| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # |
| 4 | # Copyright 2025 the LLAMA4, Meta Inc., vLLM, and HuggingFace Inc. team. |
| 5 | # All rights reserved. |
| 6 | # |
| 7 | # |
| 8 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 9 | # you may not use this file except in compliance with the License. |
| 10 | # You may obtain a copy of the License at |
| 11 | # |
| 12 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 13 | # |
| 14 | # Unless required by applicable law or agreed to in writing, software |
| 15 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 16 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 17 | # See the License for the specific language governing permissions and |
| 18 | # limitations under the License. |
| 19 | import math |
| 20 | from collections.abc import Hashable, Iterable, Mapping |
| 21 | from typing import Annotated, Any, Literal |
| 22 | |
| 23 | import torch |
| 24 | from torch import nn |
| 25 | from transformers import BatchFeature, Llama4Config, Llama4VisionConfig |
| 26 | from transformers.image_utils import SizeDict |
| 27 | from transformers.models.llama4 import Llama4Processor |
| 28 | from transformers.models.llama4.image_processing_llama4 import ( |
| 29 | find_supported_resolutions, |
| 30 | get_best_fit, |
| 31 | ) |
| 32 | |
| 33 | from vllm.compilation.decorators import ( |
| 34 | should_torch_compile_mm_encoder, |
| 35 | support_torch_compile, |
| 36 | ) |
| 37 | from vllm.config import VllmConfig, set_current_vllm_config |
| 38 | from vllm.config.multimodal import MultiModalDummyOptions |
| 39 | from vllm.distributed import get_tensor_model_parallel_world_size |
| 40 | from vllm.inputs import MultiModalDataDict |
| 41 | from vllm.model_executor.layers.attention import MMEncoderAttention |
| 42 | from vllm.model_executor.layers.fused_moe import ( |
| 43 | fused_moe_make_expert_params_mapping, |
| 44 | ) |
| 45 | from vllm.model_executor.layers.linear import ( |
| 46 | ColumnParallelLinear, |
| 47 | QKVParallelLinear, |
| 48 | ReplicatedLinear, |
| 49 | RowParallelLinear, |
| 50 | ) |
| 51 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 52 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 53 | from vllm.model_executor.model_loader.utils import initialize_model |
| 54 | from vllm.model_executor.model_loader.weight_utils import ( |
| 55 | default_weight_loader, |
| 56 | maybe_remap_moe_expert_param_name, |
| 57 | ) |
| 58 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 59 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 60 | from vllm.multimodal.inputs import ( |
| 61 | MultiModalFieldConfig, |
| 62 | MultiModalKwargsItem, |
| 63 | MultiModalKwargsItems, |
| 64 | ) |
| 65 | from vllm.multimodal.parse import ImageProcessorItems, ImageSize, MultiModalDataItems |
| 66 | from vllm.multimodal.processing import ( |
| 67 | BaseDummyInputsBuilder, |
| 68 | BaseMultiModalProcessor, |
| 69 | BaseProcessingInfo, |
| 70 | PromptReplacement, |
| 71 | PromptUpdate, |
| 72 | PromptUpdateDetails, |
| 73 | cached_encode, |
| 74 | ) |
| 75 | from vllm.sequence import IntermediateTensors |
| 76 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 77 | |
| 78 | from .interfaces import ( |
| 79 | MixtureOfExperts, |
| 80 | MultiModalEmbeddings, |
| 81 | SupportsEagle3, |
| 82 | SupportsEncoderCudaGraph, |
| 83 | SupportsLoRA, |
| 84 | SupportsMultiModal, |
| 85 | SupportsPP, |
| 86 | ) |
| 87 | from .llama4 import Llama4ForCausalLM |
| 88 | from .utils import AutoWeightsLoader, StageMissingLayer, maybe_prefix |
| 89 | from .vision import is_vit_use_data_parallel, run_dp_sharded_vision_model |
| 90 | |
| 91 | |
| 92 | class Llama4ImagePatchInputs(TensorSchema): |
| 93 | """Dimensions: |
| 94 | - batch_size: Batch size |
| 95 | - total_num_chunks: Batch size * number of chunks |
| 96 | - num_channels: Number of channels |
| 97 | - image_size: Size of each image |
| 98 | """ |
| 99 | |
| 100 | type: Literal["pixel_values"] = "pixel_values" |
| 101 | |
| 102 | pixel_values: Annotated[ |
| 103 | torch.Tensor, |
| 104 | TensorShape("total_num_chunks", "num_channels", "image_size", "image_size"), |
| 105 | ] |
| 106 | |
| 107 | patches_per_image: Annotated[torch.Tensor, TensorShape("batch_size")] |
| 108 | """ |
| 109 | The number of chunked image tiles for each image in the batch. |
| 110 | |
| 111 | This is used to split the embeddings which has the first two dimensions |
| 112 | flattened just like `pixel_values`. |
| 113 | """ |
| 114 | |
| 115 | aspect_ratios: Annotated[torch.Tensor, TensorShape("batch_size", 2)] |
| 116 | """ |
| 117 | A list of aspect ratios corresponding to the number of tiles |
| 118 | in each dimension that each image in the batch corresponds to. |
| 119 | Each aspect ratio is a pair (ratio_h, ratio_w). |
| 120 | """ |
| 121 | |
| 122 | |
| 123 | class Llama4VisionMLP(nn.Module): |
| 124 | def __init__( |
| 125 | self, |
| 126 | input_size: int, |
| 127 | intermediate_size: int, |
| 128 | output_size: int, |
| 129 | bias: bool, |
| 130 | output_activation: bool, |
| 131 | quant_config: QuantizationConfig | None = None, |
| 132 | prefix: str = "", |
| 133 | ): |
| 134 | super().__init__() |
| 135 | use_data_parallel = is_vit_use_data_parallel() |
| 136 | self.fc1 = ColumnParallelLinear( |
| 137 | input_size=input_size, |
| 138 | output_size=intermediate_size, |
| 139 | bias=bias, |
| 140 | quant_config=quant_config, |
| 141 | prefix=f"{prefix}.fc1", |
| 142 | disable_tp=use_data_parallel, |
| 143 | ) |
| 144 | self.fc2 = RowParallelLinear( |
| 145 | input_size=intermediate_size, |
| 146 | output_size=output_size, |
| 147 | bias=bias, |
| 148 | quant_config=quant_config, |
| 149 | prefix=f"{prefix}.fc2", |
| 150 | disable_tp=use_data_parallel, |
| 151 | ) |
| 152 | self.activation_fn = nn.GELU() |
| 153 | self.output_activation = output_activation |
| 154 | |
| 155 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 156 | hidden_states, _ = self.fc1(hidden_states) |
| 157 | hidden_states = self.activation_fn(hidden_states) |
| 158 | hidden_states, _ = self.fc2(hidden_states) |
| 159 | if self.output_activation: |
| 160 | return self.activation_fn(hidden_states) |
| 161 | return hidden_states |
| 162 | |
| 163 | |
| 164 | class Llama4MultiModalProjector(nn.Module): |
| 165 | def __init__( |
| 166 | self, |
| 167 | config, |
| 168 | quant_config: QuantizationConfig | None = None, |
| 169 | prefix: str = "", |
| 170 | ): |
| 171 | super().__init__() |
| 172 | self.linear_1 = ColumnParallelLinear( |
| 173 | input_size=config.vision_config.vision_output_dim, |
| 174 | output_size=config.text_config.hidden_size, |
| 175 | bias=False, |
| 176 | quant_config=quant_config, |
| 177 | gather_output=True, |
| 178 | prefix=f"{prefix}.linear_1", |
| 179 | ) |
| 180 | |
| 181 | def forward(self, image_features): |
| 182 | hidden_states, _ = self.linear_1(image_features) |
| 183 | return hidden_states |
| 184 | |
| 185 | |
| 186 | def pixel_shuffle(input_tensor, shuffle_ratio): |
| 187 | # input_tensor: [batch_size, num_patches, channels] |
| 188 | batch_size, num_patches, channels = input_tensor.shape |
| 189 | patch_size = int(math.sqrt(num_patches)) |
| 190 | |
| 191 | input_tensor = input_tensor.view(batch_size, patch_size, patch_size, -1) |
| 192 | batch_size, height, width, channels = input_tensor.size() |
| 193 | |
| 194 | reshaped_tensor = input_tensor.view( |
| 195 | batch_size, height, int(width * shuffle_ratio), int(channels / shuffle_ratio) |
| 196 | ) |
| 197 | reshaped_tensor = reshaped_tensor.permute(0, 2, 1, 3).contiguous() |
| 198 | |
| 199 | reshaped_tensor = reshaped_tensor.view( |
| 200 | batch_size, |
| 201 | int(height * shuffle_ratio), |
| 202 | int(width * shuffle_ratio), |
| 203 | int(channels / (shuffle_ratio**2)), |
| 204 | ) |
| 205 | reshaped_tensor = reshaped_tensor.permute(0, 2, 1, 3).contiguous() |
| 206 | |
| 207 | output_tensor = reshaped_tensor.view(batch_size, -1, reshaped_tensor.shape[-1]) |
| 208 | return output_tensor |
| 209 | |
| 210 | |
| 211 | class Llama4VisionPixelShuffleMLP(nn.Module): |
| 212 | def __init__( |
| 213 | self, |
| 214 | config, |
| 215 | quant_config: QuantizationConfig | None = None, |
| 216 | prefix: str = "", |
| 217 | ): |
| 218 | super().__init__() |
| 219 | self.pixel_shuffle_ratio = config.pixel_shuffle_ratio |
| 220 | self.inner_dim = int( |
| 221 | config.projector_input_dim // (self.pixel_shuffle_ratio**2) |
| 222 | ) |
| 223 | self.output_dim = config.projector_output_dim |
| 224 | self.mlp = Llama4VisionMLP( |
| 225 | input_size=config.intermediate_size, |
| 226 | intermediate_size=config.projector_input_dim, |
| 227 | output_size=config.projector_output_dim, |
| 228 | bias=config.multi_modal_projector_bias, |
| 229 | output_activation=True, |
| 230 | quant_config=quant_config, |
| 231 | prefix=f"{prefix}.mlp", |
| 232 | ) |
| 233 | |
| 234 | def forward(self, encoded_patches: torch.Tensor) -> torch.Tensor: |
| 235 | encoded_patches = pixel_shuffle(encoded_patches, self.pixel_shuffle_ratio) |
| 236 | return self.mlp(encoded_patches) |
| 237 | |
| 238 | |
| 239 | class Llama4VisionAttention(nn.Module): |
| 240 | def __init__( |
| 241 | self, |
| 242 | config: Llama4VisionConfig, |
| 243 | quant_config: QuantizationConfig | None, |
| 244 | prefix: str = "", |
| 245 | ): |
| 246 | super().__init__() |
| 247 | self.config = config |
| 248 | use_data_parallel = is_vit_use_data_parallel() |
| 249 | self.tp_size = ( |
| 250 | 1 if use_data_parallel else get_tensor_model_parallel_world_size() |
| 251 | ) |
| 252 | self.embed_dim = config.hidden_size |
| 253 | self.num_heads = config.num_attention_heads |
| 254 | self.head_dim = config.hidden_size // self.num_heads |
| 255 | assert self.num_heads % self.tp_size == 0 |
| 256 | self.num_local_heads = self.num_heads // self.tp_size |
| 257 | self.q_size = self.num_local_heads * self.head_dim |
| 258 | self.kv_size = self.num_local_heads * self.head_dim |
| 259 | self.attention_dropout = config.attention_dropout |
| 260 | self.scaling = self.head_dim**-0.5 |
| 261 | |
| 262 | self.attn = MMEncoderAttention( |
| 263 | self.num_local_heads, |
| 264 | self.head_dim, |
| 265 | self.scaling, |
| 266 | prefix=f"{prefix}.attn", |
| 267 | ) |
| 268 | |
| 269 | if use_data_parallel: |
| 270 | self.qkv_proj = ReplicatedLinear( |
| 271 | self.embed_dim, |
| 272 | self.q_size + 2 * self.kv_size, |
| 273 | bias=True, |
| 274 | quant_config=quant_config, |
| 275 | prefix=f"{prefix}.qkv_proj", |
| 276 | ) |
| 277 | self.o_proj = ReplicatedLinear( |
| 278 | self.num_heads * self.head_dim, |
| 279 | self.embed_dim, |
| 280 | bias=True, |
| 281 | quant_config=quant_config, |
| 282 | prefix=f"{prefix}.o_proj", |
| 283 | ) |
| 284 | else: |
| 285 | self.qkv_proj = QKVParallelLinear( |
| 286 | self.embed_dim, |
| 287 | self.head_dim, |
| 288 | self.num_heads, |
| 289 | bias=True, |
| 290 | quant_config=quant_config, |
| 291 | prefix=f"{prefix}.qkv_proj", |
| 292 | ) |
| 293 | self.o_proj = RowParallelLinear( |
| 294 | self.num_heads * self.head_dim, |
| 295 | self.embed_dim, |
| 296 | bias=True, |
| 297 | input_is_parallel=True, |
| 298 | quant_config=quant_config, |
| 299 | prefix=f"{prefix}.o_proj", |
| 300 | ) |
| 301 | |
| 302 | rope_parameters = { |
| 303 | "rope_type": "mllama4", |
| 304 | "rope_theta": config.rope_parameters["rope_theta"], |
| 305 | "partial_rotary_factor": 0.5, |
| 306 | } |
| 307 | |
| 308 | self.rotary_emb = get_rope( |
| 309 | head_size=self.head_dim, |
| 310 | # number of image patches |
| 311 | max_position=(config.image_size // config.patch_size) ** 2, |
| 312 | rope_parameters=rope_parameters, |
| 313 | is_neox_style=False, |
| 314 | dtype=torch.complex64, # important |
| 315 | ) |
| 316 | |
| 317 | def forward( |
| 318 | self, |
| 319 | hidden_states: torch.Tensor, |
| 320 | ) -> torch.Tensor: |
| 321 | input_shape = hidden_states.shape[:-1] |
| 322 | |
| 323 | qkv, _ = self.qkv_proj(hidden_states) |
| 324 | q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) |
| 325 | |
| 326 | q = q.view(q.shape[0], q.shape[1], self.num_local_heads, self.head_dim) |
| 327 | k = k.view(k.shape[0], k.shape[1], self.num_local_heads, self.head_dim) |
| 328 | q, k = self.rotary_emb(q, k) |
| 329 | |
| 330 | q = q.view(q.shape[0], q.shape[1], -1) |
| 331 | k = k.view(k.shape[0], k.shape[1], -1) |
| 332 | |
| 333 | attn_output = self.attn(q, k, v) |
| 334 | attn_output = attn_output.reshape(*input_shape, -1).contiguous() |
| 335 | attn_output, _ = self.o_proj(attn_output) |
| 336 | |
| 337 | return attn_output |
| 338 | |
| 339 | |
| 340 | class Llama4VisionEncoderLayer(nn.Module): |
| 341 | def __init__( |
| 342 | self, |
| 343 | config: Llama4VisionConfig, |
| 344 | quant_config: QuantizationConfig | None, |
| 345 | prefix: str = "", |
| 346 | ): |
| 347 | super().__init__() |
| 348 | self.hidden_size = config.hidden_size |
| 349 | self.num_attention_heads = config.num_attention_heads |
| 350 | self.intermediate_size = config.intermediate_size |
| 351 | |
| 352 | self.self_attn = Llama4VisionAttention( |
| 353 | config, |
| 354 | quant_config=quant_config, |
| 355 | prefix=f"{prefix}.self_attn", |
| 356 | ) |
| 357 | self.mlp = Llama4VisionMLP( |
| 358 | input_size=config.hidden_size, |
| 359 | intermediate_size=config.intermediate_size, |
| 360 | output_size=config.hidden_size, |
| 361 | bias=True, |
| 362 | output_activation=False, |
| 363 | quant_config=quant_config, |
| 364 | prefix=f"{prefix}.mlp", |
| 365 | ) |
| 366 | |
| 367 | self.input_layernorm = nn.LayerNorm(config.hidden_size) |
| 368 | self.post_attention_layernorm = nn.LayerNorm(config.hidden_size) |
| 369 | |
| 370 | def forward( |
| 371 | self, |
| 372 | hidden_state: torch.Tensor, |
| 373 | ): |
| 374 | # Self Attention |
| 375 | residual = hidden_state |
| 376 | hidden_state = self.input_layernorm(hidden_state) |
| 377 | hidden_state = self.self_attn(hidden_state) |
| 378 | hidden_state = residual + hidden_state |
| 379 | |
| 380 | # Feed forward |
| 381 | residual = hidden_state |
| 382 | hidden_state = self.post_attention_layernorm(hidden_state) |
| 383 | hidden_state = self.mlp(hidden_state) |
| 384 | hidden_state = residual + hidden_state |
| 385 | |
| 386 | outputs = (hidden_state,) |
| 387 | return outputs |
| 388 | |
| 389 | |
| 390 | class Llama4VisionEncoder(nn.Module): |
| 391 | def __init__( |
| 392 | self, |
| 393 | config: Llama4VisionConfig, |
| 394 | quant_config: QuantizationConfig | None, |
| 395 | prefix: str = "", |
| 396 | ): |
| 397 | super().__init__() |
| 398 | self.config = config |
| 399 | self.layers = nn.ModuleList( |
| 400 | [ |
| 401 | Llama4VisionEncoderLayer( |
| 402 | config, |
| 403 | quant_config=quant_config, |
| 404 | prefix=f"{prefix}.layers.{layer_idx}", |
| 405 | ) |
| 406 | for layer_idx in range(config.num_hidden_layers) |
| 407 | ] |
| 408 | ) |
| 409 | |
| 410 | def forward( |
| 411 | self, |
| 412 | hidden_states: torch.Tensor, |
| 413 | ) -> torch.Tensor: |
| 414 | r"""Args: |
| 415 | hidden_states: Input tensor of shape |
| 416 | (batch_size, sequence_length, hidden_size). |
| 417 | Hidden states from the model embeddings, representing |
| 418 | the input tokens. |
| 419 | associated vectors than the model's internal embedding |
| 420 | lookup matrix. |
| 421 | |
| 422 | """ |
| 423 | for encoder_layer in self.layers: |
| 424 | layer_outputs = encoder_layer(hidden_states) |
| 425 | hidden_states = layer_outputs[0] |
| 426 | |
| 427 | return hidden_states |
| 428 | |
| 429 | |
| 430 | class Llama4UnfoldConvolution(nn.Module): |
| 431 | def __init__( |
| 432 | self, |
| 433 | config: Llama4VisionConfig, |
| 434 | quant_config: QuantizationConfig | None = None, |
| 435 | prefix: str = "", |
| 436 | ): |
| 437 | super().__init__() |
| 438 | kernel_size = config.patch_size |
| 439 | if isinstance(kernel_size, int): |
| 440 | kernel_size = (kernel_size, kernel_size) |
| 441 | self.unfold = torch.nn.Unfold(kernel_size=kernel_size, stride=config.patch_size) |
| 442 | use_data_parallel = is_vit_use_data_parallel() |
| 443 | self.linear = ColumnParallelLinear( |
| 444 | input_size=config.num_channels * kernel_size[0] * kernel_size[1], |
| 445 | output_size=config.hidden_size, |
| 446 | bias=False, |
| 447 | gather_output=True, |
| 448 | quant_config=quant_config, |
| 449 | prefix=f"{prefix}.linear", |
| 450 | disable_tp=use_data_parallel, |
| 451 | ) |
| 452 | |
| 453 | def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: |
| 454 | hidden_states = self.unfold(hidden_states) |
| 455 | hidden_states = hidden_states.permute(0, 2, 1) |
| 456 | hidden_states, _ = self.linear(hidden_states) |
| 457 | return hidden_states |
| 458 | |
| 459 | |
| 460 | @support_torch_compile( |
| 461 | dynamic_arg_dims={"images_flattened": 0}, |
| 462 | enable_if=should_torch_compile_mm_encoder, |
| 463 | is_encoder=True, |
| 464 | ) |
| 465 | class Llama4VisionModel(nn.Module): |
| 466 | def __init__( |
| 467 | self, |
| 468 | config: Llama4VisionConfig, |
| 469 | quant_config: QuantizationConfig | None = None, |
| 470 | prefix: str = "", |
| 471 | ): |
| 472 | super().__init__() |
| 473 | self.config = config |
| 474 | self.image_size = config.image_size |
| 475 | self.patch_size = config.patch_size |
| 476 | self.hidden_size = config.hidden_size |
| 477 | self.num_channels = config.num_channels |
| 478 | |
| 479 | self.num_patches = (self.image_size // self.patch_size) ** 2 + 1 |
| 480 | self.scale = config.hidden_size**-0.5 |
| 481 | |
| 482 | self.patch_embedding = Llama4UnfoldConvolution( |
| 483 | config, |
| 484 | quant_config=quant_config, |
| 485 | prefix=f"{prefix}.patch_embedding", |
| 486 | ) |
| 487 | |
| 488 | self.class_embedding = nn.Parameter(self.scale * torch.randn(self.hidden_size)) |
| 489 | self.positional_embedding_vlm = nn.Parameter( |
| 490 | self.scale * torch.randn(self.num_patches, self.hidden_size) |
| 491 | ) |
| 492 | |
| 493 | # layer norms |
| 494 | self.layernorm_pre = nn.LayerNorm(self.hidden_size, eps=1e-5) |
| 495 | self.layernorm_post = nn.LayerNorm(self.hidden_size, eps=1e-5) |
| 496 | |
| 497 | # encoders |
| 498 | self.model = Llama4VisionEncoder( |
| 499 | config, |
| 500 | quant_config=quant_config, |
| 501 | prefix=f"{prefix}.model", |
| 502 | ) |
| 503 | |
| 504 | self.vision_adapter = Llama4VisionPixelShuffleMLP( |
| 505 | config, |
| 506 | quant_config, |
| 507 | prefix=f"{prefix}.vision_adapter", |
| 508 | ) |
| 509 | |
| 510 | def forward( |
| 511 | self, |
| 512 | images_flattened: torch.Tensor, |
| 513 | ) -> torch.Tensor: |
| 514 | # Patch embedding |
| 515 | hidden_state = self.patch_embedding(images_flattened) |
| 516 | num_tiles, num_patches, hidden_dim = hidden_state.shape |
| 517 | |
| 518 | # Add cls token |
| 519 | class_embedding = self.class_embedding.expand( |
| 520 | hidden_state.shape[0], 1, hidden_state.shape[-1] |
| 521 | ) |
| 522 | hidden_state = torch.cat([hidden_state, class_embedding], dim=1) |
| 523 | num_patches += 1 |
| 524 | |
| 525 | # Position embeddings |
| 526 | hidden_state = hidden_state.reshape( |
| 527 | num_tiles, |
| 528 | 1, |
| 529 | num_patches, |
| 530 | hidden_dim, |
| 531 | ) |
| 532 | positional_embedding = self.positional_embedding_vlm.to( |
| 533 | dtype=hidden_state.dtype, device=hidden_state.device |
| 534 | ) |
| 535 | hidden_state = hidden_state + positional_embedding |
| 536 | hidden_state = self.layernorm_pre(hidden_state) |
| 537 | hidden_state = hidden_state.view(num_tiles, -1, hidden_dim) |
| 538 | |
| 539 | # Apply encoder |
| 540 | hidden_state = self.model(hidden_state) |
| 541 | hidden_state = self.layernorm_post(hidden_state) |
| 542 | |
| 543 | # Remove CLS token output |
| 544 | hidden_state = hidden_state[:, :-1, :] |
| 545 | |
| 546 | # now, we use Llama4VisionPixelShuffle + mlp to project embeddings |
| 547 | hidden_state = self.vision_adapter(hidden_state) |
| 548 | |
| 549 | return hidden_state |
| 550 | |
| 551 | |
| 552 | class Mllama4ProcessingInfo(BaseProcessingInfo): |
| 553 | def get_hf_config(self) -> Llama4Config: |
| 554 | return self.ctx.get_hf_config(Llama4Config) |
| 555 | |
| 556 | def get_hf_processor(self, **kwargs: object) -> Llama4Processor: |
| 557 | return self.ctx.get_hf_processor(Llama4Processor, **kwargs) |
| 558 | |
| 559 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 560 | # Although vLLM can support more images from an infra capability |
| 561 | # perspective, we do not recommend using >10 images in practice. |
| 562 | return {"image": None} |
| 563 | |
| 564 | @staticmethod |
| 565 | def get_patch_per_chunk(vision_config: Llama4VisionConfig) -> int: |
| 566 | image_size = vision_config.image_size |
| 567 | patch_size = vision_config.patch_size |
| 568 | |
| 569 | assert image_size % patch_size == 0, ( |
| 570 | f"chunk size {image_size} should be multiple of " |
| 571 | ) |
| 572 | f"patch_size {patch_size}" |
| 573 | |
| 574 | ds_ratio = int(round(1.0 / (vision_config.pixel_shuffle_ratio**2))) |
| 575 | return (image_size // patch_size) ** 2 // ds_ratio |
| 576 | |
| 577 | def get_max_num_tiles(self) -> int: |
| 578 | image_processor = self.get_hf_processor().image_processor |
| 579 | return image_processor.max_patches |
| 580 | |
| 581 | def get_image_size_with_most_features(self) -> ImageSize: |
| 582 | vision_config = self.get_hf_config().vision_config |
| 583 | image_size = vision_config.image_size |
| 584 | # Result in the max possible feature size (h:w = 16:1) |
| 585 | return ImageSize(height=self.get_max_num_tiles() * image_size, width=image_size) |
| 586 | |
| 587 | |
| 588 | class Mllama4MultiModalProcessor(BaseMultiModalProcessor[Mllama4ProcessingInfo]): |
| 589 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 590 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 591 | |
| 592 | def _postprocess_hf_mm_data( |
| 593 | self, |
| 594 | mm_data: Mapping[str, object], |
| 595 | hf_processor_mm_kwargs: Mapping[str, object], |
| 596 | processed_data: BatchFeature, |
| 597 | ) -> BatchFeature: |
| 598 | if not mm_data: |
| 599 | return processed_data |
| 600 | |
| 601 | processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 602 | image_processor = processor.image_processor |
| 603 | vision_config = self.info.get_hf_config().vision_config |
| 604 | |
| 605 | if processed_data.get("pixel_values") is not None: |
| 606 | assert "images" in mm_data, ( |
| 607 | "images expected to be in mm_data when pixel_values is present" |
| 608 | ) |
| 609 | |
| 610 | images = mm_data["images"] |
| 611 | mm_items = self.info.parse_mm_data({"image": images}, validate=False) |
| 612 | parsed_images = mm_items.get_items("image", ImageProcessorItems) |
| 613 | |
| 614 | tile_size = vision_config.image_size |
| 615 | possible_resolutions = find_supported_resolutions( |
| 616 | max_num_chunks=self.info.get_max_num_tiles(), |
| 617 | patch_size=SizeDict(height=tile_size, width=tile_size), |
| 618 | ) |
| 619 | best_fit_sizes = [] |
| 620 | for image in parsed_images: |
| 621 | assert image is not None |
| 622 | best_fit_sizes.append( |
| 623 | get_best_fit( |
| 624 | (image.size[1], image.size[0]), |
| 625 | torch.tensor(possible_resolutions), |
| 626 | resize_to_max_canvas=image_processor.resize_to_max_canvas, |
| 627 | ) |
| 628 | ) |
| 629 | # TODO tile height/width do not necessarily need to match |
| 630 | aspect_ratios = [ |
| 631 | (image_size[0] // tile_size, image_size[1] // tile_size) |
| 632 | for image_size in best_fit_sizes |
| 633 | ] |
| 634 | patches_per_image = [ |
| 635 | 1 if r_h * r_w == 1 else 1 + r_h * r_w for (r_h, r_w) in aspect_ratios |
| 636 | ] |
| 637 | |
| 638 | processed_data["aspect_ratios"] = torch.tensor(aspect_ratios) |
| 639 | processed_data["patches_per_image"] = torch.tensor(patches_per_image) |
| 640 | |
| 641 | return processed_data |
| 642 | |
| 643 | def _get_mm_fields_config( |
| 644 | self, |
| 645 | hf_inputs: BatchFeature, |
| 646 | hf_processor_mm_kwargs: Mapping[str, object], |
| 647 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 648 | patches_per_image = hf_inputs.get("patches_per_image", torch.empty(0)) |
| 649 | return dict( |
| 650 | pixel_values=MultiModalFieldConfig.flat_from_sizes( |
| 651 | "image", patches_per_image |
| 652 | ), |
| 653 | patches_per_image=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 654 | aspect_ratios=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 655 | ) |
| 656 | |
| 657 | def _get_prompt_updates( |
| 658 | self, |
| 659 | mm_items: MultiModalDataItems, |
| 660 | hf_processor_mm_kwargs: Mapping[str, object], |
| 661 | out_mm_kwargs: MultiModalKwargsItems, |
| 662 | ) -> list[PromptUpdate]: |
| 663 | config = self.info.get_hf_config() |
| 664 | vision_config = config.vision_config |
| 665 | |
| 666 | num_patches_per_chunk = self.info.get_patch_per_chunk(vision_config) |
| 667 | hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 668 | img_patch_token = hf_processor.img_patch_token |
| 669 | |
| 670 | tokenizer = self.info.get_tokenizer() |
| 671 | img_patch_token_ids = cached_encode( |
| 672 | tokenizer, img_patch_token, add_special_tokens=False |
| 673 | ) |
| 674 | |
| 675 | def get_replacement(item_idx: int): |
| 676 | out_item = out_mm_kwargs["image"][item_idx] |
| 677 | aspect_ratio = out_item["aspect_ratios"].data |
| 678 | |
| 679 | repl = hf_processor._prompt_split_image( |
| 680 | aspect_ratio=aspect_ratio, |
| 681 | num_patches_per_chunk=num_patches_per_chunk, |
| 682 | ) |
| 683 | |
| 684 | repl_ids = cached_encode(tokenizer, repl, add_special_tokens=False) |
| 685 | return PromptUpdateDetails.select_token_ids(repl_ids, img_patch_token_ids) |
| 686 | |
| 687 | return [ |
| 688 | PromptReplacement( |
| 689 | modality="image", |
| 690 | target=[hf_processor.image_token_id], |
| 691 | replacement=get_replacement, |
| 692 | ) |
| 693 | ] |
| 694 | |
| 695 | |
| 696 | class Mllama4DummyInputsBuilder(BaseDummyInputsBuilder[Mllama4ProcessingInfo]): |
| 697 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 698 | num_images = mm_counts.get("image", 0) |
| 699 | |
| 700 | processor = self.info.get_hf_processor() |
| 701 | image_token = processor.fake_image_token |
| 702 | |
| 703 | return image_token * num_images |
| 704 | |
| 705 | def get_dummy_mm_data( |
| 706 | self, |
| 707 | seq_len: int, |
| 708 | mm_counts: Mapping[str, int], |
| 709 | mm_options: MultiModalDummyOptions, |
| 710 | ) -> MultiModalDataDict: |
| 711 | (target_width, target_height) = self.info.get_image_size_with_most_features() |
| 712 | |
| 713 | return { |
| 714 | "image": self._get_dummy_images( |
| 715 | width=target_width, |
| 716 | height=target_height, |
| 717 | num_images=mm_counts.get("image", 0), |
| 718 | overrides=mm_options.get("image"), |
| 719 | ) |
| 720 | } |
| 721 | |
| 722 | |
| 723 | @MULTIMODAL_REGISTRY.register_processor( |
| 724 | Mllama4MultiModalProcessor, |
| 725 | info=Mllama4ProcessingInfo, |
| 726 | dummy_inputs=Mllama4DummyInputsBuilder, |
| 727 | ) |
| 728 | class Llama4ForConditionalGeneration( |
| 729 | nn.Module, |
| 730 | SupportsMultiModal, |
| 731 | SupportsPP, |
| 732 | MixtureOfExperts, |
| 733 | SupportsEncoderCudaGraph, |
| 734 | SupportsEagle3, |
| 735 | SupportsLoRA, |
| 736 | ): |
| 737 | packed_modules_mapping = { |
| 738 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 739 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 740 | } |
| 741 | |
| 742 | supports_encoder_tp_data = True |
| 743 | supports_tower_connector_lora = True |
| 744 | |
| 745 | @classmethod |
| 746 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 747 | if modality.startswith("image"): |
| 748 | return "<|image|>" |
| 749 | |
| 750 | raise ValueError("Only image modality is supported") |
| 751 | |
| 752 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 753 | super().__init__() |
| 754 | config = vllm_config.model_config.hf_config |
| 755 | quant_config = vllm_config.quant_config |
| 756 | multimodal_config = vllm_config.model_config.multimodal_config |
| 757 | assert multimodal_config is not None |
| 758 | self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data" |
| 759 | |
| 760 | self.vllm_config = vllm_config |
| 761 | self.config = config |
| 762 | self.quant_config = quant_config |
| 763 | self.multimodal_config = multimodal_config |
| 764 | |
| 765 | with self._mark_tower_model(vllm_config, "image"): |
| 766 | with set_current_vllm_config(vllm_config): |
| 767 | self.vision_model = Llama4VisionModel( |
| 768 | config=config.vision_config, |
| 769 | quant_config=None, |
| 770 | prefix=maybe_prefix(prefix, "vision_model"), |
| 771 | ) |
| 772 | |
| 773 | self.multi_modal_projector = Llama4MultiModalProjector( |
| 774 | config=self.config, |
| 775 | quant_config=None, |
| 776 | prefix=maybe_prefix(prefix, "multi_modal_projector"), |
| 777 | ) |
| 778 | |
| 779 | with self._mark_language_model(vllm_config): |
| 780 | self.language_model = initialize_model( |
| 781 | vllm_config=vllm_config.with_hf_config( |
| 782 | config.text_config, ["LlamaForCausalLM"] |
| 783 | ), |
| 784 | prefix=maybe_prefix(prefix, "language_model"), |
| 785 | model_class=Llama4ForCausalLM, |
| 786 | ) |
| 787 | |
| 788 | self.make_empty_intermediate_tensors = ( |
| 789 | self.language_model.make_empty_intermediate_tensors |
| 790 | ) |
| 791 | |
| 792 | # Set MoE hyperparameters |
| 793 | self.num_expert_groups = 1 |
| 794 | self.num_logical_experts = self.language_model.num_logical_experts |
| 795 | self.num_physical_experts = self.language_model.num_physical_experts |
| 796 | self.num_local_physical_experts = self.language_model.num_local_physical_experts |
| 797 | self.num_routed_experts = self.language_model.num_routed_experts |
| 798 | self.num_shared_experts = self.language_model.num_shared_experts |
| 799 | self.num_redundant_experts = self.language_model.num_redundant_experts |
| 800 | self.moe_layers = self.language_model.moe_layers |
| 801 | self.num_moe_layers = len(self.moe_layers) |
| 802 | |
| 803 | def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None: |
| 804 | # Delegate to underlying language model (Llama4ForCausalLM) |
| 805 | assert hasattr(self.language_model, "set_aux_hidden_state_layers") |
| 806 | self.language_model.set_aux_hidden_state_layers(layers) |
| 807 | |
| 808 | def get_eagle3_default_aux_hidden_state_layers(self) -> tuple[int, ...]: |
| 809 | # Delegate to underlying language model (Llama4ForCausalLM) |
| 810 | assert hasattr( |
| 811 | self.language_model, "get_eagle3_default_aux_hidden_state_layers" |
| 812 | ) |
| 813 | return self.language_model.get_eagle3_default_aux_hidden_state_layers() |
| 814 | |
| 815 | def update_physical_experts_metadata( |
| 816 | self, num_physical_experts: int, num_local_physical_experts: int |
| 817 | ): |
| 818 | self.language_model.update_physical_experts_metadata( |
| 819 | num_physical_experts, num_local_physical_experts |
| 820 | ) |
| 821 | |
| 822 | def get_image_patches_per_chunk(self) -> int: |
| 823 | return Mllama4ProcessingInfo.get_patch_per_chunk(self.config.vision_config) |
| 824 | |
| 825 | def encode_image_chunks( |
| 826 | self, |
| 827 | pixel_values: torch.Tensor, |
| 828 | *, |
| 829 | use_data_parallel: bool, |
| 830 | ) -> torch.Tensor: |
| 831 | if use_data_parallel: |
| 832 | vision_embeddings = run_dp_sharded_vision_model( |
| 833 | pixel_values, self.vision_model |
| 834 | ) |
| 835 | else: |
| 836 | vision_embeddings = self.vision_model(pixel_values) |
| 837 | |
| 838 | return self.multi_modal_projector(vision_embeddings) |
| 839 | |
| 840 | def get_encoder_cudagraph_config(self): |
| 841 | from vllm.v1.worker.encoder_cudagraph_defs import ( |
| 842 | EncoderCudaGraphConfig, |
| 843 | ) |
| 844 | |
| 845 | return EncoderCudaGraphConfig( |
| 846 | modalities=["image"], |
| 847 | buffer_keys=["pixel_values"], |
| 848 | out_hidden_size=self.config.text_config.hidden_size, |
| 849 | ) |
| 850 | |
| 851 | def get_input_modality( |
| 852 | self, |
| 853 | mm_kwargs: dict[str, Any], |
| 854 | ) -> str: |
| 855 | return "image" |
| 856 | |
| 857 | def get_encoder_cudagraph_budget_range( |
| 858 | self, |
| 859 | vllm_config: VllmConfig, |
| 860 | ) -> tuple[int, int]: |
| 861 | min_budget = self.get_image_patches_per_chunk() |
| 862 | max_budget = min( |
| 863 | vllm_config.scheduler_config.max_num_batched_tokens, |
| 864 | self.vllm_config.model_config.max_model_len, |
| 865 | ) |
| 866 | return (min_budget, max_budget) |
| 867 | |
| 868 | def get_encoder_cudagraph_item_specs( |
| 869 | self, |
| 870 | mm_kwargs: dict[str, Any], |
| 871 | ): |
| 872 | from vllm.v1.worker.encoder_cudagraph_defs import EncoderItemSpec |
| 873 | |
| 874 | patches_per_chunk = self.get_image_patches_per_chunk() |
| 875 | return [ |
| 876 | EncoderItemSpec( |
| 877 | input_size=num_chunks, |
| 878 | output_tokens=num_chunks * patches_per_chunk, |
| 879 | ) |
| 880 | for num_chunks in mm_kwargs["patches_per_image"].tolist() |
| 881 | ] |
| 882 | |
| 883 | def select_encoder_cudagraph_items( |
| 884 | self, |
| 885 | mm_kwargs: dict[str, Any], |
| 886 | indices: list[int], |
| 887 | ) -> dict[str, Any]: |
| 888 | pixel_values = mm_kwargs["pixel_values"] |
| 889 | patches_per_image = mm_kwargs["patches_per_image"] |
| 890 | |
| 891 | if len(indices) == 0: |
| 892 | return { |
| 893 | "pixel_values": pixel_values[:0], |
| 894 | "patches_per_image": patches_per_image[:0], |
| 895 | } |
| 896 | |
| 897 | cum_chunks = [0] |
| 898 | for num_chunks in patches_per_image.tolist(): |
| 899 | cum_chunks.append(cum_chunks[-1] + num_chunks) |
| 900 | |
| 901 | selected_pixel_values = torch.cat( |
| 902 | [pixel_values[cum_chunks[i] : cum_chunks[i + 1]] for i in indices], |
| 903 | dim=0, |
| 904 | ) |
| 905 | |
| 906 | return { |
| 907 | "pixel_values": selected_pixel_values, |
| 908 | "patches_per_image": patches_per_image[indices], |
| 909 | } |
| 910 | |
| 911 | def prepare_encoder_cudagraph_capture_inputs( |
| 912 | self, |
| 913 | token_budget: int, |
| 914 | max_batch_size: int, |
| 915 | max_frames_per_batch: int, |
| 916 | device: torch.device, |
| 917 | dtype: torch.dtype, |
| 918 | path: str = "default", |
| 919 | axis_keys: tuple[Hashable, ...] | None = None, |
| 920 | ): |
| 921 | from vllm.v1.worker.encoder_cudagraph_defs import ( |
| 922 | EncoderCudaGraphCaptureInputs, |
| 923 | ) |
| 924 | |
| 925 | vision_config = self.config.vision_config |
| 926 | patches_per_chunk = self.get_image_patches_per_chunk() |
| 927 | chunks_per_capture = max( |
| 928 | 1, (token_budget + patches_per_chunk - 1) // patches_per_chunk |
| 929 | ) |
| 930 | dummy_pixel_values = torch.randn( |
| 931 | chunks_per_capture, |
| 932 | vision_config.num_channels, |
| 933 | vision_config.image_size, |
| 934 | vision_config.image_size, |
| 935 | device=device, |
| 936 | dtype=dtype, |
| 937 | ) |
| 938 | |
| 939 | return EncoderCudaGraphCaptureInputs( |
| 940 | values={"pixel_values": dummy_pixel_values}, |
| 941 | ) |
| 942 | |
| 943 | def prepare_encoder_cudagraph_replay_buffers( |
| 944 | self, |
| 945 | mm_kwargs: dict[str, Any], |
| 946 | max_batch_size: int, |
| 947 | max_frames_per_batch: int, |
| 948 | path: str = "default", |
| 949 | ): |
| 950 | from vllm.v1.worker.encoder_cudagraph_defs import ( |
| 951 | EncoderCudaGraphReplayBuffers, |
| 952 | ) |
| 953 | |
| 954 | return EncoderCudaGraphReplayBuffers( |
| 955 | values={"pixel_values": mm_kwargs["pixel_values"]}, |
| 956 | ) |
| 957 | |
| 958 | def encoder_cudagraph_forward( |
| 959 | self, |
| 960 | inputs: dict[str, torch.Tensor], |
| 961 | path: str = "default", |
| 962 | ) -> torch.Tensor: |
| 963 | return self.encode_image_chunks( |
| 964 | inputs["pixel_values"], |
| 965 | use_data_parallel=False, |
| 966 | ).flatten(0, 1) |
| 967 | |
| 968 | def encoder_eager_forward( |
| 969 | self, |
| 970 | mm_kwargs: dict[str, Any], |
| 971 | path: str = "default", |
| 972 | ) -> torch.Tensor: |
| 973 | return self.encode_image_chunks( |
| 974 | mm_kwargs["pixel_values"], |
| 975 | use_data_parallel=False, |
| 976 | ).flatten(0, 1) |
| 977 | |
| 978 | def _parse_and_validate_image_input( |
| 979 | self, **kwargs: object |
| 980 | ) -> Llama4ImagePatchInputs | None: |
| 981 | # total_num_chunks, channel, image_size, image_size |
| 982 | pixel_values = kwargs.pop("pixel_values", None) |
| 983 | if pixel_values is None: |
| 984 | return None |
| 985 | |
| 986 | patches_per_image = kwargs.pop("patches_per_image") |
| 987 | aspect_ratios = kwargs.pop("aspect_ratios") |
| 988 | |
| 989 | return Llama4ImagePatchInputs( |
| 990 | type="pixel_values", |
| 991 | pixel_values=pixel_values, |
| 992 | patches_per_image=patches_per_image, |
| 993 | aspect_ratios=aspect_ratios, |
| 994 | ) |
| 995 | |
| 996 | def _process_image_input( |
| 997 | self, image_input: Llama4ImagePatchInputs |
| 998 | ) -> MultiModalEmbeddings: |
| 999 | assert self.vision_model and self.multi_modal_projector |
| 1000 | pixel_values = image_input["pixel_values"] |
| 1001 | patches_per_image = image_input["patches_per_image"].tolist() |
| 1002 | |
| 1003 | vision_embeddings_flat = self.encode_image_chunks( |
| 1004 | pixel_values, |
| 1005 | use_data_parallel=self.use_data_parallel, |
| 1006 | ) |
| 1007 | |
| 1008 | return [ |
| 1009 | img.flatten(0, 1) |
| 1010 | for img in vision_embeddings_flat.split(patches_per_image, dim=0) |
| 1011 | ] |
| 1012 | |
| 1013 | def embed_multimodal(self, **kwargs) -> MultiModalEmbeddings: |
| 1014 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 1015 | if image_input is None: |
| 1016 | return [] |
| 1017 | |
| 1018 | return self._process_image_input(image_input) |
| 1019 | |
| 1020 | def forward( |
| 1021 | self, |
| 1022 | input_ids: torch.Tensor | None, |
| 1023 | positions: torch.Tensor, |
| 1024 | intermediate_tensors: IntermediateTensors | None = None, |
| 1025 | inputs_embeds: torch.Tensor | None = None, |
| 1026 | **kwargs: object, |
| 1027 | ) -> torch.Tensor | IntermediateTensors: |
| 1028 | if intermediate_tensors is not None: |
| 1029 | inputs_embeds = None |
| 1030 | |
| 1031 | return self.language_model( |
| 1032 | input_ids, positions, intermediate_tensors, inputs_embeds |
| 1033 | ) |
| 1034 | |
| 1035 | def compute_logits( |
| 1036 | self, |
| 1037 | hidden_states: torch.Tensor, |
| 1038 | ) -> torch.Tensor | None: |
| 1039 | return self.language_model.compute_logits(hidden_states) |
| 1040 | |
| 1041 | def _consolidate_qkv_weights( |
| 1042 | self, weights: Iterable[tuple[str, torch.Tensor]] |
| 1043 | ) -> Iterable[tuple[str, torch.Tensor]]: |
| 1044 | qkv_idx_mappings = { |
| 1045 | ".self_attn.q_proj": 0, |
| 1046 | ".self_attn.k_proj": 1, |
| 1047 | ".self_attn.v_proj": 2, |
| 1048 | } |
| 1049 | qkv_weights = {} |
| 1050 | for name, loaded_weight in weights: |
| 1051 | for weight_name, idx in qkv_idx_mappings.items(): |
| 1052 | if weight_name not in name: |
| 1053 | continue |
| 1054 | new_name = name.replace(weight_name, ".self_attn.qkv_proj") |
| 1055 | if new_name not in qkv_weights: |
| 1056 | qkv_weights[new_name] = [None] * 3 |
| 1057 | qkv_weights[new_name][idx] = loaded_weight |
| 1058 | break |
| 1059 | else: |
| 1060 | yield name, loaded_weight |
| 1061 | for key, weight in qkv_weights.items(): |
| 1062 | qkv_weight = torch.cat(weight, dim=0) |
| 1063 | yield key, qkv_weight |
| 1064 | |
| 1065 | def _rename_weight_for_modelopt_checkpoint(self, name: str) -> str: |
| 1066 | """Rename weights from ModelOpt llama4 fp8 checkpoints to vLLM |
| 1067 | format.""" |
| 1068 | if name.startswith("model.") or name.startswith("language_model.model."): |
| 1069 | renamed = ( |
| 1070 | name.replace("model.", "language_model.model.", 1) |
| 1071 | if name.startswith("model.") |
| 1072 | else name |
| 1073 | ) |
| 1074 | # Handle expert scale parameters with flat naming |
| 1075 | if "feed_forward.experts." in name and ( |
| 1076 | "_input_scale" in name or "_weight_scale" in name |
| 1077 | ): |
| 1078 | # Map checkpoint naming to vLLM's expected naming |
| 1079 | if "down_proj_input_scale" in renamed: |
| 1080 | return renamed.replace("down_proj_input_scale", "w2_input_scale") |
| 1081 | elif "down_proj_weight_scale" in renamed: |
| 1082 | return renamed.replace("down_proj_weight_scale", "w2_weight_scale") |
| 1083 | elif "gate_up_proj_input_scale" in renamed: |
| 1084 | return renamed.replace( |
| 1085 | "gate_up_proj_input_scale", "w13_input_scale" |
| 1086 | ) |
| 1087 | elif "gate_up_proj_weight_scale" in renamed: |
| 1088 | return renamed.replace( |
| 1089 | "gate_up_proj_weight_scale", "w13_weight_scale" |
| 1090 | ) |
| 1091 | return renamed |
| 1092 | |
| 1093 | # Handle attention scale parameters |
| 1094 | elif "self_attn." in name and (".k_scale" in name or ".v_scale" in name): |
| 1095 | if ".k_proj.k_scale" in renamed: |
| 1096 | return renamed.replace(".k_proj.k_scale", ".attn.k_scale") |
| 1097 | elif ".v_proj.v_scale" in renamed: |
| 1098 | return renamed.replace(".v_proj.v_scale", ".attn.v_scale") |
| 1099 | return renamed |
| 1100 | |
| 1101 | # Standard model.* to language_model.model.* renaming |
| 1102 | return renamed |
| 1103 | |
| 1104 | elif name.startswith("lm_head.weight"): |
| 1105 | return name.replace("lm_head.weight", "language_model.lm_head.weight") |
| 1106 | |
| 1107 | return name |
| 1108 | |
| 1109 | def _load_other_weights( |
| 1110 | self, |
| 1111 | other_weights: Iterable[tuple[str, torch.Tensor]], |
| 1112 | params_dict: dict, |
| 1113 | stacked_params_mapping: list, |
| 1114 | ) -> set[str]: |
| 1115 | """Load non-language-model weights with stacking support.""" |
| 1116 | updated_params = set() |
| 1117 | |
| 1118 | if self.use_data_parallel: |
| 1119 | other_weights = self._consolidate_qkv_weights(other_weights) |
| 1120 | |
| 1121 | for name, loaded_weight in other_weights: |
| 1122 | # Try stacked parameter mapping first |
| 1123 | for param_name, weight_name, shard_id in stacked_params_mapping: |
| 1124 | if weight_name not in name or self.use_data_parallel: |
| 1125 | continue |
| 1126 | name = name.replace(weight_name, param_name) |
| 1127 | param = params_dict[name] |
| 1128 | updated_params.add(name) |
| 1129 | weight_loader = param.weight_loader |
| 1130 | weight_loader(param, loaded_weight, shard_id) |
| 1131 | break |
| 1132 | else: |
| 1133 | # Use regular weight loading |
| 1134 | param = params_dict[name] |
| 1135 | weight_loader = getattr(param, "weight_loader", default_weight_loader) |
| 1136 | weight_loader(param, loaded_weight) |
| 1137 | updated_params.add(name) |
| 1138 | |
| 1139 | return updated_params |
| 1140 | |
| 1141 | def get_expert_mapping(self) -> list[tuple[str, str, int, str]]: |
| 1142 | # Params for weights, fp8 weight scales, fp8 activation scales |
| 1143 | # (param_name, weight_name, expert_id, shard_id) |
| 1144 | return fused_moe_make_expert_params_mapping( |
| 1145 | self, |
| 1146 | ckpt_gate_proj_name="gate_proj", |
| 1147 | ckpt_down_proj_name="down_proj", |
| 1148 | ckpt_up_proj_name="up_proj", |
| 1149 | num_experts=self.config.text_config.num_local_experts, |
| 1150 | num_redundant_experts=self.num_redundant_experts, |
| 1151 | ) |
| 1152 | |
| 1153 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 1154 | stacked_params_mapping = [ |
| 1155 | # (param_name, shard_name, shard_id) |
| 1156 | (".self_attn.qkv_proj", ".self_attn.q_proj", "q"), |
| 1157 | (".self_attn.qkv_proj", ".self_attn.k_proj", "k"), |
| 1158 | (".self_attn.qkv_proj", ".self_attn.v_proj", "v"), |
| 1159 | # Shared expert gate_up_proj stacking |
| 1160 | (".shared_expert.gate_up_proj", ".shared_expert.gate_proj", 0), |
| 1161 | (".shared_expert.gate_up_proj", ".shared_expert.up_proj", 1), |
| 1162 | # Feed forward gate_up_proj stacking (for non-MoE layers if any) |
| 1163 | (".feed_forward.gate_up_proj", ".feed_forward.gate_proj", 0), |
| 1164 | (".feed_forward.gate_up_proj", ".feed_forward.up_proj", 1), |
| 1165 | ] |
| 1166 | params_dict = dict(self.named_parameters()) |
| 1167 | updated_params: set[str] = set() |
| 1168 | |
| 1169 | # Stream thelanguage-model weights straight into |
| 1170 | # AutoWeightsLoader so each tensor is loaded and released as we iterate, |
| 1171 | # instead of materializing the whole checkpoint in host memory first. |
| 1172 | # Only the small vision/projector and scalar expert-scale groups are |
| 1173 | # buffered. |
| 1174 | other_weights: list[tuple[str, torch.Tensor]] = [] |
| 1175 | expert_scale_weights: list[tuple[str, torch.Tensor]] = [] |
| 1176 | |
| 1177 | def regular_language_model_weights() -> Iterable[tuple[str, torch.Tensor]]: |
| 1178 | """Rename weights and separate them into language_model and other |
| 1179 | weights. |
| 1180 | |
| 1181 | Yields the (large) language_model weights for streaming; the small |
| 1182 | groups (vision/projector and scalar expert scales) are buffered into |
| 1183 | the lists above. |
| 1184 | """ |
| 1185 | for name, weight in weights: |
| 1186 | renamed = self._rename_weight_for_modelopt_checkpoint(name) |
| 1187 | |
| 1188 | attr = renamed.split(".", 1)[0] |
| 1189 | if isinstance(getattr(self, attr), StageMissingLayer): |
| 1190 | continue |
| 1191 | |
| 1192 | if not renamed.startswith("language_model."): |
| 1193 | other_weights.append((renamed, weight)) |
| 1194 | continue |
| 1195 | |
| 1196 | # Handle expert scale parameters that need broadcasting. |
| 1197 | # ModelOpt checkpoints use a single value tensor scalar for BMM |
| 1198 | # style experts, vLLM expects the scale to be broadcasted across |
| 1199 | # all experts. |
| 1200 | if ( |
| 1201 | "feed_forward.experts." in renamed |
| 1202 | and "scale" in renamed |
| 1203 | and ".shared_expert" not in renamed |
| 1204 | ): |
| 1205 | renamed = maybe_remap_moe_expert_param_name(renamed, params_dict) |
| 1206 | if renamed in params_dict: |
| 1207 | param = params_dict[renamed] |
| 1208 | if ( |
| 1209 | hasattr(param, "data") |
| 1210 | and param.data.numel() > 1 |
| 1211 | and weight.numel() == 1 |
| 1212 | ): |
| 1213 | # Broadcast single value to all experts |
| 1214 | param.data.fill_(weight.item()) |
| 1215 | updated_params.add(renamed) |
| 1216 | continue |
| 1217 | |
| 1218 | expert_scale_weights.append((renamed, weight)) |
| 1219 | continue |
| 1220 | |
| 1221 | yield renamed, weight |
| 1222 | |
| 1223 | loader = AutoWeightsLoader(self) |
| 1224 | # AutoWeightsLoader consumes its input lazily and runs to exhaustion, |
| 1225 | # so other_weights / expert_scale_weights are fully populated as a side |
| 1226 | # effect by the time this returns. |
| 1227 | loaded_language_model_params = loader.load_weights( |
| 1228 | regular_language_model_weights() |
| 1229 | ) |
| 1230 | assert loaded_language_model_params is not None |
| 1231 | updated_params.update(loaded_language_model_params) |
| 1232 | |
| 1233 | if expert_scale_weights: |
| 1234 | loaded_expert_scale_params = loader.load_weights(expert_scale_weights) |
| 1235 | if loaded_expert_scale_params: |
| 1236 | updated_params.update(loaded_expert_scale_params) |
| 1237 | |
| 1238 | updated_params.update( |
| 1239 | self._load_other_weights(other_weights, params_dict, stacked_params_mapping) |
| 1240 | ) |
| 1241 | |
| 1242 | return updated_params |
| 1243 | |
| 1244 | def get_mm_mapping(self) -> MultiModelKeys: |
| 1245 | """Get the module prefix in multimodal models.""" |
| 1246 | return MultiModelKeys.from_string_field( |
| 1247 | language_model="language_model", |
| 1248 | connector=[ |
| 1249 | "multi_modal_projector.", |
| 1250 | "vision_model.vision_adapter.", |
| 1251 | ], |
| 1252 | tower_model="vision_model.", |
| 1253 | ) |
| 1254 | |
| 1255 | def get_mm_lora_token_counts( |
| 1256 | self, |
| 1257 | *, |
| 1258 | modality: str, |
| 1259 | mm_kwargs: MultiModalKwargsItem | None, |
| 1260 | num_mm_embeds: int, |
| 1261 | ) -> tuple[int, int | None]: |
| 1262 | del modality, mm_kwargs |
| 1263 | vision_config = self.config.vision_config |
| 1264 | patches_per_chunk = Mllama4ProcessingInfo.get_patch_per_chunk(vision_config) |
| 1265 | if num_mm_embeds <= 0 or patches_per_chunk <= 0: |
| 1266 | return 0, 0 |
| 1267 | raw_patches = (vision_config.image_size // vision_config.patch_size) ** 2 |
| 1268 | num_chunks = num_mm_embeds // patches_per_chunk |
| 1269 | return num_chunks * (raw_patches + 1), num_chunks * patches_per_chunk |