image-match/python/py/Lib/site-packages/transformers/models/glm46v/modeling_glm46v.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/glm46v/modular_glm46v.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_glm46v.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2025 the HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.


import itertools
from dataclasses import dataclass
from typing import Any

import torch
import torch.nn as nn

from ...cache_utils import Cache
from ...generation import GenerationMixin
from ...modeling_outputs import BaseModelOutputWithPooling, ModelOutput
from ...modeling_utils import PreTrainedModel
from ...processing_utils import Unpack
from ...utils import (
    TransformersKwargs,
    auto_docstring,
    can_return_tuple,
    torch_compilable_check,
)
from ..auto import AutoModel
from .configuration_glm46v import Glm46VConfig


@auto_docstring
class Glm46VPreTrainedModel(PreTrainedModel):
    config: Glm46VConfig
    base_model_prefix = "model"
    input_modalities = ("image", "video", "text")
    supports_gradient_checkpointing = True
    _no_split_modules = None
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn = True
    _supports_sdpa = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = None


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Llava outputs, with hidden states and attentions.
    """
)
class Glm46VModelOutputWithPast(ModelOutput):
    r"""
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
        The rope index difference between sequence length and multimodal rope.
    """

    last_hidden_state: torch.FloatTensor | None = None
    past_key_values: Cache | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None
    attentions: tuple[torch.FloatTensor] | None = None
    rope_deltas: torch.LongTensor | None = None


@auto_docstring
class Glm46VModel(Glm46VPreTrainedModel):
    base_model_prefix = "model"
    # Reference: fix gemma3 grad acc #37208
    accepts_loss_kwargs = False
    _no_split_modules = None

    def __init__(self, config):
        super().__init__(config)
        self.visual = AutoModel.from_config(config.vision_config)
        self.language_model = AutoModel.from_config(config.text_config)
        self.rope_deltas = None  # cache rope_deltas here

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.language_model.set_input_embeddings(value)

    def get_vision_position_ids(
        self,
        start_position: int,
        grid_thw: list[int, int, int] | torch.Tensor,
        temp_merge_size: int = 1,
        spatial_merge_size: int = 1,
        time_interval: int = 1,
        device: str | torch.device | None = None,
    ):
        """
        Compute 3D positional indices for vision tokens derived from a single image or video input.

        The positions are generated from the input grid defined by temporal (T), height (H), and
        width (W) dimensions. Temporal and spatial dimensions can be downscaled according to the
        merge sizes used in the vision backbone. The resulting positions are offset by `start_position`.

        Args:
            start_position (`int`):
                Offset added to all computed positional indices.
            grid_thw (`Sequence[int]` or `torch.Tensor` of shape `(3,)`):
                The (T, H, W) grid representing the feature layout of the current image or video after patch embedding.
            temp_merge_size (`int`, *optional*):
                Factor by which the temporal dimension is reduced in the backbone. The temporal grid size is divided
                by this value. Defaults to 1.
            spatial_merge_size (`int`, *optional*):
                Factor by which the spatial dimensions (H and W) are reduced in the backbone. Both H and W are divided
                by this value. Defaults to 1.
            time_interval (`int`, *optional*):
                Spacing factor applied between consecutive temporal position indices.Defaults to 1.
            device (`str` or `torch.device`, *optional*):
                Device on which the resulting tensor is allocated. If `None`, uses the current default device.

        Returns:
            torch.LongTensor of shape (3, sequence_length):
                Positional indices for temporal, height, and width dimensions,
                flattened into sequence form and offset by `start_position`.
        """
        llm_grid_t, llm_grid_h, llm_grid_w = (
            grid_thw[0].item() // temp_merge_size,
            grid_thw[1].item() // spatial_merge_size,
            grid_thw[2].item() // spatial_merge_size,
        )

        image_seq_length = llm_grid_h * llm_grid_w * llm_grid_t
        position_width = torch.arange(start_position, start_position + llm_grid_w, device=device).repeat(
            llm_grid_h * llm_grid_t
        )
        position_height = torch.arange(start_position, start_position + llm_grid_h, device=device).repeat_interleave(
            llm_grid_w * llm_grid_t
        )
        position_temporal = torch.full((image_seq_length,), start_position, device=device, dtype=torch.long)
        position_temporal = position_temporal * time_interval
        vision_position_ids = torch.stack([position_temporal, position_height, position_width], dim=0)

        return vision_position_ids

    def get_rope_index(
        self,
        input_ids: torch.LongTensor,
        mm_token_type_ids: torch.IntTensor,
        image_grid_thw: torch.LongTensor | None = None,
        video_grid_thw: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Calculate the 3D rope index based on image and video's sizes. The utility expects a `vision + text`
        sequence and will error out otherwise. For pure text sequence, please rely on model's auto-inferred
        position ids. In a mixed vision + text sequence, vision tokens use 3D RoPE (temporal, height, width)
        while text tokens use standard 1D RoPE.

        Example:
            Temporal patches: 3; Height patches: 2; Width patches: 2
            Each vision input results in (temporal x height × width) positions. Here: 3 x 2 × 2 = 12 positions total.

            Temporal position IDs are spaced by:
                `interval = tokens_per_second * temporal_patch_size / fps`

                If fps = 1; tokens_per_second = 25; temporal_patch_size = 2, temporal IDs increase by 50 for each temporal patch:
                `[0, 0, 0, 0, 50, 50, 50, 50, 100, 100, 100, 100]`

            Height IDs repeat per row: `[0, 0, 1, 1, ...]`
            Width IDs alternate per column: `[0, 1, 0, 1, ...]`
            Text tokens follow standard 1D RoPE and the position IDs grow consequently with a step of `1`

        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
                it.
            mm_token_type_ids (`torch.IntTensor` of shape `(batch_size, sequence_length)`):
                Token type ids matching each modality to a different value in the input sequence, i.e. text (0), image (1), video (2).
            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
                The temporal, height and width of feature shape of each image in LLM.
            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
                The temporal, height and width of feature shape of each video in LLM.
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

        Returns:
            position_ids (`torch.LongTensor` of shape `(3, batch_size, sequence_length)`)
            mrope_position_deltas (`torch.Tensor` of shape `(batch_size)`)
        """
        spatial_merge_size = self.config.vision_config.spatial_merge_size

        mrope_position_deltas = []
        position_ids = torch.zeros(
            3,
            input_ids.shape[0],
            input_ids.shape[1],
            dtype=input_ids.dtype,
            device=input_ids.device,
        )
        grid_iters = {
            1: iter(image_grid_thw) if image_grid_thw is not None else None,
            2: iter(video_grid_thw) if video_grid_thw is not None else None,
        }

        for batch_idx, current_input_ids in enumerate(input_ids):
            input_token_type = mm_token_type_ids[batch_idx]
            if attention_mask is not None:
                current_input_ids = current_input_ids[attention_mask[batch_idx].bool()]
                input_token_type = input_token_type[attention_mask[batch_idx].bool()]

            input_type_group = []
            for key, group in itertools.groupby(enumerate(input_token_type.tolist()), lambda x: x[1]):
                group = list(group)
                start_index = group[0][0]
                end_index = group[-1][0] + 1
                input_type_group.append((key, start_index, end_index))

            current_pos = 0
            video_group_index = 0
            llm_pos_ids_list = []
            for modality_type, start_idx, end_idx in input_type_group:
                # text == 0
                if modality_type == 0:
                    text_len = end_idx - start_idx
                    llm_pos_ids_list.append(
                        torch.arange(text_len, device=input_ids.device).view(1, -1).expand(3, -1) + current_pos
                    )
                    current_pos += text_len
                # image == 1, video == 2
                else:
                    # GLM46V splits video into segments per frame but there's only one `grid_thw`
                    # per whole video. We can't exhaus the iterator and have to re-use the grid
                    # while processing the same video!
                    if modality_type == 2:
                        if video_group_index == 0:
                            grid_thw = next(grid_iters[modality_type])
                        video_group_index += 1
                        video_group_index = 0 if video_group_index >= grid_thw[0] else video_group_index
                    else:
                        grid_thw = next(grid_iters[modality_type])

                    # Videos are processed per frame separately, each temporal grid is always `1`
                    temp_merge_size = grid_thw[0]
                    vision_position_ids = self.get_vision_position_ids(
                        current_pos, grid_thw, temp_merge_size, spatial_merge_size, device=input_ids.device
                    )
                    llm_pos_ids_list.append(vision_position_ids)
                    current_pos += max(grid_thw[1], grid_thw[2]) // spatial_merge_size
            llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
            if attention_mask is not None:
                position_ids[:, batch_idx, attention_mask[batch_idx].bool()] = llm_positions.to(position_ids.device)
            else:
                position_ids[:, batch_idx] = llm_positions.to(position_ids.device)
            mrope_position_deltas.append(llm_positions.max() + 1 - len(current_input_ids))
        mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1)
        return position_ids, mrope_position_deltas

    @can_return_tuple
    @auto_docstring
    def get_video_features(
        self,
        pixel_values_videos: torch.FloatTensor,
        video_grid_thw: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
            The tensors corresponding to the input videos.
        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of feature shape of each video in LLM.
        """
        pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
        # reshape video_grid_thw -> [b, 3] -> [1, h, w] * frames
        temp_frames_hw = []
        video_grid_thw_list = video_grid_thw.tolist()
        for t, h, w in video_grid_thw_list:
            repeated_row = torch.tensor([1, h, w]).unsqueeze(0).repeat(t, 1)
            temp_frames_hw.append(repeated_row)
        flattened_video_grid_thw = torch.cat(temp_frames_hw, dim=0)
        vision_outputs = self.visual(
            pixel_values_videos, grid_thw=flattened_video_grid_thw, return_dict=True, **kwargs
        )
        split_sizes = (video_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
        video_embeds = torch.split(vision_outputs.pooler_output, split_sizes)
        vision_outputs.pooler_output = video_embeds

        return vision_outputs

    @can_return_tuple
    @auto_docstring
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        image_grid_thw: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
            The tensors corresponding to the input images.
        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of feature shape of each image in LLM.
        """
        pixel_values = pixel_values.type(self.visual.dtype)
        vision_outputs = self.visual(pixel_values, grid_thw=image_grid_thw, **kwargs)
        split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
        image_embeds = torch.split(vision_outputs.pooler_output, split_sizes)
        vision_outputs.pooler_output = image_embeds

        return vision_outputs

    def get_placeholder_mask(
        self,
        input_ids: torch.LongTensor,
        inputs_embeds: torch.FloatTensor,
        image_features: torch.FloatTensor | None = None,
        video_features: torch.FloatTensor | None = None,
    ):
        """
        Obtains multimodal placeholder mask from `input_ids` or `inputs_embeds`, and checks that the placeholder token count is
        equal to the length of multimodal features. If the lengths are different, an error is raised.
        """
        if input_ids is None:
            special_image_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.image_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            special_image_mask = special_image_mask.all(-1)
            special_video_mask = inputs_embeds == self.get_input_embeddings()(
                torch.tensor(self.config.video_token_id, dtype=torch.long, device=inputs_embeds.device)
            )
            special_video_mask = special_video_mask.all(-1)
        else:
            # GLM-4.1V and GLM-4.5V special_video_mask is special_image_mask
            special_image_mask = input_ids == self.config.image_token_id
            special_video_mask = input_ids == self.config.image_token_id

        n_image_tokens = special_image_mask.sum()
        special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        if image_features is not None:
            torch_compilable_check(
                inputs_embeds[special_image_mask].numel() == image_features.numel(),
                f"Image features and image tokens do not match, tokens: {n_image_tokens}, features: {image_features.shape[0]}",
            )

        n_video_tokens = special_video_mask.sum()
        special_video_mask = special_video_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
        if video_features is not None:
            torch_compilable_check(
                inputs_embeds[special_video_mask].numel() == video_features.numel(),
                f"Video features and video tokens do not match, tokens: {n_video_tokens}, features: {video_features.shape[0]}",
            )
        return special_image_mask, special_video_mask

    def compute_3d_position_ids(
        self,
        input_ids: torch.Tensor | None,
        inputs_embeds: torch.Tensor | None,
        image_grid_thw: torch.Tensor | None = None,
        video_grid_thw: torch.Tensor | None = None,
        attention_mask: torch.Tensor | None = None,
        past_key_values: torch.Tensor | None = None,
        mm_token_type_ids: torch.IntTensor | None = None,
    ) -> torch.Tensor | None:
        past_key_values_length = 0 if past_key_values is None else past_key_values.get_seq_length()
        has_multimodal = image_grid_thw is not None or video_grid_thw is not None
        if has_multimodal and mm_token_type_ids is None and input_ids is not None:
            raise ValueError(
                "Multimodal data was passed (via `image_grid_thw` or `video_grid_thw`) but `mm_token_type_ids` is "
                "missing. Please pass `mm_token_type_ids` to the model so that multimodal RoPE (M-RoPE) can be "
                "computed correctly. `mm_token_type_ids` is returned by the processor alongside `input_ids`."
            )
        can_compute_mrope = input_ids is not None and mm_token_type_ids is not None and has_multimodal

        if can_compute_mrope and (self.rope_deltas is None or past_key_values_length == 0):
            position_ids, rope_deltas = self.get_rope_index(
                input_ids,
                image_grid_thw=image_grid_thw,
                video_grid_thw=video_grid_thw,
                attention_mask=attention_mask,
                mm_token_type_ids=mm_token_type_ids,
            )
            self.rope_deltas = rope_deltas
        # Use pre-calculated rope-deltas to infer correct 3D position ids during incremental
        # generation (past_key_values_length > 0) or when only inputs_embeds is provided (no input_ids
        # to recompute from). Skip when input_ids is provided without past_key_values to avoid shape
        # mismatches from stale rope_deltas (e.g., training forward pass after generation).
        elif self.rope_deltas is not None and (past_key_values_length > 0 or input_ids is None):
            batch_size, seq_length, _ = inputs_embeds.shape
            if attention_mask is not None:
                position_ids = attention_mask.long().cumsum(-1) - 1
                position_ids = position_ids.masked_fill(attention_mask == 0, 0)
                position_ids = position_ids.view(1, batch_size, -1).repeat(3, 1, 1).to(inputs_embeds.device)
            else:
                position_ids = torch.arange(past_key_values_length, past_key_values_length + seq_length)
                position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1).to(inputs_embeds.device)
            delta = self.rope_deltas.repeat_interleave(batch_size // self.rope_deltas.shape[0], dim=0)
            position_ids = position_ids + delta.to(device=inputs_embeds.device)
        else:
            # Can't build correct 3D positions. Let the model infer it
            position_ids = None
        return position_ids

    @auto_docstring
    @can_return_tuple
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        pixel_values: torch.Tensor | None = None,
        pixel_values_videos: torch.FloatTensor | None = None,
        image_grid_thw: torch.LongTensor | None = None,
        video_grid_thw: torch.LongTensor | None = None,
        rope_deltas: torch.LongTensor | None = None,
        mm_token_type_ids: torch.IntTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | Glm46VModelOutputWithPast:
        r"""
        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of feature shape of each image in LLM.
        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of feature shape of each video in LLM.
        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
            The rope index difference between sequence length and multimodal rope.
        """
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.get_input_embeddings()(input_ids)

        if pixel_values is not None:
            image_embeds = self.get_image_features(pixel_values, image_grid_thw, return_dict=True).pooler_output
            image_embeds = torch.cat(image_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
            image_mask, _ = self.get_placeholder_mask(input_ids, inputs_embeds, image_features=image_embeds)
            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)

        if pixel_values_videos is not None:
            video_embeds = self.get_video_features(pixel_values_videos, video_grid_thw, return_dict=True).pooler_output
            video_embeds = torch.cat(video_embeds, dim=0).to(inputs_embeds.device, inputs_embeds.dtype)
            _, video_mask = self.get_placeholder_mask(input_ids, inputs_embeds, video_features=video_embeds)
            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)

        if position_ids is None:
            position_ids = self.compute_3d_position_ids(
                input_ids=input_ids,
                image_grid_thw=image_grid_thw,
                video_grid_thw=video_grid_thw,
                inputs_embeds=inputs_embeds,
                attention_mask=attention_mask,
                past_key_values=past_key_values,
                mm_token_type_ids=mm_token_type_ids,
            )

        outputs = self.language_model(
            input_ids=None,
            position_ids=position_ids,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            **kwargs,
        )

        return Glm46VModelOutputWithPast(
            **outputs,
            rope_deltas=self.rope_deltas,
        )


@dataclass
@auto_docstring(
    custom_intro="""
    Base class for Glm46V causal language model (or autoregressive) outputs.
    """
)
class Glm46VCausalLMOutputWithPast(ModelOutput):
    r"""
    loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
        Language modeling loss (for next-token prediction).
    logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
        Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
    past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
        It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).

        Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
        `past_key_values` input) to speed up sequential decoding.
    rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
        The rope index difference between sequence length and multimodal rope.
    """

    loss: torch.FloatTensor | None = None
    logits: torch.FloatTensor | None = None
    past_key_values: Cache | None = None
    hidden_states: tuple[torch.FloatTensor] | None = None
    attentions: tuple[torch.FloatTensor] | None = None
    rope_deltas: torch.LongTensor | None = None


class Glm46VForConditionalGeneration(Glm46VPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
    # Reference: fix gemma3 grad acc #37208
    accepts_loss_kwargs = False

    def __init__(self, config):
        super().__init__(config)
        self.model = Glm46VModel(config)
        self.lm_head = nn.Linear(config.text_config.hidden_size, config.text_config.vocab_size, bias=False)

        self.post_init()

    def get_input_embeddings(self):
        return self.model.get_input_embeddings()

    def set_input_embeddings(self, value):
        self.model.set_input_embeddings(value)

    @auto_docstring
    def get_video_features(
        self,
        pixel_values_videos: torch.FloatTensor,
        video_grid_thw: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
            The tensors corresponding to the input videos.
        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of feature shape of each video in LLM.
        """
        return self.model.get_video_features(
            pixel_values_videos=pixel_values_videos, video_grid_thw=video_grid_thw, **kwargs
        )

    @auto_docstring
    def get_image_features(
        self,
        pixel_values: torch.FloatTensor,
        image_grid_thw: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPooling:
        r"""
        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
            The tensors corresponding to the input images.
        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of feature shape of each image in LLM.
        """
        return self.model.get_image_features(pixel_values=pixel_values, image_grid_thw=image_grid_thw, **kwargs)

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        pixel_values: torch.Tensor | None = None,
        pixel_values_videos: torch.FloatTensor | None = None,
        image_grid_thw: torch.LongTensor | None = None,
        video_grid_thw: torch.LongTensor | None = None,
        mm_token_type_ids: torch.IntTensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | Glm46VCausalLMOutputWithPast:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
            The temporal, height and width of feature shape of each image in LLM.
        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
            The temporal, height and width of feature shape of each video in LLM.

        Example:

        ```python
        >>> from PIL import Image
        >>> import httpx
        >>> from io import BytesIO
        >>> from transformers import AutoProcessor, Glm46VForConditionalGeneration

        >>> model = Glm46VForConditionalGeneration.from_pretrained("zai-org/GLM-4.1V-9B-Thinking")
        >>> processor = AutoProcessor.from_pretrained("zai-org/GLM-4.1V-9B-Thinking")

        >>> messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image", "url": "https://www.ilankelman.org/stopsigns/australia.jpg"},
                    {"type": "text", "text": "What is shown in this image?"},
                ],
            },
        ]
        >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
        >>> with httpx.stream("GET", url) as response:
        ...     image = Image.open(BytesIO(response.read()))

        >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
        ```"""
        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            pixel_values_videos=pixel_values_videos,
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
            mm_token_type_ids=mm_token_type_ids,
            position_ids=position_ids,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            **kwargs,
        )

        hidden_states = outputs[0]

        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size)

        return Glm46VCausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            rope_deltas=outputs.rope_deltas,
        )

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        position_ids=None,
        use_cache=True,
        pixel_values=None,
        pixel_values_videos=None,
        image_grid_thw=None,
        video_grid_thw=None,
        is_first_iteration=False,
        **kwargs,
    ):
        # Overwritten -- in specific circumstances we don't want to forward image inputs to the model

        model_inputs = super().prepare_inputs_for_generation(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            position_ids=position_ids,
            pixel_values=pixel_values,
            pixel_values_videos=pixel_values_videos,
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
            use_cache=use_cache,
            is_first_iteration=is_first_iteration,
            **kwargs,
        )

        if not is_first_iteration and use_cache:
            model_inputs["pixel_values"] = None
            model_inputs["pixel_values_videos"] = None

        return model_inputs

    def _prepare_position_ids_for_generation(self, inputs_tensor, model_kwargs):
        # Overwritten -- requires 3D position ids

        text_positions = super()._prepare_position_ids_for_generation(inputs_tensor, model_kwargs)

        # Early exit in case we are continuing generation from past kv
        past_length = 0
        if (cache := model_kwargs.get("past_key_values")) is not None:
            past_length = cache.get_seq_length()
        if past_length != 0 and self.model.rope_deltas is not None:
            position_ids = text_positions[None, ...] + self.model.rope_deltas
            return position_ids

        # Otherwise compute 3d position ids for vision tokens and concat with text position ids
        if "input_ids" in model_kwargs and model_kwargs["input_ids"].shape[1] > 0:
            inputs_tensor = model_kwargs["input_ids"]

        is_input_ids = len(inputs_tensor.shape) == 2 and inputs_tensor.dtype in [torch.int, torch.long]
        if (
            is_input_ids
            and model_kwargs.get("mm_token_type_ids") is not None
            and (model_kwargs.get("image_grid_thw") is not None or model_kwargs.get("video_grid_thw") is not None)
        ):
            model_kwargs = {k: v for k, v in model_kwargs.items() if k != "input_ids"}
            vision_positions, rope_deltas = self.model.get_rope_index(inputs_tensor, **model_kwargs)
            self.model.rope_deltas = rope_deltas
        else:
            vision_positions = text_positions.unsqueeze(0).expand(3, -1, -1)
            self.model.rope_deltas = torch.zeros(
                inputs_tensor.shape[0], 1, dtype=torch.long, device=inputs_tensor.device
            )

        # Concatenate "text + vision" positions into [4, bs, seq-len]
        text_positions = text_positions[None, ...]
        position_ids = torch.cat([text_positions, vision_positions], dim=0)

        return position_ids

    def _get_image_nums_and_video_nums(
        self,
        input_ids: torch.LongTensor | None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        """
        Get the number of images and videos for each sample to calculate the separation length of the sample tensor.
        These parameters are not passed through the processor to avoid unpredictable impacts from interface modifications.

        Args:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                Indices of input sequence tokens in the vocabulary.

        Returns:
            image_nums (`torch.LongTensor` of shape `(batch_size, num_images_sample)`)
            video_nums (`torch.LongTensor` of shape `(batch_size, num_videos_sample)`)
        """

        if inputs_embeds is not None:
            is_image = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(self.config.image_start_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            )[..., 0]
            is_video_start = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(self.config.video_start_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            )[..., 0]
            is_video_end = (
                inputs_embeds
                == self.get_input_embeddings()(
                    torch.tensor(self.config.video_end_token_id, dtype=torch.long, device=inputs_embeds.device)
                )
            )[..., 0]
        else:
            is_image = input_ids == self.config.image_start_token_id
            is_video_start = input_ids == self.config.video_start_token_id
            is_video_end = input_ids == self.config.video_end_token_id

        # Cumulative sum to track if we're inside a video span
        # We'll assume well-formed video tags (i.e. matching starts and ends)
        video_level = torch.cumsum(is_video_start.int() - is_video_end.int(), dim=1)
        inside_video = video_level > 0  # shape (batch_size, seq_length)

        # Mask out image tokens that are inside video spans
        standalone_images = is_image & (~inside_video)

        # Count per batch
        image_counts = standalone_images.sum(dim=1)
        video_counts = is_video_start.sum(dim=1)

        return image_counts, video_counts

    def _expand_inputs_for_generation(
        self,
        expand_size: int = 1,
        is_encoder_decoder: bool = False,
        input_ids: torch.LongTensor | None = None,
        **model_kwargs,
    ) -> tuple[torch.LongTensor, dict[str, Any]]:
        # Overwritten -- Support for expanding tensors without a batch size dimension
        # e.g., pixel_values, image_grid_thw, pixel_values_videos, video_grid_thw, second_per_grid_t
        # pixel_values.shape[0] is sum(seqlen_images for samples)
        # image_grid_thw.shape[0] is sum(num_images for samples)

        if expand_size == 1:
            return input_ids, model_kwargs

        visual_keys = ["pixel_values", "image_grid_thw", "pixel_values_videos", "video_grid_thw", "second_per_grid_ts"]

        def _expand_dict_for_generation_visual(dict_to_expand):
            image_grid_thw = model_kwargs.get("image_grid_thw", None)
            video_grid_thw = model_kwargs.get("video_grid_thw", None)
            image_nums, video_nums = self._get_image_nums_and_video_nums(
                input_ids, inputs_embeds=model_kwargs.get("inputs_embeds", None)
            )

            def _repeat_interleave_samples(x, lengths, repeat_times):
                samples = torch.split(x, lengths)
                repeat_args = [repeat_times] + [1] * (x.dim() - 1)
                result = torch.cat([sample.repeat(*repeat_args) for sample in samples], dim=0)
                return result

            for key in dict_to_expand:
                if key == "pixel_values":
                    # split images into samples
                    samples = torch.split(image_grid_thw, list(image_nums))
                    # compute the sequence length of images for each sample
                    lengths = [torch.prod(sample, dim=1).sum() for sample in samples]
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
                    )
                elif key == "image_grid_thw":
                    # get the num of images for each sample
                    lengths = list(image_nums)
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
                    )
                elif key == "pixel_values_videos":
                    samples = torch.split(video_grid_thw, list(video_nums))
                    lengths = [torch.prod(sample, dim=1).sum() for sample in samples]
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
                    )
                elif key == "video_grid_thw":
                    lengths = list(video_nums)
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=lengths, repeat_times=expand_size
                    )
                elif key == "second_per_grid_ts":
                    dict_to_expand[key] = _repeat_interleave_samples(
                        dict_to_expand[key], lengths=list(video_nums), repeat_times=expand_size
                    )
            return dict_to_expand

        def _expand_dict_for_generation(dict_to_expand):
            for key in dict_to_expand:
                if key == "position_ids" and dict_to_expand[key].ndim == 3:
                    dict_to_expand[key] = dict_to_expand[key].repeat_interleave(expand_size, dim=1)
                elif (
                    dict_to_expand[key] is not None
                    and isinstance(dict_to_expand[key], torch.Tensor)
                    and key not in visual_keys
                ):
                    dict_to_expand[key] = dict_to_expand[key].repeat_interleave(expand_size, dim=0)
            return dict_to_expand

        model_kwargs = _expand_dict_for_generation_visual(model_kwargs)

        if input_ids is not None:
            input_ids = input_ids.repeat_interleave(expand_size, dim=0)

        model_kwargs = _expand_dict_for_generation(model_kwargs)

        if is_encoder_decoder:
            if model_kwargs.get("encoder_outputs") is None:
                raise ValueError("If `is_encoder_decoder` is True, make sure that `encoder_outputs` is defined.")
            model_kwargs["encoder_outputs"] = _expand_dict_for_generation(model_kwargs["encoder_outputs"])

        return input_ids, model_kwargs


__all__ = ["Glm46VModel", "Glm46VPreTrainedModel", "Glm46VForConditionalGeneration"]