vllm/vllm/model_executor/models/vision.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import itertools
import math
from abc import ABC, abstractmethod
from collections.abc import Callable
from typing import Final, Generic, Literal, Protocol, TypeAlias, TypeVar

import torch
from transformers import PretrainedConfig

from vllm.attention.backends.registry import AttentionBackendEnum
from vllm.config import VllmConfig
from vllm.distributed import (
    get_tensor_model_parallel_rank,
    get_tensor_model_parallel_world_size,
    tensor_model_parallel_all_gather,
)
from vllm.logger import init_logger
from vllm.platforms import current_platform

logger = init_logger(__name__)

_C = TypeVar("_C", bound=PretrainedConfig)


class _RootConfig(Protocol[_C]):
    vision_config: _C


class VisionEncoderInfo(ABC, Generic[_C]):
    def __init__(self, hf_config: _RootConfig[_C]) -> None:
        super().__init__()

        self.hf_config = hf_config
        self.vision_config = hf_config.vision_config

    @abstractmethod
    def get_num_image_tokens(
        self,
        *,
        image_width: int,
        image_height: int,
    ) -> int:
        raise NotImplementedError

    @abstractmethod
    def get_image_size(self) -> int:
        raise NotImplementedError

    @abstractmethod
    def get_patch_size(self) -> int:
        raise NotImplementedError

    @abstractmethod
    def get_patch_grid_length(self) -> int:
        raise NotImplementedError


class VisionLanguageConfig(Protocol):
    vision_config: Final[PretrainedConfig]


def get_vision_encoder_info(hf_config: VisionLanguageConfig) -> VisionEncoderInfo:
    # Avoid circular imports
    from .clip import CLIPEncoderInfo, CLIPVisionConfig
    from .pixtral import PixtralHFEncoderInfo, PixtralVisionConfig
    from .siglip import SiglipEncoderInfo, SiglipVisionConfig

    if isinstance(hf_config.vision_config, CLIPVisionConfig):
        return CLIPEncoderInfo(hf_config)
    if isinstance(hf_config.vision_config, PixtralVisionConfig):
        return PixtralHFEncoderInfo(hf_config)
    if isinstance(hf_config.vision_config, SiglipVisionConfig):
        return SiglipEncoderInfo(hf_config)

    msg = f"Unsupported vision config: {type(hf_config.vision_config)}"
    raise NotImplementedError(msg)


def get_vit_attn_backend(
    head_size: int,
    dtype: torch.dtype,
    *,
    attn_backend_override: AttentionBackendEnum | None = None,
) -> AttentionBackendEnum:
    """
    Get the available attention backend for Vision Transformer.
    """
    if attn_backend_override is not None:
        return attn_backend_override

    # Lazy import to avoid circular dependency
    from vllm.attention.selector import get_env_variable_attn_backend

    selected_backend: AttentionBackendEnum | None = get_env_variable_attn_backend()
    if selected_backend is not None:
        return selected_backend

    return current_platform.get_vit_attn_backend(head_size, dtype)


def should_torch_compile_mm_vit(vllm_config: VllmConfig) -> bool:
    """Callable to be passed to `@support_torch_compile`'s `enable_if` argument."""
    return vllm_config.compilation_config.compile_mm_encoder


VisionFeatureSelectStrategyStr = Literal["class", "default", "full"]

VisionFeatureSelectStrategy: TypeAlias = (
    VisionFeatureSelectStrategyStr | Callable[[torch.Tensor], torch.Tensor]
)


def _get_vision_feature_selector(
    strategy: VisionFeatureSelectStrategy | str,
) -> Callable[[torch.Tensor], torch.Tensor]:
    if callable(strategy):
        return strategy

    # https://github.com/huggingface/transformers/blob/cd74917ffc3e8f84e4a886052c5ab32b7ac623cc/src/transformers/models/clip/modeling_clip.py#L762
    if strategy == "class":
        return lambda feats: feats[:, :1, :]

    # https://github.com/huggingface/transformers/blob/4a02bc7004285bdb12cc033e87ad2578ce2fa900/src/transformers/models/llava/modeling_llava.py#L196
    if strategy == "default":
        return lambda feats: feats[:, 1:, :]

    if strategy == "full":
        return lambda feats: feats

    raise ValueError(f"Unexpected feature select strategy: {strategy!r}")


def get_num_selected_vision_tokens(
    num_vision_tokens: int,
    strategy: VisionFeatureSelectStrategy | str,
) -> int:
    if callable(strategy):
        dummy_features = torch.empty(1, num_vision_tokens, 64)  # [B, L, D]
        dummy_selected_features = strategy(dummy_features)
        return dummy_selected_features.shape[1]

    if strategy == "class":
        return 1

    if strategy == "default":
        return num_vision_tokens - 1

    if strategy == "full":
        return num_vision_tokens

    raise ValueError(f"Unexpected feature select strategy: {strategy!r}")


def resolve_visual_encoder_outputs(
    encoder_outputs: torch.Tensor | list[torch.Tensor],
    post_layer_norm: torch.nn.LayerNorm | None,
    *,
    select_layers: list[int] | None = None,
    max_possible_layers: int | None = None,
    feature_select_strategy: VisionFeatureSelectStrategy | None = None,
) -> torch.Tensor:
    """Given the outputs a visual encoder module that may correspond to the
    output of the last layer, or a list of hidden states to be stacked,
    handle post normalization and resolve it into a single output tensor.

    Args:
        encoder_outputs: Output of encoder's last layer or all hidden states.
        post_layer_norm: Post norm to apply to the output of the encoder.
        select_layers: Optional layer indices to grab from the encoder
            outputs; if provided, encoder outputs must be a list.
        max_possible_layers: Total layers in the fully loaded visual encoder.
        feature_select_strategy: Defines how to select the hidden states
            from each layer.
    """
    if select_layers is None:
        if not isinstance(encoder_outputs, torch.Tensor):
            raise ValueError(
                "Expected only a single encoder output when "
                "`select_layers` is not provided"
            )

        if feature_select_strategy is not None:
            select_features = _get_vision_feature_selector(feature_select_strategy)
            encoder_outputs = select_features(encoder_outputs)

        if post_layer_norm is not None:
            return post_layer_norm(encoder_outputs)

        return encoder_outputs

    if max_possible_layers is None:
        raise ValueError(
            "`max_possible_layers` must be provided alongside `select_layers`"
        )

    # Get the hidden states corresponding to the layer indices.
    # Negative values are relative to the full visual encoder,
    # so offset them depending on how many layers were loaded.
    # NOTE: this assumes that encoder_outputs is a list containing
    # the inputs to the visual encoder, followed by the hidden states
    # of each layer.
    num_loaded_layers = len(encoder_outputs) - 1
    offset = max_possible_layers - num_loaded_layers
    hs_pool = [
        encoder_outputs[layer_idx]
        if layer_idx >= 0
        else encoder_outputs[layer_idx + offset]
        for layer_idx in select_layers
    ]

    if feature_select_strategy is not None:
        select_features = _get_vision_feature_selector(feature_select_strategy)
        hs_pool = [select_features(hs) for hs in hs_pool]

    # Apply post-norm on the final hidden state if we are using it
    uses_last_layer = select_layers[-1] in (max_possible_layers - 1, -1)
    if post_layer_norm is not None and uses_last_layer:
        hs_pool[-1] = post_layer_norm(hs_pool[-1])

    return torch.cat(hs_pool, dim=-1)


def run_dp_sharded_vision_model(
    image_input: torch.Tensor, vision_model: torch.nn.Module
) -> torch.Tensor:
    """Run a vision model with data parallelism (DP) sharding. The function
    will shard the input image tensor on the first dimension and run the vision
    model

    Args:
        image_input (torch.Tensor): Image input tensor.
        vision_model (torch.nn.Module): Vision model.
    Returns:
        torch.Tensor: Output image embeddings
    """

    num_chunks = image_input.shape[0]
    mp_world_size = get_tensor_model_parallel_world_size()
    num_chunks_per_rank = (num_chunks + mp_world_size - 1) // mp_world_size
    num_padded_chunks = num_chunks_per_rank * mp_world_size - num_chunks
    pad = (0,) * (2 * (image_input.dim() - 1)) + (0, num_padded_chunks)
    image_input_padded = torch.nn.functional.pad(image_input, pad)
    rank = get_tensor_model_parallel_rank()
    image_input_per_rank = image_input_padded[
        rank * num_chunks_per_rank : (rank + 1) * num_chunks_per_rank, ...
    ]

    vision_embeddings = vision_model(image_input_per_rank)
    # Ensure tensor is contiguous before all_gather
    vision_embeddings = vision_embeddings.contiguous()
    vision_embeddings = tensor_model_parallel_all_gather(vision_embeddings, dim=0)
    vision_embeddings = vision_embeddings[:num_chunks, ...]
    return vision_embeddings


def get_load_balance_assignment(
    sizes: list[int],
    num_gpus: int = 2,
) -> tuple[list[int], list[int], list[int]]:
    """
    Generate load balancing assignment and metadata
    for distributing data across GPUs.
    The load is determined by the total image sizes,
    not the number of images.

    Args:
        sizes: The size of each image
        num_gpus: Number of GPUs to balance across

    Returns:
        shuffle_indices:
            Indices to reorder data for balanced loading
        gpu_sample_counts:
            Number of samples assigned to each GPU
        grouped_sizes_per_gpu:
            Total size assigned to each GPU

    Example:
        ```
        sizes = [1000, 100, 200, 50]
        num_gpus = 2
        ```

    """

    n_samples = len(sizes)

    # Handle edge cases
    if n_samples == 0:
        return [], [0] * num_gpus, [0] * num_gpus

    # Use greedy algorithm - balance by total size, not sample count
    gpu_assignments = [list[int]() for _ in range(num_gpus)]
    gpu_loads = [0] * num_gpus  # This tracks total SIZE, not sample count

    # Sort indices by size (largest first for better load balancing)
    # sizes = [1000, 100, 200, 50]
    # large_to_small_indices = [0, 2, 1, 3]
    large_to_small_indices = sorted(
        range(n_samples), key=lambda i: sizes[i], reverse=True
    )

    for idx in large_to_small_indices:
        # Find GPU with minimum current load (by total size)
        min_gpu = min(range(num_gpus), key=lambda i: gpu_loads[i])
        gpu_assignments[min_gpu].append(idx)
        gpu_loads[min_gpu] += sizes[idx]

    # Create shuffle indices and counts
    shuffle_indices = list[int]()
    gpu_sample_counts = list[int]()
    for gpu_id in range(num_gpus):
        # GPU_0 = [1000] = [0]
        # GPU_1 = [200, 100, 50] = [2, 1, 3]
        # shuffle_indices = [0, 2, 1, 3]
        shuffle_indices.extend(gpu_assignments[gpu_id])
        # GPU_0 = [1]
        # GPU_1 = [3]
        # gpu_sample_counts = [1, 3]
        gpu_sample_counts.append(len(gpu_assignments[gpu_id]))

    return (shuffle_indices, gpu_sample_counts, gpu_loads)


def run_dp_sharded_mrope_vision_model(
    vision_model: torch.nn.Module,
    pixel_values: torch.Tensor,
    grid_thw_list: list[list[int]],
    *,
    rope_type: Literal["rope_3d", "rope_2d"],
) -> tuple[torch.Tensor, ...]:
    """Run a vision model with data parallelism (DP) sharding.
    The function will shard the input image tensor on the
    first dimension and run the vision model.
    This function is used to run the vision model with mrope.

    Args:
        vision_model (torch.nn.Module): Vision model.
        pixel_values (torch.Tensor): Image/Video input tensor.
        grid_thw_list: List of grid dimensions for each image
        rope_type: Type of rope used in the vision model.
                   Different rope types have different dimension to do ViT.
                   "rope_3d" for 3D rope (e.g., Qwen2.5-VL)
                   "rope_2d" for 2D rope (e.g., Kimi-VL)
    Returns:
        torch.Tensor: Output image embeddings

    Example:
        ```
        vision_model.out_hidden_size = 64
        vision_model.spatial_merge_size = 2
        pixel_values.shape = (1350, channel)
        grid_thw_list = [[1, 10, 100], [1, 10, 10], [1, 10, 20], [1, 50]]
        tp_size = 2
        ```

    """
    tp_size = get_tensor_model_parallel_world_size()

    # GPU_0 tp_rank_local = 0
    # GPU_1 tp_rank_local = 1
    tp_rank_local = get_tensor_model_parallel_rank()

    # patches_per_image = [1000, 100, 200, 50]
    patches_per_image = [math.prod(grid_thw) for grid_thw in grid_thw_list]
    # patches_per_image = [0, 1000, 1100, 1300, 1350]
    cum_patches_per_image = [0, *itertools.accumulate(patches_per_image)]

    # Get load balancing assignment with all metadata
    # image_to_tp_rank = [0, 2, 1, 3]
    # gpu_sample_counts = [1, 3]
    # grouped_pixel_values_len = [1000, 350]
    (image_to_tp_rank, gpu_sample_counts, grouped_pixel_values_len) = (
        get_load_balance_assignment(patches_per_image, tp_size)
    )

    # cu_gpu_sample_counts = [0, 1, 4]
    cum_gpu_sample_counts = [0, *itertools.accumulate(gpu_sample_counts)]

    # GPU_0 image_idxs_local = [0]
    # GPU_1 image_idxs_local = [2, 1, 3]
    image_idxs_local = image_to_tp_rank[
        cum_gpu_sample_counts[tp_rank_local] : cum_gpu_sample_counts[tp_rank_local + 1]
    ]

    # Get the pixel values for the local images based on the image_idxs_local
    if len(image_idxs_local) > 0:
        pixel_values_local = torch.cat(
            [
                pixel_values[cum_patches_per_image[i] : cum_patches_per_image[i + 1]]
                for i in image_idxs_local
            ]
        )
    else:
        # Handle case where this rank has no images
        pixel_values_local = torch.empty(
            (0, pixel_values.shape[1]),
            device=pixel_values.device,
            dtype=pixel_values.dtype,
        )
    # embed_dim_reduction_factor = 2 * 2
    if rope_type == "rope_2d":
        embed_dim_reduction_factor = (
            vision_model.merge_kernel_size[0] * vision_model.merge_kernel_size[1]
        )
    else:
        embed_dim_reduction_factor = (
            vision_model.spatial_merge_size * vision_model.spatial_merge_size
        )

    # Find the max length across all ranks
    # The output embedding of every DP rank has to be
    # padded to this length for tensor_model_parallel_all_gather
    # to work
    max_len_per_rank = max(grouped_pixel_values_len) // embed_dim_reduction_factor
    local_grid_thw_list = [grid_thw_list[i] for i in image_idxs_local]

    # Run the vision model on the local pixel_values_local
    if rope_type == "rope_2d":
        if pixel_values_local.shape[0] > 0:
            image_embeds_local = vision_model(
                pixel_values_local, torch.tensor(local_grid_thw_list)
            )
            if isinstance(image_embeds_local, list):
                image_embeds_local = torch.cat(image_embeds_local, dim=0)
        else:
            out_dim = getattr(vision_model.config, "hidden_size", None)
            image_embeds_local = torch.empty(
                (0, embed_dim_reduction_factor, out_dim),
                device=pixel_values.device,
                dtype=pixel_values.dtype,
            )
    else:
        if pixel_values_local.shape[0] > 0:
            image_embeds_local = vision_model(pixel_values_local, local_grid_thw_list)
        else:
            # Handle empty case
            image_embeds_local = torch.empty(
                (0, vision_model.out_hidden_size),
                device=pixel_values.device,
                dtype=pixel_values.dtype,
            )

    # Pad the output based on max_len_per_rank
    # for tensor_model_parallel_all_gather to work
    current_len = image_embeds_local.shape[0]
    if current_len < max_len_per_rank:
        padding_size = max_len_per_rank - current_len
        if rope_type == "rope_2d":
            padding = torch.empty(
                (
                    padding_size,
                    image_embeds_local.shape[1],
                    image_embeds_local.shape[2],
                ),
                dtype=image_embeds_local.dtype,
                device=image_embeds_local.device,
            )
        else:
            padding = torch.empty(
                (padding_size, image_embeds_local.shape[1]),
                dtype=image_embeds_local.dtype,
                device=image_embeds_local.device,
            )
        image_embeds_local_padded = torch.cat([image_embeds_local, padding], dim=0)
    else:
        image_embeds_local_padded = image_embeds_local

    # Do all_gather to collect embeddings from all ranks
    gathered_embeds = tensor_model_parallel_all_gather(image_embeds_local_padded, dim=0)

    # Remove padding and reconstruct per-rank embeddings
    rank_embeddings = list[torch.Tensor]()
    for rank in range(tp_size):
        start_idx = rank * max_len_per_rank
        end_idx = start_idx + (
            grouped_pixel_values_len[rank] // embed_dim_reduction_factor
        )
        rank_embeddings.append(gathered_embeds[start_idx:end_idx])

    patches_per_output_image = [
        (patch_size // embed_dim_reduction_factor) for patch_size in patches_per_image
    ]

    # Reconstruct embeddings in the original order
    original_order_embeddings = [None] * len(grid_thw_list)
    current_idx = 0
    for rank in range(tp_size):
        count = gpu_sample_counts[rank]
        if count > 0:
            # Get images assigned to this rank in shuffled order
            # GPU_0 = image_idxs_local  [0]
            # GPU_1 = image_idxs_local  [2, 1, 3]
            rank_images = image_to_tp_rank[current_idx : current_idx + count]

            rank_embed = rank_embeddings[rank]
            # Split rank embeddings back to individual images
            embed_start = 0
            for img_idx in rank_images:
                img_patches = patches_per_output_image[img_idx]
                original_order_embeddings[img_idx] = rank_embed[
                    embed_start : embed_start + img_patches
                ]
                embed_start += img_patches
            current_idx += count
    out_embeddings = tuple(
        embed for embed in original_order_embeddings if embed is not None
    )
    assert len(out_embeddings) == len(original_order_embeddings), (
        "Found unassigned embeddings"
    )
    return out_embeddings


def get_llm_pos_ids_for_vision(
    start_idx: int,
    vision_idx: int,
    spatial_merge_size: int,
    t_index: list[int],
    grid_hs: torch.Tensor,
    grid_ws: torch.Tensor,
) -> torch.Tensor:
    llm_pos_ids_list = []
    llm_grid_h = grid_hs[vision_idx] // spatial_merge_size
    llm_grid_w = grid_ws[vision_idx] // spatial_merge_size
    h_index = (
        torch.arange(llm_grid_h)
        .view(1, -1, 1)
        .expand(len(t_index), -1, llm_grid_w)
        .flatten()
    )
    w_index = (
        torch.arange(llm_grid_w)
        .view(1, 1, -1)
        .expand(len(t_index), llm_grid_h, -1)
        .flatten()
    )
    t_index_tensor = (
        torch.Tensor(t_index)
        .to(llm_grid_h.device)
        .view(-1, 1)
        .expand(-1, llm_grid_h * llm_grid_w)
        .long()
        .flatten()
    )
    _llm_pos_ids = torch.stack([t_index_tensor, h_index, w_index])
    llm_pos_ids_list.append(_llm_pos_ids + start_idx)
    llm_pos_ids = torch.cat(llm_pos_ids_list, dim=1)
    return llm_pos_ids


# Due to a performance regression with Conv3D in PyTorch2.9, we reshape
# Conv3D weights to Linear weights for better performance.
# See: https://github.com/vllm-project/vllm/issues/27406
# and https://github.com/pytorch/pytorch/issues/166122
# FIXME(Isotr0py): Revert the PR introduces this workaround
# (https://github.com/vllm-project/vllm/pull/27418),
# once the performance issue is resolved in PyTorch.
def conv3d_to_linear_weight(conv3d_weight: torch.Tensor) -> torch.Tensor:
    """
    Reshape Conv3D weight to Linear weight. Only work when kernel_size==stride.
    """
    out_channels, in_channels, kt, kh, kw = conv3d_weight.shape
    linear_weight = conv3d_weight.reshape(out_channels, in_channels * kt * kh * kw)
    return linear_weight