qwen2_5_vl_custom ¶

def forward(
    self,
    hidden_states: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.LongTensor] = None,
    past_key_value: Optional[Tuple[torch.Tensor]] = None,
    output_attentions: Optional[bool] = False,
    use_cache: Optional[bool] = False,
    cache_position: Optional[torch.LongTensor] = None,
    position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
    cp_mesh: DeviceMesh | None = None,
    **kwargs,
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
    """
    Args:
        hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
        attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
            `(batch, sequence_length)` where padding elements are indicated by 0.
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under
            returned tensors for more detail.
        use_cache (`bool`, *optional*):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
            (see `past_key_values`).
        past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
            Indices depicting the position of the input sequence tokens in the sequence.
        position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
            Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
            with `head_dim` being the embedding dimension of each attention head.
        cp_mesh (DeviceMesh, optional): Device mesh over which context parallelism is done.
        kwargs (`dict`, *optional*):
            Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
            into the model
    """

    residual = hidden_states

    hidden_states = self.input_layernorm(hidden_states)
    hidden_states, self_attn_weights, present_key_value = self.self_attn(
        hidden_states,
        attention_mask,
        position_ids,
        past_key_value,
        output_attentions,
        use_cache,
        cache_position,
        position_embeddings,
        cp_mesh=cp_mesh,
    )
    hidden_states = residual + hidden_states

    residual = hidden_states
    hidden_states = self.post_attention_layernorm(hidden_states)
    hidden_states = self.mlp(hidden_states)
    hidden_states = residual + hidden_states

    outputs = (hidden_states,)

    if output_attentions:
        outputs += (self_attn_weights,)

    if use_cache:
        outputs += (present_key_value,)
    return outputs

def forward(
    self,
    hidden_states: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.LongTensor] = None,
    past_key_value: Optional[Cache] = None,
    output_attentions: bool = False,
    use_cache: bool = False,
    cache_position: Optional[torch.LongTensor] = None,
    position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
    cp_mesh: DeviceMesh | None = None,
):
    """
    Args:
        cp_mesh (DeviceMesh, optional): Device mesh over which context parallelism is done.
    """
    bsz, q_len, _ = hidden_states.size()

    query_states = self.q_proj(hidden_states)
    key_states = self.k_proj(hidden_states)
    value_states = self.v_proj(hidden_states)
    query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
    key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
    value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)

    # Because the input can be padded, the absolute sequence length depends on the max position id.
    cos, sin = position_embeddings
    query_states, key_states = apply_multimodal_rotary_pos_emb(
        query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
    )

    if past_key_value is not None:
        cache_kwargs = {
            "sin": sin,
            "cos": cos,
            "cache_position": cache_position,
        }  # Specific to RoPE models
        key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

    # repeat k/v heads if n_kv_heads < n_heads
    key_states = repeat_kv(key_states, self.num_key_value_groups)
    value_states = repeat_kv(value_states, self.num_key_value_groups)
    dropout_rate = 0.0 if not self.training else self.attention_dropout

    # In PEFT, usually we cast the layer norms in float32 for training stability reasons
    # therefore the input hidden states gets silently casted in float32. Hence, we need
    # cast them back in float16 just to be sure everything works as expected.
    input_dtype = query_states.dtype
    if input_dtype == torch.float32:
        if torch.is_autocast_enabled():
            target_dtype = torch.get_autocast_gpu_dtype()
        # Handle the case where the model is quantized
        elif hasattr(self.config, "_pre_quantization_dtype"):
            target_dtype = self.config._pre_quantization_dtype
        else:
            target_dtype = self.q_proj.weight.dtype

        logger.warning_once(
            f"The input hidden states seems to be silently casted in float32, this might be related to"
            f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
            f" {target_dtype}."
        )

        query_states = query_states.to(target_dtype)
        key_states = key_states.to(target_dtype)
        value_states = value_states.to(target_dtype)

    # Reashape to the expected shape for Flash Attention
    query_states = query_states.transpose(1, 2)
    key_states = key_states.transpose(1, 2)
    value_states = value_states.transpose(1, 2)

    if (
        self.config.use_sliding_window
        and getattr(self.config, "sliding_window", None) is not None
        and self.layer_idx >= self.config.max_window_layers
    ):
        sliding_window = self.config.sliding_window
    else:
        sliding_window = None
    if cp_mesh is not None:
        raise NotImplementedError(
            "CP is not supported for flash attention2, _flash_attention_forward will produce wrong output if query_states is sharded"
        )

    attn_output = _flash_attention_forward(
        query_states,
        key_states,
        value_states,
        attention_mask,
        q_len,
        dropout=dropout_rate,
        sliding_window=sliding_window,
        is_causal=self.is_causal,
        use_top_left_mask=self._flash_attn_uses_top_left_mask,
    )

    attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
    attn_output = self.o_proj(attn_output)

    if not output_attentions:
        attn_weights = None

    return attn_output, attn_weights, past_key_value

def forward(
    self,
    input_ids: torch.LongTensor = None,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.LongTensor] = None,
    past_key_values: Optional[List[torch.FloatTensor]] = None,
    inputs_embeds: Optional[torch.FloatTensor] = None,
    labels: Optional[torch.LongTensor] = None,
    use_cache: Optional[bool] = None,
    output_attentions: Optional[bool] = None,
    output_hidden_states: Optional[bool] = None,
    return_dict: Optional[bool] = None,
    pixel_values: Optional[torch.Tensor] = None,
    pixel_values_videos: Optional[torch.FloatTensor] = None,
    image_grid_thw: Optional[torch.LongTensor] = None,
    video_grid_thw: Optional[torch.LongTensor] = None,
    rope_deltas: Optional[torch.LongTensor] = None,
    cache_position: Optional[torch.LongTensor] = None,
    second_per_grid_ts: Optional[torch.Tensor] = None,
) -> Union[Tuple, Qwen2_5_VLCausalLMOutputWithPast]:
    r"""
    Args:
        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]`.

    Returns:

    Example:

    ```python
    >>> from PIL import Image
    >>> import requests
    >>> from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

    >>> model = Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
    >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")

    >>> messages = [
        {
            "role": "user",
            "content": [
                {"type": "image"},
                {"type": "text", "text": "What is shown in this image?"},
            ],
        },
    ]
    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
    >>> image = Image.open(requests.get(url, stream=True).raw)

    >>> 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 ..."
    ```"""

    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
    output_hidden_states = (
        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
    )
    return_dict = return_dict if return_dict is not None else self.config.use_return_dict

    if inputs_embeds is None:
        inputs_embeds = self.model.embed_tokens(input_ids)
        # This is a trick to handle TP for LLM but no TP for vision encoder, we need to convert DTensor to regular tensor later
        is_inputs_embeds_dtensor = isinstance(inputs_embeds, DTensor)  # This is True for TP>1, False for TP=1
        if is_inputs_embeds_dtensor:
            target_device_mesh = inputs_embeds.device_mesh
            target_placements = inputs_embeds.placements
            inputs_embeds = inputs_embeds.full_tensor()

        if pixel_values is not None:
            pixel_values = pixel_values.type(self.visual.dtype)
            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
            n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
            n_image_features = image_embeds.shape[0]
            if n_image_tokens != n_image_features:
                raise ValueError(
                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
                )

            mask = input_ids == self.config.image_token_id
            mask_unsqueezed = mask.unsqueeze(-1)
            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
            image_mask = mask_expanded.to(inputs_embeds.device)

            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)

        if pixel_values_videos is not None:
            pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
            video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
            n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
            n_video_features = video_embeds.shape[0]
            if n_video_tokens != n_video_features:
                raise ValueError(
                    f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
                )

            mask = input_ids == self.config.video_token_id
            mask_unsqueezed = mask.unsqueeze(-1)
            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
            video_mask = mask_expanded.to(inputs_embeds.device)

            video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
        if is_inputs_embeds_dtensor:
            inputs_embeds = (
                DTensor.from_local(inputs_embeds, device_mesh=target_device_mesh)
                .redistribute(placements=target_placements)
                .to_local()
            )
        if attention_mask is not None:
            attention_mask = attention_mask.to(inputs_embeds.device)

    # if we get 4D attention mask we cannot calculate rope deltas anymore.
    if position_ids is None and (attention_mask is None or attention_mask.ndim == 2):
        # calculate RoPE index once per generation in the pre-fill stage only
        if (
            (cache_position is not None and cache_position[0] == 0)
            or self.rope_deltas is None
            or (past_key_values is None or past_key_values.get_seq_length() == 0)
        ):
            position_ids, rope_deltas = self.get_rope_index(
                input_ids,
                image_grid_thw,
                video_grid_thw,
                second_per_grid_ts,
                attention_mask,
            )
            self.rope_deltas = rope_deltas
        # then use the prev pre-calculated rope-deltas to get the correct position ids
        else:
            batch_size, seq_length, _ = inputs_embeds.shape
            delta = (
                (cache_position[0] + self.rope_deltas).to(inputs_embeds.device) if cache_position is not None else 0
            )
            position_ids = torch.arange(seq_length, device=inputs_embeds.device)
            position_ids = position_ids.view(1, -1).expand(batch_size, -1)
            if cache_position is not None:  # otherwise `deltas` is an int `0`
                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
            position_ids = position_ids.add(delta)
            position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)

    outputs = self.model(
        input_ids=None,
        position_ids=position_ids,
        attention_mask=attention_mask,
        past_key_values=past_key_values,
        inputs_embeds=inputs_embeds,
        use_cache=use_cache,
        output_attentions=output_attentions,
        output_hidden_states=output_hidden_states,
        return_dict=return_dict,
        cache_position=cache_position,
    )

    hidden_states = outputs[0]
    logits = self.lm_head(hidden_states)

    loss = None
    if labels is not None:
        # Upcast to float if we need to compute the loss to avoid potential precision issues
        logits = logits.float()
        # Shift so that tokens < n predict n
        shift_logits = logits[..., :-1, :].contiguous()
        shift_labels = labels[..., 1:].contiguous()
        # Flatten the tokens
        loss_fct = CrossEntropyLoss()
        shift_logits = shift_logits.view(-1, self.config.vocab_size)
        shift_labels = shift_labels.view(-1)
        # Enable model parallelism
        shift_labels = shift_labels.to(shift_logits.device)
        loss = loss_fct(shift_logits, shift_labels)

    if not return_dict:
        output = (logits,)  # + outputs[1:]
        return (loss,) + output if loss is not None else output

    return Qwen2_5_VLCausalLMOutputWithPast(
        loss=loss,
        logits=logits,
        past_key_values=outputs.past_key_values,
        hidden_states=outputs.hidden_states,
        attentions=outputs.attentions,
        rope_deltas=self.rope_deltas,
    )

def get_rope_index(
    self,
    input_ids: Optional[torch.LongTensor] = None,
    image_grid_thw: Optional[torch.LongTensor] = None,
    video_grid_thw: Optional[torch.LongTensor] = None,
    second_per_grid_ts: Optional[torch.Tensor] = None,
    attention_mask: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
    """
    Calculate the 3D rope index based on image and video's temporal, height and width in LLM.

    Explanation:
        Each embedding sequence contains vision embedding and text embedding or just contains text embedding.

        For pure text embedding sequence, the rotary position embedding has no difference with modern LLMs.
        Examples:
            input_ids: [T T T T T], here T is for text.
            temporal position_ids: [0, 1, 2, 3, 4]
            height position_ids: [0, 1, 2, 3, 4]
            width position_ids: [0, 1, 2, 3, 4]

        For vision and text embedding sequence, we calculate 3D rotary position embedding for vision part
        and 1D rotary position embeddin for text part.
        Examples:
            Temporal (Time): 3 patches, representing different segments of the video in time.
            Height: 2 patches, dividing each frame vertically.
            Width: 2 patches, dividing each frame horizontally.
            We also have some important parameters:
            fps (Frames Per Second): The video's frame rate, set to 1. This means one frame is processed each second.
            tokens_per_second: This is a crucial parameter. It dictates how many "time-steps" or "temporal tokens" are conceptually packed into a one-second interval of the video. In this case, we have 25 tokens per second. So each second of the video will be represented with 25 separate time points. It essentially defines the temporal granularity.
            temporal_patch_size: The number of frames that compose one temporal patch. Here, it's 2 frames.
            interval: The step size for the temporal position IDs, calculated as tokens_per_second * temporal_patch_size / fps. In this case, 25 * 2 / 1 = 50. This means that each temporal patch will be have a difference of 50 in the temporal position IDs.
            input_ids: [V V V V V V V V V V V V T T T T T], here V is for vision.
            vision temporal position_ids: [0, 0, 0, 0, 50, 50, 50, 50, 100, 100, 100, 100]
            vision height position_ids: [0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1]
            vision width position_ids: [0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1]
            text temporal position_ids: [101, 102, 103, 104, 105]
            text height position_ids: [101, 102, 103, 104, 105]
            text width position_ids: [101, 102, 103, 104, 105]
            Here we calculate the text start position_ids as the max vision position_ids plus 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.
        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.
        second_per_grid_ts (`torch.Tensor` of shape `(num_videos)`, *optional*):
            The time interval (in seconds) for each grid along the temporal dimension in the 3D position IDs.
        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
    image_token_id = self.config.image_token_id
    video_token_id = self.config.video_token_id
    vision_start_token_id = self.config.vision_start_token_id
    mrope_position_deltas = []
    if input_ids is not None and (image_grid_thw is not None or video_grid_thw is not None):
        total_input_ids = input_ids
        if attention_mask is None:
            attention_mask = torch.ones_like(total_input_ids)
        position_ids = torch.ones(
            3,
            input_ids.shape[0],
            input_ids.shape[1],
            dtype=input_ids.dtype,
            device=input_ids.device,
        )
        image_index, video_index = 0, 0
        attention_mask = attention_mask.to(total_input_ids.device)
        for i, input_ids in enumerate(total_input_ids):
            input_ids = input_ids[attention_mask[i] == 1]
            image_nums, video_nums = 0, 0
            vision_start_indices = torch.argwhere(input_ids == vision_start_token_id).squeeze(1)
            vision_tokens = input_ids[vision_start_indices + 1]
            image_nums = (vision_tokens == image_token_id).sum()
            video_nums = (vision_tokens == video_token_id).sum()
            input_tokens = input_ids.tolist()
            llm_pos_ids_list: list = []
            st = 0
            remain_images, remain_videos = image_nums, video_nums
            for _ in range(image_nums + video_nums):
                if image_token_id in input_tokens and remain_images > 0:
                    ed_image = input_tokens.index(image_token_id, st)
                else:
                    ed_image = len(input_tokens) + 1
                if video_token_id in input_tokens and remain_videos > 0:
                    ed_video = input_tokens.index(video_token_id, st)
                else:
                    ed_video = len(input_tokens) + 1
                if ed_image < ed_video:
                    t, h, w = (
                        image_grid_thw[image_index][0],
                        image_grid_thw[image_index][1],
                        image_grid_thw[image_index][2],
                    )
                    second_per_grid_t = 0
                    image_index += 1
                    remain_images -= 1
                    ed = ed_image

                else:
                    t, h, w = (
                        video_grid_thw[video_index][0],
                        video_grid_thw[video_index][1],
                        video_grid_thw[video_index][2],
                    )
                    second_per_grid_t = second_per_grid_ts[video_index] if second_per_grid_ts is not None else 1.0
                    video_index += 1
                    remain_videos -= 1
                    ed = ed_video
                llm_grid_t, llm_grid_h, llm_grid_w = (
                    t.item(),
                    h.item() // spatial_merge_size,
                    w.item() // spatial_merge_size,
                )
                text_len = ed - st

                st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                llm_pos_ids_list.append(
                    torch.arange(text_len, device=input_ids.device).view(1, -1).expand(3, -1) + st_idx
                )

                range_tensor = torch.arange(llm_grid_t).view(-1, 1)
                expanded_range = range_tensor.expand(-1, llm_grid_h * llm_grid_w)

                time_tensor = expanded_range * second_per_grid_t * self.config.vision_config.tokens_per_second

                time_tensor_long = time_tensor.long()
                t_index = time_tensor_long.flatten()

                h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten()
                w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten()
                llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + text_len + st_idx)
                st = ed + llm_grid_t * llm_grid_h * llm_grid_w

            if st < len(input_tokens):
                st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
                text_len = len(input_tokens) - st
                llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)

            llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
            position_ids[..., i, attention_mask[i] == 1] = llm_positions.to(position_ids.device)
            mrope_position_deltas.append(llm_positions.max() + 1 - len(total_input_ids[i]))
        mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1)
        return position_ids, mrope_position_deltas
    else:
        if attention_mask is not None:
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            position_ids = position_ids.unsqueeze(0).expand(3, -1, -1).to(attention_mask.device)
            max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0]
            mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
        else:
            position_ids = (
                torch.arange(input_ids.shape[1], device=input_ids.device)
                .view(1, 1, -1)
                .expand(3, input_ids.shape[0], -1)
            )
            mrope_position_deltas = torch.zeros(
                [input_ids.shape[0], 1],
                device=input_ids.device,
                dtype=input_ids.dtype,
            )

        return position_ids, mrope_position_deltas

def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor) -> torch.Tensor:
    """
    Args:
        hidden_states (`torch.Tensor` of shape `(batch_size, seq_len, hidden_size)`):
            The final hidden states of the model.
        grid_thw (`torch.Tensor` of shape `(num_images_or_videos, 3)`):
            The temporal, height and width of feature shape of each image in LLM.

    Returns:
        `torch.Tensor`: hidden_states.
    """
    hidden_states = self.patch_embed(hidden_states)
    rotary_pos_emb = self.rot_pos_emb(grid_thw)
    window_index, cu_window_seqlens = self.get_window_index(grid_thw)
    cu_window_seqlens = torch.tensor(
        cu_window_seqlens,
        device=hidden_states.device,
        dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
    )
    cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)

    seq_len, _ = hidden_states.size()
    hidden_states = hidden_states.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
    hidden_states = hidden_states[window_index, :, :]
    hidden_states = hidden_states.reshape(seq_len, -1)
    rotary_pos_emb = rotary_pos_emb.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
    rotary_pos_emb = rotary_pos_emb[window_index, :, :]
    rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)

    cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
        dim=0,
        # Select dtype based on the following factors:
        #  - FA2 requires that cu_seqlens_q must have dtype int32
        #  - torch.onnx.export requires that cu_seqlens_q must have same dtype as grid_thw
        # See https://github.com/huggingface/transformers/pull/34852 for more information
        dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
    )
    cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)

    for layer_num, blk in enumerate(self.blocks):
        cu_seqlens_now = cu_seqlens if layer_num in self.fullatt_block_indexes else cu_window_seqlens
        if self.gradient_checkpointing and self.training:
            hidden_states = self._gradient_checkpointing_func(
                blk.__call__, hidden_states, cu_seqlens_now, rotary_pos_emb
            )
        else:
            hidden_states = blk(
                hidden_states,
                cu_seqlens=cu_seqlens_now,
                rotary_pos_emb=rotary_pos_emb,
            )

    hidden_states = self.merger(hidden_states)
    reverse_indices = torch.argsort(window_index)
    hidden_states = hidden_states[reverse_indices, :]

    return hidden_states