| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """vLLM support for microsoft/Phi-4-reasoning-vision-15B. |
| 4 | |
| 5 | Architecture: Siglip2 vision tower + MLP projector + Phi3 language model. |
| 6 | """ |
| 7 | |
| 8 | import math |
| 9 | from collections.abc import Iterable, Mapping, Sequence |
| 10 | from typing import Annotated, Any, Literal |
| 11 | |
| 12 | import torch |
| 13 | import torch.nn as nn |
| 14 | from transformers import BatchFeature, PreTrainedConfig, Siglip2VisionConfig |
| 15 | |
| 16 | from vllm.config import VllmConfig |
| 17 | from vllm.config.multimodal import MultiModalDummyOptions |
| 18 | from vllm.inputs import MultiModalDataDict |
| 19 | from vllm.logger import init_logger |
| 20 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 21 | from vllm.multimodal.inputs import ( |
| 22 | MultiModalFieldConfig, |
| 23 | MultiModalKwargsItems, |
| 24 | ) |
| 25 | from vllm.multimodal.parse import ( |
| 26 | ImageSize, |
| 27 | MultiModalDataItems, |
| 28 | ) |
| 29 | from vllm.multimodal.processing import ( |
| 30 | BaseDummyInputsBuilder, |
| 31 | PromptReplacement, |
| 32 | PromptUpdate, |
| 33 | ) |
| 34 | from vllm.multimodal.processing.processor import ( |
| 35 | BaseMultiModalProcessor, |
| 36 | BaseProcessingInfo, |
| 37 | cached_encode, |
| 38 | ) |
| 39 | from vllm.sequence import IntermediateTensors |
| 40 | from vllm.utils.tensor_schema import TensorSchema, TensorShape |
| 41 | |
| 42 | from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP |
| 43 | from .lfm2_siglip2 import Siglip2Model |
| 44 | from .llava import LlavaMultiModalProjector |
| 45 | from .utils import ( |
| 46 | AutoWeightsLoader, |
| 47 | WeightsMapper, |
| 48 | init_vllm_registered_model, |
| 49 | maybe_prefix, |
| 50 | ) |
| 51 | |
| 52 | logger = init_logger(__name__) |
| 53 | |
| 54 | IMAGE_TOKEN_INDEX = -200 |
| 55 | DEFAULT_IMAGE_TOKEN = "<image>" |
| 56 | |
| 57 | # The HF processor replaces "<image>" with IMAGE_TOKEN_INDEX (-200) in input_ids. |
| 58 | # Negative token IDs cause OverflowError during decoding, so we remap to a real |
| 59 | # in-vocabulary token. The Phi-4-reasoning-vision tokenizer ships with reserved |
| 60 | # dummy tokens (<|dummy_0|> … <|dummy_83|>); we reuse the first one as the |
| 61 | # image placeholder. This mirrors how Phi-3-vision uses its dedicated <|image|> |
| 62 | # token (ID 32044). |
| 63 | _IMAGE_TOKEN_ID = 100256 # <|dummy_0|> in the Phi-4 tokenizer |
| 64 | |
| 65 | |
| 66 | # --------------------------------------------------------------------------- |
| 67 | # Processing |
| 68 | # --------------------------------------------------------------------------- |
| 69 | |
| 70 | |
| 71 | class Phi4SiglipProcessingInfo(BaseProcessingInfo): |
| 72 | def get_supported_mm_limits(self) -> Mapping[str, int | None]: |
| 73 | return {"image": None} |
| 74 | |
| 75 | def _get_vision_config(self) -> dict: |
| 76 | return self.get_hf_config().vision_config # type: ignore[attr-defined] |
| 77 | |
| 78 | def _get_patch_size(self) -> int: |
| 79 | vc = self._get_vision_config() |
| 80 | if isinstance(vc, dict): |
| 81 | return vc.get("patch_size", 16) |
| 82 | return getattr(vc, "patch_size", 16) |
| 83 | |
| 84 | def _get_max_num_patches(self) -> int: |
| 85 | return getattr(self.get_hf_config(), "max_num_patches", 3600) |
| 86 | |
| 87 | def _get_min_num_patches(self) -> int: |
| 88 | return getattr(self.get_hf_config(), "min_num_patches", 256) |
| 89 | |
| 90 | def get_num_image_tokens( |
| 91 | self, |
| 92 | *, |
| 93 | image_width: int, |
| 94 | image_height: int, |
| 95 | ) -> int: |
| 96 | patch_size = self._get_patch_size() |
| 97 | min_patches = self._get_min_num_patches() |
| 98 | max_patches = self._get_max_num_patches() |
| 99 | |
| 100 | num_patches_h = image_height // patch_size |
| 101 | num_patches_w = image_width // patch_size |
| 102 | num_patches = max(num_patches_h * num_patches_w, 1) |
| 103 | num_patches = max(min(num_patches, max_patches), min_patches) |
| 104 | return num_patches |
| 105 | |
| 106 | def get_image_size_with_most_features(self) -> ImageSize: |
| 107 | patch_size = self._get_patch_size() |
| 108 | max_patches = self._get_max_num_patches() |
| 109 | side = int(math.sqrt(max_patches)) * patch_size |
| 110 | return ImageSize(width=side, height=side) |
| 111 | |
| 112 | def get_mm_max_tokens_per_item( |
| 113 | self, seq_len: int, mm_counts: Mapping[str, int] |
| 114 | ) -> Mapping[str, int]: |
| 115 | return {"image": self._get_max_num_patches()} |
| 116 | |
| 117 | |
| 118 | class Phi4SiglipDummyInputsBuilder( |
| 119 | BaseDummyInputsBuilder[Phi4SiglipProcessingInfo], |
| 120 | ): |
| 121 | def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: |
| 122 | num_images = mm_counts.get("image", 0) |
| 123 | return DEFAULT_IMAGE_TOKEN * num_images |
| 124 | |
| 125 | def get_dummy_mm_data( |
| 126 | self, |
| 127 | seq_len: int, |
| 128 | mm_counts: Mapping[str, int], |
| 129 | mm_options: MultiModalDummyOptions, |
| 130 | ) -> MultiModalDataDict: |
| 131 | size = self.info.get_image_size_with_most_features() |
| 132 | return { |
| 133 | "image": self._get_dummy_images( |
| 134 | width=size.width, |
| 135 | height=size.height, |
| 136 | num_images=mm_counts.get("image", 0), |
| 137 | overrides=mm_options.get("image"), |
| 138 | ), |
| 139 | } |
| 140 | |
| 141 | |
| 142 | class Phi4SiglipMultiModalProcessor( |
| 143 | BaseMultiModalProcessor[Phi4SiglipProcessingInfo], |
| 144 | ): |
| 145 | def _get_hf_mm_text(self, mm_counts: Mapping[str, int]) -> str: |
| 146 | return self.dummy_inputs.get_dummy_text(mm_counts) |
| 147 | |
| 148 | def _get_mm_fields_config( |
| 149 | self, |
| 150 | hf_inputs: BatchFeature, |
| 151 | hf_processor_mm_kwargs: Mapping[str, object], |
| 152 | ) -> Mapping[str, MultiModalFieldConfig]: |
| 153 | return dict( |
| 154 | pixel_values=MultiModalFieldConfig.batched("image"), |
| 155 | pixel_attention_mask=MultiModalFieldConfig.batched("image"), |
| 156 | spatial_shapes=MultiModalFieldConfig.batched("image", keep_on_cpu=True), |
| 157 | ) |
| 158 | |
| 159 | def _get_prompt_updates( |
| 160 | self, |
| 161 | mm_items: MultiModalDataItems, |
| 162 | hf_processor_mm_kwargs: Mapping[str, Any], |
| 163 | out_mm_kwargs: MultiModalKwargsItems, |
| 164 | ) -> Sequence[PromptUpdate]: |
| 165 | tokenizer = self.info.get_tokenizer() |
| 166 | |
| 167 | def get_replacement(item_idx: int): |
| 168 | # Read the actual patch grid from the NaFlex processor's |
| 169 | # spatial_shapes output (same pattern as LFM2-VL). This avoids |
| 170 | # predicting from raw image dimensions, which can diverge from |
| 171 | # the NaFlex resize/tile logic. |
| 172 | out_item = out_mm_kwargs["image"][item_idx] |
| 173 | spatial_shapes = out_item["spatial_shapes"].data |
| 174 | assert isinstance(spatial_shapes, torch.Tensor) |
| 175 | num_tokens = int(spatial_shapes.prod().item()) |
| 176 | return [_IMAGE_TOKEN_ID] * num_tokens |
| 177 | |
| 178 | return [ |
| 179 | PromptReplacement( |
| 180 | modality="image", |
| 181 | target=cached_encode( |
| 182 | tokenizer, DEFAULT_IMAGE_TOKEN, add_special_tokens=False |
| 183 | ), |
| 184 | replacement=get_replacement, |
| 185 | ), |
| 186 | ] |
| 187 | |
| 188 | |
| 189 | # --------------------------------------------------------------------------- |
| 190 | # Input schemas |
| 191 | # --------------------------------------------------------------------------- |
| 192 | |
| 193 | |
| 194 | class Phi4SiglipImagePixelInputs(TensorSchema): |
| 195 | """Dimensions: |
| 196 | - bn: Batch size * number of images |
| 197 | - d: Max number of patches (padded across images in the batch) |
| 198 | - fd: Features per patch (patch_size * patch_size * channels) |
| 199 | """ |
| 200 | |
| 201 | type: Literal["pixel_values"] = "pixel_values" |
| 202 | pixel_values: Annotated[torch.Tensor, TensorShape("bn", "d", "fd")] |
| 203 | pixel_attention_mask: Annotated[torch.Tensor, TensorShape("bn", "d")] |
| 204 | spatial_shapes: Annotated[torch.Tensor, TensorShape("bn", 2)] |
| 205 | |
| 206 | |
| 207 | # --------------------------------------------------------------------------- |
| 208 | # Model |
| 209 | # --------------------------------------------------------------------------- |
| 210 | |
| 211 | |
| 212 | @MULTIMODAL_REGISTRY.register_processor( |
| 213 | Phi4SiglipMultiModalProcessor, |
| 214 | info=Phi4SiglipProcessingInfo, |
| 215 | dummy_inputs=Phi4SiglipDummyInputsBuilder, |
| 216 | ) |
| 217 | class Phi4ForCausalLMV(nn.Module, SupportsMultiModal, SupportsPP): |
| 218 | hf_to_vllm_mapper = WeightsMapper( |
| 219 | orig_to_new_prefix={ |
| 220 | "model.vision_tower.vision_tower.vision_model.head.": None, |
| 221 | "model.vision_tower.vision_tower.": "vision_tower.", |
| 222 | "model.mm_projector.0.": "multi_modal_projector.linear_1.", |
| 223 | "model.mm_projector.2.": "multi_modal_projector.linear_2.", |
| 224 | "lm_head.": "language_model.lm_head.", |
| 225 | "model.": "language_model.model.", |
| 226 | }, |
| 227 | ) |
| 228 | |
| 229 | @classmethod |
| 230 | def get_placeholder_str(cls, modality: str, i: int) -> str | None: |
| 231 | if modality.startswith("image"): |
| 232 | return DEFAULT_IMAGE_TOKEN |
| 233 | raise ValueError("Only image modality is supported") |
| 234 | |
| 235 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None: |
| 236 | super().__init__() |
| 237 | |
| 238 | config: PreTrainedConfig = vllm_config.model_config.hf_config |
| 239 | quant_config = vllm_config.quant_config |
| 240 | self.config = config |
| 241 | |
| 242 | vision_config_dict: dict = getattr(config, "vision_config", {}) |
| 243 | if isinstance(vision_config_dict, dict): |
| 244 | if "patch_size" not in vision_config_dict: |
| 245 | vision_config_dict["patch_size"] = 16 |
| 246 | siglip2_config = Siglip2VisionConfig(**vision_config_dict) |
| 247 | else: |
| 248 | siglip2_config = vision_config_dict |
| 249 | |
| 250 | vision_hidden_size: int = config.mm_hidden_size # type: ignore[attr-defined] |
| 251 | text_hidden_size: int = config.hidden_size # type: ignore[attr-defined] |
| 252 | |
| 253 | with self._mark_tower_model(vllm_config, "image"): |
| 254 | layer_idx = -2 |
| 255 | num_hidden_layers = siglip2_config.num_hidden_layers + layer_idx + 1 |
| 256 | |
| 257 | self.vision_tower = Siglip2Model( |
| 258 | siglip2_config, |
| 259 | quant_config=quant_config, |
| 260 | num_hidden_layers_override=num_hidden_layers, |
| 261 | require_post_norm=False, |
| 262 | prefix=maybe_prefix(prefix, "vision_tower"), |
| 263 | ) |
| 264 | self.multi_modal_projector = LlavaMultiModalProjector( |
| 265 | vision_hidden_size=vision_hidden_size, |
| 266 | text_hidden_size=text_hidden_size, |
| 267 | projector_hidden_act="gelu", |
| 268 | multimodal_projector_bias=True, |
| 269 | quant_config=quant_config, |
| 270 | prefix=maybe_prefix(prefix, "multi_modal_projector"), |
| 271 | ) |
| 272 | |
| 273 | with self._mark_language_model(vllm_config): |
| 274 | self.language_model = init_vllm_registered_model( |
| 275 | vllm_config=vllm_config, |
| 276 | hf_config=config, |
| 277 | prefix=maybe_prefix(prefix, "language_model"), |
| 278 | architectures=["Phi3ForCausalLM"], |
| 279 | ) |
| 280 | |
| 281 | self.make_empty_intermediate_tensors = ( |
| 282 | self.language_model.make_empty_intermediate_tensors |
| 283 | ) |
| 284 | |
| 285 | self.configure_mm_token_handling( |
| 286 | vocab_size=config.vocab_size, # type: ignore[attr-defined] |
| 287 | mm_token_ids=[_IMAGE_TOKEN_ID], |
| 288 | ) |
| 289 | |
| 290 | def _packed_from_padded( |
| 291 | self, |
| 292 | pixel_values: torch.Tensor, |
| 293 | pixel_attention_mask: torch.Tensor, |
| 294 | spatial_shapes: torch.Tensor, |
| 295 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: |
| 296 | """Convert padded NaFlex tensors to packed format for Siglip2Model.""" |
| 297 | valid_counts = pixel_attention_mask.sum(dim=1).to(torch.int32) |
| 298 | pixel_values_packed = pixel_values[pixel_attention_mask.bool()] |
| 299 | cu_seqlens = torch.zeros( |
| 300 | len(valid_counts) + 1, |
| 301 | dtype=torch.int32, |
| 302 | device=pixel_values.device, |
| 303 | ) |
| 304 | cu_seqlens[1:] = valid_counts.cumsum(0) |
| 305 | max_seqlen = valid_counts.max() |
| 306 | return ( |
| 307 | pixel_values_packed, |
| 308 | spatial_shapes, |
| 309 | cu_seqlens, |
| 310 | max_seqlen, |
| 311 | ) |
| 312 | |
| 313 | def _parse_and_validate_image_input( |
| 314 | self, **kwargs: object |
| 315 | ) -> Phi4SiglipImagePixelInputs | None: |
| 316 | pixel_values = kwargs.pop("pixel_values", None) |
| 317 | pixel_attention_mask = kwargs.pop("pixel_attention_mask", None) |
| 318 | spatial_shapes = kwargs.pop("spatial_shapes", None) |
| 319 | if pixel_values is None: |
| 320 | return None |
| 321 | |
| 322 | return Phi4SiglipImagePixelInputs( |
| 323 | type="pixel_values", |
| 324 | pixel_values=pixel_values, |
| 325 | pixel_attention_mask=pixel_attention_mask, |
| 326 | spatial_shapes=spatial_shapes, |
| 327 | ) |
| 328 | |
| 329 | def _process_image_input( |
| 330 | self, image_input: Phi4SiglipImagePixelInputs |
| 331 | ) -> MultiModalEmbeddings: |
| 332 | pixel_values = image_input["pixel_values"] |
| 333 | pixel_attention_mask = image_input["pixel_attention_mask"] |
| 334 | spatial_shapes = image_input["spatial_shapes"] |
| 335 | |
| 336 | ( |
| 337 | pixel_values_packed, |
| 338 | spatial_shapes_packed, |
| 339 | cu_seqlens, |
| 340 | max_seqlen, |
| 341 | ) = self._packed_from_padded(pixel_values, pixel_attention_mask, spatial_shapes) |
| 342 | |
| 343 | vision_features = self.vision_tower( |
| 344 | pixel_values_packed=pixel_values_packed, |
| 345 | spatial_shapes=spatial_shapes_packed, |
| 346 | cu_seqlens=cu_seqlens, |
| 347 | max_seqlen=max_seqlen, |
| 348 | select_layers=[-2], |
| 349 | ) |
| 350 | |
| 351 | if vision_features.dim() == 3: |
| 352 | vision_features = vision_features.squeeze(0) |
| 353 | |
| 354 | image_features = self.multi_modal_projector(vision_features) |
| 355 | |
| 356 | valid_counts = pixel_attention_mask.sum(dim=1).tolist() |
| 357 | return torch.split(image_features, [int(c) for c in valid_counts]) |
| 358 | |
| 359 | def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: |
| 360 | image_input = self._parse_and_validate_image_input(**kwargs) |
| 361 | if image_input is None: |
| 362 | return [] |
| 363 | |
| 364 | return self._process_image_input(image_input) |
| 365 | |
| 366 | def forward( |
| 367 | self, |
| 368 | input_ids: torch.Tensor | None, |
| 369 | positions: torch.Tensor, |
| 370 | intermediate_tensors: IntermediateTensors | None = None, |
| 371 | inputs_embeds: torch.Tensor | None = None, |
| 372 | **kwargs: object, |
| 373 | ) -> torch.Tensor | IntermediateTensors: |
| 374 | if intermediate_tensors is not None: |
| 375 | inputs_embeds = None |
| 376 | |
| 377 | hidden_states = self.language_model.model( |
| 378 | input_ids, |
| 379 | positions, |
| 380 | intermediate_tensors, |
| 381 | inputs_embeds=inputs_embeds, |
| 382 | ) |
| 383 | return hidden_states |
| 384 | |
| 385 | def compute_logits( |
| 386 | self, |
| 387 | hidden_states: torch.Tensor, |
| 388 | ) -> torch.Tensor | None: |
| 389 | return self.language_model.compute_logits(hidden_states) |
| 390 | |
| 391 | def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: |
| 392 | loader = AutoWeightsLoader(self) |
| 393 | return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) |