_size, |
| 1078 | eps=config.layer_norm_eps, |
| 1079 | ) |
| 1080 | |
| 1081 | def forward( |
| 1082 | self, |
| 1083 | positions: torch.Tensor, |
| 1084 | hidden_states: torch.Tensor, |
| 1085 | residual: torch.Tensor | None, |
| 1086 | **kwargs: object, |
| 1087 | ) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]: |
| 1088 | # Self Attention |
| 1089 | if residual is None: |
| 1090 | residual = hidden_states |
| 1091 | hidden_states = self.input_layernorm(hidden_states) |
| 1092 | else: |
| 1093 | hidden_states, residual = self.input_layernorm(hidden_states, residual) |
| 1094 | hidden_states = self.self_attn( |
| 1095 | positions=positions, |
| 1096 | hidden_states=hidden_states, |
| 1097 | **kwargs, |
| 1098 | ) |
| 1099 | |
| 1100 | hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) |
| 1101 | hidden_states = self.mlp(hidden_states) |
| 1102 | return hidden_states, residual |
| 1103 | |
| 1104 | |
| 1105 | class Molmo2DecoderNormAfterLayer(Molmo2DecoderLayer): |
| 1106 | def forward( |
| 1107 | self, |
| 1108 | positions: torch.Tensor, |
| 1109 | hidden_states: torch.Tensor, |
| 1110 | residual: torch.Tensor | None, |
| 1111 | **kwargs: object, |
| 1112 | ) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]: |
| 1113 | # Self Attention |
| 1114 | residual = hidden_states |
| 1115 | hidden_states = self.self_attn( |
| 1116 | positions=positions, |
| 1117 | hidden_states=hidden_states, |
| 1118 | **kwargs, |
| 1119 | ) |
| 1120 | |
| 1121 | hidden_states = self.input_layernorm(hidden_states) |
| 1122 | hidden_states = hidden_states + residual |
| 1123 | residual = hidden_states |
| 1124 | |
| 1125 | hidden_states = self.mlp(hidden_states) |
| 1126 | hidden_states = self.post_attention_layernorm(hidden_states) |
| 1127 | hidden_states = hidden_states + residual |
| 1128 | residual = None |
| 1129 | return hidden_states, residual |
| 1130 | |
| 1131 | |
| 1132 | @support_torch_compile |
| 1133 | class Molmo2TextModel(nn.Module, SupportsQuant): |
| 1134 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 1135 | super().__init__() |
| 1136 | |
| 1137 | config = vllm_config.model_config.hf_config |
| 1138 | cache_config = vllm_config.cache_config |
| 1139 | quant_config = vllm_config.quant_config |
| 1140 | |
| 1141 | self.config = config |
| 1142 | |
| 1143 | if hasattr(config, "text_config"): |
| 1144 | hf_text_config = config.text_config |
| 1145 | else: |
| 1146 | hf_text_config = config.llm_config |
| 1147 | |
| 1148 | kwargs = {} |
| 1149 | for field in fields(TextConfig): |
| 1150 | kwargs[field.name] = getattr(hf_text_config, field.name) |
| 1151 | text_config = TextConfig(**kwargs) |
| 1152 | |
| 1153 | self.embedding_size = text_config.vocab_size |
| 1154 | self.embedding_size += text_config.additional_vocab_size or 0 |
| 1155 | self.embed_tokens = VocabParallelEmbedding( |
| 1156 | self.embedding_size, |
| 1157 | text_config.hidden_size, |
| 1158 | quant_config=quant_config, |
| 1159 | ) |
| 1160 | |
| 1161 | decoder_layer = ( |
| 1162 | Molmo2DecoderNormAfterLayer |
| 1163 | if text_config.norm_after |
| 1164 | else Molmo2DecoderLayer |
| 1165 | ) |
| 1166 | self.start_layer, self.end_layer, self.layers = make_layers( |
| 1167 | text_config.num_hidden_layers, |
| 1168 | lambda prefix: decoder_layer( |
| 1169 | text_config, |
| 1170 | hf_text_config.rope_parameters, |
| 1171 | cache_config=cache_config, |
| 1172 | quant_config=quant_config, |
| 1173 | prefix=prefix, |
| 1174 | ), |
| 1175 | prefix=f"{prefix}.layers", |
| 1176 | ) |
| 1177 | |
| 1178 | self.norm = RMSNorm(text_config.hidden_size, eps=text_config.layer_norm_eps) |
| 1179 | |
| 1180 | self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( |
| 1181 | ["hidden_states", "residual"], |
| 1182 | text_config.hidden_size, |
| 1183 | ) |
| 1184 | |
| 1185 | def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 1186 | return self.embed_tokens(input_ids) |
| 1187 | |
| 1188 | def forward( |
| 1189 | self, |
| 1190 | input_ids: torch.Tensor | None, |
| 1191 | positions: torch.Tensor, |
| 1192 | intermediate_tensors: IntermediateTensors | None = None, |
| 1193 | inputs_embeds: torch.Tensor | None = None, |
| 1194 | **kwargs: object, |
| 1195 | ) -> torch.Tensor: |
| 1196 | if get_pp_group().is_first_rank: |
| 1197 | if inputs_embeds is not None: |
| 1198 | hidden_states = inputs_embeds |
| 1199 | else: |
| 1200 | hidden_states = self.embed_tokens(input_ids) |
| 1201 | residual = None |
| 1202 | else: |
| 1203 | assert intermediate_tensors is not None |
| 1204 | hidden_states = intermediate_tensors["hidden_states"] |
| 1205 | residual = intermediate_tensors["residual"] |
| 1206 | |
| 1207 | # Apply blocks one-by-one. |
| 1208 | for layer in islice(self.layers, self.start_layer, self.end_layer): |
| 1209 | hidden_states, residual = layer( |
| 1210 | positions, |
| 1211 | hidden_states, |
| 1212 | residual, |
| 1213 | **kwargs, |
| 1214 | ) |
| 1215 | if not get_pp_group().is_last_rank: |
| 1216 | return IntermediateTensors( |
| 1217 | {"hidden_states": hidden_states, "residual": residual} |
| 1218 | ) |
| 1219 | if residual is not None: |
| 1220 | hidden_states, _ = self.norm(hidden_states, residual) |
| 1221 | else: |
| 1222 | hidden_states = self.norm(hidden_states) |
| 1223 | return hidden_states |
| 1224 | |
| 1225 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1226 | return AutoWeightsLoader(self).load_weights(weights) |
| 1227 | |
| 1228 | |
| 1229 | def get_patches_grid_size( |
| 1230 | *, |
| 1231 | image_h: int, |
| 1232 | image_w: int, |
| 1233 | patch_size: int, |
| 1234 | pool_h: int, |
| 1235 | pool_w: int, |
| 1236 | ) -> tuple[int, int]: |
| 1237 | patch_h = image_h // patch_size |
| 1238 | patch_w = image_w // patch_size |
| 1239 | h_pad = round_down(patch_h + pool_h - 1, pool_h) - patch_h |
| 1240 | w_pad = round_down(patch_w + pool_w - 1, pool_w) - patch_w |
| 1241 | nrows = (patch_h + h_pad) // pool_h |
| 1242 | ncols = (patch_w + w_pad) // pool_w |
| 1243 | |
| 1244 | return nrows, ncols |
| 1245 | |
| 1246 | |
| 1247 | def get_candidate_tilings(max_num: int) -> list[tuple[int, int]]: |
| 1248 | tilings = [ |
| 1249 | (i, j) |
| 1250 | for i in range(1, max_num + 1) |
| 1251 | for j in range(1, max_num + 1) |
| 1252 | if i * j <= max_num |
| 1253 | ] |
| 1254 | return sorted(tilings, key=lambda x: (x[0] * x[1], x[0])) |
| 1255 | |
| 1256 | |
| 1257 | def select_tiling( |
| 1258 | *, |
| 1259 | height: int, |
| 1260 | width: int, |
| 1261 | patch_size: int, |
| 1262 | max_num_patches: int, |
| 1263 | ): |
| 1264 | tilings = get_candidate_tilings(max_num_patches) |
| 1265 | candidate_tilings = np.array(tilings, dtype=np.int32) |
| 1266 | candidate_resolutions = candidate_tilings * patch_size |
| 1267 | |
| 1268 | original_size = np.array([height, width], dtype=np.float32) |
| 1269 | required_scale_d = candidate_resolutions.astype(np.float32) / original_size |
| 1270 | required_scale = required_scale_d.min(axis=-1, keepdims=True) |
| 1271 | |
| 1272 | if (required_scale < 1).all(): |
| 1273 | ix = required_scale.argmax() |
| 1274 | else: |
| 1275 | ix = np.where(required_scale < 1.0, 10e9, required_scale).argmin() |
| 1276 | |
| 1277 | return candidate_tilings[ix] |
| 1278 | |
| 1279 | |
| 1280 | def get_image_size(image: ImageInput) -> ImageSize: |
| 1281 | if isinstance(image, Image): |
| 1282 | return ImageSize(*image.size) |
| 1283 | elif isinstance(image, (np.ndarray, torch.Tensor)): |
| 1284 | assert image.ndim == 3 |
| 1285 | h, w, c = image.shape |
| 1286 | assert c in [1, 3] |
| 1287 | return ImageSize(w, h) |
| 1288 | else: |
| 1289 | raise ValueError(f"Unknown image type: {type(image)}") |
| 1290 | |
| 1291 | |
| 1292 | def exif_transpose( |
| 1293 | images: ImageInput | None, |
| 1294 | ) -> ImageInput | None: |
| 1295 | if images is None: |
| 1296 | return None |
| 1297 | if images is not None and isinstance(images, (list, tuple)): |
| 1298 | images = [ |
| 1299 | exif_transpose(img) if isinstance(img, Image) else img for img in images |
| 1300 | ] |
| 1301 | elif images is not None and isinstance(images, Image): |
| 1302 | images = normalize_image(images) |
| 1303 | return images |
| 1304 | |
| 1305 | |
| 1306 | def build_flat_image_bool_length( |
| 1307 | image_grids: torch.LongTensor, |
| 1308 | hf_config: PreTrainedConfig, |
| 1309 | image_use_col_tokens: bool = True, |
| 1310 | use_single_crop_col_tokens: bool | None = None, |
| 1311 | use_single_crop_start_token: bool = True, |
| 1312 | ) -> tuple[torch.LongTensor, torch.LongTensor]: |
| 1313 | image_patch_id = hf_config.image_patch_id |
| 1314 | low_res_image_start_id = hf_config.low_res_image_start_token_id |
| 1315 | image_start_id = hf_config.image_start_token_id |
| 1316 | image_col_id = hf_config.image_col_id |
| 1317 | image_end_id = hf_config.image_end_token_id |
| 1318 | |
| 1319 | device = image_grids.device |
| 1320 | B = image_grids.shape[0] |
| 1321 | |
| 1322 | resized_h = image_grids[:, 0] |
| 1323 | resized_w = image_grids[:, 1] |
| 1324 | h = image_grids[:, 2] |
| 1325 | w = image_grids[:, 3] |
| 1326 | |
| 1327 | low_res_use_col_tokens = ( |
| 1328 | image_use_col_tokens |
| 1329 | if use_single_crop_col_tokens is None |
| 1330 | else use_single_crop_col_tokens |
| 1331 | ) |
| 1332 | low_res_extra = int(low_res_use_col_tokens) |
| 1333 | high_res_extra = int(image_use_col_tokens) |
| 1334 | |
| 1335 | lengths = ( |
| 1336 | resized_h * (resized_w + low_res_extra) + h * (w + high_res_extra) + 4 |
| 1337 | ) # [B] |
| 1338 | total_len = int(lengths.sum().item()) |
| 1339 | |
| 1340 | flat = torch.empty(total_len, dtype=torch.long, device=device) |
| 1341 | |
| 1342 | offset = 0 |
| 1343 | for i in range(B): |
| 1344 | resized_h_i, resized_w_i, h_i, w_i = image_grids[i].tolist() |
| 1345 | L_i = int(lengths[i].item()) |
| 1346 | |
| 1347 | idx = offset |
| 1348 | |
| 1349 | flat[idx] = ( |
| 1350 | low_res_image_start_id if use_single_crop_start_token else image_start_id |
| 1351 | ) |
| 1352 | idx += 1 |
| 1353 | |
| 1354 | low_res_block_len = resized_w_i + low_res_extra |
| 1355 | if low_res_block_len > 0 and resized_h_i > 0: |
| 1356 | line = torch.empty(low_res_block_len, dtype=torch.long, device=device) |
| 1357 | if resized_w_i > 0: |
| 1358 | line[:resized_w_i] = image_patch_id |
| 1359 | if low_res_use_col_tokens: |
| 1360 | line[resized_w_i] = image_col_id |
| 1361 | |
| 1362 | block = line.repeat(resized_h_i) |
| 1363 | flat[idx : idx + resized_h_i * low_res_block_len] = block |
| 1364 | idx += resized_h_i * low_res_block_len |
| 1365 | |
| 1366 | flat[idx] = image_end_id |
| 1367 | idx += 1 |
| 1368 | |
| 1369 | flat[idx] = image_start_id |
| 1370 | idx += 1 |
| 1371 | |
| 1372 | block_len = w_i + high_res_extra |
| 1373 | if block_len > 0 and h_i > 0: |
| 1374 | line = torch.empty(block_len, dtype=torch.long, device=device) |
| 1375 | if w_i > 0: |
| 1376 | line[:w_i] = image_patch_id |
| 1377 | if image_use_col_tokens: |
| 1378 | line[w_i] = image_col_id |
| 1379 | |
| 1380 | block = line.repeat(h_i) |
| 1381 | flat[idx : idx + h_i * block_len] = block |
| 1382 | idx += h_i * block_len |
| 1383 | |
| 1384 | flat[idx] = image_end_id |
| 1385 | idx += 1 |
| 1386 | |
| 1387 | assert idx - offset == L_i |
| 1388 | |
| 1389 | offset += L_i |
| 1390 | |
| 1391 | return flat, lengths |
| 1392 | |
| 1393 | |
| 1394 | def build_flat_video_bool_length( |
| 1395 | video_grids: torch.LongTensor, |
| 1396 | hf_config: PreTrainedConfig, |
| 1397 | ) -> tuple[torch.LongTensor, torch.LongTensor]: |
| 1398 | image_patch_id = hf_config.image_patch_id |
| 1399 | frame_start_id = hf_config.frame_start_token_id |
| 1400 | frame_end_id = hf_config.frame_end_token_id |
| 1401 | |
| 1402 | device = video_grids.device |
| 1403 | B = video_grids.shape[0] |
| 1404 | |
| 1405 | t = video_grids[:, 0] |
| 1406 | resized_h = video_grids[:, 1] |
| 1407 | resized_w = video_grids[:, 2] |
| 1408 | |
| 1409 | P = resized_h * resized_w |
| 1410 | block_len = P + 2 |
| 1411 | lengths = t * block_len |
| 1412 | |
| 1413 | total_len = int(lengths.sum().item()) |
| 1414 | flat = torch.empty(total_len, dtype=torch.long, device=device) |
| 1415 | |
| 1416 | offset = 0 |
| 1417 | for i in range(B): |
| 1418 | ti = int(t[i].item()) |
| 1419 | Pi = int(P[i].item()) |
| 1420 | Li = int(lengths[i].item()) |
| 1421 | |
| 1422 | block = torch.empty(Pi + 2, dtype=torch.long, device=device) |
| 1423 | block[0] = frame_start_id |
| 1424 | if Pi > 0: |
| 1425 | block[1 : 1 + Pi] = image_patch_id |
| 1426 | block[-1] = frame_end_id |
| 1427 | |
| 1428 | seq = block.repeat(ti) |
| 1429 | |
| 1430 | flat[offset : offset + Li] = seq |
| 1431 | offset += Li |
| 1432 | |
| 1433 | return flat, lengths |
| 1434 | |
| 1435 | |
| 1436 | def get_candidate_target_fps( |
| 1437 | video_fps: int | float, |
| 1438 | sampling_fps: int | float, |
| 1439 | max_fps: int | float = _MAX_VIDEO_FPS, |
| 1440 | ) -> list[float]: |
| 1441 | """Return the subset of `video_fps` factors that remain multiples |
| 1442 | of `sampling_fps`. |
| 1443 | |
| 1444 | Examples: |
| 1445 | >>> get_candidate_target_fps(video_fps=6, sampling_fps=2) |
| 1446 | [2, 6] |
| 1447 | >>> get_candidate_target_fps(video_fps=5, sampling_fps=1) |
| 1448 | [1, 5] |
| 1449 | >>> get_candidate_target_fps(video_fps=2, sampling_fps=2) |
| 1450 | [2] |
| 1451 | >>> get_candidate_target_fps(video_fps=5, sampling_fps=2) |
| 1452 | Traceback (most recent call last): |
| 1453 | ... |
| 1454 | ValueError: sampling_fps=2 must divide video_fps=5 to produce |
| 1455 | consistent frame steps. |
| 1456 | |
| 1457 | """ |
| 1458 | video_fps = int(video_fps) |
| 1459 | sampling_fps = int(sampling_fps) |
| 1460 | max_fps = int(max_fps) |
| 1461 | |
| 1462 | if sampling_fps is None: |
| 1463 | raise ValueError("sampling_fps must be provided") |
| 1464 | if video_fps <= 0 or sampling_fps <= 0: |
| 1465 | raise ValueError( |
| 1466 | "video_fps and sampling_fps must be positive " |
| 1467 | f"(got {video_fps}, {sampling_fps})" |
| 1468 | ) |
| 1469 | if video_fps % sampling_fps != 0: |
| 1470 | raise ValueError( |
| 1471 | f"sampling_fps={sampling_fps} must divide video_fps={video_fps}." |
| 1472 | ) |
| 1473 | |
| 1474 | candidates = [] |
| 1475 | for candidate in range(sampling_fps, video_fps + 1, sampling_fps): |
| 1476 | if candidate > max_fps: |
| 1477 | break |
| 1478 | if video_fps % candidate == 0: |
| 1479 | candidates.append(float(candidate)) |
| 1480 | |
| 1481 | return candidates |
| 1482 | |
| 1483 | |
| 1484 | def get_target_fps( |
| 1485 | video_fps: float, |
| 1486 | max_frames: int, |
| 1487 | total_frames: int, |
| 1488 | frame_sample_mode: str, |
| 1489 | candidate_target_fps: list[float], |
| 1490 | ) -> float | None: |
| 1491 | """Get the target fps that best spans the video and has the most frames sampled.""" |
| 1492 | num_frames_sampled = 0 |
| 1493 | selected_target_fps = None |
| 1494 | for target_fps in candidate_target_fps: |
| 1495 | step_size = max(int(video_fps / target_fps), 1) |
| 1496 | num_frames_sampled_at_fps = int(total_frames / step_size) |
| 1497 | if num_frames_sampled == 0: |
| 1498 | if ( |
| 1499 | "uniform" in frame_sample_mode |
| 1500 | and num_frames_sampled_at_fps > max_frames |
| 1501 | ): |
| 1502 | break |
| 1503 | selected_target_fps = target_fps |
| 1504 | num_frames_sampled = num_frames_sampled_at_fps |
| 1505 | |
| 1506 | else: |
| 1507 | # the candidate sampling fps increases so frame count can't decrease |
| 1508 | assert num_frames_sampled <= num_frames_sampled_at_fps |
| 1509 | if num_frames_sampled_at_fps > max_frames: |
| 1510 | # choose the sampling fps that spans the video |
| 1511 | continue |
| 1512 | |
| 1513 | elif num_frames_sampled_at_fps > num_frames_sampled: |
| 1514 | # both are less than max_frames; choose the one with higher |
| 1515 | # density of frames sampled |
| 1516 | selected_target_fps = target_fps |
| 1517 | num_frames_sampled = num_frames_sampled_at_fps |
| 1518 | return selected_target_fps |
| 1519 | |
| 1520 | |
| 1521 | def get_frame_times_and_chosen_fps( |
| 1522 | selected_target_fps, total_frames, max_frames, video_fps |
| 1523 | ): |
| 1524 | if selected_target_fps is None: |
| 1525 | frame_indices = np.linspace( |
| 1526 | 0, total_frames, max_frames, endpoint=False, dtype=int |
| 1527 | ) |
| 1528 | else: |
| 1529 | step_size = max(int(video_fps / selected_target_fps), 1) |
| 1530 | frame_indices = np.arange(0, total_frames, step_size) |
| 1531 | if len(frame_indices) > max_frames: |
| 1532 | frame_indices = frame_indices[:max_frames] |
| 1533 | return selected_target_fps, frame_indices |
| 1534 | |
| 1535 | |
| 1536 | class Molmo2ProcessingInfo(BaseProcessingInfo): |
| 1537 | def get_data_parser(self): |
| 1538 | return MultiModalDataParser( |
| 1539 | video_needs_metadata=True, |
| 1540 | expected_hidden_size=self._get_expected_hidden_size(), |
| 1541 | ) |
| 1542 | |
| 1543 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 1544 | return {"image": None, "video": 1} |
| 1545 | |
| 1546 | def select_tiling( |
| 1547 | self, |
| 1548 | *, |
| 1549 | image_width: int, |
| 1550 | image_height: int, |
| 1551 | image_processor: BaseImageProcessor, |
| 1552 | ) -> tuple[int, int]: |
| 1553 | max_crops = image_processor.max_crops |
| 1554 | left_margin, right_margin = image_processor.overlap_margins |
| 1555 | base_image_input_d = image_processor.patch_size |
| 1556 | |
| 1557 | total_margin_pixels = base_image_input_d * (right_margin + left_margin) |
| 1558 | crop_patches = image_processor.size["height"] // base_image_input_d |
| 1559 | crop_window_patches = crop_patches - (right_margin + left_margin) |
| 1560 | crop_window_size = crop_window_patches * base_image_input_d |
| 1561 | tiling_h, tiling_w = select_tiling( |
| 1562 | height=image_height - total_margin_pixels, |
| 1563 | width=image_width - total_margin_pixels, |
| 1564 | patch_size=crop_window_size, |
| 1565 | max_num_patches=max_crops, |
| 1566 | ) |
| 1567 | |
| 1568 | return tiling_w, tiling_h |
| 1569 | |
| 1570 | def get_base_grid_size( |
| 1571 | self, |
| 1572 | image_processor: BaseImageProcessor | BaseVideoProcessor, |
| 1573 | ) -> tuple[int, int]: |
| 1574 | nrows, ncols = get_patches_grid_size( |
| 1575 | image_h=image_processor.size["height"], |
| 1576 | image_w=image_processor.size["width"], |
| 1577 | patch_size=image_processor.patch_size, |
| 1578 | pool_h=image_processor.pooling_size[0], |
| 1579 | pool_w=image_processor.pooling_size[1], |
| 1580 | ) |
| 1581 | |
| 1582 | return ncols, nrows |
| 1583 | |
| 1584 | def get_patches_grid_size( |
| 1585 | self, |
| 1586 | *, |
| 1587 | image_width: int, |
| 1588 | image_height: int, |
| 1589 | image_processor: BaseImageProcessor, |
| 1590 | ) -> tuple[int, int]: |
| 1591 | left_margin, right_margin = image_processor.overlap_margins |
| 1592 | base_image_input_d = image_processor.patch_size |
| 1593 | |
| 1594 | total_margin_pixels = base_image_input_d * (right_margin + left_margin) |
| 1595 | crop_patches = image_processor.size["height"] // base_image_input_d |
| 1596 | crop_window_patches = crop_patches - (right_margin + left_margin) |
| 1597 | crop_window_size = crop_window_patches * base_image_input_d |
| 1598 | |
| 1599 | tiling_w, tiling_h = self.select_tiling( |
| 1600 | image_height=image_height, |
| 1601 | image_width=image_width, |
| 1602 | image_processor=image_processor, |
| 1603 | ) |
| 1604 | |
| 1605 | nrows, ncols = get_patches_grid_size( |
| 1606 | image_h=tiling_h * crop_window_size + total_margin_pixels, |
| 1607 | image_w=tiling_w * crop_window_size + total_margin_pixels, |
| 1608 | patch_size=base_image_input_d, |
| 1609 | pool_h=image_processor.pooling_size[0], |
| 1610 | pool_w=image_processor.pooling_size[1], |
| 1611 | ) |
| 1612 | |
| 1613 | return ncols, nrows |
| 1614 | |
| 1615 | def get_num_image_tokens( |
| 1616 | self, |
| 1617 | *, |
| 1618 | image_height: int, |
| 1619 | image_width: int, |
| 1620 | processor: ProcessorMixin, |
| 1621 | ) -> int: |
| 1622 | image_processor = processor.image_processor |
| 1623 | |
| 1624 | resize_ncols, resize_nrows = self.get_base_grid_size(image_processor) |
| 1625 | # start/end tokens + image patch token + col tokens |
| 1626 | if processor.use_single_crop_col_tokens is not None: |
| 1627 | use_col_tokens = processor.use_single_crop_col_tokens |
| 1628 | else: |
| 1629 | use_col_tokens = processor.image_use_col_tokens |
| 1630 | extra = 2 + resize_nrows * (resize_ncols + int(use_col_tokens)) |
| 1631 | overlap_ncols, overlap_nrows = self.get_patches_grid_size( |
| 1632 | image_height=image_height, |
| 1633 | image_width=image_width, |
| 1634 | image_processor=image_processor, |
| 1635 | ) |
| 1636 | joint = 2 + overlap_nrows * ( |
| 1637 | overlap_ncols + int(processor.image_use_col_tokens) |
| 1638 | ) |
| 1639 | |
| 1640 | return extra + joint |
| 1641 | |
| 1642 | def get_num_video_tokens( |
| 1643 | self, |
| 1644 | *, |
| 1645 | num_frames: int, |
| 1646 | processor: ProcessorMixin, |
| 1647 | ) -> int: |
| 1648 | video_processor = processor.video_processor |
| 1649 | |
| 1650 | resize_ncols, resize_nrows = self.get_base_grid_size(video_processor) |
| 1651 | # start/end tokens |
| 1652 | extra = 2 + resize_nrows * (resize_ncols + int(processor.video_use_col_tokens)) |
| 1653 | return num_frames * extra |
| 1654 | |
| 1655 | def get_image_size_with_most_features(self) -> ImageSize: |
| 1656 | processor = self.get_hf_processor() |
| 1657 | image_processor = processor.image_processor |
| 1658 | |
| 1659 | left_margin, right_margin = image_processor.overlap_margins |
| 1660 | base_image_input_d = image_processor.patch_size |
| 1661 | |
| 1662 | total_margin_pixels = base_image_input_d * (right_margin + left_margin) |
| 1663 | crop_patches = image_processor.size["height"] // base_image_input_d |
| 1664 | crop_window_patches = crop_patches - (right_margin + left_margin) |
| 1665 | crop_window_size = crop_window_patches * base_image_input_d |
| 1666 | |
| 1667 | tilings = get_candidate_tilings(image_processor.max_crops) |
| 1668 | largest_feature_size, largest_feature_pinpoint = 0, None |
| 1669 | |
| 1670 | for hr, wr in tilings: |
| 1671 | height = hr * crop_window_size + total_margin_pixels |
| 1672 | width = wr * crop_window_size + total_margin_pixels |
| 1673 | |
| 1674 | feat_size = self.get_num_image_tokens( |
| 1675 | image_height=height, |
| 1676 | image_width=width, |
| 1677 | processor=processor, |
| 1678 | ) |
| 1679 | if feat_size > largest_feature_size: |
| 1680 | largest_feature_size = feat_size |
| 1681 | largest_feature_pinpoint = ImageSize(width=width, height=height) |
| 1682 | |
| 1683 | if largest_feature_size == 0 or largest_feature_pinpoint is None: |
| 1684 | raise ValueError("Cannot have a largest feature size of 0!") |
| 1685 | |
| 1686 | return largest_feature_pinpoint |
| 1687 | |
| 1688 | def _get_max_video_frames( |
| 1689 | self, |
| 1690 | max_tokens: int, |
| 1691 | processor: ProcessorMixin, |
| 1692 | ) -> int: |
| 1693 | num_tokens_per_frame = self.get_num_video_tokens( |
| 1694 | num_frames=1, |
| 1695 | processor=processor, |
| 1696 | ) |
| 1697 | max_frames = max_tokens // num_tokens_per_frame |
| 1698 | return max(max_frames, 1) |
| 1699 | |
| 1700 | def get_num_frames_with_most_features( |
| 1701 | self, |
| 1702 | seq_len: int, |
| 1703 | mm_counts: Mapping[str, int], |
| 1704 | ) -> int: |
| 1705 | processor = self.get_hf_processor() |
| 1706 | video_processor = processor.video_processor |
| 1707 | |
| 1708 | num_frames = video_processor.num_frames |
| 1709 | max_videos = mm_counts.get("video", 0) |
| 1710 | max_total_frames = self._get_max_video_frames(seq_len, processor) |
| 1711 | max_frames_per_video = min( |
| 1712 | max_total_frames // max(max_videos, 1), |
| 1713 | num_frames, |
| 1714 | ) |
| 1715 | return max(max_frames_per_video, 1) |
| 1716 | |
| 1717 | def _sample_frames( |
| 1718 | self, |
| 1719 | total_num_frames: int, |
| 1720 | video_fps: float, |
| 1721 | duration: float, |
| 1722 | frame_sample_mode: str, |
| 1723 | num_frames: int, |
| 1724 | max_fps: int, |
| 1725 | sampling_fps: int, |
| 1726 | ) -> np.ndarray: |
| 1727 | if frame_sample_mode == "uniform_last_frame" and max_fps is not None: |
| 1728 | if total_num_frames <= 2: |
| 1729 | indices = np.arange(total_num_frames).astype(int) |
| 1730 | elif duration > (num_frames - 1) / max_fps: # -1 to include the last frame |
| 1731 | # uniform fallback |
| 1732 | indices = np.linspace( |
| 1733 | 0, |
| 1734 | total_num_frames - 1, |
| 1735 | num=min(num_frames, total_num_frames), |
| 1736 | endpoint=True, |
| 1737 | ).astype(int) |
| 1738 | else: |
| 1739 | float_indices = np.arange( |
| 1740 | 0.0, |
| 1741 | stop=total_num_frames - 1, |
| 1742 | step=float(video_fps / max_fps), |
| 1743 | ) |
| 1744 | if np.round(float_indices[-1]) != total_num_frames - 1: |
| 1745 | float_indices = np.concatenate( |
| 1746 | [float_indices, [total_num_frames - 1]], axis=0 |
| 1747 | ) |
| 1748 | indices = np.round(float_indices).astype(int) |
| 1749 | assert indices[-1] < total_num_frames |
| 1750 | assert len(float_indices) <= num_frames |
| 1751 | elif frame_sample_mode == "uniform_last_frame": |
| 1752 | indices = np.linspace( |
| 1753 | 0, |
| 1754 | total_num_frames - 1, |
| 1755 | num=min(num_frames, total_num_frames), |
| 1756 | endpoint=True, |
| 1757 | ).astype(int) |
| 1758 | elif frame_sample_mode == "fps": |
| 1759 | candidate_target_fps = get_candidate_target_fps(video_fps, sampling_fps) |
| 1760 | selected_target_fps = get_target_fps( |
| 1761 | video_fps, |
| 1762 | num_frames, |
| 1763 | total_num_frames, |
| 1764 | frame_sample_mode, |
| 1765 | candidate_target_fps, |
| 1766 | ) |
| 1767 | _, indices = get_frame_times_and_chosen_fps( |
| 1768 | selected_target_fps, |
| 1769 | total_num_frames, |
| 1770 | num_frames, |
| 1771 | video_fps, |
| 1772 | ) |
| 1773 | else: |
| 1774 | raise NotImplementedError(frame_sample_mode) |
| 1775 | |
| 1776 | return indices |
| 1777 | |
| 1778 | def _get_video_second_idx( |
| 1779 | self, |
| 1780 | metadata: dict[str, Any], |
| 1781 | do_sample_frames: bool | None = None, |
| 1782 | ) -> list[float]: |
| 1783 | processor = self.get_hf_processor() |
| 1784 | video_processor = processor.video_processor |
| 1785 | |
| 1786 | # metadata["fps"] refers to the true fps of the input video. |
| 1787 | video_fps = metadata["fps"] |
| 1788 | frames_indices = metadata.get("frames_indices") |
| 1789 | if do_sample_frames is None: |
| 1790 | do_sample_frames = metadata.get("do_sample_frames", False) |
| 1791 | |
| 1792 | if do_sample_frames: |
| 1793 | # Frame-based sampling is applied in HF video processor |
| 1794 | total_num_frames = metadata["total_num_frames"] |
| 1795 | duration = total_num_frames / video_fps |
| 1796 | frame_sample_mode = video_processor.frame_sample_mode |
| 1797 | num_frames = video_processor.num_frames |
| 1798 | max_fps = video_processor.max_fps |
| 1799 | sampling_fps = video_processor.sampling_fps |
| 1800 | frames_indices = self._sample_frames( |
| 1801 | total_num_frames, |
| 1802 | video_fps, |
| 1803 | duration, |
| 1804 | frame_sample_mode, |
| 1805 | num_frames, |
| 1806 | max_fps, |
| 1807 | sampling_fps, |
| 1808 | ) |
| 1809 | else: |
| 1810 | # Time-based sampling is done in vllm molmo2 video loader or molmo_utils |
| 1811 | assert frames_indices is not None |
| 1812 | timestamps = [frame_idx / video_fps for frame_idx in frames_indices] |
| 1813 | return timestamps |
| 1814 | |
| 1815 | |
| 1816 | class Molmo2DummyInputsBuilder(BaseDummyInputsBuilder[Molmo2ProcessingInfo]): |
| 1817 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 1818 | num_images = mm_counts.get("image", 0) |
| 1819 | num_videos = mm_counts.get("video", 0) |
| 1820 | |
| 1821 | image_placeholder_token = IMAGE_PROMPT |
| 1822 | video_placeholder_token = VIDEO_PROMPT |
| 1823 | |
| 1824 | if num_images == 1: |
| 1825 | image_string = image_placeholder_token |
| 1826 | else: |
| 1827 | image_string = "".join( |
| 1828 | [f"Image {i + 1}" + image_placeholder_token for i in range(num_images)] |
| 1829 | ) |
| 1830 | |
| 1831 | return image_string + video_placeholder_token * num_videos |
| 1832 | |
| 1833 | def get_dummy_mm_data( |
| 1834 | self, |
| 1835 | seq_len: int, |
| 1836 | mm_counts: Mapping[str, int], |
| 1837 | mm_options: MultiModalDummyOptions, |
| 1838 | ) -> MultiModalDataDict: |
| 1839 | num_images = mm_counts.get("image", 0) |
| 1840 | num_videos = mm_counts.get("video", 0) |
| 1841 | |
| 1842 | dummy_images = [] |
| 1843 | dummy_videos = [] |
| 1844 | |
| 1845 | if num_images > 0: |
| 1846 | target_width, target_height = self.info.get_image_size_with_most_features() |
| 1847 | |
| 1848 | dummy_images = self._get_dummy_images( |
| 1849 | width=target_width, |
| 1850 | height=target_height, |
| 1851 | num_images=num_images, |
| 1852 | overrides=mm_options.get("image"), |
| 1853 | ) |
| 1854 | |
| 1855 | if num_videos > 0: |
| 1856 | processor = self.info.get_hf_processor() |
| 1857 | video_size = processor.video_processor.size |
| 1858 | target_num_frames = self.info.get_num_frames_with_most_features( |
| 1859 | seq_len, mm_counts |
| 1860 | ) |
| 1861 | |
| 1862 | video_overrides = mm_options.get("video") |
| 1863 | |
| 1864 | if video_overrides: |
| 1865 | num_frames_override = video_overrides.num_frames |
| 1866 | if num_frames_override: |
| 1867 | if num_frames_override > target_num_frames: |
| 1868 | logger.warning( |
| 1869 | "video.num_frames override (%d) exceeds model's " |
| 1870 | "maximum number of frames (%d), will be ignored", |
| 1871 | num_frames_override, |
| 1872 | target_num_frames, |
| 1873 | ) |
| 1874 | if num_frames_override < 2: |
| 1875 | logger.warning( |
| 1876 | "video.num_frames override (%d) cannot be less " |
| 1877 | "than 2, will be ignored", |
| 1878 | num_frames_override, |
| 1879 | ) |
| 1880 | target_num_frames = min(target_num_frames, num_frames_override) |
| 1881 | target_num_frames = max(target_num_frames, 2) |
| 1882 | |
| 1883 | dummy_videos = self._get_dummy_videos( |
| 1884 | width=video_size["width"], |
| 1885 | height=video_size["height"], |
| 1886 | num_frames=target_num_frames, |
| 1887 | num_videos=num_videos, |
| 1888 | ) |
| 1889 | |
| 1890 | return { |
| 1891 | "image": dummy_images, |
| 1892 | "video": dummy_videos, |
| 1893 | } |
| 1894 | |
| 1895 | def _get_dummy_videos( |
| 1896 | self, |
| 1897 | *, |
| 1898 | width: int, |
| 1899 | height: int, |
| 1900 | num_frames: int, |
| 1901 | num_videos: int, |
| 1902 | overrides: VideoDummyOptions | None = None, |
| 1903 | ) -> list[VideoItem]: |
| 1904 | videos = super()._get_dummy_videos( |
| 1905 | width=width, |
| 1906 | height=height, |
| 1907 | num_frames=num_frames, |
| 1908 | num_videos=num_videos, |
| 1909 | overrides=overrides, |
| 1910 | ) |
| 1911 | videos = [v.copy() for v in videos] |
| 1912 | |
| 1913 | video_items: list[VideoItem] = [] |
| 1914 | for video in videos: |
| 1915 | video_num_frames = video.shape[0] |
| 1916 | video_metadata: dict[str, int | float | str | bool | list[int]] = { |
| 1917 | "fps": 2.0, |
| 1918 | "duration": video_num_frames / 2.0, |
| 1919 | "total_num_frames": video_num_frames, |
| 1920 | "frames_indices": list(range(video_num_frames)), |
| 1921 | "video_backend": "decord", |
| 1922 | "do_sample_frames": False, |
| 1923 | "height": height, |
| 1924 | "width": width, |
| 1925 | } |
| 1926 | video_items.append((video, video_metadata)) |
| 1927 | |
| 1928 | return video_items |
| 1929 | |
| 1930 | |
| 1931 | class Molmo2MultiModalProcessor(BaseMultiModalProcessor[Molmo2ProcessingInfo]): |
| 1932 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 1933 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 1934 | |
| 1935 | def _postprocess_prompt(self, prompt: list[int]) -> list[int]: |
| 1936 | processor = self.info.get_hf_processor() |
| 1937 | tokenizer = processor.tokenizer |
| 1938 | bos_token_id = tokenizer.bos_token_id or tokenizer.eos_token_id |
| 1939 | |
| 1940 | if len(prompt) == 0 or prompt[0] != bos_token_id: |
| 1941 | # Prepend the bos token to the prompt tokens |
| 1942 | prompt = [bos_token_id] + prompt |
| 1943 | |
| 1944 | return prompt |
| 1945 | |
| 1946 | def _apply_hf_processor_main( |
| 1947 | self, |
| 1948 | mm_items: MultiModalDataItems, |
| 1949 | hf_kwargs: Mapping[str, object], |
| 1950 | ) -> BatchFeature: |
| 1951 | hf_data, hf_kwargs, passthrough_data = self._get_hf_mm_inputs( |
| 1952 | mm_items, hf_kwargs |
| 1953 | ) |
| 1954 | |
| 1955 | if not hf_data: |
| 1956 | return self._finalize_hf_mm_data(hf_data, hf_kwargs, passthrough_data) |
| 1957 | |
| 1958 | prompt_text = hf_data.pop("text") |
| 1959 | assert isinstance(prompt_text, str) |
| 1960 | |
| 1961 | hf_config = self.info.get_hf_config() |
| 1962 | hf_processor = self.info.get_hf_processor(**hf_kwargs) |
| 1963 | |
| 1964 | def patched_call(text=None, images=None, videos=None, **kwargs) -> BatchFeature: |
| 1965 | res = hf_processor(text=text, images=images, videos=videos, **kwargs) |
| 1966 | |
| 1967 | # Molmo2Processor.insert_bos results in float outputs |
| 1968 | # if the input text is empty |
| 1969 | if not text: |
| 1970 | res["input_ids"] = res["input_ids"].long() |
| 1971 | |
| 1972 | return res |
| 1973 | |
| 1974 | tokenizer = hf_processor.tokenizer |
| 1975 | image_processor = hf_processor.image_processor |
| 1976 | |
| 1977 | if videos := hf_data.pop("videos", []): |
| 1978 | assert isinstance(videos, Sequence) |
| 1979 | bos_token_id = tokenizer.bos_token_id or tokenizer.eos_token_id |
| 1980 | |
| 1981 | pixel_values_videos_lst = [] |
| 1982 | video_token_pooling_lst = [] |
| 1983 | video_num_crops_lst = [] |
| 1984 | video_num_pooled_patches_lst = [] |
| 1985 | video_num_patches_lst = [] |
| 1986 | video_tokens_lst = [] |
| 1987 | num_video_tokens_lst = [] |
| 1988 | |
| 1989 | for item in videos: |
| 1990 | assert isinstance(item, tuple) and len(item) == 2 |
| 1991 | video_array, metadata = item |
| 1992 | assert isinstance(metadata, Mapping) |
| 1993 | |
| 1994 | # NOTE: metadata.frames_indices indicates |
| 1995 | # the sampled frames indices of pre-sampled videos, which is |
| 1996 | # used to calculate the timestamps. Make sure that |
| 1997 | # do_sample_frames in hf_kwargs is false for |
| 1998 | # presampled videos. |
| 1999 | |
| 2000 | # NOTE: a copy of hf_kwargs is created to update |
| 2001 | # do_sample_frames, otherwise mm_hash for the object will be |
| 2002 | # incorrect. |
| 2003 | video_mm_kwargs = dict(**hf_kwargs) |
| 2004 | if "do_sample_frames" not in video_mm_kwargs: |
| 2005 | # molmo_utils already has "do_sample_frames" in |
| 2006 | # hf_kwargs, don't overwrite it. |
| 2007 | video_mm_kwargs["do_sample_frames"] = metadata.get( |
| 2008 | "do_sample_frames", False |
| 2009 | ) |
| 2010 | |
| 2011 | metadata = VideoMetadata( |
| 2012 | **{k: metadata[k] for k in metadata if k != "do_sample_frames"} |
| 2013 | ) |
| 2014 | |
| 2015 | video_mm_data = dict() |
| 2016 | video_mm_data["videos"] = [[video_array]] |
| 2017 | video_mm_data["video_metadata"] = [[metadata]] |
| 2018 | |
| 2019 | video_outputs = self.info.ctx.call_hf_processor( |
| 2020 | patched_call, |
| 2021 | dict(text=VIDEO_PROMPT, **video_mm_data), |
| 2022 | video_mm_kwargs, |
| 2023 | ) |
| 2024 | |
| 2025 | input_ids = video_outputs.pop("input_ids") |
| 2026 | if input_ids[0, 0] == bos_token_id: |
| 2027 | input_ids = input_ids[:, 1:] |
| 2028 | |
| 2029 | video_string = tokenizer.batch_decode(input_ids)[0] |
| 2030 | prompt_text = prompt_text.replace(VIDEO_PROMPT, video_string, 1) |
| 2031 | |
| 2032 | video_grids = video_outputs.pop("video_grids") |
| 2033 | assert video_grids[:, 0].sum() == len( |
| 2034 | video_outputs["pixel_values_videos"] |
| 2035 | ) |
| 2036 | |
| 2037 | video_outputs["video_num_crops"] = video_grids[:, 0] |
| 2038 | video_outputs["video_num_pooled_patches"] = video_grids.prod(dim=1) |
| 2039 | n_patches = video_outputs["pixel_values_videos"].shape[1] |
| 2040 | video_outputs["video_num_patches"] = ( |
| 2041 | video_outputs["video_num_crops"] * n_patches |
| 2042 | ) |
| 2043 | (video_outputs["video_tokens"], video_outputs["num_video_tokens"]) = ( |
| 2044 | build_flat_video_bool_length(video_grids, hf_config) |
| 2045 | ) |
| 2046 | |
| 2047 | pixel_values_videos_lst.append(video_outputs["pixel_values_videos"]) |
| 2048 | video_token_pooling_lst.append(video_outputs["video_token_pooling"]) |
| 2049 | video_num_crops_lst.append(video_outputs["video_num_crops"]) |
| 2050 | video_num_pooled_patches_lst.append( |
| 2051 | video_outputs["video_num_pooled_patches"] |
| 2052 | ) |
| 2053 | video_num_patches_lst.append(video_outputs["video_num_patches"]) |
| 2054 | video_tokens_lst.append(video_outputs["video_tokens"]) |
| 2055 | num_video_tokens_lst.append(video_outputs["num_video_tokens"]) |
| 2056 | |
| 2057 | all_video_outputs = dict( |
| 2058 | pixel_values_videos=torch.cat(pixel_values_videos_lst), |
| 2059 | video_token_pooling=torch.cat(video_token_pooling_lst), |
| 2060 | video_num_crops=torch.cat(video_num_crops_lst), |
| 2061 | video_num_pooled_patches=torch.cat(video_num_pooled_patches_lst), |
| 2062 | video_num_patches=torch.cat(video_num_patches_lst), |
| 2063 | video_tokens=torch.cat(video_tokens_lst), |
| 2064 | num_video_tokens=torch.cat(num_video_tokens_lst), |
| 2065 | ) |
| 2066 | else: |
| 2067 | all_video_outputs = dict() |
| 2068 | |
| 2069 | processed_data = self.info.ctx.call_hf_processor( |
| 2070 | patched_call, |
| 2071 | dict(text=prompt_text, **hf_data), |
| 2072 | hf_kwargs, |
| 2073 | ) |
| 2074 | |
| 2075 | if (images := hf_data.get("images")) is not None: |
| 2076 | mm_items = self.info.parse_mm_data({"image": images}, validate=False) |
| 2077 | parsed_images = mm_items.get_items("image", ImageProcessorItems) |
| 2078 | image_sizes = [ |
| 2079 | parsed_images.get_image_size(i) for i in range(len(parsed_images)) |
| 2080 | ] |
| 2081 | |
| 2082 | # For each image: tiling_h * tiling_w + global view |
| 2083 | tilings = [ |
| 2084 | self.info.select_tiling( |
| 2085 | image_width=image_size.width, |
| 2086 | image_height=image_size.height, |
| 2087 | image_processor=image_processor, |
| 2088 | ) |
| 2089 | for image_size in image_sizes |
| 2090 | ] |
| 2091 | num_crops = torch.tensor(tilings).prod(-1) + 1 |
| 2092 | assert sum(num_crops) == len(processed_data["pixel_values"]) |
| 2093 | assert sum(num_crops) == processed_data["image_num_crops"].sum().item() |
| 2094 | |
| 2095 | image_grids = processed_data.pop("image_grids") |
| 2096 | image_num_pooled_patches = image_grids[:, :2].prod(dim=1) + image_grids[ |
| 2097 | :, 2: |
| 2098 | ].prod(dim=1) |
| 2099 | |
| 2100 | processed_data["image_num_pooled_patches"] = image_num_pooled_patches |
| 2101 | n_patches = processed_data["pixel_values"].shape[1] |
| 2102 | processed_data["image_num_patches"] = ( |
| 2103 | processed_data["image_num_crops"] * n_patches |
| 2104 | ) |
| 2105 | ( |
| 2106 | processed_data["image_tokens"], |
| 2107 | processed_data["num_image_tokens"], |
| 2108 | ) = build_flat_image_bool_length( |
| 2109 | image_grids, |
| 2110 | hf_config, |
| 2111 | image_use_col_tokens=hf_processor.image_use_col_tokens, |
| 2112 | use_single_crop_col_tokens=hf_processor.use_single_crop_col_tokens, |
| 2113 | use_single_crop_start_token=hf_processor.use_single_crop_start_token, |
| 2114 | ) |
| 2115 | |
| 2116 | processed_data.update(all_video_outputs) |
| 2117 | processed_data.pop("input_ids") |
| 2118 | |
| 2119 | return self._finalize_hf_mm_data( |
| 2120 | hf_data, hf_kwargs, passthrough_data, processed_data |
| 2121 | ) |
| 2122 | |
| 2123 | def _get_mm_fields_config( |
| 2124 | self, |
| 2125 | hf_inputs: BatchFeature, |
| 2126 | hf_processor_mm_kwargs: Mapping[str, object], |
| 2127 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 2128 | image_num_crops = hf_inputs.get("image_num_crops", torch.empty(0)) |
| 2129 | image_num_pooled_patches = hf_inputs.get( |
| 2130 | "image_num_pooled_patches", torch.empty(0) |
| 2131 | ) |
| 2132 | video_num_crops = hf_inputs.get("video_num_crops", torch.empty(0)) |
| 2133 | video_num_pooled_patches = hf_inputs.get( |
| 2134 | "video_num_pooled_patches", torch.empty(0) |
| 2135 | ) |
| 2136 | num_image_tokens = hf_inputs.get("num_image_tokens", torch.empty(0)) |
| 2137 | num_video_tokens = hf_inputs.get("num_video_tokens", torch.empty(0)) |
| 2138 | |
| 2139 | return dict( |
| 2140 | pixel_values=MultiModalFieldConfig.flat_from_sizes( |
| 2141 | "image", image_num_crops |
| 2142 | ), |
| 2143 | image_token_pooling=MultiModalFieldConfig.flat_from_sizes( |
| 2144 | "image", image_num_pooled_patches |
| 2145 | ), |
| 2146 | image_num_crops=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 2147 | image_num_pooled_patches=MultiModalFieldConfig.batched( |
| 2148 | "image", keep_on_cpu=True |
| 2149 | ), |
| 2150 | image_num_patches=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 2151 | image_tokens=MultiModalFieldConfig.flat_from_sizes( |
| 2152 | "image", num_image_tokens |
| 2153 | ), |
| 2154 | num_image_tokens=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 2155 | pixel_values_videos=MultiModalFieldConfig.flat_from_sizes( |
| 2156 | "video", video_num_crops |
| 2157 | ), |
| 2158 | video_token_pooling=MultiModalFieldConfig.flat_from_sizes( |
| 2159 | "video", video_num_pooled_patches |
| 2160 | ), |
| 2161 | video_num_crops=MultiModalFieldConfig.batched("video", keep_on_cpu=True), |
| 2162 | video_num_pooled_patches=MultiModalFieldConfig.batched( |
| 2163 | "video", keep_on_cpu=True |
| 2164 | ), |
| 2165 | video_num_patches=MultiModalFieldConfig.batched("video", keep_on_cpu=True), |
| 2166 | video_tokens=MultiModalFieldConfig.flat_from_sizes( |
| 2167 | "video", num_video_tokens |
| 2168 | ), |
| 2169 | num_video_tokens=MultiModalFieldConfig.batched("video", keep_on_cpu=True), |
| 2170 | ) |
| 2171 | |
| 2172 | def _get_prompt_updates( |
| 2173 | self, |
| 2174 | mm_items: MultiModalDataItems, |
| 2175 | hf_processor_mm_kwargs: Mapping[str, object], |
| 2176 | out_mm_kwargs: MultiModalKwargsItems, |
| 2177 | ) -> Sequence[PromptUpdate]: |
| 2178 | hf_config = self.info.get_hf_config() |
| 2179 | img_patch_id = hf_config.image_patch_id |
| 2180 | img_col_id = hf_config.image_col_id |
| 2181 | img_start_id = hf_config.image_start_token_id |
| 2182 | img_end_id = hf_config.image_end_token_id |
| 2183 | low_res_im_start_id = hf_config.low_res_image_start_token_id |
| 2184 | frame_start_id = hf_config.frame_start_token_id |
| 2185 | frame_end_id = hf_config.frame_end_token_id |
| 2186 | im_low_res_id = hf_config.image_low_res_id |
| 2187 | |
| 2188 | emb_tok_ids = [ |
| 2189 | img_patch_id, |
| 2190 | img_col_id, |
| 2191 | img_start_id, |
| 2192 | low_res_im_start_id, |
| 2193 | frame_start_id, |
| 2194 | img_end_id, |
| 2195 | frame_end_id, |
| 2196 | im_low_res_id, |
| 2197 | ] |
| 2198 | |
| 2199 | processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 2200 | image_use_col_tokens = processor.image_use_col_tokens |
| 2201 | use_single_crop_col_tokens = processor.use_single_crop_col_tokens |
| 2202 | use_single_crop_start_token = processor.use_single_crop_start_token |
| 2203 | video_use_col_tokens = processor.video_use_col_tokens |
| 2204 | use_frame_special_tokens = processor.use_frame_special_tokens |
| 2205 | |
| 2206 | tokenizer = processor.tokenizer |
| 2207 | vocab = tokenizer.get_vocab() |
| 2208 | |
| 2209 | image_processor = processor.image_processor |
| 2210 | video_processor = processor.video_processor |
| 2211 | |
| 2212 | def get_image_replacement_molmo2(item_idx: int): |
| 2213 | images = mm_items.get_items("image", ImageProcessorItems) |
| 2214 | image = images.get(item_idx) |
| 2215 | image = exif_transpose(image) |
| 2216 | |
| 2217 | resize_ncols, resize_nrows = self.info.get_base_grid_size(image_processor) |
| 2218 | if use_single_crop_col_tokens is not None: |
| 2219 | use_col_tokens = use_single_crop_col_tokens |
| 2220 | else: |
| 2221 | use_col_tokens = image_use_col_tokens |
| 2222 | if use_single_crop_start_token: |
| 2223 | start_id = low_res_im_start_id |
| 2224 | else: |
| 2225 | start_id = img_start_id |
| 2226 | extra_row = [img_patch_id] * resize_ncols + [img_col_id] * int( |
| 2227 | use_col_tokens |
| 2228 | ) |
| 2229 | extra_joint = [start_id] + extra_row * resize_nrows + [img_end_id] |
| 2230 | |
| 2231 | image_size = get_image_size(image) |
| 2232 | |
| 2233 | ncols, nrows = self.info.get_patches_grid_size( |
| 2234 | image_height=image_size.height, |
| 2235 | image_width=image_size.width, |
| 2236 | image_processor=image_processor, |
| 2237 | ) |
| 2238 | |
| 2239 | joint_row = [img_patch_id] * ncols + [img_col_id] * int( |
| 2240 | image_use_col_tokens |
| 2241 | ) |
| 2242 | joint = [img_start_id] + joint_row * nrows + [img_end_id] |
| 2243 | img_token_ids = extra_joint + joint |
| 2244 | |
| 2245 | return PromptUpdateDetails.select_token_ids(img_token_ids, emb_tok_ids) |
| 2246 | |
| 2247 | def get_video_replacement_molmo2(item_idx: int): |
| 2248 | video, metadata = mm_items["video"][item_idx] |
| 2249 | do_sample_frames = hf_processor_mm_kwargs.get("do_sample_frames") |
| 2250 | assert do_sample_frames is None or isinstance(do_sample_frames, bool) |
| 2251 | |
| 2252 | timestamps = self.info._get_video_second_idx(metadata, do_sample_frames) |
| 2253 | ncols, nrows = self.info.get_base_grid_size(video_processor) |
| 2254 | |
| 2255 | if use_frame_special_tokens: |
| 2256 | start_id = frame_start_id |
| 2257 | end_id = frame_end_id |
| 2258 | else: |
| 2259 | start_id = img_start_id |
| 2260 | end_id = img_end_id |
| 2261 | |
| 2262 | img_token_ids = [] |
| 2263 | |
| 2264 | for frame_idx, frame_time in enumerate(timestamps): |
| 2265 | prev_space = " " if frame_idx > 0 else "" |
| 2266 | frame_prefix = ( |
| 2267 | prev_space + f"{frame_time:.1f} " |
| 2268 | ) # explicit whitespace before/after image tokens |
| 2269 | |
| 2270 | img_token_ids += tokenizer.encode( |
| 2271 | frame_prefix, |
| 2272 | add_special_tokens=False, |
| 2273 | ) |
| 2274 | |
| 2275 | joint_row = [img_patch_id] * ncols + [img_col_id] * int( |
| 2276 | video_use_col_tokens |
| 2277 | ) |
| 2278 | joint = [start_id] + nrows * joint_row + [end_id] |
| 2279 | img_token_ids += joint |
| 2280 | |
| 2281 | return PromptUpdateDetails.select_token_ids(img_token_ids, emb_tok_ids) |
| 2282 | |
| 2283 | return [ |
| 2284 | PromptReplacement( |
| 2285 | modality=modality, |
| 2286 | target=[target], |
| 2287 | replacement=replacement_fn, |
| 2288 | ) |
| 2289 | for modality, target, replacement_fn in zip( |
| 2290 | ["image", "video"], |
| 2291 | [vocab[IMAGE_PROMPT], vocab[VIDEO_PROMPT]], |
| 2292 | [get_image_replacement_molmo2, get_video_replacement_molmo2], |
| 2293 | ) |
| 2294 | ] |
| 2295 | |
| 2296 | |
| 2297 | @MULTIMODAL_REGISTRY.register_processor( |
| 2298 | Molmo2MultiModalProcessor, |
| 2299 | info=Molmo2ProcessingInfo, |
| 2300 | dummy_inputs=Molmo2DummyInputsBuilder, |
| 2301 | ) |
| 2302 | class Molmo2ForConditionalGeneration( |
| 2303 | nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA, SupportsQuant |
| 2304 | ): |
| 2305 | hf_to_vllm_mapper = WeightsMapper( |
| 2306 | orig_to_new_substr={ |
| 2307 | # vision backbone mapping |
| 2308 | "image_pooling_2d.wq": "image_pooling_2d.q_proj", |
| 2309 | "image_pooling_2d.wk": "image_pooling_2d.k_proj", |
| 2310 | "image_pooling_2d.wv": "image_pooling_2d.v_proj", |
| 2311 | "image_pooling_2d.wo": "image_pooling_2d.o_proj", |
| 2312 | "image_projector.w1": "image_projector.gate_proj", |
| 2313 | "image_projector.w3": "image_projector.up_proj", |
| 2314 | "image_projector.w2": "image_projector.down_proj", |
| 2315 | # language backbone mapping |
| 2316 | "att_proj": "qkv_proj", |
| 2317 | "attn_out": "o_proj", |
| 2318 | "q_norm": "q_norm", |
| 2319 | "k_norm": "k_norm", |
| 2320 | "ff_proj": "up_gate_proj", |
| 2321 | "ff_out": "down_proj", |
| 2322 | "attn_norm": "input_layernorm", |
| 2323 | "ff_norm": "post_attention_layernorm", |
| 2324 | }, |
| 2325 | orig_to_new_prefix={ |
| 2326 | # vision backbone mapping |
| 2327 | "model.vision_backbone.": "vision_backbone.", |
| 2328 | # language backbone mapping |
| 2329 | "model.transformer.blocks.": "model.layers.", |
| 2330 | "model.transformer.ln_f.": "model.norm.", |
| 2331 | }, |
| 2332 | ) |
| 2333 | |
| 2334 | packed_modules_mapping = { |
| 2335 | "qkv_proj": ["qkv_proj"], |
| 2336 | "up_gate_proj": ["up_gate_proj"], # language model |
| 2337 | "merged_qkv": ["wq", "wk", "wv"], # vision backbone |
| 2338 | "merged_kv": ["k_proj", "v_proj"], # image_pooling_2d |
| 2339 | "merged_linear": ["gate_proj", "up_proj"], # image_projector |
| 2340 | } |
| 2341 | |
| 2342 | @classmethod |
| 2343 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 2344 | if modality.startswith("image"): |
| 2345 | return IMAGE_PROMPT |
| 2346 | if modality.startswith("video"): |
| 2347 | return VIDEO_PROMPT |
| 2348 | |
| 2349 | raise ValueError("Only image or video modality is supported") |
| 2350 | |
| 2351 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 2352 | super().__init__() |
| 2353 | config = vllm_config.model_config.hf_config |
| 2354 | quant_config = vllm_config.quant_config |
| 2355 | multimodal_config = vllm_config.model_config.multimodal_config |
| 2356 | self.config = config |
| 2357 | self.multimodal_config = multimodal_config |
| 2358 | |
| 2359 | kwargs = {} |
| 2360 | for field in fields(VitConfig): |
| 2361 | kwargs[field.name] = getattr(config.vit_config, field.name) |
| 2362 | vit_config = VitConfig(**kwargs) |
| 2363 | |
| 2364 | kwargs = {} |
| 2365 | for field in fields(AdapterConfig): |
| 2366 | kwargs[field.name] = getattr(config.adapter_config, field.name) |
| 2367 | adapter_config = AdapterConfig(**kwargs) |
| 2368 | |
| 2369 | with self._mark_tower_model(vllm_config, {"image", "video"}): |
| 2370 | self.vision_backbone = Molmo2VisionBackbone( |
| 2371 | vit_config, |
| 2372 | adapter_config, |
| 2373 | quant_config, |
| 2374 | prefix=maybe_prefix(prefix, "vision_backbone"), |
| 2375 | ) |
| 2376 | |
| 2377 | with self._mark_language_model(vllm_config): |
| 2378 | self.model = Molmo2TextModel( |
| 2379 | vllm_config=vllm_config, |
| 2380 | prefix=maybe_prefix(prefix, "model"), |
| 2381 | ) |
| 2382 | |
| 2383 | self.img_patch_id = config.image_patch_id |
| 2384 | |
| 2385 | if hasattr(config, "text_config"): |
| 2386 | hf_text_config = config.text_config |
| 2387 | else: |
| 2388 | hf_text_config = config.llm_config |
| 2389 | |
| 2390 | self.lm_head = ParallelLMHead( |
| 2391 | hf_text_config.vocab_size, |
| 2392 | hf_text_config.hidden_size, |
| 2393 | quant_config=quant_config, |
| 2394 | ) |
| 2395 | self.logits_processor = LogitsProcessor(hf_text_config.vocab_size) |
| 2396 | |
| 2397 | self.make_empty_intermediate_tensors = ( |
| 2398 | self.model.make_empty_intermediate_tensors |
| 2399 | ) |
| 2400 | |
| 2401 | @property |
| 2402 | def dtype(self): |
| 2403 | return next(self.parameters()).dtype |
| 2404 | |
| 2405 | def _parse_and_validate_image_input( |
| 2406 | self, |
| 2407 | **kwargs: object, |
| 2408 | ) -> Molmo2ImageInputs | None: |
| 2409 | pixel_values = kwargs.pop("pixel_values", None) |
| 2410 | if pixel_values is None: |
| 2411 | return None |
| 2412 | |
| 2413 | token_pooling = kwargs.pop("image_token_pooling", None) |
| 2414 | num_pooled_patches = kwargs.pop("image_num_pooled_patches", None) |
| 2415 | num_patches = kwargs.pop("image_num_patches", None) |
| 2416 | image_tokens = kwargs.pop("image_tokens", None) |
| 2417 | num_image_tokens = kwargs.pop("num_image_tokens", None) |
| 2418 | |
| 2419 | assert isinstance(token_pooling, torch.Tensor) |
| 2420 | assert isinstance(num_pooled_patches, torch.Tensor) |
| 2421 | assert isinstance(num_patches, torch.Tensor) |
| 2422 | assert isinstance(image_tokens, torch.Tensor) |
| 2423 | assert isinstance(num_image_tokens, torch.Tensor) |
| 2424 | |
| 2425 | accum_patches = [0] + num_patches.cumsum(dim=0)[:-1].tolist() |
| 2426 | patch_offset = 0 |
| 2427 | new_token_pooling = token_pooling.clone() |
| 2428 | for i, n_tensor in enumerate(num_pooled_patches): |
| 2429 | n = int(n_tensor) |
| 2430 | cur_slice = token_pooling[patch_offset : patch_offset + n] |
| 2431 | index_offset = int(accum_patches[i]) |
| 2432 | new_token_pooling[patch_offset : patch_offset + n] = torch.where( |
| 2433 | cur_slice >= 0, |
| 2434 | cur_slice + index_offset, |
| 2435 | cur_slice, |
| 2436 | ) |
| 2437 | patch_offset += n |
| 2438 | |
| 2439 | return Molmo2ImageInputs( |
| 2440 | pixel_values=pixel_values, |
| 2441 | token_pooling=new_token_pooling, |
| 2442 | num_pooled_patches=num_pooled_patches, |
| 2443 | image_tokens=image_tokens, |
| 2444 | num_image_tokens=num_image_tokens, |
| 2445 | ) |
| 2446 | |
| 2447 | def _parse_and_validate_video_input( |
| 2448 | self, |
| 2449 | **kwargs: object, |
| 2450 | ) -> Molmo2VideoInputs | None: |
| 2451 | pixel_values_videos = kwargs.pop("pixel_values_videos", None) |
| 2452 | if pixel_values_videos is None: |
| 2453 | return None |
| 2454 | |
| 2455 | token_pooling = kwargs.pop("video_token_pooling", None) |
| 2456 | num_pooled_patches = kwargs.pop("video_num_pooled_patches", None) |
| 2457 | num_patches = kwargs.pop("video_num_patches", None) |
| 2458 | video_tokens = kwargs.pop("video_tokens", None) |
| 2459 | num_video_tokens = kwargs.pop("num_video_tokens", None) |
| 2460 | |
| 2461 | assert isinstance(token_pooling, torch.Tensor) |
| 2462 | assert isinstance(num_pooled_patches, torch.Tensor) |
| 2463 | assert isinstance(num_patches, torch.Tensor) |
| 2464 | assert isinstance(video_tokens, torch.Tensor) |
| 2465 | assert isinstance(num_video_tokens, torch.Tensor) |
| 2466 | |
| 2467 | accum_patches = [0] + num_patches.cumsum(dim=0)[:-1].tolist() |
| 2468 | patch_offset = 0 |
| 2469 | new_token_pooling = token_pooling.clone() |
| 2470 | for i, n_tensor in enumerate(num_pooled_patches): |
| 2471 | n = int(n_tensor) |
| 2472 | cur_slice = token_pooling[patch_offset : patch_offset + n] |
| 2473 | index_offset = int(accum_patches[i]) |
| 2474 | new_token_pooling[patch_offset : patch_offset + n] = torch.where( |
| 2475 | cur_slice >= 0, |
| 2476 | cur_slice + index_offset, |
| 2477 | cur_slice, |
| 2478 | ) |
| 2479 | patch_offset += n |
| 2480 | |
| 2481 | return Molmo2VideoInputs( |
| 2482 | pixel_values_videos=pixel_values_videos, |
| 2483 | token_pooling=new_token_pooling, |
| 2484 | num_pooled_patches=num_pooled_patches, |
| 2485 | video_tokens=video_tokens, |
| 2486 | num_video_tokens=num_video_tokens, |
| 2487 | ) |
| 2488 | |
| 2489 | def _parse_and_validate_multimodal_inputs( |
| 2490 | self, **kwargs: object |
| 2491 | ) -> Molmo2MultiModalInputs: |
| 2492 | modalities: Molmo2MultiModalInputs = {} |
| 2493 | |
| 2494 | for input_key in kwargs: |
| 2495 | if input_key in ("pixel_values",) and "images" not in modalities: |
| 2496 | modalities["images"] = self._parse_and_validate_image_input(**kwargs) |
| 2497 | if input_key in ("pixel_values_videos",) and "videos" not in modalities: |
| 2498 | modalities["videos"] = self._parse_and_validate_video_input(**kwargs) |
| 2499 | return modalities |
| 2500 | |
| 2501 | def _process_image_input( |
| 2502 | self, |
| 2503 | image_input: Molmo2ImageInputs, |
| 2504 | ) -> tuple[torch.Tensor, ...]: |
| 2505 | pixel_values = image_input["pixel_values"] |
| 2506 | token_pooling = image_input["token_pooling"] |
| 2507 | num_pooled_patches = image_input["num_pooled_patches"] |
| 2508 | image_tokens = image_input["image_tokens"] |
| 2509 | num_image_tokens = image_input["num_image_tokens"] |
| 2510 | |
| 2511 | image_features_flat = self.vision_backbone( |
| 2512 | images=pixel_values.unsqueeze(0), |
| 2513 | token_pooling=token_pooling.unsqueeze(0), |
| 2514 | ) |
| 2515 | |
| 2516 | assert len(image_features_flat) == num_pooled_patches.sum() |
| 2517 | image_features_list = image_features_flat.split( |
| 2518 | num_pooled_patches.tolist(), dim=0 |
| 2519 | ) |
| 2520 | image_tokens_list = image_tokens.split(num_image_tokens.tolist(), dim=0) |
| 2521 | out = [] |
| 2522 | for image_features_i, image_tokens_i in zip( |
| 2523 | image_features_list, image_tokens_list |
| 2524 | ): |
| 2525 | out_features = self.get_language_model().embed_input_ids(image_tokens_i) |
| 2526 | is_image_patch = image_tokens_i == self.img_patch_id |
| 2527 | out_features[is_image_patch] = image_features_i |
| 2528 | out.append(out_features) |
| 2529 | return tuple(out) |
| 2530 | |
| 2531 | def _process_video_input( |
| 2532 | self, |
| 2533 | video_input: Molmo2VideoInputs, |
| 2534 | ) -> tuple[torch.Tensor, ...]: |
| 2535 | pixel_values_videos = video_input["pixel_values_videos"] |
| 2536 | token_pooling = video_input["token_pooling"] |
| 2537 | num_pooled_patches = video_input["num_pooled_patches"] |
| 2538 | video_tokens = video_input["video_tokens"] |
| 2539 | num_video_tokens = video_input["num_video_tokens"] |
| 2540 | |
| 2541 | image_features_flat = self.vision_backbone( |
| 2542 | images=pixel_values_videos.unsqueeze(0), |
| 2543 | token_pooling=token_pooling.unsqueeze(0), |
| 2544 | ) |
| 2545 | |
| 2546 | assert len(image_features_flat) == num_pooled_patches.sum() |
| 2547 | image_features_list = image_features_flat.split( |
| 2548 | num_pooled_patches.tolist(), dim=0 |
| 2549 | ) |
| 2550 | video_tokens_list = video_tokens.split(num_video_tokens.tolist(), dim=0) |
| 2551 | out = [] |
| 2552 | for image_features_i, video_tokens_i in zip( |
| 2553 | image_features_list, video_tokens_list |
| 2554 | ): |
| 2555 | out_features = self.get_language_model().embed_input_ids(video_tokens_i) |
| 2556 | is_image_patch = video_tokens_i == self.img_patch_id |
| 2557 | out_features[is_image_patch] = image_features_i |
| 2558 | out.append(out_features) |
| 2559 | return tuple(out) |
| 2560 | |
| 2561 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings | None: |
| 2562 | modalities = self._parse_and_validate_multimodal_inputs(**kwargs) |
| 2563 | if not modalities: |
| 2564 | return [] |
| 2565 | |
| 2566 | multimodal_embeddings: tuple[torch.Tensor, ...] = () |
| 2567 | |
| 2568 | for modality in modalities: |
| 2569 | if modality == "images": |
| 2570 | image_input = modalities["images"] |
| 2571 | assert image_input is not None |
| 2572 | image_embeddings = self._process_image_input(image_input) |
| 2573 | multimodal_embeddings += image_embeddings |
| 2574 | if modality == "videos": |
| 2575 | video_input = modalities["videos"] |
| 2576 | assert video_input is not None |
| 2577 | video_embeddings = self._process_video_input(video_input) |
| 2578 | multimodal_embeddings += video_embeddings |
| 2579 | |
| 2580 | return multimodal_embeddings |
| 2581 | |
| 2582 | def embed_input_ids( |
| 2583 | self, |
| 2584 | input_ids: torch.Tensor, |
| 2585 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 2586 | *, |
| 2587 | is_multimodal: torch.Tensor | None = None, |
| 2588 | ) -> torch.Tensor: |
| 2589 | inputs_embeds = self._embed_text_input_ids( |
| 2590 | input_ids, |
| 2591 | self.get_language_model().embed_input_ids, |
| 2592 | is_multimodal=is_multimodal, |
| 2593 | ) |
| 2594 | |
| 2595 | if multimodal_embeddings is None or len(multimodal_embeddings) == 0: |
| 2596 | return inputs_embeds |
| 2597 | |
| 2598 | if is_multimodal is None: |
| 2599 | raise ValueError( |
| 2600 | "`embed_input_ids` now requires `is_multimodal` arg, " |
| 2601 | "please update your model runner according to " |
| 2602 | "https://github.com/vllm-project/vllm/pull/16229." |
| 2603 | ) |
| 2604 | |
| 2605 | inputs_embeds = _merge_multimodal_embeddings( |
| 2606 | inputs_embeds=inputs_embeds, |
| 2607 | multimodal_embeddings=multimodal_embeddings, |
| 2608 | is_multimodal=is_multimodal, |
| 2609 | ) |
| 2610 | return inputs_embeds |
| 2611 | |
| 2612 | def forward( |
| 2613 | self, |
| 2614 | input_ids: torch.LongTensor, |
| 2615 | positions: torch.LongTensor, |
| 2616 | intermediate_tensors: IntermediateTensors | None = None, |
| 2617 | inputs_embeds: torch.Tensor | None = None, |
| 2618 | **kwargs: object, |
| 2619 | ) -> torch.Tensor: |
| 2620 | if intermediate_tensors is not None: |
| 2621 | inputs_embeds = None |
| 2622 | |
| 2623 | hidden_states = self.model( |
| 2624 | input_ids, |
| 2625 | positions, |
| 2626 | intermediate_tensors, |
| 2627 | inputs_embeds=inputs_embeds, |
| 2628 | **kwargs, |
| 2629 | ) |
| 2630 | |
| 2631 | return hidden_states |
| 2632 | |
| 2633 | def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 2634 | logits = self.logits_processor(self.lm_head, hidden_states) |
| 2635 | return logits |
| 2636 | |
| 2637 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): |
| 2638 | loader = AutoWeightsLoader(self) |
| 2639 | weights = _get_weights_with_merged_embedding(weights) |
| 2640 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |
| 2641 | |
| 2642 | def get_mm_mapping(self) -> MultiModelKeys: |
| 2643 | """Get the module prefix in multimodal models.""" |
| 2644 | return MultiModelKeys.from_string_field( |
| 2645 | language_model="model", |
| 2646 | connector="vision_backbone.image_projector", |
| 2647 | tower_model="vision_backbone", |
| 2648 | ) |
| 2649 | |
| 2650 | |
| 2651 | def _get_weights_with_merged_embedding( |
| 2652 | weights: Iterable[tuple[str, torch.Tensor]], |
| 2653 | ) -> Iterable[tuple[str, torch.Tensor]]: |
| 2654 | embedding_weights = {} |
| 2655 | for name, weight in weights: |
| 2656 | if "wte.embedding" in name: |
| 2657 | embedding_weights["embedding"] = weight |
| 2658 | elif "wte.new_embedding" in name: |
| 2659 | embedding_weights["new_embedding"] = weight |
| 2660 | else: |
| 2661 | yield (name, weight) |
| 2662 | # this is compatible with most of quantization, |
| 2663 | # because they won't quantize embed_tokens |
| 2664 | if "embedding" not in embedding_weights or "new_embedding" not in embedding_weights: |
| 2665 | raise ValueError( |
| 2666 | "Checkpoint is missing 'wte.embedding' or " |
| 2667 | "'wte.new_embedding' weights required for Molmo2." |
| 2668 | ) |
| 2669 | |
| 2670 | embedding_weights = torch.cat( |
| 2671 | [embedding_weights["embedding"], embedding_weights["new_embedding"]], |
| 2672 | dim=0, |
| 2673 | ) |
| 2674 | yield ("model.embed_tokens.weight", embedding_weights) |