| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """vLLM implementation of Granite 4 Vision. |
| 4 | |
| 5 | Uses GraniteForCausalLM as the language backbone with SigLIP vision encoder |
| 6 | and deepstack feature injection via WindowQFormer projectors. |
| 7 | |
| 8 | LoRA support: use --enable-lora --default-mm-loras for LM-only LoRA adapters. |
| 9 | """ |
| 10 | |
| 11 | import math |
| 12 | from collections.abc import Iterable, Mapping |
| 13 | from fractions import Fraction |
| 14 | from itertools import islice |
| 15 | |
| 16 | import torch |
| 17 | import torch.nn as nn |
| 18 | from transformers import BatchFeature |
| 19 | from transformers.models.blip_2.configuration_blip_2 import Blip2QFormerConfig |
| 20 | from transformers.models.llava_next.modeling_llava_next import ( |
| 21 | get_anyres_image_grid_shape, |
| 22 | image_size_to_num_patches, |
| 23 | unpad_image, |
| 24 | ) |
| 25 | |
| 26 | from vllm.compilation.decorators import support_torch_compile |
| 27 | from vllm.config import CacheConfig, VllmConfig |
| 28 | from vllm.distributed.parallel_state import get_pp_group |
| 29 | from vllm.logger import init_logger |
| 30 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 31 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 32 | from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead |
| 33 | from vllm.model_executor.models.granite import GraniteForCausalLM, GraniteModel |
| 34 | from vllm.model_executor.models.interfaces import ( |
| 35 | MultiModalEmbeddings, |
| 36 | SupportsLoRA, |
| 37 | SupportsMultiModal, |
| 38 | SupportsPP, |
| 39 | ) |
| 40 | from vllm.model_executor.models.llava import LlavaDummyInputsBuilder |
| 41 | from vllm.model_executor.models.llava_next import ( |
| 42 | BaseLlavaNextMultiModalProcessor, |
| 43 | LlavaNextImageEmbeddingInputs, |
| 44 | LlavaNextImageInputs, |
| 45 | LlavaNextImagePixelInputs, |
| 46 | LlavaNextProcessingInfo, |
| 47 | ) |
| 48 | from vllm.model_executor.models.module_mapping import MultiModelKeys |
| 49 | from vllm.model_executor.models.siglip import SiglipVisionModel |
| 50 | from vllm.model_executor.models.utils import ( |
| 51 | AutoWeightsLoader, |
| 52 | PPMissingLayer, |
| 53 | WeightsMapper, |
| 54 | maybe_prefix, |
| 55 | ) |
| 56 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 57 | from vllm.multimodal.inputs import MultiModalFieldConfig |
| 58 | from vllm.sequence import IntermediateTensors |
| 59 | |
| 60 | from .blip2 import Blip2QFormerModel |
| 61 | |
| 62 | logger = init_logger(__name__) |
| 63 | |
| 64 | |
| 65 | # --------------------------------------------------------------------------- |
| 66 | # Downsampler modules (translated from HF downsampling.py) |
| 67 | # --------------------------------------------------------------------------- |
| 68 | |
| 69 | |
| 70 | class InterpolateDownsampler: |
| 71 | """Spatial downsampling via area interpolation.""" |
| 72 | |
| 73 | def __init__(self, config, mode="area"): |
| 74 | self.orig_image_side = ( |
| 75 | config.vision_config.image_size // config.vision_config.patch_size |
| 76 | ) |
| 77 | self.new_image_side = int( |
| 78 | self.orig_image_side * Fraction(config.downsample_rate) |
| 79 | ) |
| 80 | self.mode = mode |
| 81 | |
| 82 | def __call__(self, image_features: torch.Tensor) -> torch.Tensor: |
| 83 | batch_size, _, dim = image_features.size() |
| 84 | up_shape = [batch_size, self.orig_image_side, self.orig_image_side, dim] |
| 85 | large = image_features.view(up_shape).permute(0, 3, 1, 2) |
| 86 | small = torch.nn.functional.interpolate( |
| 87 | large, |
| 88 | size=(self.new_image_side, self.new_image_side), |
| 89 | mode=self.mode, |
| 90 | ) |
| 91 | return small.permute(0, 2, 3, 1).flatten(1, 2) |
| 92 | |
| 93 | |
| 94 | class SpatialOffsetDownsampler: |
| 95 | """Sample one position from each 2x2 block (offset 0-3 = TL/TR/BL/BR).""" |
| 96 | |
| 97 | def __init__(self, config, offset: int = 0): |
| 98 | self.orig_image_side = ( |
| 99 | config.vision_config.image_size // config.vision_config.patch_size |
| 100 | ) |
| 101 | self.new_image_side = self.orig_image_side // 2 |
| 102 | offsets = [(0, 0), (0, 1), (1, 0), (1, 1)] |
| 103 | self.offset_h, self.offset_w = offsets[offset] |
| 104 | |
| 105 | def __call__(self, image_features: torch.Tensor) -> torch.Tensor: |
| 106 | B, _, C = image_features.shape |
| 107 | features_2d = image_features.reshape( |
| 108 | B, self.orig_image_side, self.orig_image_side, C |
| 109 | ) |
| 110 | n = self.new_image_side |
| 111 | blocks = features_2d.reshape(B, n, 2, n, 2, C) |
| 112 | sampled = blocks[:, :, self.offset_h, :, self.offset_w, :] |
| 113 | return sampled.reshape(B, -1, C) |
| 114 | |
| 115 | |
| 116 | class WindowQFormerDownsampler(nn.Module): |
| 117 | """Window-based QFormer downsampler (matches HF downsampling.py exactly).""" |
| 118 | |
| 119 | def __init__( |
| 120 | self, |
| 121 | config, |
| 122 | quant_config: QuantizationConfig | None = None, |
| 123 | cache_config: CacheConfig | None = None, |
| 124 | spatial_offset: int | None = None, |
| 125 | prefix: str = "", |
| 126 | ): |
| 127 | super().__init__() |
| 128 | llm_hidden_size = config.text_config.hidden_size |
| 129 | vision_hidden_size = config.vision_config.hidden_size |
| 130 | |
| 131 | self.dropout = nn.Dropout(config.projector_dropout) |
| 132 | |
| 133 | self.downsampler: SpatialOffsetDownsampler | InterpolateDownsampler |
| 134 | if spatial_offset is not None: |
| 135 | self.downsampler = SpatialOffsetDownsampler(config, offset=spatial_offset) |
| 136 | else: |
| 137 | self.downsampler = InterpolateDownsampler(config) |
| 138 | |
| 139 | qformer_config = Blip2QFormerConfig( |
| 140 | hidden_size=vision_hidden_size, |
| 141 | num_attention_heads=vision_hidden_size // 64, |
| 142 | intermediate_size=3072, |
| 143 | num_hidden_layers=1, |
| 144 | encoder_hidden_size=vision_hidden_size, |
| 145 | cross_attention_frequency=1, |
| 146 | max_position_embeddings=2048, |
| 147 | use_qformer_text_input=False, |
| 148 | ) |
| 149 | self.qformer = Blip2QFormerModel( |
| 150 | qformer_config, |
| 151 | quant_config=quant_config, |
| 152 | cache_config=cache_config, |
| 153 | prefix=maybe_prefix(prefix, "qformer"), |
| 154 | ) |
| 155 | |
| 156 | self.image_side = ( |
| 157 | config.vision_config.image_size // config.vision_config.patch_size |
| 158 | ) |
| 159 | q, w = config.downsample_rate.split("/") |
| 160 | self.query_side, self.window_side = int(q), int(w) |
| 161 | self.query_length = self.query_side**2 |
| 162 | |
| 163 | embed_std = 1 / math.sqrt(vision_hidden_size) |
| 164 | self.norm = nn.LayerNorm(vision_hidden_size, eps=1e-6) |
| 165 | self.query = nn.Parameter( |
| 166 | torch.randn(1, self.query_length, vision_hidden_size) * embed_std |
| 167 | ) |
| 168 | self.image_positions = nn.Parameter( |
| 169 | torch.randn(1, self.window_side**2, vision_hidden_size) * embed_std |
| 170 | ) |
| 171 | self.out_linear = nn.Linear(vision_hidden_size, llm_hidden_size, bias=True) |
| 172 | |
| 173 | def _win(self, x: torch.Tensor, side: int, win: int) -> torch.Tensor: |
| 174 | """(B, side*side, C) → (B*n*n, win*win, C) where n=side//win.""" |
| 175 | B, _, C = x.shape |
| 176 | n = side // win |
| 177 | return ( |
| 178 | x.view(B, side, side, C) |
| 179 | .view(B, n, win, n, win, C) |
| 180 | .transpose(2, 3) |
| 181 | .flatten(0, 2) |
| 182 | .flatten(1, 2) |
| 183 | ) |
| 184 | |
| 185 | def _unwin(self, xw: torch.Tensor, n: int, win: int) -> torch.Tensor: |
| 186 | """(B*n*n, win*win, C) → (B, (n*win)^2, C).""" |
| 187 | Bnn, _, C = xw.shape |
| 188 | B = Bnn // (n * n) |
| 189 | side = n * win |
| 190 | return ( |
| 191 | xw.view(B, n, n, win, win, C) |
| 192 | .transpose(2, 3) |
| 193 | .contiguous() |
| 194 | .view(B, side, side, C) |
| 195 | .flatten(1, 2) |
| 196 | ) |
| 197 | |
| 198 | def forward(self, image_features: torch.Tensor) -> torch.Tensor: |
| 199 | B, HW, C = image_features.shape |
| 200 | assert self.image_side * self.image_side == HW |
| 201 | n = self.image_side // self.window_side |
| 202 | |
| 203 | image_features = self.norm(image_features) |
| 204 | enc = self._win(image_features, self.image_side, self.window_side) |
| 205 | |
| 206 | downsampled = self.downsampler(image_features) |
| 207 | new_side = n * self.query_side |
| 208 | downsampled_w = self._win(downsampled, new_side, self.query_side) |
| 209 | |
| 210 | query_embeds = self.query + downsampled_w |
| 211 | encoder_embeds = self.dropout(enc + self.image_positions) |
| 212 | out_w = self.qformer( |
| 213 | query_embeds=query_embeds, |
| 214 | encoder_hidden_states=encoder_embeds, |
| 215 | ) |
| 216 | |
| 217 | out = self._unwin(out_w, n=n, win=self.query_side) |
| 218 | out = self.dropout(out) |
| 219 | return self.out_linear(out) |
| 220 | |
| 221 | |
| 222 | # --------------------------------------------------------------------------- |
| 223 | # LLM subclasses with deepstack injection in the layer loop |
| 224 | # --------------------------------------------------------------------------- |
| 225 | |
| 226 | |
| 227 | @support_torch_compile( |
| 228 | dynamic_arg_dims={ |
| 229 | "input_ids": 0, |
| 230 | "positions": 0, |
| 231 | "intermediate_tensors": 0, |
| 232 | "inputs_embeds": 0, |
| 233 | "deepstack_input_embeds": 0, |
| 234 | } |
| 235 | ) |
| 236 | class Granite4VisionLLMModel(GraniteModel): |
| 237 | """GraniteModel with deepstack feature injection in the layer loop.""" |
| 238 | |
| 239 | def forward( |
| 240 | self, |
| 241 | input_ids: torch.Tensor | None, |
| 242 | positions: torch.Tensor, |
| 243 | intermediate_tensors: IntermediateTensors | None = None, |
| 244 | inputs_embeds: torch.Tensor | None = None, |
| 245 | deepstack_input_embeds: IntermediateTensors | None = None, |
| 246 | ) -> torch.Tensor | IntermediateTensors: |
| 247 | if get_pp_group().is_first_rank: |
| 248 | if inputs_embeds is not None: |
| 249 | hidden_states = inputs_embeds |
| 250 | else: |
| 251 | hidden_states = self.embed_input_ids(input_ids) |
| 252 | hidden_states = hidden_states * self.config.embedding_multiplier |
| 253 | else: |
| 254 | assert intermediate_tensors is not None |
| 255 | hidden_states = intermediate_tensors["hidden_states"] |
| 256 | # Recover deepstack features forwarded from the previous PP rank. |
| 257 | if deepstack_input_embeds is None: |
| 258 | ds_keys = [ |
| 259 | k for k in intermediate_tensors.tensors if k.startswith("ds_") |
| 260 | ] |
| 261 | if ds_keys: |
| 262 | deepstack_input_embeds = IntermediateTensors( |
| 263 | {k: intermediate_tensors[k] for k in ds_keys} |
| 264 | ) |
| 265 | |
| 266 | for layer_idx, layer in islice( |
| 267 | enumerate(self.layers), self.start_layer, self.end_layer |
| 268 | ): |
| 269 | if deepstack_input_embeds is not None: |
| 270 | key = f"ds_{layer_idx}" |
| 271 | if key in deepstack_input_embeds.tensors: |
| 272 | feat = deepstack_input_embeds[key] |
| 273 | # Resize to match hidden_states in case of CUDA graph padding |
| 274 | num_tokens = hidden_states.size(0) |
| 275 | buf_len = feat.shape[0] |
| 276 | if buf_len != num_tokens: |
| 277 | feat = torch.nn.functional.pad( |
| 278 | feat[:num_tokens], |
| 279 | (0, 0, 0, max(0, num_tokens - buf_len)), |
| 280 | ) |
| 281 | hidden_states = hidden_states + feat |
| 282 | hidden_states = layer(positions, hidden_states) |
| 283 | |
| 284 | if not get_pp_group().is_last_rank: |
| 285 | # Forward hidden_states and any deepstack features for later ranks. |
| 286 | it = {"hidden_states": hidden_states} |
| 287 | if deepstack_input_embeds is not None: |
| 288 | remaining = { |
| 289 | k: v |
| 290 | for k, v in deepstack_input_embeds.tensors.items() |
| 291 | if int(k.split("_")[1]) >= self.end_layer |
| 292 | } |
| 293 | it.update(remaining) |
| 294 | return IntermediateTensors(it) |
| 295 | |
| 296 | hidden_states = self.norm(hidden_states) |
| 297 | return hidden_states |
| 298 | |
| 299 | |
| 300 | class Granite4VisionLLMForCausalLM(GraniteForCausalLM): |
| 301 | """GraniteForCausalLM backed by Granite4VisionLLMModel.""" |
| 302 | |
| 303 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 304 | nn.Module.__init__(self) |
| 305 | config = vllm_config.model_config.hf_config |
| 306 | quant_config = vllm_config.quant_config |
| 307 | self.config = config |
| 308 | self.quant_config = quant_config |
| 309 | self.model = Granite4VisionLLMModel( |
| 310 | vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model") |
| 311 | ) |
| 312 | if get_pp_group().is_last_rank: |
| 313 | self.lm_head = ParallelLMHead( |
| 314 | config.vocab_size, |
| 315 | config.hidden_size, |
| 316 | quant_config=quant_config, |
| 317 | prefix=maybe_prefix(prefix, "lm_head"), |
| 318 | ) |
| 319 | if config.tie_word_embeddings: |
| 320 | self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens) |
| 321 | logit_scale = getattr(config, "logit_scale", 1.0) |
| 322 | if hasattr(config, "logits_scaling"): |
| 323 | logit_scale /= config.logits_scaling |
| 324 | self.logits_processor = LogitsProcessor( |
| 325 | config.vocab_size, scale=logit_scale |
| 326 | ) |
| 327 | else: |
| 328 | self.lm_head = PPMissingLayer() |
| 329 | |
| 330 | def make_empty_intermediate_tensors( |
| 331 | self, batch_size: int, dtype: torch.dtype, device: torch.device |
| 332 | ) -> IntermediateTensors: |
| 333 | tensors = super().make_empty_intermediate_tensors(batch_size, dtype, device) |
| 334 | # Include deepstack buffers so non-first PP ranks receive them. |
| 335 | # _ds_layer_indices is set directly on this instance by the outer model. |
| 336 | for llm_layer in getattr(self, "_ds_layer_indices", []): |
| 337 | tensors.tensors[f"ds_{llm_layer}"] = torch.zeros( |
| 338 | (batch_size, self.config.hidden_size), dtype=dtype, device=device |
| 339 | ) |
| 340 | return tensors |
| 341 | |
| 342 | |
| 343 | # --------------------------------------------------------------------------- |
| 344 | # Processing info / processor (reuses LlavaNext patterns) |
| 345 | # --------------------------------------------------------------------------- |
| 346 | |
| 347 | |
| 348 | class Granite4VisionProcessingInfo(LlavaNextProcessingInfo): |
| 349 | def get_hf_config(self): |
| 350 | return self.ctx.get_hf_config() |
| 351 | |
| 352 | def get_hf_processor(self, **kwargs): |
| 353 | return self.ctx.get_hf_processor(**kwargs) |
| 354 | |
| 355 | def get_num_image_tokens( |
| 356 | self, |
| 357 | *, |
| 358 | image_width: int, |
| 359 | image_height: int, |
| 360 | ) -> int: |
| 361 | hf_config = self.get_hf_config() |
| 362 | vision_encoder_info = self.get_vision_encoder_info() |
| 363 | |
| 364 | # After QFormer downsampling, patch grid is scaled by downsample_rate |
| 365 | ds_rate = Fraction(hf_config.downsample_rate) |
| 366 | patch_grid = vision_encoder_info.get_patch_grid_length() # 24 for 384/16 |
| 367 | downsampled_grid = int(patch_grid * ds_rate) # 12 for rate 4/8 |
| 368 | |
| 369 | # Base feature: downsampled_grid^2 |
| 370 | base_feature_size = downsampled_grid * downsampled_grid |
| 371 | |
| 372 | num_patch_height, num_patch_width = get_anyres_image_grid_shape( |
| 373 | image_size=(image_height, image_width), |
| 374 | grid_pinpoints=hf_config.image_grid_pinpoints, |
| 375 | patch_size=vision_encoder_info.get_image_size(), |
| 376 | ) |
| 377 | |
| 378 | ( |
| 379 | unpadded_feature_size, |
| 380 | newline_feature_size, |
| 381 | ) = self._get_num_unpadded_features( |
| 382 | original_height=image_height, |
| 383 | original_width=image_width, |
| 384 | npatches=downsampled_grid, |
| 385 | num_patch_height=num_patch_height, |
| 386 | num_patch_width=num_patch_width, |
| 387 | ) |
| 388 | |
| 389 | return unpadded_feature_size + newline_feature_size + base_feature_size |
| 390 | |
| 391 | |
| 392 | class Granite4VisionMultiModalProcessor( |
| 393 | BaseLlavaNextMultiModalProcessor[Granite4VisionProcessingInfo] |
| 394 | ): |
| 395 | def _get_mm_fields_config( |
| 396 | self, |
| 397 | hf_inputs: BatchFeature, |
| 398 | hf_processor_mm_kwargs: Mapping[str, object], |
| 399 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 400 | return dict( |
| 401 | pixel_values=MultiModalFieldConfig.batched("image"), |
| 402 | image_sizes=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 403 | ) |
| 404 | |
| 405 | |
| 406 | # --------------------------------------------------------------------------- |
| 407 | # Top-level model |
| 408 | # --------------------------------------------------------------------------- |
| 409 | |
| 410 | |
| 411 | @MULTIMODAL_REGISTRY.register_processor( |
| 412 | Granite4VisionMultiModalProcessor, |
| 413 | info=Granite4VisionProcessingInfo, |
| 414 | dummy_inputs=LlavaDummyInputsBuilder, |
| 415 | ) |
| 416 | class Granite4VisionForConditionalGeneration( |
| 417 | nn.Module, SupportsLoRA, SupportsMultiModal, SupportsPP |
| 418 | ): |
| 419 | """vLLM implementation of Granite 4 Vision. |
| 420 | |
| 421 | Architecture: |
| 422 | - SigLIP vision tower -> WindowQFormerDownsampler projectors |
| 423 | - Deepstack: 4 vision layers projected and injected at 4 LLM layers |
| 424 | - Spatial: 4 offset groups from last vision layer injected at 4 more LLM layers |
| 425 | - Granite language backbone with embedding_multiplier |
| 426 | - logits_scaling via LogitsProcessor |
| 427 | |
| 428 | The outer model runs the LLM layer loop directly (like HF does) to inject |
| 429 | deepstack features. This avoids wrapping the inner model and keeps weight |
| 430 | loading simple. |
| 431 | |
| 432 | LoRA support: |
| 433 | - Full merge: --hf-overrides '{"adapter_path": "path/to/lora"}' merges |
| 434 | LM-only LoRA deltas at load time (W += scaling * B @ A). |
| 435 | - Native LoRA: --enable-lora --default-mm-loras '{"image": "path/to/lora"}' |
| 436 | lets vLLM runtime serve LM LoRA per-request. |
| 437 | Both modes expect a LM-only adapter (no modules_to_save). |
| 438 | """ |
| 439 | |
| 440 | # LoRA class attributes (matches GraniteForCausalLM) |
| 441 | packed_modules_mapping = { |
| 442 | "qkv_proj": ["q_proj", "k_proj", "v_proj"], |
| 443 | "gate_up_proj": ["gate_proj", "up_proj"], |
| 444 | } |
| 445 | embedding_modules = {} |
| 446 | |
| 447 | # Weight mapping: HF checkpoint -> vLLM parameter names |
| 448 | # HF: model.language_model.layers.0... |
| 449 | # vLLM: language_model.model.layers.0... |
| 450 | # (because GraniteForCausalLM.model = GraniteModel) |
| 451 | hf_to_vllm_mapper = WeightsMapper( |
| 452 | orig_to_new_prefix={ |
| 453 | "model.language_model.": "language_model.model.", |
| 454 | "model.layerwise_projectors.": "layerwise_projectors.", |
| 455 | "model.spatial_projectors.": "spatial_projectors.", |
| 456 | "model.image_newline": "image_newline", |
| 457 | "model.vision_tower.": "vision_tower.", |
| 458 | "lm_head.": "language_model.lm_head.", |
| 459 | } |
| 460 | ) |
| 461 | |
| 462 | @classmethod |
| 463 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 464 | if modality.startswith("image"): |
| 465 | return "<image>" |
| 466 | raise ValueError(f"Only image modality is supported, got {modality}") |
| 467 | |
| 468 | def get_mm_mapping(self) -> MultiModelKeys: |
| 469 | return MultiModelKeys.from_string_field( |
| 470 | language_model="language_model", |
| 471 | connector=["layerwise_projectors", "spatial_projectors"], |
| 472 | tower_model="vision_tower", |
| 473 | ) |
| 474 | |
| 475 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 476 | super().__init__() |
| 477 | |
| 478 | config = vllm_config.model_config.hf_config |
| 479 | quant_config = vllm_config.quant_config |
| 480 | self.config = config |
| 481 | self.vllm_config = vllm_config |
| 482 | |
| 483 | # ----- Vision tower + projectors (marked as tower) ----- |
| 484 | with self._mark_tower_model(vllm_config, "image"): |
| 485 | # Do NOT use init_vision_tower_for_llava here — it truncates the |
| 486 | # encoder to vision_feature_layer depth. Deepstack needs ALL hidden |
| 487 | # states (deepstack_layer_map uses negative indices into the full |
| 488 | # encoder output list). |
| 489 | self.vision_tower = SiglipVisionModel( |
| 490 | config.vision_config, |
| 491 | quant_config=quant_config, |
| 492 | require_post_norm=False, |
| 493 | prefix=maybe_prefix(prefix, "vision_tower"), |
| 494 | ) |
| 495 | |
| 496 | # image_newline parameter |
| 497 | if config.use_image_newline_parameter: |
| 498 | self.image_newline = nn.Parameter( |
| 499 | torch.empty(config.text_config.hidden_size) |
| 500 | ) |
| 501 | else: |
| 502 | self.image_newline = None |
| 503 | |
| 504 | cache_config = vllm_config.cache_config |
| 505 | |
| 506 | # Deepstack projectors: one per (vision_layer, llm_layer) pair |
| 507 | self.layerwise_projectors = nn.ModuleList( |
| 508 | [ |
| 509 | WindowQFormerDownsampler( |
| 510 | config, |
| 511 | quant_config=quant_config, |
| 512 | cache_config=cache_config, |
| 513 | prefix=maybe_prefix(prefix, f"layerwise_projectors.{i}"), |
| 514 | ) |
| 515 | for i in range(len(config.deepstack_layer_map)) |
| 516 | ] |
| 517 | ) |
| 518 | |
| 519 | # Spatial projectors: 4 offset groups |
| 520 | self.spatial_projectors = None |
| 521 | if config.use_spatial_sampling: |
| 522 | self.spatial_projectors = nn.ModuleList( |
| 523 | [ |
| 524 | WindowQFormerDownsampler( |
| 525 | config, |
| 526 | quant_config=quant_config, |
| 527 | cache_config=cache_config, |
| 528 | spatial_offset=i, |
| 529 | prefix=maybe_prefix(prefix, f"spatial_projectors.{i}"), |
| 530 | ) |
| 531 | for i in range(4) |
| 532 | ] |
| 533 | ) |
| 534 | |
| 535 | # ----- Language model (marked as LM) ----- |
| 536 | with self._mark_language_model(vllm_config): |
| 537 | self.language_model = Granite4VisionLLMForCausalLM( |
| 538 | vllm_config=vllm_config.with_hf_config(config.text_config), |
| 539 | prefix=maybe_prefix(prefix, "language_model"), |
| 540 | ) |
| 541 | |
| 542 | self.make_empty_intermediate_tensors = ( |
| 543 | self.language_model.make_empty_intermediate_tensors |
| 544 | ) |
| 545 | |
| 546 | # Store config values we need |
| 547 | self._deepstack_layer_map = config.deepstack_layer_map # [[-19, 9], ...] |
| 548 | self._use_spatial_sampling = getattr(config, "use_spatial_sampling", False) |
| 549 | self._spatial_vision_layer = getattr(config, "spatial_vision_layer", -1) |
| 550 | self._spatial_target_layers = getattr(config, "spatial_target_layers", []) |
| 551 | self._vision_feature_select_strategy = getattr( |
| 552 | config, "vision_feature_select_strategy", "full" |
| 553 | ) |
| 554 | self._downsample_rate = Fraction(config.downsample_rate) |
| 555 | |
| 556 | # Ordered list of LLM layer indices for each deepstack level. |
| 557 | # Pre-populated from config so it's available during CUDA graph capture |
| 558 | # (before any embed_multimodal call). |
| 559 | self._ds_layer_indices: list[int] = [ |
| 560 | llm_layer for _, llm_layer in config.deepstack_layer_map |
| 561 | ] + list(getattr(config, "spatial_target_layers", [])) |
| 562 | |
| 563 | # Share ds_layer_indices with the LLM causal model so |
| 564 | # make_empty_intermediate_tensors includes the correct keys |
| 565 | # (its self.config is text_config, no deepstack_layer_map). |
| 566 | self.language_model._ds_layer_indices = self._ds_layer_indices |
| 567 | |
| 568 | # Pre-allocated persistent GPU buffers for deepstack features. |
| 569 | # Written via .copy_() in embed_input_ids(), read by forward() via a |
| 570 | # slice. Because the buffer address is fixed, CUDA graph replay sees |
| 571 | # the updated values written just before each prefill. |
| 572 | # Shape: (max_num_batched_tokens, lm_hidden_size) per level. |
| 573 | n_layerwise = len(config.deepstack_layer_map) |
| 574 | n_spatial = len(getattr(config, "spatial_target_layers", [])) |
| 575 | num_ds_levels = n_layerwise + n_spatial |
| 576 | lm_hidden = config.text_config.hidden_size |
| 577 | max_tokens = vllm_config.scheduler_config.max_num_batched_tokens |
| 578 | # Allocated on CPU first; moved to GPU in embed_input_ids on first use. |
| 579 | self._ds_buffers: list[torch.Tensor] = [ |
| 580 | torch.zeros(max_tokens, lm_hidden) for _ in range(num_ds_levels) |
| 581 | ] |
| 582 | self._ds_num_tokens: int = 0 # tokens written in last embed_input_ids call |
| 583 | |
| 584 | # ----- Vision feature extraction ----- |
| 585 | |
| 586 | def _get_vision_hidden_states( |
| 587 | self, pixel_values: torch.Tensor |
| 588 | ) -> list[torch.Tensor]: |
| 589 | """Run vision tower and return all hidden states (including input embeddings). |
| 590 | |
| 591 | Uses SiglipEncoder's built-in return_all_hidden_states support. |
| 592 | Returns list[Tensor] where index 0 = embeddings, index i = after layer i-1. |
| 593 | """ |
| 594 | vt = self.vision_tower |
| 595 | vm = vt.vision_model if hasattr(vt, "vision_model") else vt |
| 596 | |
| 597 | hidden_states = vm.embeddings(pixel_values) |
| 598 | all_hidden_states = vm.encoder( |
| 599 | inputs_embeds=hidden_states, |
| 600 | return_all_hidden_states=True, |
| 601 | ) |
| 602 | return all_hidden_states |
| 603 | |
| 604 | def _pack_and_unpad_image_features( |
| 605 | self, |
| 606 | image_features: list[torch.Tensor] | tuple[torch.Tensor, ...], |
| 607 | image_sizes: torch.Tensor, |
| 608 | ) -> list[torch.Tensor]: |
| 609 | """Reshape, unpad, and pack image features. |
| 610 | |
| 611 | Matches HF Granite4VisionModel.pack_and_unpad_image_features exactly. |
| 612 | """ |
| 613 | config = self.config |
| 614 | ds_rate = self._downsample_rate |
| 615 | new_image_features = [] |
| 616 | |
| 617 | for image_idx, image_feature in enumerate(image_features): |
| 618 | if image_feature.shape[0] > 1: |
| 619 | # Multi-patch: first is base, rest are high-res |
| 620 | base_image_feature = image_feature[0] |
| 621 | image_feature = image_feature[1:] |
| 622 | |
| 623 | height = width = ( |
| 624 | config.vision_config.image_size // config.vision_config.patch_size |
| 625 | ) |
| 626 | # After QFormer downsampling |
| 627 | height = int(height * ds_rate) |
| 628 | width = int(width * ds_rate) |
| 629 | |
| 630 | num_patch_height, num_patch_width = get_anyres_image_grid_shape( |
| 631 | image_sizes[image_idx], |
| 632 | config.image_grid_pinpoints, |
| 633 | config.vision_config.image_size, |
| 634 | ) |
| 635 | |
| 636 | image_feature = image_feature.view( |
| 637 | num_patch_height, num_patch_width, height, width, -1 |
| 638 | ) |
| 639 | image_feature = ( |
| 640 | image_feature.permute(4, 0, 2, 1, 3) |
| 641 | .contiguous() |
| 642 | .flatten(1, 2) |
| 643 | .flatten(2, 3) |
| 644 | ) |
| 645 | image_feature = unpad_image(image_feature, image_sizes[image_idx]) |
| 646 | |
| 647 | if self.image_newline is not None: |
| 648 | image_feature = torch.cat( |
| 649 | ( |
| 650 | image_feature, |
| 651 | self.image_newline[:, None, None] |
| 652 | .expand(*image_feature.shape[:-1], 1) |
| 653 | .to(image_feature.device, image_feature.dtype), |
| 654 | ), |
| 655 | dim=-1, |
| 656 | ) |
| 657 | |
| 658 | image_feature = image_feature.flatten(1, 2).transpose(0, 1) |
| 659 | image_feature = torch.cat((base_image_feature, image_feature), dim=0) |
| 660 | else: |
| 661 | image_feature = image_feature[0] |
| 662 | if self.image_newline is not None: |
| 663 | image_feature = torch.cat( |
| 664 | (image_feature, self.image_newline[None].to(image_feature)), |
| 665 | dim=0, |
| 666 | ) |
| 667 | |
| 668 | new_image_features.append(image_feature) |
| 669 | |
| 670 | return new_image_features |
| 671 | |
| 672 | def _get_all_layer_features( |
| 673 | self, |
| 674 | pixel_values: torch.Tensor, |
| 675 | image_sizes: torch.Tensor, |
| 676 | ) -> tuple[list[int], list[torch.Tensor]]: |
| 677 | """Extract deepstack + spatial features for all levels. |
| 678 | |
| 679 | Returns: |
| 680 | llm_layer_indices: ordered list of target LLM layer indices |
| 681 | per_image_packed: one tensor per image, shape |
| 682 | (num_tokens_i, lm_hidden_size * num_levels), |
| 683 | all levels packed on dim=-1. |
| 684 | |
| 685 | Packing on dim=-1 means the framework's token-level slicing for |
| 686 | chunked prefill preserves all levels intact. |
| 687 | |
| 688 | """ |
| 689 | select_strategy = self._vision_feature_select_strategy |
| 690 | |
| 691 | image_num_patches = [ |
| 692 | image_size_to_num_patches( |
| 693 | image_size=imsize, |
| 694 | grid_pinpoints=self.config.image_grid_pinpoints, |
| 695 | patch_size=self.config.vision_config.image_size, |
| 696 | ) |
| 697 | for imsize in image_sizes |
| 698 | ] |
| 699 | |
| 700 | if pixel_values.dim() == 5: |
| 701 | pixel_values = torch.cat( |
| 702 | [pv[:np_] for pv, np_ in zip(pixel_values, image_num_patches)], |
| 703 | dim=0, |
| 704 | ) |
| 705 | |
| 706 | all_hidden_states = self._get_vision_hidden_states(pixel_values) |
| 707 | |
| 708 | # Collect per-level: (llm_layer, [per_image_tensor, ...]) |
| 709 | levels: list[tuple[int, list[torch.Tensor]]] = [] |
| 710 | |
| 711 | for proj_idx, (vision_layer, llm_layer) in enumerate(self._deepstack_layer_map): |
| 712 | selected = all_hidden_states[vision_layer] |
| 713 | if select_strategy == "default": |
| 714 | selected = selected[:, 1:] |
| 715 | projected = self.layerwise_projectors[proj_idx](selected) |
| 716 | per_image = self._pack_and_unpad_image_features( |
| 717 | torch.split(projected, image_num_patches, dim=0), image_sizes |
| 718 | ) |
| 719 | levels.append((llm_layer, per_image)) |
| 720 | |
| 721 | if self._use_spatial_sampling and self.spatial_projectors is not None: |
| 722 | spatial_hidden = all_hidden_states[self._spatial_vision_layer] |
| 723 | if select_strategy == "default": |
| 724 | spatial_hidden = spatial_hidden[:, 1:] |
| 725 | for group_idx, llm_layer in enumerate(self._spatial_target_layers): |
| 726 | projected = self.spatial_projectors[group_idx](spatial_hidden) |
| 727 | per_image = self._pack_and_unpad_image_features( |
| 728 | torch.split(projected, image_num_patches, dim=0), image_sizes |
| 729 | ) |
| 730 | levels.append((llm_layer, per_image)) |
| 731 | |
| 732 | llm_layer_indices = [llm_layer for llm_layer, _ in levels] |
| 733 | num_images = len(image_sizes) |
| 734 | per_image_packed = [ |
| 735 | torch.cat([levels[lvl][1][img] for lvl in range(len(levels))], dim=-1) |
| 736 | for img in range(num_images) |
| 737 | ] |
| 738 | |
| 739 | return llm_layer_indices, per_image_packed |
| 740 | |
| 741 | # ----- Multimodal interface ----- |
| 742 | |
| 743 | def _parse_and_validate_image_input( |
| 744 | self, **kwargs: object |
| 745 | ) -> LlavaNextImageInputs | None: |
| 746 | pixel_values = kwargs.pop("pixel_values", None) |
| 747 | image_sizes = kwargs.pop("image_sizes", None) |
| 748 | image_embeds = kwargs.pop("image_embeds", None) |
| 749 | |
| 750 | if pixel_values is None and image_embeds is None: |
| 751 | return None |
| 752 | |
| 753 | if pixel_values is not None: |
| 754 | expected_h = expected_w = self.config.vision_config.image_size |
| 755 | return LlavaNextImagePixelInputs( |
| 756 | type="pixel_values", |
| 757 | pixel_values=pixel_values, |
| 758 | image_sizes=image_sizes, |
| 759 | resolve_bindings={"h": expected_h, "w": expected_w}, |
| 760 | ) |
| 761 | |
| 762 | if image_embeds is not None: |
| 763 | return LlavaNextImageEmbeddingInputs( |
| 764 | type="image_embeds", |
| 765 | data=image_embeds, |
| 766 | ) |
| 767 | |
| 768 | raise AssertionError("Unreachable") |
| 769 | |
| 770 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 771 | """Run vision tower and return per-image packed feature tensors. |
| 772 | |
| 773 | Each returned tensor has shape (num_tokens_i, lm_hidden_size * num_levels) |
| 774 | with all deepstack levels packed on dim=-1. The framework caches these |
| 775 | tensors and slices along dim=0 for chunked prefill — all levels survive |
| 776 | intact because slicing is token-wise, not feature-wise. |
| 777 | |
| 778 | embed_input_ids() splits the packed tensor back into per-level buffers. |
| 779 | """ |
| 780 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 781 | if image_input is None: |
| 782 | return [] |
| 783 | |
| 784 | if image_input["type"] == "image_embeds": |
| 785 | return [image_input["data"]] |
| 786 | |
| 787 | pixel_values = image_input["pixel_values"] |
| 788 | image_sizes = image_input.get("image_sizes") |
| 789 | |
| 790 | if isinstance(pixel_values, list): |
| 791 | pixel_values = torch.cat(pixel_values, dim=0) |
| 792 | |
| 793 | llm_layer_indices, per_image_packed = self._get_all_layer_features( |
| 794 | pixel_values, image_sizes |
| 795 | ) |
| 796 | self._ds_layer_indices = llm_layer_indices |
| 797 | return per_image_packed |
| 798 | |
| 799 | def embed_input_ids( |
| 800 | self, |
| 801 | input_ids: torch.Tensor, |
| 802 | multimodal_embeddings: MultiModalEmbeddings | None = None, |
| 803 | *, |
| 804 | is_multimodal: torch.Tensor | None = None, |
| 805 | handle_oov_mm_token: bool = True, |
| 806 | ) -> torch.Tensor: |
| 807 | """Merge text and vision embeddings, apply embedding_multiplier. |
| 808 | |
| 809 | HF flow: |
| 810 | 1. inputs_embeds = embed_tokens(input_ids) |
| 811 | 2. inputs_embeds.masked_fill(vision_mask, 0.0) |
| 812 | 3. hidden_states = inputs_embeds * embedding_multiplier |
| 813 | 4. layer loop injects deepstack features at target layers |
| 814 | |
| 815 | multimodal_embeddings contains packed tensors from embed_multimodal(): |
| 816 | shape (num_tokens_i, lm_hidden_size * num_levels). We split on dim=-1 |
| 817 | to get per-level features, build batch-sized buffers (zero at text |
| 818 | positions), and store in self._ds_features for forward(). |
| 819 | """ |
| 820 | lm_inner = self.language_model.model |
| 821 | |
| 822 | has_vision = ( |
| 823 | multimodal_embeddings is not None |
| 824 | and is_multimodal is not None |
| 825 | and len(multimodal_embeddings) > 0 |
| 826 | and is_multimodal.any() |
| 827 | ) |
| 828 | |
| 829 | if not has_vision: |
| 830 | self._ds_num_tokens = 0 |
| 831 | embeds = lm_inner.embed_input_ids(input_ids) |
| 832 | return embeds * lm_inner.config.embedding_multiplier |
| 833 | |
| 834 | # 1. Text embeddings |
| 835 | text_embeds = lm_inner.embed_input_ids(input_ids) |
| 836 | |
| 837 | # 2. Zero image positions (matches HF masked_fill(vision_mask, 0.0)) |
| 838 | text_embeds[is_multimodal] = 0.0 |
| 839 | |
| 840 | # 3. Apply embedding_multiplier |
| 841 | inputs_embeds = text_embeds * lm_inner.config.embedding_multiplier |
| 842 | |
| 843 | # 4. Split packed tensors into per-level features and build buffers. |
| 844 | # multimodal_embeddings is a list of per-image packed tensors |
| 845 | # (possibly a chunk slice from the framework's encoder cache). |
| 846 | # Concatenate along token dim → (total_mm_tokens, lm_h * num_levels). |
| 847 | N, lm_h = inputs_embeds.shape |
| 848 | assert multimodal_embeddings is not None |
| 849 | all_packed = torch.cat( |
| 850 | [t.to(dtype=inputs_embeds.dtype) for t in multimodal_embeddings], |
| 851 | dim=0, |
| 852 | ) |
| 853 | level_features = all_packed.split(lm_h, dim=-1) # num_levels tensors |
| 854 | |
| 855 | # Ensure persistent buffers are on the right device/dtype (first call). |
| 856 | buf0 = self._ds_buffers[0] |
| 857 | if buf0.device != inputs_embeds.device or buf0.dtype != inputs_embeds.dtype: |
| 858 | self._ds_buffers = [ |
| 859 | b.to(device=inputs_embeds.device, dtype=inputs_embeds.dtype) |
| 860 | for b in self._ds_buffers |
| 861 | ] |
| 862 | |
| 863 | for level_idx in range(len(self._ds_layer_indices)): |
| 864 | target = self._ds_buffers[level_idx][:N] |
| 865 | target.zero_() |
| 866 | target[is_multimodal] = level_features[level_idx] |
| 867 | |
| 868 | self._ds_num_tokens = N |
| 869 | return inputs_embeds |
| 870 | |
| 871 | # ----- Forward ----- |
| 872 | |
| 873 | def forward( |
| 874 | self, |
| 875 | input_ids: torch.Tensor, |
| 876 | positions: torch.Tensor, |
| 877 | intermediate_tensors: IntermediateTensors | None = None, |
| 878 | inputs_embeds: torch.Tensor | None = None, |
| 879 | **kwargs: object, |
| 880 | ) -> torch.Tensor | IntermediateTensors: |
| 881 | if intermediate_tensors is not None: |
| 882 | inputs_embeds = None |
| 883 | |
| 884 | # Build IntermediateTensors from pre-allocated persistent buffers. |
| 885 | # Always pass deepstack when inputs_embeds is non-None (prefill path), |
| 886 | # including during CUDA graph capture (buffers are zero → no-op injection). |
| 887 | # This ensures the graph captures the injection code path. |
| 888 | if ( |
| 889 | inputs_embeds is not None |
| 890 | and get_pp_group().is_first_rank |
| 891 | and self._ds_layer_indices |
| 892 | ): |
| 893 | n = inputs_embeds.size(0) |
| 894 | ds: IntermediateTensors | None = IntermediateTensors( |
| 895 | { |
| 896 | f"ds_{llm_layer}": self._ds_buffers[lvl][:n] |
| 897 | for lvl, llm_layer in enumerate(self._ds_layer_indices) |
| 898 | } |
| 899 | ) |
| 900 | else: |
| 901 | ds = None |
| 902 | |
| 903 | hidden_states = self.language_model.model( |
| 904 | input_ids=input_ids, |
| 905 | positions=positions, |
| 906 | intermediate_tensors=intermediate_tensors, |
| 907 | inputs_embeds=inputs_embeds, |
| 908 | deepstack_input_embeds=ds, |
| 909 | ) |
| 910 | |
| 911 | # Clear buffers after use so stale features don't leak into the next request. |
| 912 | if ( |
| 913 | inputs_embeds is not None |
| 914 | and get_pp_group().is_first_rank |
| 915 | and self._ds_num_tokens > 0 |
| 916 | ): |
| 917 | n = self._ds_num_tokens |
| 918 | for buf in self._ds_buffers: |
| 919 | buf[:n].zero_() |
| 920 | self._ds_num_tokens = 0 |
| 921 | |
| 922 | return hidden_states |
| 923 | |
| 924 | def compute_logits( |
| 925 | self, |
| 926 | hidden_states: torch.Tensor, |
| 927 | ) -> torch.Tensor | None: |
| 928 | # GraniteForCausalLM.compute_logits uses |
| 929 | # LogitsProcessor(scale=1/logits_scaling) |
| 930 | return self.language_model.compute_logits(hidden_states) |
| 931 | |
| 932 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 933 | loader = AutoWeightsLoader(self) |
| 934 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |