paligemma.py

# SPDX-License-Identifier: Apache-2.0

from typing import (Iterable, Literal, Mapping, Optional, Set, Tuple,
                    TypedDict, Union)

import torch
from torch import nn
from transformers import BatchFeature, PaliGemmaConfig

from vllm.config import VllmConfig
from vllm.logger import init_logger
from vllm.model_executor.layers.sampler import SamplerOutput
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
                                    MultiModalInputs, MultiModalKwargs,
                                    NestedTensors)
from vllm.multimodal.parse import MultiModalDataItems
from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                        BaseProcessingInfo, PromptIndexTargets,
                                        PromptInsertion, PromptReplacement,
                                        PromptUpdateDetails)
from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
from vllm.sequence import IntermediateTensors

from .interfaces import SupportsMultiModal, SupportsPP
from .siglip import SiglipVisionModel, get_max_siglip_image_tokens
from .utils import (AutoWeightsLoader, init_vllm_registered_model,
                    maybe_prefix, merge_multimodal_embeddings)

logger = init_logger(__name__)


class PaliGemmaImagePixelInputs(TypedDict):
    type: Literal["pixel_values"]
    data: torch.Tensor
    """Shape: `(batch_size * num_images, num_channels, height, width)`"""


class PaliGemmaImageEmbeddingInputs(TypedDict):
    type: Literal["image_embeds"]
    data: torch.Tensor
    """Shape: `(batch_size * num_images, image_feature_size, hidden_size)`

    `hidden_size` must match the hidden size of language model backbone.
    """


PaliGemmaImageInputs = Union[PaliGemmaImagePixelInputs,
                             PaliGemmaImageEmbeddingInputs]


class PaliGemmaMultiModalProjector(nn.Module):

    def __init__(self, vision_hidden_size: int, projection_dim: int):
        super().__init__()

        self.linear = nn.Linear(vision_hidden_size, projection_dim, bias=True)

    def forward(self, image_features: torch.Tensor) -> torch.Tensor:
        hidden_states = self.linear(image_features)
        return hidden_states


class PaliGemmaProcessingInfo(BaseProcessingInfo):

    def get_hf_config(self):
        return self.ctx.get_hf_config(PaliGemmaConfig)

    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
        return {"image": 1}

    def get_mm_max_tokens_per_item(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> Mapping[str, int]:
        return {"image": self.get_num_image_tokens()}

    def get_num_image_tokens(self) -> int:
        hf_config = self.get_hf_config()
        vision_config = hf_config.vision_config
        return get_max_siglip_image_tokens(vision_config)


class PaliGemmaDummyInputsBuilder(
        BaseDummyInputsBuilder[PaliGemmaProcessingInfo]):

    def get_dummy_processor_inputs(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> ProcessorInputs:
        hf_config = self.info.get_hf_config()
        vision_config = hf_config.vision_config
        max_image_size = vision_config.image_size

        num_images = mm_counts.get("image", 0)

        mm_data = {
            "image":
            self._get_dummy_images(width=max_image_size,
                                   height=max_image_size,
                                   num_images=num_images)
        }

        return ProcessorInputs(
            prompt_text="",
            mm_data=mm_data,
        )


class PaliGemmaMultiModalProcessor(
        BaseMultiModalProcessor[PaliGemmaProcessingInfo]):

    def _call_hf_processor(
        self,
        prompt: str,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
    ) -> BatchFeature:
        tokenizer = self.info.get_tokenizer()
        if not mm_data:
            prompt_ids = tokenizer.encode(prompt)
            return BatchFeature(dict(input_ids=[prompt_ids]), tensor_type="pt")

        return super()._call_hf_processor(
            prompt=prompt,
            mm_data=mm_data,
            mm_kwargs=mm_kwargs,
        )

    def _get_mm_fields_config(
        self,
        hf_inputs: BatchFeature,
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
        return dict(pixel_values=MultiModalFieldConfig.batched("image"))

    def _get_prompt_updates(
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
        out_mm_kwargs: MultiModalKwargs,
    ) -> list[PromptReplacement]:
        hf_config = self.info.get_hf_config()
        image_token_id = hf_config.image_token_index

        tokenizer = self.info.get_tokenizer()
        num_image_tokens = self.info.get_num_image_tokens()
        image_tokens = [image_token_id] * num_image_tokens

        bos_token_id = tokenizer.bos_token_id
        assert isinstance(bos_token_id, int)

        # Paligemma 1 and 2 have different tokenizer.add_bos_token
        # Insert <image>*n + <bos> after <bos> for Paligemma 1
        # Insert <image>*n + <bos> for Paligemma 2
        return [
            PromptInsertion(
                modality="image",
                target=PromptIndexTargets.prefix(
                    [bos_token_id] if tokenizer.add_bos_token else []),
                insertion=PromptUpdateDetails(
                    full=image_tokens + [bos_token_id],
                    features=image_tokens,
                ),
            )
        ]

    def apply(
        self,
        prompt: Union[str, list[int]],
        mm_data: MultiModalDataDict,
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> MultiModalInputs:
        mm_inputs = super().apply(prompt, mm_data, hf_processor_mm_kwargs)
        prompt_token_ids = mm_inputs["prompt_token_ids"]

        tokenizer = self.info.get_tokenizer()
        newline_prompt = "\n"
        newline_token_id = tokenizer.encode(newline_prompt)[-1]  # 108
        # Force to add newline at the end of prompt for paligemma's format
        # This step can NOT be replacemented by current PromptUpdate methods
        if len(prompt_token_ids) and prompt_token_ids[-1] != newline_token_id:
            prompt_token_ids.append(newline_token_id)
            mm_inputs["prompt_token_ids"] = prompt_token_ids
            mm_inputs["prompt"] += newline_prompt

        return mm_inputs


@MULTIMODAL_REGISTRY.register_processor(
    PaliGemmaMultiModalProcessor,
    info=PaliGemmaProcessingInfo,
    dummy_inputs=PaliGemmaDummyInputsBuilder)
class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal,
                                        SupportsPP):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        multimodal_config = vllm_config.model_config.multimodal_config
        self.config = config
        self.multimodal_config = multimodal_config

        self.vision_tower = SiglipVisionModel(config.vision_config,
                                              quant_config,
                                              prefix=maybe_prefix(
                                                  prefix, "vision_tower"))
        self.multi_modal_projector = PaliGemmaMultiModalProjector(
            vision_hidden_size=config.vision_config.hidden_size,
            projection_dim=config.vision_config.projection_dim)

        self.quant_config = quant_config

        if config.text_config.model_type == "gemma":
            config.text_config.architectures = ["GemmaForCausalLM"]
        else:
            config.text_config.architectures = ["Gemma2ForCausalLM"]
        self.language_model = init_vllm_registered_model(
            vllm_config=vllm_config,
            hf_config=config.text_config,
            prefix=maybe_prefix(prefix, "language_model"),
        )
        logit_scale = getattr(config, "logit_scale", 1.0)
        self.language_model.logits_processor.scale *= logit_scale

        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors)

    @property
    def sampler(self):
        return self.language_model.sampler

    def _validate_pixel_values(self, data: torch.Tensor) -> torch.Tensor:
        h = w = self.config.vision_config.image_size
        expected_dims = (3, h, w)
        actual_dims = tuple(data.shape[1:])

        if actual_dims != expected_dims:
            expected_expr = ("batch_size", *map(str, expected_dims))
            raise ValueError(
                f"The expected shape of pixel values is {expected_expr}. "
                f"You supplied {tuple(data.shape)}.")

        return data

    def _parse_and_validate_image_input(
            self, **kwargs: object) -> Optional[PaliGemmaImageInputs]:
        pixel_values = kwargs.pop("pixel_values", None)
        image_embeds = kwargs.pop("image_embeds", None)

        if pixel_values is None and image_embeds is None:
            return None

        if pixel_values is not None:
            if not isinstance(pixel_values, torch.Tensor):
                raise ValueError("Incorrect type of pixel values. "
                                 f"Got type: {type(pixel_values)}")

            # Remove the N dimension until multiple images are supported.
            pixel_values = pixel_values.squeeze(1)

            return PaliGemmaImagePixelInputs(
                type="pixel_values",
                data=self._validate_pixel_values(pixel_values),
            )

        if image_embeds is not None:
            if not isinstance(image_embeds, torch.Tensor):
                raise ValueError("Incorrect type of image embeddings. "
                                 f"Got type: {type(image_embeds)}")

            # Remove the N dimension until multiple images are supported.
            image_embeds = image_embeds.squeeze(1)

            return PaliGemmaImageEmbeddingInputs(
                type="image_embeds",
                data=image_embeds,
            )

        raise AssertionError("This line should be unreachable.")

    def _image_pixels_to_features(
        self,
        vision_tower: SiglipVisionModel,
        pixel_values: torch.Tensor,
    ) -> torch.Tensor:

        target_dtype = vision_tower.get_input_embeddings().weight.dtype
        image_features = vision_tower(pixel_values.to(dtype=target_dtype))

        return image_features

    def _process_image_input(
        self,
        image_input: PaliGemmaImageInputs,
    ) -> torch.Tensor:

        if image_input["type"] == "image_embeds":
            return image_input["data"]

        assert self.vision_tower is not None
        pixel_values = image_input["data"]
        image_features = self._image_pixels_to_features(
            self.vision_tower,
            pixel_values,
        )

        return self.multi_modal_projector(image_features)

    def get_multimodal_embeddings(
        self, **kwargs
    ) -> Union[list[torch.Tensor], torch.Tensor, tuple[torch.Tensor, ...]]:
        image_input = self._parse_and_validate_image_input(**kwargs)
        if image_input is None:
            return None
        vision_embeddings = self._process_image_input(image_input)
        # https://github.com/huggingface/transformers/blob/main/src/transformers/models/paligemma/modeling_paligemma.py#L294 # noqa
        vision_embeddings = vision_embeddings * (self.config.hidden_size**-0.5)
        return vision_embeddings

    def get_input_embeddings(
        self,
        input_ids: torch.Tensor,
        multimodal_embeddings: Optional[NestedTensors] = None,
    ) -> torch.Tensor:
        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
        if multimodal_embeddings is not None:
            inputs_embeds = merge_multimodal_embeddings(
                input_ids, inputs_embeds, multimodal_embeddings,
                self.config.image_token_index)
        return inputs_embeds

    def forward(self,
                input_ids: torch.Tensor,
                positions: torch.Tensor,
                intermediate_tensors: Optional[IntermediateTensors] = None,
                inputs_embeds: Optional[torch.Tensor] = None,
                **kwargs: object) -> Union[SamplerOutput, IntermediateTensors]:
        if intermediate_tensors is not None:
            inputs_embeds = None

        # NOTE: In v1, inputs_embeds is always generated at model runner, this
        # condition is for v0 compatibility.
        elif inputs_embeds is None:
            vision_embeddings = self.get_multimodal_embeddings(**kwargs)
            inputs_embeds = self.get_input_embeddings(input_ids,
                                                      vision_embeddings)
            input_ids = None

        hidden_states = self.language_model.model(input_ids,
                                                  positions,
                                                  intermediate_tensors,
                                                  inputs_embeds=inputs_embeds)

        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        return self.language_model.compute_logits(hidden_states,
                                                  sampling_metadata)

    def sample(
        self,
        logits: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
        return self.language_model.sample(logits, sampling_metadata)

    def load_weights(self, weights: Iterable[Tuple[str,
                                                   torch.Tensor]]) -> Set[str]:
        loader = AutoWeightsLoader(self)
        return loader.load_weights(weights)