| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Adapted from |
| 5 | # https://github.com/zai-org/CogAgent |
| 6 | """Inference-only CogAgent model compatible with THUDM weights.""" |
| 7 | |
| 8 | from argparse import Namespace |
| 9 | from collections.abc import Iterator, Mapping, Sequence |
| 10 | from typing import Annotated, Literal |
| 11 | |
| 12 | import numpy as np |
| 13 | import torch |
| 14 | from torch import nn |
| 15 | from torch.nn import LayerNorm |
| 16 | from transformers import BatchFeature |
| 17 | |
| 18 | from vllm.config import VllmConfig |
| 19 | from vllm.config.multimodal import MultiModalDummyOptions |
| 20 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 21 | from vllm.inputs import MultiModalDataDict |
| 22 | from vllm.model_executor.layers.activation import SiluAndMul, get_act_fn |
| 23 | from vllm.model_executor.layers.attention import MMEncoderAttention |
| 24 | from vllm.model_executor.layers.conv import Conv2dLayer |
| 25 | from vllm.model_executor.layers.linear import ( |
| 26 | ColumnParallelLinear, |
| 27 | MergedColumnParallelLinear, |
| 28 | QKVParallelLinear, |
| 29 | ReplicatedLinear, |
| 30 | RowParallelLinear, |
| 31 | ) |
| 32 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 33 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 34 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 35 | from vllm.multimodal.inputs import ( |
| 36 | MultiModalFeatureSpec, |
| 37 | MultiModalFieldConfig, |
| 38 | MultiModalKwargsItems, |
| 39 | ) |
| 40 | from vllm.multimodal.parse import MultiModalDataItems |
| 41 | from vllm.multimodal.processing import ( |
| 42 | BaseDummyInputsBuilder, |
| 43 | BaseMultiModalProcessor, |
| 44 | BaseProcessingInfo, |
| 45 | PromptReplacement, |
| 46 | PromptUpdate, |
| 47 | ) |
| 48 | from vllm.sequence import IntermediateTensors |
| 49 | from vllm.transformers_utils.configs.chatglm import ChatGLMConfig |
| 50 | from vllm.transformers_utils.processors.glm4v import ( |
| 51 | GLM4VImageProcessorFast, |
| 52 | GLM4VProcessor, |
| 53 | ) |
| 54 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 55 | |
| 56 | from .chatglm import ChatGLMBaseModel, ChatGLMModel, GLMTransformer |
| 57 | from .interfaces import ( |
| 58 | MultiModalEmbeddings, |
| 59 | SupportsLoRA, |
| 60 | SupportsMRoPE, |
| 61 | SupportsMultiModal, |
| 62 | SupportsPP, |
| 63 | ) |
| 64 | from .utils import WeightsMapper |
| 65 | |
| 66 | |
| 67 | class GLMVImagePixelInputs(TensorSchema): |
| 68 | """Dimensions: |
| 69 | - b: Batch size |
| 70 | - c: Number of channels (3) |
| 71 | - h: Height of image |
| 72 | - w: Width of image |
| 73 | """ |
| 74 | |
| 75 | type: Literal["pixel_values"] = "pixel_values" |
| 76 | data: Annotated[torch.Tensor, TensorShape("b", 3, "h", "w")] |
| 77 | |
| 78 | |
| 79 | class EVA2CLIPPatchEmbedding(nn.Module): |
| 80 | def __init__(self, config): |
| 81 | super().__init__() |
| 82 | self.proj = Conv2dLayer( |
| 83 | config.in_channels, |
| 84 | config.hidden_size, |
| 85 | kernel_size=config.patch_size, |
| 86 | stride=config.patch_size, |
| 87 | ) |
| 88 | self.cls_embedding = nn.Parameter(torch.zeros(1, config.hidden_size)) |
| 89 | self.position_embedding = nn.Embedding(config.num_positions, config.hidden_size) |
| 90 | |
| 91 | def forward(self, images: torch.Tensor) -> torch.Tensor: |
| 92 | """Parameters |
| 93 | images : torch.Tensor |
| 94 | Input image tensor with shape (B, C, H, W) |
| 95 | |
| 96 | Returns |
| 97 | ------- |
| 98 | torch.Tensor |
| 99 | Transformed tensor with shape (B, L, D) |
| 100 | |
| 101 | """ |
| 102 | images = images.to(device=self.proj.weight.device, dtype=self.proj.weight.dtype) |
| 103 | x = self.proj(images) |
| 104 | x = x.flatten(2).transpose(1, 2) |
| 105 | cls_token = self.cls_embedding.expand(x.shape[0], -1, -1) |
| 106 | x = torch.cat((cls_token, x), dim=1) |
| 107 | x += self.position_embedding.weight.unsqueeze(0) |
| 108 | return x |
| 109 | |
| 110 | |
| 111 | class EVA2CLIPAttention(nn.Module): |
| 112 | def __init__( |
| 113 | self, |
| 114 | config, |
| 115 | quant_config: QuantizationConfig | None = None, |
| 116 | prefix: str = "", |
| 117 | ): |
| 118 | super().__init__() |
| 119 | self.hidden_size = config.hidden_size |
| 120 | self.tp_size = get_tensor_model_parallel_world_size() |
| 121 | self.num_heads_per_rank = config.num_heads // self.tp_size |
| 122 | self.head_dim = config.hidden_size // config.num_heads |
| 123 | self.scale = self.head_dim**-0.5 |
| 124 | |
| 125 | self.query_key_value = QKVParallelLinear( |
| 126 | config.hidden_size, |
| 127 | self.head_dim, |
| 128 | config.num_heads, |
| 129 | quant_config=quant_config, |
| 130 | prefix=f"{prefix}.query_key_value", |
| 131 | ) |
| 132 | self.dense = RowParallelLinear( |
| 133 | config.hidden_size, |
| 134 | config.hidden_size, |
| 135 | quant_config=quant_config, |
| 136 | prefix=f"{prefix}.dense", |
| 137 | ) |
| 138 | |
| 139 | self.attn = MMEncoderAttention( |
| 140 | self.num_heads_per_rank, |
| 141 | self.head_dim, |
| 142 | self.scale, |
| 143 | prefix=f"{prefix}.attn", |
| 144 | ) |
| 145 | self.output_dropout = torch.nn.Dropout(config.dropout_prob) |
| 146 | |
| 147 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 148 | qkv, _ = self.query_key_value(x) # B, L, 3 * H * D |
| 149 | q, k, v = qkv.chunk(3, dim=-1) |
| 150 | |
| 151 | out = self.attn(q, k, v) |
| 152 | output, _ = self.dense(out) |
| 153 | output = self.output_dropout(output) |
| 154 | return output |
| 155 | |
| 156 | |
| 157 | class EVA2CLIPMLP(nn.Module): |
| 158 | def __init__( |
| 159 | self, |
| 160 | config, |
| 161 | quant_config: QuantizationConfig | None = None, |
| 162 | prefix: str = "", |
| 163 | ): |
| 164 | super().__init__() |
| 165 | self.config = config |
| 166 | self.activation_fn = get_act_fn(config.hidden_act) |
| 167 | self.fc1 = ColumnParallelLinear( |
| 168 | config.hidden_size, |
| 169 | config.intermediate_size, |
| 170 | quant_config=quant_config, |
| 171 | prefix=f"{prefix}.fc1", |
| 172 | ) |
| 173 | self.fc2 = RowParallelLinear( |
| 174 | config.intermediate_size, |
| 175 | config.hidden_size, |
| 176 | quant_config=quant_config, |
| 177 | prefix=f"{prefix}.fc2", |
| 178 | ) |
| 179 | |
| 180 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 181 | x, _ = self.fc1(x) |
| 182 | x = self.activation_fn(x) |
| 183 | x, _ = self.fc2(x) |
| 184 | return x |
| 185 | |
| 186 | |
| 187 | class EVA2CLIPTransformerLayer(nn.Module): |
| 188 | def __init__( |
| 189 | self, |
| 190 | config, |
| 191 | quant_config: QuantizationConfig | None = None, |
| 192 | prefix: str = "", |
| 193 | ): |
| 194 | super().__init__() |
| 195 | self.input_layernorm = LayerNorm(config.hidden_size, eps=config.layer_norm_eps) |
| 196 | self.attention = EVA2CLIPAttention( |
| 197 | config, quant_config=quant_config, prefix=f"{prefix}.attention" |
| 198 | ) |
| 199 | self.mlp = EVA2CLIPMLP( |
| 200 | config, quant_config=quant_config, prefix=f"{prefix}.mlp" |
| 201 | ) |
| 202 | self.post_attention_layernorm = LayerNorm( |
| 203 | config.hidden_size, eps=config.layer_norm_eps |
| 204 | ) |
| 205 | |
| 206 | def forward(self, hidden_states): |
| 207 | attention_input = hidden_states |
| 208 | attention_output = self.input_layernorm(self.attention(attention_input)) |
| 209 | hidden_states = attention_input + attention_output |
| 210 | mlp_input = hidden_states |
| 211 | mlp_output = self.post_attention_layernorm(self.mlp(mlp_input)) |
| 212 | output = mlp_input + mlp_output |
| 213 | return output |
| 214 | |
| 215 | |
| 216 | class EVA2CLIPTransformer(nn.Module): |
| 217 | def __init__( |
| 218 | self, |
| 219 | config, |
| 220 | quant_config: QuantizationConfig | None = None, |
| 221 | prefix: str = "", |
| 222 | ): |
| 223 | super().__init__() |
| 224 | self.layers = nn.ModuleList( |
| 225 | [ |
| 226 | EVA2CLIPTransformerLayer( |
| 227 | config, |
| 228 | quant_config=quant_config, |
| 229 | prefix=f"{prefix}.layers.{layer_idx}", |
| 230 | ) |
| 231 | for layer_idx in range(config.num_hidden_layers) |
| 232 | ] |
| 233 | ) |
| 234 | |
| 235 | def forward(self, hidden_states): |
| 236 | for layer_module in self.layers: |
| 237 | hidden_states = layer_module(hidden_states) |
| 238 | return hidden_states |
| 239 | |
| 240 | |
| 241 | class EVA2CLIPGLU(nn.Module): |
| 242 | def __init__( |
| 243 | self, |
| 244 | config, |
| 245 | in_features, |
| 246 | quant_config: QuantizationConfig | None = None, |
| 247 | prefix: str = "", |
| 248 | ): |
| 249 | """The original implementation is the same as: |
| 250 | ```python |
| 251 | self.dense_h_to_4h = ColumnParallelLinear( |
| 252 | config.hidden_size, |
| 253 | config.ffn_hidden_size, |
| 254 | bias=False, |
| 255 | quant_config=quant_config, |
| 256 | ) |
| 257 | |
| 258 | self.gate_proj = ColumnParallelLinear( |
| 259 | config.hidden_size, |
| 260 | config.ffn_hidden_size, |
| 261 | bias=False, |
| 262 | quant_config=quant_config, |
| 263 | ) |
| 264 | ``` |
| 265 | ``` |
| 266 | gate_proj_output, _ = self.gate_proj(x) |
| 267 | dense_h_to_4h_output, _ = self.dense_h_to_4h(x) |
| 268 | x = torch.cat([gate_proj_output, dense_h_to_4h_output], dim=-1) |
| 269 | ``` |
| 270 | |
| 271 | We merge two ColumnParallelLinear into one MergedColumnParallelLinear: |
| 272 | ``` |
| 273 | self.merged_proj = MergedColumnParallelLinear( |
| 274 | config.hidden_size, |
| 275 | [config.ffn_hidden_size] * 2, |
| 276 | bias=False, |
| 277 | quant_config=quant_config, |
| 278 | ) |
| 279 | ``` |
| 280 | ``` |
| 281 | x, _ = self.merged_proj(x) |
| 282 | ``` |
| 283 | """ |
| 284 | super().__init__() |
| 285 | self.linear_proj = ReplicatedLinear( |
| 286 | in_features, |
| 287 | config.hidden_size, |
| 288 | bias=False, |
| 289 | quant_config=quant_config, |
| 290 | prefix=f"{prefix}.linear_proj", |
| 291 | ) |
| 292 | self.norm1 = nn.LayerNorm(config.hidden_size) |
| 293 | self.act1 = nn.GELU() |
| 294 | self.act2 = SiluAndMul() |
| 295 | |
| 296 | self.merged_proj = MergedColumnParallelLinear( |
| 297 | config.hidden_size, |
| 298 | [config.ffn_hidden_size] * 2, |
| 299 | bias=False, |
| 300 | quant_config=quant_config, |
| 301 | prefix=f"{prefix}.merged_proj", |
| 302 | ) |
| 303 | |
| 304 | self.dense_4h_to_h = RowParallelLinear( |
| 305 | config.ffn_hidden_size, |
| 306 | config.hidden_size, |
| 307 | bias=False, |
| 308 | quant_config=quant_config, |
| 309 | prefix=f"{prefix}.dense_4h_to_h", |
| 310 | ) |
| 311 | |
| 312 | def forward(self, x): |
| 313 | x, _ = self.linear_proj(x) |
| 314 | x = self.act1(self.norm1(x)) |
| 315 | x, _ = self.merged_proj(x) |
| 316 | x = self.act2(x) |
| 317 | x, _ = self.dense_4h_to_h(x) |
| 318 | return x |
| 319 | |
| 320 | |
| 321 | class EVA2CLIPModel(nn.Module): |
| 322 | def __init__( |
| 323 | self, |
| 324 | config, |
| 325 | quant_config: QuantizationConfig | None = None, |
| 326 | prefix: str = "", |
| 327 | ): |
| 328 | super().__init__() |
| 329 | vision_config = Namespace(**config.vision_config) |
| 330 | self.patch_embedding = EVA2CLIPPatchEmbedding(vision_config) |
| 331 | self.transformer = EVA2CLIPTransformer( |
| 332 | vision_config, quant_config=quant_config, prefix=f"{prefix}.transformer" |
| 333 | ) |
| 334 | self.linear_proj = EVA2CLIPGLU( |
| 335 | config, |
| 336 | in_features=config.hidden_size, |
| 337 | quant_config=quant_config, |
| 338 | prefix=f"{prefix}.linear_proj", |
| 339 | ) |
| 340 | self.conv = Conv2dLayer( |
| 341 | in_channels=vision_config.hidden_size, |
| 342 | out_channels=config.hidden_size, |
| 343 | kernel_size=2, |
| 344 | stride=2, |
| 345 | ) |
| 346 | self.boi = nn.Parameter(torch.zeros(1, 1, config.hidden_size)) |
| 347 | self.eoi = nn.Parameter(torch.zeros(1, 1, config.hidden_size)) |
| 348 | self.scaling_factor = vision_config.scaling_factor |
| 349 | |
| 350 | def forward(self, images: torch.Tensor) -> torch.Tensor: |
| 351 | """Parameters |
| 352 | images : torch.Tensor |
| 353 | Input image tensor with shape (B, C, H, W) |
| 354 | |
| 355 | Returns |
| 356 | ------- |
| 357 | torch.Tensor |
| 358 | Transformed tensor with shape (B, L, D) |
| 359 | |
| 360 | """ |
| 361 | x = self.patch_embedding(images) |
| 362 | x = self.transformer(x) |
| 363 | x = x[:, 1:] |
| 364 | |
| 365 | b, s, h = x.shape |
| 366 | grid_size = int(s**0.5) |
| 367 | x = x.view(b, grid_size, grid_size, h).permute(0, 3, 1, 2) |
| 368 | x = self.conv(x) |
| 369 | |
| 370 | x = x.flatten(2).transpose(1, 2) |
| 371 | x = self.linear_proj(x) |
| 372 | boi = self.boi.expand(x.shape[0], -1, -1) |
| 373 | eoi = self.eoi.expand(x.shape[0], -1, -1) |
| 374 | x = torch.cat((boi, x, eoi), dim=1) |
| 375 | x = x / self.scaling_factor |
| 376 | return x |
| 377 | |
| 378 | |
| 379 | class GLM4VModel(ChatGLMModel): |
| 380 | hf_to_vllm_mapper = ChatGLMModel.hf_to_vllm_mapper | WeightsMapper( |
| 381 | orig_to_new_stacked={ |
| 382 | # weight_name: (param_name, shard_id) |
| 383 | # Vision GLU projections |
| 384 | "linear_proj.gate_proj": ("linear_proj.merged_proj", 0), |
| 385 | "linear_proj.dense_h_to_4h": ("linear_proj.merged_proj", 1), |
| 386 | } |
| 387 | ) |
| 388 | |
| 389 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 390 | super().__init__(vllm_config=vllm_config, prefix=prefix) |
| 391 | |
| 392 | quant_config = vllm_config.quant_config |
| 393 | |
| 394 | self.vision = EVA2CLIPModel( |
| 395 | self.config, quant_config, prefix=f"{prefix}.vision" |
| 396 | ) |
| 397 | |
| 398 | |
| 399 | class GLM4VProcessingInfo(BaseProcessingInfo): |
| 400 | def get_hf_config(self): |
| 401 | return self.ctx.get_hf_config(ChatGLMConfig) |
| 402 | |
| 403 | def get_image_processor(self, **kwargs): |
| 404 | config = self.get_hf_config() |
| 405 | vision_config = config.vision_config |
| 406 | |
| 407 | image_size = vision_config["image_size"] |
| 408 | kwargs = self.ctx.get_merged_mm_kwargs(kwargs) |
| 409 | kwargs.setdefault("size", {"width": image_size, "height": image_size}) |
| 410 | |
| 411 | return GLM4VImageProcessorFast(**kwargs) |
| 412 | |
| 413 | def get_hf_processor(self, **kwargs: object) -> GLM4VProcessor: |
| 414 | return GLM4VProcessor( |
| 415 | tokenizer=self.get_tokenizer(), |
| 416 | image_processor=self.get_image_processor(**kwargs), |
| 417 | ) |
| 418 | |
| 419 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 420 | return {"image": 1} |
| 421 | |
| 422 | def get_num_image_tokens(self) -> int: |
| 423 | hf_config = self.get_hf_config() |
| 424 | vision_config = hf_config.vision_config |
| 425 | |
| 426 | image_size = vision_config["image_size"] |
| 427 | patch_size = vision_config["patch_size"] |
| 428 | grid_length = image_size // patch_size // 2 |
| 429 | return grid_length * grid_length |
| 430 | |
| 431 | def get_num_image_feature_tokens(self) -> int: |
| 432 | # EVA2CLIPModel has embeddings for boi and eoi tokens as well |
| 433 | return self.get_num_image_tokens() + 2 |
| 434 | |
| 435 | |
| 436 | class GLM4VDummyInputsBuilder(BaseDummyInputsBuilder[GLM4VProcessingInfo]): |
| 437 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 438 | num_images = mm_counts.get("image", 0) |
| 439 | |
| 440 | base_text = "<|begin_of_image|><|endoftext|><|end_of_image|>" |
| 441 | |
| 442 | return base_text * num_images |
| 443 | |
| 444 | def get_dummy_mm_data( |
| 445 | self, |
| 446 | seq_len: int, |
| 447 | mm_counts: Mapping[str, int], |
| 448 | mm_options: MultiModalDummyOptions, |
| 449 | ) -> MultiModalDataDict: |
| 450 | hf_config = self.info.get_hf_config() |
| 451 | vision_config = hf_config.vision_config |
| 452 | |
| 453 | target_width = target_height = vision_config["image_size"] |
| 454 | |
| 455 | return { |
| 456 | "image": self._get_dummy_images( |
| 457 | width=target_width, |
| 458 | height=target_height, |
| 459 | num_images=mm_counts.get("image", 0), |
| 460 | overrides=mm_options.get("image"), |
| 461 | ) |
| 462 | } |
| 463 | |
| 464 | |
| 465 | class GLM4VMultiModalProcessor(BaseMultiModalProcessor[GLM4VProcessingInfo]): |
| 466 | def _get_mm_fields_config( |
| 467 | self, |
| 468 | hf_inputs: BatchFeature, |
| 469 | hf_processor_mm_kwargs: Mapping[str, object], |
| 470 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 471 | return dict(pixel_values=MultiModalFieldConfig.batched("image")) |
| 472 | |
| 473 | def _get_prompt_updates( |
| 474 | self, |
| 475 | mm_items: MultiModalDataItems, |
| 476 | hf_processor_mm_kwargs: Mapping[str, object], |
| 477 | out_mm_kwargs: MultiModalKwargsItems, |
| 478 | ) -> Sequence[PromptUpdate]: |
| 479 | hf_config = self.info.get_hf_config() |
| 480 | |
| 481 | boi_token_id = hf_config.boi_token_id |
| 482 | image_token_id = hf_config.pad_token_id |
| 483 | eoi_token_id = hf_config.eoi_token_id |
| 484 | |
| 485 | def get_replacement(item_idx: int): |
| 486 | num_image_tokens = self.info.get_num_image_tokens() |
| 487 | image_tokens = [image_token_id] * num_image_tokens |
| 488 | |
| 489 | return [boi_token_id] + image_tokens + [eoi_token_id] |
| 490 | |
| 491 | return [ |
| 492 | PromptReplacement( |
| 493 | modality="image", |
| 494 | target=[boi_token_id, image_token_id, eoi_token_id], |
| 495 | replacement=get_replacement, |
| 496 | ), |
| 497 | ] |
| 498 | |
| 499 | |
| 500 | @MULTIMODAL_REGISTRY.register_processor( |
| 501 | GLM4VMultiModalProcessor, |
| 502 | info=GLM4VProcessingInfo, |
| 503 | dummy_inputs=GLM4VDummyInputsBuilder, |
| 504 | ) |
| 505 | class GLM4VForCausalLM( |
| 506 | ChatGLMBaseModel, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE |
| 507 | ): |
| 508 | # NOTE: we must bring this to the surface because GLM4VModel.hf_to_vllm_mapper |
| 509 | # contains non-stacking related mappings which LoRA/BnB needs to know about |
| 510 | hf_to_vllm_mapper = GLM4VModel.hf_to_vllm_mapper |
| 511 | packed_modules_mapping = { |
| 512 | "query_key_value": ["query_key_value"], |
| 513 | "dense_h_to_4h": ["dense_h_to_4h"], |
| 514 | "merged_proj": ["gate_proj", "dense_h_to_4h"], |
| 515 | } |
| 516 | |
| 517 | def get_mm_mapping(self) -> MultiModelKeys: |
| 518 | """Get the module prefix in multimodal models.""" |
| 519 | return MultiModelKeys.from_string_field( |
| 520 | language_model="transformer.encoder", |
| 521 | connector="transformer.vision.linear_proj", |
| 522 | tower_model="transformer.vision.transformer", |
| 523 | ) |
| 524 | |
| 525 | @classmethod |
| 526 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 527 | if modality.startswith("image"): |
| 528 | return "<|begin_of_image|><|endoftext|><|end_of_image|>" |
| 529 | |
| 530 | raise ValueError("Only image modality is supported") |
| 531 | |
| 532 | def __init__( |
| 533 | self, |
| 534 | *, |
| 535 | vllm_config: VllmConfig, |
| 536 | prefix: str = "", |
| 537 | transformer_type: type[GLM4VModel] = GLM4VModel, |
| 538 | ) -> None: |
| 539 | with self._mark_composite_model( |
| 540 | vllm_config, |
| 541 | language_targets=GLMTransformer, |
| 542 | tower_targets={"image": EVA2CLIPModel}, |
| 543 | ): |
| 544 | super().__init__( |
| 545 | vllm_config=vllm_config, |
| 546 | prefix=prefix, |
| 547 | transformer_type=transformer_type, |
| 548 | ) |
| 549 | |
| 550 | self.transformer: GLM4VModel |
| 551 | |
| 552 | def _parse_and_validate_image_input( |
| 553 | self, **kwargs: object |
| 554 | ) -> GLMVImagePixelInputs | None: |
| 555 | pixel_values = kwargs.pop("pixel_values", None) |
| 556 | |
| 557 | if pixel_values is not None: |
| 558 | expected_h = expected_w = self.config.vision_config["image_size"] |
| 559 | return GLMVImagePixelInputs( |
| 560 | type="pixel_values", |
| 561 | data=pixel_values, |
| 562 | resolve_bindings={"h": expected_h, "w": expected_w}, |
| 563 | ) |
| 564 | |
| 565 | return None |
| 566 | |
| 567 | def _process_image_input(self, image_input: GLMVImagePixelInputs) -> torch.Tensor: |
| 568 | pixel_values = image_input["data"].to(dtype=self.config.dtype) |
| 569 | |
| 570 | return self.transformer.vision(pixel_values) |
| 571 | |
| 572 | def iter_mm_grid_thw( |
| 573 | self, mm_features: list[MultiModalFeatureSpec] |
| 574 | ) -> Iterator[tuple[int, int, int, int]]: |
| 575 | hf_config = self.config |
| 576 | spatial_merge_size = hf_config.vision_config.spatial_merge_size |
| 577 | for mm_feature in sorted(mm_features, key=lambda f: f.mm_position.offset): |
| 578 | offset = mm_feature.mm_position.offset |
| 579 | if mm_feature.modality == "image": |
| 580 | feature_data = mm_feature.data |
| 581 | assert feature_data is not None |
| 582 | grid_item = feature_data.get("image_grid_thw") |
| 583 | assert grid_item is not None |
| 584 | grid_data = grid_item.data |
| 585 | assert isinstance(grid_data, torch.Tensor) |
| 586 | t, h, w = grid_data.tolist() |
| 587 | assert t == 1, f"Image must have 1 frame, got {t}" |
| 588 | yield offset, t, h // spatial_merge_size, w // spatial_merge_size |
| 589 | else: |
| 590 | # glm4v only supports image modality |
| 591 | raise ValueError(f"Unsupported modality: {mm_feature.modality}") |
| 592 | |
| 593 | def get_mrope_input_positions( |
| 594 | self, |
| 595 | input_tokens: list[int], |
| 596 | mm_features: list[MultiModalFeatureSpec], |
| 597 | ) -> tuple[torch.Tensor, int]: |
| 598 | llm_pos_ids_list: list = [] |
| 599 | st = 0 |
| 600 | for ( |
| 601 | offset, |
| 602 | llm_grid_t, |
| 603 | llm_grid_h, |
| 604 | llm_grid_w, |
| 605 | ) in self.iter_mm_grid_thw(mm_features): |
| 606 | text_len = offset - st |
| 607 | st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0 |
| 608 | llm_pos_ids_list.append( |
| 609 | np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx |
| 610 | ) |
| 611 | grid_indices = np.indices((llm_grid_t, llm_grid_h, llm_grid_w)).reshape( |
| 612 | 3, -1 |
| 613 | ) |
| 614 | llm_pos_ids_list.append(grid_indices + text_len + st_idx) |
| 615 | # EVA2CLIPModel has embeddings for boi and eoi tokens as well |
| 616 | st = offset + 1 + llm_grid_t * llm_grid_h * llm_grid_w + 1 |
| 617 | |
| 618 | if st < len(input_tokens): |
| 619 | text_len = len(input_tokens) - st |
| 620 | st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0 |
| 621 | llm_pos_ids_list.append( |
| 622 | np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx |
| 623 | ) |
| 624 | |
| 625 | llm_positions = np.concatenate(llm_pos_ids_list, axis=1).reshape(3, -1) |
| 626 | mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item() |
| 627 | return torch.from_numpy(llm_positions), mrope_position_delta |
| 628 | |
| 629 | def embed_input_ids( |
| 630 | self, |
| 631 | input_ids: torch.Tensor, |
| 632 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 633 | is_multimodal: torch.Tensor | None = None, |
| 634 | ) -> torch.Tensor: |
| 635 | return SupportsMultiModal.embed_input_ids( |
| 636 | self, |
| 637 | input_ids, |
| 638 | multimodal_embeddings, |
| 639 | is_multimodal=is_multimodal, |
| 640 | ) |
| 641 | |
| 642 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 643 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 644 | if image_input is None: |
| 645 | return [] |
| 646 | |
| 647 | vision_embeddings = self._process_image_input(image_input) |
| 648 | return vision_embeddings |
| 649 | |
| 650 | def forward( |
| 651 | self, |
| 652 | input_ids: torch.Tensor | None, |
| 653 | positions: torch.Tensor, |
| 654 | intermediate_tensors: IntermediateTensors | None = None, |
| 655 | inputs_embeds: torch.Tensor | None = None, |
| 656 | **kwargs: object, |
| 657 | ) -> torch.Tensor | IntermediateTensors: |
| 658 | if intermediate_tensors is not None: |
| 659 | inputs_embeds = None |
| 660 | |
| 661 | hidden_states = self.transformer( |
| 662 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 663 | ) |
| 664 | |
| 665 | return hidden_states |