#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/informer/modular_informer.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_informer.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# coding=utf-8
# Copyright 2023 Amazon and The HuggingFace Inc. 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.

from typing import Callable, Optional, Union

import numpy as np
import torch
from torch import nn

from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache
from ...modeling_attn_mask_utils import (
    _prepare_4d_attention_mask,
    _prepare_4d_attention_mask_for_sdpa,
    _prepare_4d_causal_attention_mask,
    _prepare_4d_causal_attention_mask_for_sdpa,
)
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPastAndCrossAttentions,
    SampleTSPredictionOutput,
    Seq2SeqTSModelOutput,
    Seq2SeqTSPredictionOutput,
)
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...time_series_utils import NegativeBinomialOutput, NormalOutput, StudentTOutput
from ...utils import auto_docstring, is_torch_flex_attn_available, logging
from ...utils.deprecation import deprecate_kwarg
from .configuration_informer import InformerConfig


if is_torch_flex_attn_available():
    from ...integrations.flex_attention import make_flex_block_causal_mask


logger = logging.get_logger(__name__)


class InformerFeatureEmbedder(nn.Module):
    """
    Embed a sequence of categorical features.

    Args:
        cardinalities (`list[int]`):
            List of cardinalities of the categorical features.
        embedding_dims (`list[int]`):
            List of embedding dimensions of the categorical features.
    """

    def __init__(self, cardinalities: list[int], embedding_dims: list[int]) -> None:
        super().__init__()

        self.num_features = len(cardinalities)
        self.embedders = nn.ModuleList([nn.Embedding(c, d) for c, d in zip(cardinalities, embedding_dims)])

    def forward(self, features: torch.Tensor) -> torch.Tensor:
        if self.num_features > 1:
            # we slice the last dimension, giving an array of length
            # self.num_features with shape (N,T) or (N)
            cat_feature_slices = torch.chunk(features, self.num_features, dim=-1)
        else:
            cat_feature_slices = [features]

        return torch.cat(
            [
                embed(cat_feature_slice.squeeze(-1))
                for embed, cat_feature_slice in zip(self.embedders, cat_feature_slices)
            ],
            dim=-1,
        )


class InformerStdScaler(nn.Module):
    """
    Standardize features by calculating the mean and scaling along the first dimension, and then normalizes it by
    subtracting from the mean and dividing by the standard deviation.
    """

    def __init__(self, config: InformerConfig):
        super().__init__()
        self.dim = config.scaling_dim if hasattr(config, "scaling_dim") else 1
        self.keepdim = config.keepdim if hasattr(config, "keepdim") else True
        self.minimum_scale = config.minimum_scale if hasattr(config, "minimum_scale") else 1e-5

    def forward(
        self, data: torch.Tensor, observed_indicator: torch.Tensor
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        Parameters:
            data (`torch.Tensor` of shape `(batch_size, sequence_length, num_input_channels)`):
                input for Batch norm calculation
            observed_indicator (`torch.BoolTensor` of shape `(batch_size, sequence_length, num_input_channels)`):
                Calculating the scale on the observed indicator.
        Returns:
            tuple of `torch.Tensor` of shapes
                (`(batch_size, sequence_length, num_input_channels)`,`(batch_size, 1, num_input_channels)`,
                `(batch_size, 1, num_input_channels)`)
        """
        denominator = observed_indicator.sum(self.dim, keepdim=self.keepdim)
        denominator = denominator.clamp_min(1.0)
        loc = (data * observed_indicator).sum(self.dim, keepdim=self.keepdim) / denominator

        variance = (((data - loc) * observed_indicator) ** 2).sum(self.dim, keepdim=self.keepdim) / denominator
        scale = torch.sqrt(variance + self.minimum_scale)
        return (data - loc) / scale, loc, scale


class InformerMeanScaler(nn.Module):
    """
    Computes a scaling factor as the weighted average absolute value along the first dimension, and scales the data
    accordingly.
    """

    def __init__(self, config: InformerConfig):
        super().__init__()
        self.dim = config.scaling_dim if hasattr(config, "scaling_dim") else 1
        self.keepdim = config.keepdim if hasattr(config, "keepdim") else True
        self.minimum_scale = config.minimum_scale if hasattr(config, "minimum_scale") else 1e-10
        self.default_scale = config.default_scale if hasattr(config, "default_scale") else None

    def forward(
        self, data: torch.Tensor, observed_indicator: torch.Tensor
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        Parameters:
            data (`torch.Tensor` of shape `(batch_size, sequence_length, num_input_channels)`):
                input for Batch norm calculation
            observed_indicator (`torch.BoolTensor` of shape `(batch_size, sequence_length, num_input_channels)`):
                Calculating the scale on the observed indicator.
        Returns:
            tuple of `torch.Tensor` of shapes
                (`(batch_size, sequence_length, num_input_channels)`,`(batch_size, 1, num_input_channels)`,
                `(batch_size, 1, num_input_channels)`)
        """
        ts_sum = (data * observed_indicator).abs().sum(self.dim, keepdim=True)
        num_observed = observed_indicator.sum(self.dim, keepdim=True)

        scale = ts_sum / torch.clamp(num_observed, min=1)

        # If `default_scale` is provided, we use it, otherwise we use the scale
        # of the batch.
        if self.default_scale is None:
            batch_sum = ts_sum.sum(dim=0)
            batch_observations = torch.clamp(num_observed.sum(0), min=1)
            default_scale = torch.squeeze(batch_sum / batch_observations)
        else:
            default_scale = self.default_scale * torch.ones_like(scale)

        # apply default scale where there are no observations
        scale = torch.where(num_observed > 0, scale, default_scale)

        # ensure the scale is at least `self.minimum_scale`
        scale = torch.clamp(scale, min=self.minimum_scale)
        scaled_data = data / scale

        if not self.keepdim:
            scale = scale.squeeze(dim=self.dim)

        return scaled_data, torch.zeros_like(scale), scale


class InformerNOPScaler(nn.Module):
    """
    Assigns a scaling factor equal to 1 along the first dimension, and therefore applies no scaling to the input data.
    """

    def __init__(self, config: InformerConfig):
        super().__init__()
        self.dim = config.scaling_dim if hasattr(config, "scaling_dim") else 1
        self.keepdim = config.keepdim if hasattr(config, "keepdim") else True

    def forward(
        self, data: torch.Tensor, observed_indicator: Optional[torch.Tensor] = None
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        Parameters:
            data (`torch.Tensor` of shape `(batch_size, sequence_length, num_input_channels)`):
                input for Batch norm calculation
        Returns:
            tuple of `torch.Tensor` of shapes
                (`(batch_size, sequence_length, num_input_channels)`,`(batch_size, 1, num_input_channels)`,
                `(batch_size, 1, num_input_channels)`)
        """
        scale = torch.ones_like(data, requires_grad=False).mean(dim=self.dim, keepdim=self.keepdim)
        loc = torch.zeros_like(data, requires_grad=False).mean(dim=self.dim, keepdim=self.keepdim)
        return data, loc, scale


class InformerSinusoidalPositionalEmbedding(nn.Embedding):
    """This module produces sinusoidal positional embeddings of any length."""

    def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None) -> None:
        super().__init__(num_positions, embedding_dim)

    def _init_weight(self):
        """
        Identical to the XLM create_sinusoidal_embeddings except features are not interleaved. The cos features are in
        the 2nd half of the vector. [dim // 2:]
        """
        n_pos, dim = self.weight.shape
        position_enc = np.array(
            [[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)]
        )
        out = torch.empty(n_pos, dim, dtype=self.weight.dtype, requires_grad=False)
        sentinel = dim // 2 if dim % 2 == 0 else (dim // 2) + 1
        out[:, 0:sentinel] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
        out[:, sentinel:] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
        self.weight = nn.Parameter(out, requires_grad=False)

    @torch.no_grad()
    def forward(
        self, input_ids_shape: torch.Size, past_key_values_length: int = 0, position_ids: Optional[torch.Tensor] = None
    ) -> torch.Tensor:
        """`input_ids_shape` is expected to be [bsz x seqlen]."""
        if position_ids is None:
            bsz, seq_len = input_ids_shape[:2]
            position_ids = torch.arange(
                past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
            )
        return super().forward(position_ids)


class InformerValueEmbedding(nn.Module):
    def __init__(self, feature_size, d_model):
        super().__init__()
        self.value_projection = nn.Linear(in_features=feature_size, out_features=d_model, bias=False)

    def forward(self, x):
        return self.value_projection(x)


@auto_docstring
class InformerPreTrainedModel(PreTrainedModel):
    config: InformerConfig
    base_model_prefix = "model"
    main_input_name = "past_values"
    supports_gradient_checkpointing = True

    def _init_weights(self, module: nn.Module):
        super()._init_weights(module)
        if isinstance(module, InformerSinusoidalPositionalEmbedding):
            module._init_weight()

    # Copied from transformers.models.bart.modeling_bart.BartPreTrainedModel._update_full_mask
    def _update_full_mask(
        self,
        attention_mask: Union[torch.Tensor, None],
        inputs_embeds: torch.Tensor,
    ):
        if attention_mask is not None:
            if self.config._attn_implementation == "flash_attention_2":
                attention_mask = attention_mask if 0 in attention_mask else None
            elif self.config._attn_implementation == "sdpa":
                # output_attentions=True & head_mask can not be supported when using SDPA, fall back to
                # the manual implementation that requires a 4D causal mask in all cases.
                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
                attention_mask = _prepare_4d_attention_mask_for_sdpa(attention_mask, inputs_embeds.dtype)
            elif self.config._attn_implementation == "flex_attention":
                if isinstance(attention_mask, torch.Tensor):
                    attention_mask = make_flex_block_causal_mask(attention_mask, is_causal=False)
            else:
                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
                attention_mask = _prepare_4d_attention_mask(attention_mask, inputs_embeds.dtype)

        return attention_mask

    # Copied from transformers.models.musicgen.modeling_musicgen.MusicgenDecoder._update_causal_mask
    def _update_causal_mask(
        self,
        attention_mask: Union[torch.Tensor, None],
        input_shape: torch.Size,
        inputs_embeds: torch.Tensor,
        past_key_values_length: int,
    ):
        if self.config._attn_implementation == "flash_attention_2":
            # 2d mask is passed through the layers
            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
        elif self.config._attn_implementation == "sdpa":
            # output_attentions=True & cross_attn_head_mask can not be supported when using SDPA, and we fall back on
            # the manual implementation that requires a 4D causal mask in all cases.
            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
                attention_mask,
                input_shape,
                inputs_embeds,
                past_key_values_length,
            )
        elif self.config._attn_implementation == "flex_attention":
            if isinstance(attention_mask, torch.Tensor):
                attention_mask = make_flex_block_causal_mask(attention_mask)
            # Other attention flavors support in-built causal (when `mask is None`)
            # while we need to create our specific block mask regardless
            elif attention_mask is None:
                attention_mask = make_flex_block_causal_mask(
                    torch.ones(
                        size=(input_shape),
                        device=inputs_embeds.device,
                    )
                )
        else:
            # 4d mask is passed through the layers
            attention_mask = _prepare_4d_causal_attention_mask(
                attention_mask, input_shape, inputs_embeds, past_key_values_length
            )

        return attention_mask

    # Copied from transformers.models.musicgen.modeling_musicgen.MusicgenDecoder._update_cross_attn_mask
    def _update_cross_attn_mask(
        self,
        encoder_hidden_states: Union[torch.Tensor, None],
        encoder_attention_mask: Union[torch.Tensor, None],
        input_shape: torch.Size,
        inputs_embeds: torch.Tensor,
    ):
        # expand encoder attention mask
        if encoder_hidden_states is not None and encoder_attention_mask is not None:
            if self.config._attn_implementation == "flash_attention_2":
                encoder_attention_mask = encoder_attention_mask if 0 in encoder_attention_mask else None
            elif self.config._attn_implementation == "sdpa":
                # output_attentions=True & cross_attn_head_mask can not be supported when using SDPA, and we fall back on
                # the manual implementation that requires a 4D causal mask in all cases.
                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
                encoder_attention_mask = _prepare_4d_attention_mask_for_sdpa(
                    encoder_attention_mask,
                    inputs_embeds.dtype,
                    tgt_len=input_shape[-1],
                )
            elif self.config._attn_implementation == "flex_attention":
                if isinstance(encoder_attention_mask, torch.Tensor):
                    encoder_attention_mask = make_flex_block_causal_mask(
                        encoder_attention_mask,
                        query_length=input_shape[-1],
                        is_causal=False,
                    )
            else:
                # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
                encoder_attention_mask = _prepare_4d_attention_mask(
                    encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
                )

        return encoder_attention_mask


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: Optional[float] = None,
    dropout: float = 0.0,
    head_mask: Optional[torch.Tensor] = None,
    **kwargs,
):
    if scaling is None:
        scaling = query.size(-1) ** -0.5

    attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling
    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1)

    if head_mask is not None:
        attn_weights = attn_weights * head_mask.view(1, -1, 1, 1)

    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


class InformerAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
        is_causal: bool = False,
        config: Optional[InformerConfig] = None,
        layer_idx: Optional[int] = None,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads
        self.config = config

        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
                f" and `num_heads`: {num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.is_decoder = is_decoder
        self.is_causal = is_causal
        self.layer_idx = layer_idx
        if layer_idx is None and self.is_decoder:
            logger.warning_once(
                f"Instantiating a decoder {self.__class__.__name__} without passing `layer_idx` is not recommended and "
                "will lead to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_values: Optional[Cache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
        cache_position: Optional[torch.Tensor] = None,
        # TODO: we need a refactor so that the different attention modules can get their specific kwargs
        # ATM, we have mixed things encoder, decoder, and encoder-decoder attn
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None

        # determine input shapes
        bsz, tgt_len = hidden_states.shape[:-1]
        src_len = key_value_states.shape[1] if is_cross_attention else tgt_len

        q_input_shape = (bsz, tgt_len, -1, self.head_dim)
        kv_input_shape = (bsz, src_len, -1, self.head_dim)

        # get query proj
        query_states = self.q_proj(hidden_states).view(*q_input_shape).transpose(1, 2)

        is_updated = False
        if past_key_values is not None:
            if isinstance(past_key_values, EncoderDecoderCache):
                is_updated = past_key_values.is_updated.get(self.layer_idx)
                if is_cross_attention:
                    # after the first generated id, we can subsequently re-use all key/value_states from cache
                    curr_past_key_value = past_key_values.cross_attention_cache
                else:
                    curr_past_key_value = past_key_values.self_attention_cache
            else:
                curr_past_key_value = past_key_values

        current_states = key_value_states if is_cross_attention else hidden_states
        if is_cross_attention and past_key_values is not None and is_updated:
            # reuse k,v, cross_attentions
            key_states = curr_past_key_value.layers[self.layer_idx].keys
            value_states = curr_past_key_value.layers[self.layer_idx].values
        else:
            key_states = self.k_proj(current_states)
            value_states = self.v_proj(current_states)
            key_states = key_states.view(*kv_input_shape).transpose(1, 2)
            value_states = value_states.view(*kv_input_shape).transpose(1, 2)

            if past_key_values is not None:
                # save all key/value_states to cache to be re-used for fast auto-regressive generation
                cache_position = cache_position if not is_cross_attention else None
                key_states, value_states = curr_past_key_value.update(
                    key_states, value_states, self.layer_idx, {"cache_position": cache_position}
                )
                # set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls
                if is_cross_attention and isinstance(past_key_values, EncoderDecoderCache):
                    past_key_values.is_updated[self.layer_idx] = True

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.dropout,
            scaling=self.scaling,
            output_attentions=output_attentions,
            head_mask=layer_head_mask,
            **kwargs,
        )

        attn_output = attn_output.reshape(bsz, tgt_len, -1).contiguous()
        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights


class InformerProbSparseAttention(nn.Module):
    """Probabilistic Attention mechanism to select the "active"
    queries rather than the "lazy" queries and provides a sparse Transformer thus mitigating the quadratic compute and
    memory requirements of vanilla attention"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        sampling_factor: int = 5,
        bias: bool = True,
        layer_idx: Optional[int] = None,
    ):
        super().__init__()
        self.factor = sampling_factor
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads

        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
                f" and `num_heads`: {num_heads})."
            )
        self.scaling = self.head_dim**-0.5
        self.is_decoder = is_decoder
        self.layer_idx = layer_idx

        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)

    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_values: Optional[Cache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        output_attentions: bool = False,
        cache_position: Optional[torch.Tensor] = None,
    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None

        bsz, tgt_len, _ = hidden_states.size()
        src_len = key_value_states.shape[1] if is_cross_attention else tgt_len
        kv_input_shape = (bsz, src_len, -1, self.head_dim)

        # get query proj
        query_states = self.q_proj(hidden_states) * self.scaling

        is_updated = False
        if past_key_values is not None:
            if isinstance(past_key_values, EncoderDecoderCache):
                is_updated = past_key_values.is_updated.get(self.layer_idx)
                if is_cross_attention:
                    # after the first generated id, we can subsequently re-use all key/value_states from cache
                    curr_past_key_value = past_key_values.cross_attention_cache
                else:
                    curr_past_key_value = past_key_values.self_attention_cache
            else:
                curr_past_key_value = past_key_values

        current_states = key_value_states if is_cross_attention else hidden_states
        if is_cross_attention and past_key_values is not None and is_updated:
            # reuse k,v, cross_attentions
            key_states = curr_past_key_value.layers[self.layer_idx].keys
            value_states = curr_past_key_value.layers[self.layer_idx].values
        else:
            key_states = self.k_proj(current_states)
            value_states = self.v_proj(current_states)
            key_states = key_states.view(*kv_input_shape).transpose(1, 2)
            value_states = value_states.view(*kv_input_shape).transpose(1, 2)

            if past_key_values is not None:
                # save all key/value_states to cache to be re-used for fast auto-regressive generation
                cache_position = cache_position if not is_cross_attention else None
                key_states, value_states = curr_past_key_value.update(
                    key_states, value_states, self.layer_idx, {"cache_position": cache_position}
                )
                # set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls
                if is_cross_attention and isinstance(past_key_values, EncoderDecoderCache):
                    past_key_values.is_updated[self.layer_idx] = True

        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
        key_states = key_states.reshape(*proj_shape)
        value_states = value_states.reshape(*proj_shape)

        key_states_time_length = key_states.size(1)  # L_K
        log_key_states_time_length = np.ceil(np.log1p(key_states_time_length)).astype("int").item()  # log_L_K

        query_states_time_length = query_states.size(1)  # L_Q
        log_query_states_time_length = np.ceil(np.log1p(query_states_time_length)).astype("int").item()  # log_L_Q

        u_part = min(self.factor * query_states_time_length * log_key_states_time_length, key_states_time_length)
        u = min(self.factor * log_query_states_time_length, query_states_time_length)

        if key_states_time_length > 0:
            index_sample = torch.randint(0, key_states_time_length, (u_part,))
            k_sample = key_states[:, index_sample, :]
        else:
            k_sample = key_states

        queries_keys_sample = torch.bmm(query_states, k_sample.transpose(1, 2))  # Q_K_sampled

        # find the Top_k query with sparsity measurement
        if u > 0:
            sparsity_measurement = queries_keys_sample.max(dim=-1)[0] - torch.div(
                queries_keys_sample.sum(dim=-1), key_states_time_length
            )  # M
            top_u_sparsity_measurement = sparsity_measurement.topk(u, sorted=False)[1]  # M_top

            # calculate q_reduce: query_states[:, top_u_sparsity_measurement]
            dim_for_slice = torch.arange(query_states.size(0)).unsqueeze(-1)
            q_reduce = query_states[dim_for_slice, top_u_sparsity_measurement]
        else:
            q_reduce = query_states
            top_u_sparsity_measurement = None

        # Use q_reduce to calculate attention weights
        attn_weights = torch.bmm(q_reduce, key_states.transpose(1, 2))

        src_len = key_states.size(1)
        if attn_weights.size() != (bsz * self.num_heads, u, src_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz * self.num_heads, u, src_len)}, but is"
                f" {attn_weights.size()}"
            )

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
                )
            prob_mask = attention_mask.expand(bsz, self.num_heads, tgt_len, src_len).reshape(
                bsz * self.num_heads, tgt_len, src_len
            )

            if top_u_sparsity_measurement is not None:
                dim_for_slice = torch.arange(prob_mask.size(0)).unsqueeze(-1)
                prob_mask = prob_mask[dim_for_slice, top_u_sparsity_measurement, :]

            attn_weights = attn_weights.view(bsz, self.num_heads, u, src_len) + prob_mask.view(
                bsz, self.num_heads, u, src_len
            )
            attn_weights = attn_weights.view(bsz * self.num_heads, u, src_len)

        attn_weights = nn.functional.softmax(attn_weights, dim=-1)

        if layer_head_mask is not None:
            if layer_head_mask.size() != (self.num_heads,):
                raise ValueError(
                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
                    f" {layer_head_mask.size()}"
                )
            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, u, src_len)
            attn_weights = attn_weights.view(bsz * self.num_heads, u, src_len)

        if output_attentions:
            # this operation is a bit awkward, but it's required to
            # make sure that attn_weights keeps its gradient.
            # In order to do so, attn_weights have to be reshaped
            # twice and have to be reused in the following
            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, u, src_len)
            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, u, src_len)
        else:
            attn_weights_reshaped = None

        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
        attn_output = torch.bmm(attn_probs, value_states)

        # calculate context for updating the attn_output, based on:
        # https://github.com/zhouhaoyi/Informer2020/blob/ac59c7447135473fb2aafeafe94395f884d5c7a5/models/attn.py#L74
        if self.is_decoder:
            # cast to float32 before operation to avoid overflow
            context = value_states.cumsum(dim=-2, dtype=torch.float32).to(value_states.dtype)
        else:
            v_mean_dim_time = value_states.mean(dim=-2)
            context = (
                v_mean_dim_time.unsqueeze(dim=1)
                .expand(bsz * self.num_heads, query_states_time_length, v_mean_dim_time.size(-1))
                .clone()
            )

        if top_u_sparsity_measurement is not None:
            # update context: copy the attention output to the context at top_u_sparsity_measurement index
            dim_for_slice = torch.arange(context.size(0)).unsqueeze(-1)
            context[dim_for_slice, top_u_sparsity_measurement, :] = attn_output
            attn_output = context

        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz * self.num_heads, tgt_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
        attn_output = attn_output.transpose(1, 2)

        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
        # partitioned across GPUs when using tensor-parallelism.
        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights_reshaped


# source: https://github.com/zhouhaoyi/Informer2020/blob/main/models/encoder.py
class InformerConvLayer(GradientCheckpointingLayer):
    def __init__(self, c_in):
        super().__init__()
        self.downConv = nn.Conv1d(
            in_channels=c_in,
            out_channels=c_in,
            kernel_size=3,
            padding=1,
            padding_mode="circular",
        )
        self.norm = nn.BatchNorm1d(c_in)
        self.activation = nn.ELU()
        self.maxPool = nn.MaxPool1d(kernel_size=3, stride=2, padding=1)

    def forward(self, x):
        x = self.downConv(x.permute(0, 2, 1))
        x = self.norm(x)
        x = self.activation(x)
        x = self.maxPool(x)
        x = x.transpose(1, 2)
        return x


class InformerEncoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: InformerConfig):
        super().__init__()
        self.embed_dim = config.d_model
        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout
        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

        if config.attention_type == "prob":
            self.self_attn = InformerProbSparseAttention(
                embed_dim=self.embed_dim,
                num_heads=config.encoder_attention_heads,
                dropout=config.attention_dropout,
                sampling_factor=config.sampling_factor,
            )
        else:
            self.self_attn = InformerAttention(
                embed_dim=self.embed_dim,
                num_heads=config.encoder_attention_heads,
                dropout=config.attention_dropout,
                config=config,
            )

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        attention_mask: torch.FloatTensor,
        layer_head_mask: torch.FloatTensor,
        output_attentions: Optional[bool] = False,
    ) -> tuple[torch.FloatTensor, Optional[torch.FloatTensor]]:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
        """
        residual = hidden_states
        hidden_states, attn_weights = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
        )
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        residual = hidden_states
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        hidden_states = self.final_layer_norm(hidden_states)

        if hidden_states.dtype == torch.float16 and (
            torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any()
        ):
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (attn_weights,)

        return outputs


class InformerDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: InformerConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.embed_dim = config.d_model
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout

        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.encoder_attn = InformerAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
            config=config,
            layer_idx=layer_idx,
        )
        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

        if config.attention_type == "prob":
            self.self_attn = InformerProbSparseAttention(
                embed_dim=self.embed_dim,
                num_heads=config.decoder_attention_heads,
                dropout=config.attention_dropout,
                sampling_factor=config.sampling_factor,
                is_decoder=True,
                layer_idx=layer_idx,
            )
        else:
            self.self_attn = InformerAttention(
                embed_dim=self.embed_dim,
                num_heads=config.decoder_attention_heads,
                dropout=config.attention_dropout,
                is_decoder=True,
                config=config,
                layer_idx=layer_idx,
            )

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        layer_head_mask: Optional[torch.Tensor] = None,
        cross_attn_layer_head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[Cache] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = True,
        cache_position: Optional[torch.Tensor] = None,
    ) -> 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`): attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            encoder_hidden_states (`torch.FloatTensor`):
                cross attention input to the layer of shape `(batch, seq_len, embed_dim)`
            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                `(encoder_attention_heads,)`.
            cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of
                size `(decoder_attention_heads,)`.
            past_key_values (`Cache`): cached past key and value projection states
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence. It is used to update the
                cache in the correct position and to infer the complete sequence length.
        """
        residual = hidden_states

        # Self Attention
        hidden_states, self_attn_weights = self.self_attn(
            hidden_states=hidden_states,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
            cache_position=cache_position,
        )
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        # Cross-Attention Block
        cross_attn_weights = None
        if encoder_hidden_states is not None:
            residual = hidden_states

            hidden_states, cross_attn_weights = self.encoder_attn(
                hidden_states=hidden_states,
                key_value_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
                layer_head_mask=cross_attn_layer_head_mask,
                past_key_values=past_key_values,
                output_attentions=output_attentions,
                cache_position=cache_position,
            )
            hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
            hidden_states = residual + hidden_states
            hidden_states = self.encoder_attn_layer_norm(hidden_states)

        # Fully Connected
        residual = hidden_states
        hidden_states = self.activation_fn(self.fc1(hidden_states))
        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
        hidden_states = self.fc2(hidden_states)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
        hidden_states = residual + hidden_states
        hidden_states = self.final_layer_norm(hidden_states)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights, cross_attn_weights)

        return outputs


class InformerEncoder(InformerPreTrainedModel):
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
    [`InformerEncoderLayer`].

    Args:
        config: InformerConfig
    """

    def __init__(self, config: InformerConfig):
        super().__init__(config)

        self.dropout = config.dropout
        self.layerdrop = config.encoder_layerdrop
        if config.prediction_length is None:
            raise ValueError("The `prediction_length` config needs to be specified.")

        self.value_embedding = InformerValueEmbedding(feature_size=config.feature_size, d_model=config.d_model)
        self.embed_positions = InformerSinusoidalPositionalEmbedding(
            config.context_length + config.prediction_length, config.d_model
        )
        self.layers = nn.ModuleList([InformerEncoderLayer(config) for _ in range(config.encoder_layers)])
        self.layernorm_embedding = nn.LayerNorm(config.d_model)
        self.gradient_checkpointing = False

        if config.distil:
            self.conv_layers = nn.ModuleList(
                [InformerConvLayer(config.d_model) for _ in range(config.encoder_layers - 1)]
            )
            self.conv_layers.append(None)
        else:
            self.conv_layers = [None] * config.encoder_layers
        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[tuple, BaseModelOutput]:
        r"""
        Args:
            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**.

                [What are attention masks?](../glossary#attention-mask)
            head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
                than the model's internal embedding lookup matrix.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        """
        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

        hidden_states = self.value_embedding(inputs_embeds)
        embed_pos = self.embed_positions(inputs_embeds.size())

        hidden_states = self.layernorm_embedding(hidden_states + embed_pos)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        # expand attention_mask
        if attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            attention_mask = _prepare_4d_attention_mask(attention_mask, inputs_embeds.dtype)

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None

        # check if head_mask has a correct number of layers specified if desired
        if head_mask is not None:
            if head_mask.size()[0] != (len(self.layers)):
                raise ValueError(
                    f"The head_mask should be specified for {len(self.layers)} layers, but it is for"
                    f" {head_mask.size()[0]}."
                )

        for idx, (encoder_layer, conv_layer) in enumerate(zip(self.layers, self.conv_layers)):
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description)
            to_drop = False
            if self.training:
                dropout_probability = torch.rand([])
                if dropout_probability < self.layerdrop:  # skip the layer
                    to_drop = True

            if to_drop:
                layer_outputs = (None, None)
            else:
                layer_outputs = encoder_layer(
                    hidden_states,
                    attention_mask,
                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                    output_attentions=output_attentions,
                )
                if conv_layer is not None:
                    output = conv_layer(layer_outputs[0])
                    layer_outputs = (output,) + layer_outputs[1:]

                hidden_states = layer_outputs[0]

            if output_attentions:
                all_attentions = all_attentions + (layer_outputs[1],)

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
        return BaseModelOutput(
            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
        )


class InformerDecoder(InformerPreTrainedModel):
    """
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a
    [`InformerDecoderLayer`]

    Args:
        config: InformerConfig
    """

    def __init__(self, config: InformerConfig):
        super().__init__(config)
        self.dropout = config.dropout
        self.layerdrop = config.decoder_layerdrop
        if config.prediction_length is None:
            raise ValueError("The `prediction_length` config needs to be specified.")

        self.value_embedding = InformerValueEmbedding(feature_size=config.feature_size, d_model=config.d_model)
        self.embed_positions = InformerSinusoidalPositionalEmbedding(
            config.context_length + config.prediction_length, config.d_model
        )
        self.layers = nn.ModuleList([InformerDecoderLayer(config, layer_idx=i) for i in range(config.decoder_layers)])
        self.layernorm_embedding = nn.LayerNorm(config.d_model)

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

    def forward(
        self,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[tuple, BaseModelOutputWithPastAndCrossAttentions]:
        r"""
        Args:
            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**.

                [What are attention masks?](../glossary#attention-mask)
            encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
                of the decoder.
            encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
                selected in `[0, 1]`:

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

                [What are attention masks?](../glossary#attention-mask)
            head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the cross-attention modules in the decoder to avoid performing
                cross-attention on hidden heads. Mask values selected in `[0, 1]`:

                - 1 indicates the head is **not masked**,
                - 0 indicates the head is **masked**.

            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 and in the
                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.

                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
                than the model's internal embedding lookup matrix.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more detail.
            return_dict (`bool`, *optional*):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence. It is used to update the
                cache in the correct position and to infer the complete sequence length.
        """
        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
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        input_shape = inputs_embeds.size()[:-1]
        # initialize `past_key_values`
        if use_cache and past_key_values is None:
            past_key_values = EncoderDecoderCache(DynamicCache(config=self.config), DynamicCache(config=self.config))
        if use_cache and isinstance(past_key_values, tuple):
            logger.warning_once(
                "Passing a tuple of `past_key_values` is deprecated and will be removed in Transformers v4.58.0. "
                "You should pass an instance of `EncoderDecoderCache` instead, e.g. "
                "`past_key_values=EncoderDecoderCache.from_legacy_cache(past_key_values)`."
            )
            past_key_values = EncoderDecoderCache.from_legacy_cache(past_key_values)

        past_key_values_length = past_key_values.get_seq_length() if past_key_values is not None else 0
        if cache_position is None:
            cache_position = torch.arange(
                past_key_values_length, past_key_values_length + input_shape[1], device=inputs_embeds.device
            )

        attention_mask = self._update_causal_mask(
            attention_mask,
            input_shape,
            inputs_embeds,
            past_key_values_length,
        )
        encoder_attention_mask = self._update_cross_attn_mask(
            encoder_hidden_states,
            encoder_attention_mask,
            input_shape,
            inputs_embeds,
        )

        hidden_states = self.value_embedding(inputs_embeds)
        embed_pos = self.embed_positions(inputs_embeds.size(), past_key_values_length=self.config.context_length)
        hidden_states = self.layernorm_embedding(hidden_states + embed_pos)
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None

        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
        for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
            if attn_mask is not None:
                if attn_mask.size()[0] != (len(self.layers)):
                    raise ValueError(
                        f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
                        f" {head_mask.size()[0]}."
                    )

        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://huggingface.co/papers/1909.11556 for description)
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            if self.training:
                dropout_probability = torch.rand([])
                if dropout_probability < self.layerdrop:
                    continue

            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask,
                encoder_hidden_states,  # as a positional argument for gradient checkpointing
                encoder_attention_mask=encoder_attention_mask,
                layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                cross_attn_layer_head_mask=(cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None),
                past_key_values=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                cache_position=cache_position,
            )
            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

                if encoder_hidden_states is not None:
                    all_cross_attentions += (layer_outputs[2],)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        if not return_dict:
            return tuple(
                v
                for v in [hidden_states, past_key_values, all_hidden_states, all_self_attns, all_cross_attentions]
                if v is not None
            )
        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            cross_attentions=all_cross_attentions,
        )


@auto_docstring
class InformerModel(InformerPreTrainedModel):
    def __init__(self, config: InformerConfig):
        super().__init__(config)

        if config.scaling == "mean" or config.scaling is True:
            self.scaler = InformerMeanScaler(config)
        elif config.scaling == "std":
            self.scaler = InformerStdScaler(config)
        else:
            self.scaler = InformerNOPScaler(config)

        if config.num_static_categorical_features > 0:
            self.embedder = InformerFeatureEmbedder(
                cardinalities=config.cardinality,
                embedding_dims=config.embedding_dimension,
            )

        # transformer encoder-decoder and mask initializer
        self.encoder = InformerEncoder(config)
        self.decoder = InformerDecoder(config)

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

    @property
    def _past_length(self) -> int:
        return self.config.context_length + max(self.config.lags_sequence)

    def get_lagged_subsequences(
        self, sequence: torch.Tensor, subsequences_length: int, shift: int = 0
    ) -> torch.Tensor:
        """
        Returns lagged subsequences of a given sequence. Returns a tensor of shape (N, S, C, I),
            where S = subsequences_length and I = len(indices), containing lagged subsequences. Specifically, lagged[i,
            j, :, k] = sequence[i, -indices[k]-S+j, :].

        Args:
            sequence: Tensor
                The sequence from which lagged subsequences should be extracted. Shape: (N, T, C).
            subsequences_length : int
                Length of the subsequences to be extracted.
            shift: int
                Shift the lags by this amount back.
        """
        sequence_length = sequence.shape[1]
        indices = [lag - shift for lag in self.config.lags_sequence]

        if max(indices) + subsequences_length > sequence_length:
            raise ValueError(
                f"lags cannot go further than history length, found lag {max(indices)} "
                f"while history length is only {sequence_length}"
            )

        lagged_values = []
        for lag_index in indices:
            begin_index = -lag_index - subsequences_length
            end_index = -lag_index if lag_index > 0 else None
            lagged_values.append(sequence[:, begin_index:end_index, ...])
        return torch.stack(lagged_values, dim=-1)

    def create_network_inputs(
        self,
        past_values: torch.Tensor,
        past_time_features: torch.Tensor,
        static_categorical_features: Optional[torch.Tensor] = None,
        static_real_features: Optional[torch.Tensor] = None,
        past_observed_mask: Optional[torch.Tensor] = None,
        future_values: Optional[torch.Tensor] = None,
        future_time_features: Optional[torch.Tensor] = None,
    ):
        # time feature
        time_feat = (
            torch.cat(
                (
                    past_time_features[:, self._past_length - self.config.context_length :, ...],
                    future_time_features,
                ),
                dim=1,
            )
            if future_values is not None
            else past_time_features[:, self._past_length - self.config.context_length :, ...]
        )

        # target
        if past_observed_mask is None:
            past_observed_mask = torch.ones_like(past_values)

        context = past_values[:, -self.config.context_length :]
        observed_context = past_observed_mask[:, -self.config.context_length :]
        _, loc, scale = self.scaler(context, observed_context)

        inputs = (
            (torch.cat((past_values, future_values), dim=1) - loc) / scale
            if future_values is not None
            else (past_values - loc) / scale
        )

        # static features
        if loc.ndim == 3:
            squeezed_loc = loc.squeeze(1)
            squeezed_scale = scale.squeeze(1)
        else:
            squeezed_loc = loc
            squeezed_scale = scale
        log_abs_loc = squeezed_loc.abs().log1p()
        log_scale = squeezed_scale.log()
        static_feat = torch.cat((log_abs_loc, log_scale), dim=1)

        if static_real_features is not None:
            static_feat = torch.cat((static_real_features, static_feat), dim=1)
        if static_categorical_features is not None:
            embedded_cat = self.embedder(static_categorical_features)
            static_feat = torch.cat((embedded_cat, static_feat), dim=1)
        expanded_static_feat = static_feat.unsqueeze(1).expand(-1, time_feat.shape[1], -1)

        # all features
        features = torch.cat((expanded_static_feat, time_feat), dim=-1)

        # lagged features
        subsequences_length = (
            self.config.context_length + self.config.prediction_length
            if future_values is not None
            else self.config.context_length
        )
        lagged_sequence = self.get_lagged_subsequences(sequence=inputs, subsequences_length=subsequences_length)
        lags_shape = lagged_sequence.shape
        reshaped_lagged_sequence = lagged_sequence.reshape(lags_shape[0], lags_shape[1], -1)

        if reshaped_lagged_sequence.shape[1] != time_feat.shape[1]:
            raise ValueError(
                f"input length {reshaped_lagged_sequence.shape[1]} and time feature lengths {time_feat.shape[1]} does not match"
            )

        # transformer inputs
        transformer_inputs = torch.cat((reshaped_lagged_sequence, features), dim=-1)

        return transformer_inputs, loc, scale, static_feat

    def get_encoder(self):
        return self.encoder

    @auto_docstring
    def forward(
        self,
        past_values: torch.Tensor,
        past_time_features: torch.Tensor,
        past_observed_mask: torch.Tensor,
        static_categorical_features: Optional[torch.Tensor] = None,
        static_real_features: Optional[torch.Tensor] = None,
        future_values: Optional[torch.Tensor] = None,
        future_time_features: Optional[torch.Tensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[list[torch.FloatTensor]] = None,
        past_key_values: Optional[Cache] = None,
        output_hidden_states: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        use_cache: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[Seq2SeqTSModelOutput, tuple]:
        r"""
        past_values (`torch.FloatTensor` of shape `(batch_size, sequence_length)` or `(batch_size, sequence_length, input_size)`):
            Past values of the time series, that serve as context in order to predict the future. The sequence size of
            this tensor must be larger than the `context_length` of the model, since the model will use the larger size
            to construct lag features, i.e. additional values from the past which are added in order to serve as "extra
            context".

            The `sequence_length` here is equal to `config.context_length` + `max(config.lags_sequence)`, which if no
            `lags_sequence` is configured, is equal to `config.context_length` + 7 (as by default, the largest
            look-back index in `config.lags_sequence` is 7). The property `_past_length` returns the actual length of
            the past.

            The `past_values` is what the Transformer encoder gets as input (with optional additional features, such as
            `static_categorical_features`, `static_real_features`, `past_time_features` and lags).

            Optionally, missing values need to be replaced with zeros and indicated via the `past_observed_mask`.

            For multivariate time series, the `input_size` > 1 dimension is required and corresponds to the number of
            variates in the time series per time step.
        past_time_features (`torch.FloatTensor` of shape `(batch_size, sequence_length, num_features)`):
            Required time features, which the model internally will add to `past_values`. These could be things like
            "month of year", "day of the month", etc. encoded as vectors (for instance as Fourier features). These
            could also be so-called "age" features, which basically help the model know "at which point in life" a
            time-series is. Age features have small values for distant past time steps and increase monotonically the
            more we approach the current time step. Holiday features are also a good example of time features.

            These features serve as the "positional encodings" of the inputs. So contrary to a model like BERT, where
            the position encodings are learned from scratch internally as parameters of the model, the Time Series
            Transformer requires to provide additional time features. The Time Series Transformer only learns
            additional embeddings for `static_categorical_features`.

            Additional dynamic real covariates can be concatenated to this tensor, with the caveat that these features
            must but known at prediction time.

            The `num_features` here is equal to `config.`num_time_features` + `config.num_dynamic_real_features`.
        past_observed_mask (`torch.BoolTensor` of shape `(batch_size, sequence_length)` or `(batch_size, sequence_length, input_size)`, *optional*):
            Boolean mask to indicate which `past_values` were observed and which were missing. Mask values selected in
            `[0, 1]`:

            - 1 for values that are **observed**,
            - 0 for values that are **missing** (i.e. NaNs that were replaced by zeros).
        static_categorical_features (`torch.LongTensor` of shape `(batch_size, number of static categorical features)`, *optional*):
            Optional static categorical features for which the model will learn an embedding, which it will add to the
            values of the time series.

            Static categorical features are features which have the same value for all time steps (static over time).

            A typical example of a static categorical feature is a time series ID.
        static_real_features (`torch.FloatTensor` of shape `(batch_size, number of static real features)`, *optional*):
            Optional static real features which the model will add to the values of the time series.

            Static real features are features which have the same value for all time steps (static over time).

            A typical example of a static real feature is promotion information.
        future_values (`torch.FloatTensor` of shape `(batch_size, prediction_length)` or `(batch_size, prediction_length, input_size)`, *optional*):
            Future values of the time series, that serve as labels for the model. The `future_values` is what the
            Transformer needs during training to learn to output, given the `past_values`.

            The sequence length here is equal to `prediction_length`.

            See the demo notebook and code snippets for details.

            Optionally, during training any missing values need to be replaced with zeros and indicated via the
            `future_observed_mask`.

            For multivariate time series, the `input_size` > 1 dimension is required and corresponds to the number of
            variates in the time series per time step.
        future_time_features (`torch.FloatTensor` of shape `(batch_size, prediction_length, num_features)`):
            Required time features for the prediction window, which the model internally will add to `future_values`.
            These could be things like "month of year", "day of the month", etc. encoded as vectors (for instance as
            Fourier features). These could also be so-called "age" features, which basically help the model know "at
            which point in life" a time-series is. Age features have small values for distant past time steps and
            increase monotonically the more we approach the current time step. Holiday features are also a good example
            of time features.

            These features serve as the "positional encodings" of the inputs. So contrary to a model like BERT, where
            the position encodings are learned from scratch internally as parameters of the model, the Time Series
            Transformer requires to provide additional time features. The Time Series Transformer only learns
            additional embeddings for `static_categorical_features`.

            Additional dynamic real covariates can be concatenated to this tensor, with the caveat that these features
            must but known at prediction time.

            The `num_features` here is equal to `config.`num_time_features` + `config.num_dynamic_real_features`.
        cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
            Tuple consists of `last_hidden_state`, `hidden_states` (*optional*) and `attentions` (*optional*)
            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)` (*optional*) is a sequence of
            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.

        Examples:

        ```python
        >>> from huggingface_hub import hf_hub_download
        >>> import torch
        >>> from transformers import InformerModel

        >>> file = hf_hub_download(
        ...     repo_id="hf-internal-testing/tourism-monthly-batch", filename="train-batch.pt", repo_type="dataset"
        ... )
        >>> batch = torch.load(file)

        >>> model = InformerModel.from_pretrained("huggingface/informer-tourism-monthly")

        >>> # during training, one provides both past and future values
        >>> # as well as possible additional features
        >>> outputs = model(
        ...     past_values=batch["past_values"],
        ...     past_time_features=batch["past_time_features"],
        ...     past_observed_mask=batch["past_observed_mask"],
        ...     static_categorical_features=batch["static_categorical_features"],
        ...     static_real_features=batch["static_real_features"],
        ...     future_values=batch["future_values"],
        ...     future_time_features=batch["future_time_features"],
        ... )

        >>> last_hidden_state = outputs.last_hidden_state
        ```"""
        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
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        transformer_inputs, loc, scale, static_feat = self.create_network_inputs(
            past_values=past_values,
            past_time_features=past_time_features,
            past_observed_mask=past_observed_mask,
            static_categorical_features=static_categorical_features,
            static_real_features=static_real_features,
            future_values=future_values,
            future_time_features=future_time_features,
        )

        if encoder_outputs is None:
            enc_input = transformer_inputs[:, : self.config.context_length, ...]
            encoder_outputs = self.encoder(
                inputs_embeds=enc_input,
                head_mask=head_mask,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
            encoder_outputs = BaseModelOutput(
                last_hidden_state=encoder_outputs[0],
                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
            )

        # Avoid empty tensors and instead create a zeroes tensor which
        # will be treated the same in torch, i.e. matmul with empty == all 0s
        if self.config.context_length >= transformer_inputs.shape[1]:
            bsz, _, dim = transformer_inputs.shape
            dec_input = torch.zeros(
                size=(bsz, 1, dim), device=transformer_inputs.device, dtype=transformer_inputs.dtype
            )
        else:
            dec_input = transformer_inputs[:, self.config.context_length :, ...]

        decoder_outputs = self.decoder(
            inputs_embeds=dec_input,
            attention_mask=decoder_attention_mask,
            encoder_hidden_states=encoder_outputs[0],
            head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )

        if not return_dict:
            return decoder_outputs + encoder_outputs + (loc, scale, static_feat)

        return Seq2SeqTSModelOutput(
            last_hidden_state=decoder_outputs.last_hidden_state,
            past_key_values=decoder_outputs.past_key_values,
            decoder_hidden_states=decoder_outputs.hidden_states,
            decoder_attentions=decoder_outputs.attentions,
            cross_attentions=decoder_outputs.cross_attentions,
            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
            encoder_hidden_states=encoder_outputs.hidden_states,
            encoder_attentions=encoder_outputs.attentions,
            loc=loc,
            scale=scale,
            static_features=static_feat,
        )


def weighted_average(input_tensor: torch.Tensor, weights: Optional[torch.Tensor] = None, dim=None) -> torch.Tensor:
    """
    Computes the weighted average of a given tensor across a given `dim`, masking values associated with weight zero,
    meaning instead of `nan * 0 = nan` you will get `0 * 0 = 0`.

    Args:
        input_tensor (`torch.FloatTensor`):
            Input tensor, of which the average must be computed.
        weights (`torch.FloatTensor`, *optional*):
            Weights tensor, of the same shape as `input_tensor`.
        dim (`int`, *optional*):
            The dim along which to average `input_tensor`.

    Returns:
        `torch.FloatTensor`: The tensor with values averaged along the specified `dim`.
    """
    if weights is not None:
        weighted_tensor = torch.where(weights != 0, input_tensor * weights, torch.zeros_like(input_tensor))
        sum_weights = torch.clamp(weights.sum(dim=dim) if dim else weights.sum(), min=1.0)
        return (weighted_tensor.sum(dim=dim) if dim else weighted_tensor.sum()) / sum_weights
    else:
        return input_tensor.mean(dim=dim)


def nll(input: torch.distributions.Distribution, target: torch.Tensor) -> torch.Tensor:
    """
    Computes the negative log likelihood loss from input distribution with respect to target.
    """
    return -input.log_prob(target)


@auto_docstring
class InformerForPrediction(InformerPreTrainedModel):
    def __init__(self, config: InformerConfig):
        super().__init__(config)

        self.model = InformerModel(config)
        if config.distribution_output == "student_t":
            self.distribution_output = StudentTOutput(dim=config.input_size)
        elif config.distribution_output == "normal":
            self.distribution_output = NormalOutput(dim=config.input_size)
        elif config.distribution_output == "negative_binomial":
            self.distribution_output = NegativeBinomialOutput(dim=config.input_size)
        else:
            raise ValueError(f"Unknown distribution output {config.distribution_output}")

        self.parameter_projection = self.distribution_output.get_parameter_projection(self.model.config.d_model)
        self.target_shape = self.distribution_output.event_shape

        if config.loss == "nll":
            self.loss = nll
        else:
            raise ValueError(f"Unknown loss function {config.loss}")

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

    def output_params(self, dec_output):
        return self.parameter_projection(dec_output)

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

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

    @torch.jit.ignore
    def output_distribution(self, params, loc=None, scale=None, trailing_n=None) -> torch.distributions.Distribution:
        sliced_params = params
        if trailing_n is not None:
            sliced_params = [p[:, -trailing_n:] for p in params]
        return self.distribution_output.distribution(sliced_params, loc=loc, scale=scale)

    @auto_docstring
    def forward(
        self,
        past_values: torch.Tensor,
        past_time_features: torch.Tensor,
        past_observed_mask: torch.Tensor,
        static_categorical_features: Optional[torch.Tensor] = None,
        static_real_features: Optional[torch.Tensor] = None,
        future_values: Optional[torch.Tensor] = None,
        future_time_features: Optional[torch.Tensor] = None,
        future_observed_mask: Optional[torch.Tensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[list[torch.FloatTensor]] = None,
        past_key_values: Optional[Cache] = None,
        output_hidden_states: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        use_cache: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[Seq2SeqTSModelOutput, tuple]:
        r"""
        past_values (`torch.FloatTensor` of shape `(batch_size, sequence_length)` or `(batch_size, sequence_length, input_size)`):
            Past values of the time series, that serve as context in order to predict the future. The sequence size of
            this tensor must be larger than the `context_length` of the model, since the model will use the larger size
            to construct lag features, i.e. additional values from the past which are added in order to serve as "extra
            context".

            The `sequence_length` here is equal to `config.context_length` + `max(config.lags_sequence)`, which if no
            `lags_sequence` is configured, is equal to `config.context_length` + 7 (as by default, the largest
            look-back index in `config.lags_sequence` is 7). The property `_past_length` returns the actual length of
            the past.

            The `past_values` is what the Transformer encoder gets as input (with optional additional features, such as
            `static_categorical_features`, `static_real_features`, `past_time_features` and lags).

            Optionally, missing values need to be replaced with zeros and indicated via the `past_observed_mask`.

            For multivariate time series, the `input_size` > 1 dimension is required and corresponds to the number of
            variates in the time series per time step.
        past_time_features (`torch.FloatTensor` of shape `(batch_size, sequence_length, num_features)`):
            Required time features, which the model internally will add to `past_values`. These could be things like
            "month of year", "day of the month", etc. encoded as vectors (for instance as Fourier features). These
            could also be so-called "age" features, which basically help the model know "at which point in life" a
            time-series is. Age features have small values for distant past time steps and increase monotonically the
            more we approach the current time step. Holiday features are also a good example of time features.

            These features serve as the "positional encodings" of the inputs. So contrary to a model like BERT, where
            the position encodings are learned from scratch internally as parameters of the model, the Time Series
            Transformer requires to provide additional time features. The Time Series Transformer only learns
            additional embeddings for `static_categorical_features`.

            Additional dynamic real covariates can be concatenated to this tensor, with the caveat that these features
            must but known at prediction time.

            The `num_features` here is equal to `config.`num_time_features` + `config.num_dynamic_real_features`.
        past_observed_mask (`torch.BoolTensor` of shape `(batch_size, sequence_length)` or `(batch_size, sequence_length, input_size)`, *optional*):
            Boolean mask to indicate which `past_values` were observed and which were missing. Mask values selected in
            `[0, 1]`:

            - 1 for values that are **observed**,
            - 0 for values that are **missing** (i.e. NaNs that were replaced by zeros).
        static_categorical_features (`torch.LongTensor` of shape `(batch_size, number of static categorical features)`, *optional*):
            Optional static categorical features for which the model will learn an embedding, which it will add to the
            values of the time series.

            Static categorical features are features which have the same value for all time steps (static over time).

            A typical example of a static categorical feature is a time series ID.
        static_real_features (`torch.FloatTensor` of shape `(batch_size, number of static real features)`, *optional*):
            Optional static real features which the model will add to the values of the time series.

            Static real features are features which have the same value for all time steps (static over time).

            A typical example of a static real feature is promotion information.
        future_values (`torch.FloatTensor` of shape `(batch_size, prediction_length)` or `(batch_size, prediction_length, input_size)`, *optional*):
            Future values of the time series, that serve as labels for the model. The `future_values` is what the
            Transformer needs during training to learn to output, given the `past_values`.

            The sequence length here is equal to `prediction_length`.

            See the demo notebook and code snippets for details.

            Optionally, during training any missing values need to be replaced with zeros and indicated via the
            `future_observed_mask`.

            For multivariate time series, the `input_size` > 1 dimension is required and corresponds to the number of
            variates in the time series per time step.
        future_time_features (`torch.FloatTensor` of shape `(batch_size, prediction_length, num_features)`):
            Required time features for the prediction window, which the model internally will add to `future_values`.
            These could be things like "month of year", "day of the month", etc. encoded as vectors (for instance as
            Fourier features). These could also be so-called "age" features, which basically help the model know "at
            which point in life" a time-series is. Age features have small values for distant past time steps and
            increase monotonically the more we approach the current time step. Holiday features are also a good example
            of time features.

            These features serve as the "positional encodings" of the inputs. So contrary to a model like BERT, where
            the position encodings are learned from scratch internally as parameters of the model, the Time Series
            Transformer requires to provide additional time features. The Time Series Transformer only learns
            additional embeddings for `static_categorical_features`.

            Additional dynamic real covariates can be concatenated to this tensor, with the caveat that these features
            must but known at prediction time.

            The `num_features` here is equal to `config.`num_time_features` + `config.num_dynamic_real_features`.
        future_observed_mask (`torch.BoolTensor` of shape `(batch_size, sequence_length)` or `(batch_size, sequence_length, input_size)`, *optional*):
            Boolean mask to indicate which `future_values` were observed and which were missing. Mask values selected
            in `[0, 1]`:

            - 1 for values that are **observed**,
            - 0 for values that are **missing** (i.e. NaNs that were replaced by zeros).

            This mask is used to filter out missing values for the final loss calculation.
        cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
            Tuple consists of `last_hidden_state`, `hidden_states` (*optional*) and `attentions` (*optional*)
            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)` (*optional*) is a sequence of
            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.

        Examples:

        ```python
        >>> from huggingface_hub import hf_hub_download
        >>> import torch
        >>> from transformers import InformerForPrediction

        >>> file = hf_hub_download(
        ...     repo_id="hf-internal-testing/tourism-monthly-batch", filename="train-batch.pt", repo_type="dataset"
        ... )
        >>> batch = torch.load(file)

        >>> model = InformerForPrediction.from_pretrained(
        ...     "huggingface/informer-tourism-monthly"
        ... )

        >>> # during training, one provides both past and future values
        >>> # as well as possible additional features
        >>> outputs = model(
        ...     past_values=batch["past_values"],
        ...     past_time_features=batch["past_time_features"],
        ...     past_observed_mask=batch["past_observed_mask"],
        ...     static_categorical_features=batch["static_categorical_features"],
        ...     static_real_features=batch["static_real_features"],
        ...     future_values=batch["future_values"],
        ...     future_time_features=batch["future_time_features"],
        ... )

        >>> loss = outputs.loss
        >>> loss.backward()

        >>> # during inference, one only provides past values
        >>> # as well as possible additional features
        >>> # the model autoregressively generates future values
        >>> outputs = model.generate(
        ...     past_values=batch["past_values"],
        ...     past_time_features=batch["past_time_features"],
        ...     past_observed_mask=batch["past_observed_mask"],
        ...     static_categorical_features=batch["static_categorical_features"],
        ...     static_real_features=batch["static_real_features"],
        ...     future_time_features=batch["future_time_features"],
        ... )

        >>> mean_prediction = outputs.sequences.mean(dim=1)
        ```"""

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        if future_values is not None:
            use_cache = False

        outputs = self.model(
            past_values=past_values,
            past_time_features=past_time_features,
            past_observed_mask=past_observed_mask,
            static_categorical_features=static_categorical_features,
            static_real_features=static_real_features,
            future_values=future_values,
            future_time_features=future_time_features,
            decoder_attention_mask=decoder_attention_mask,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            encoder_outputs=encoder_outputs,
            past_key_values=past_key_values,
            output_hidden_states=output_hidden_states,
            output_attentions=output_attentions,
            use_cache=use_cache,
            return_dict=return_dict,
            cache_position=cache_position,
        )

        prediction_loss = None
        params = None
        if future_values is not None:
            params = self.output_params(outputs[0])  # outputs.last_hidden_state
            # loc is 3rd last and scale is 2nd last output
            distribution = self.output_distribution(params, loc=outputs[-3], scale=outputs[-2])

            loss = self.loss(distribution, future_values)

            if future_observed_mask is None:
                future_observed_mask = torch.ones_like(future_values)

            if len(self.target_shape) == 0:
                loss_weights = future_observed_mask
            else:
                loss_weights, _ = future_observed_mask.min(dim=-1, keepdim=False)

            prediction_loss = weighted_average(loss, weights=loss_weights)

        if not return_dict:
            outputs = ((params,) + outputs[1:]) if params is not None else outputs[1:]
            return ((prediction_loss,) + outputs) if prediction_loss is not None else outputs

        return Seq2SeqTSPredictionOutput(
            loss=prediction_loss,
            params=params,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
            loc=outputs.loc,
            scale=outputs.scale,
            static_features=outputs.static_features,
        )

    @torch.no_grad()
    def generate(
        self,
        past_values: torch.Tensor,
        past_time_features: torch.Tensor,
        future_time_features: torch.Tensor,
        past_observed_mask: Optional[torch.Tensor] = None,
        static_categorical_features: Optional[torch.Tensor] = None,
        static_real_features: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
    ) -> SampleTSPredictionOutput:
        r"""
        Greedily generate sequences of sample predictions from a model with a probability distribution head.

        Parameters:
            past_values (`torch.FloatTensor` of shape `(batch_size, sequence_length)` or `(batch_size, sequence_length, input_size)`):
                Past values of the time series, that serve as context in order to predict the future. The sequence size
                of this tensor must be larger than the `context_length` of the model, since the model will use the
                larger size to construct lag features, i.e. additional values from the past which are added in order to
                serve as "extra context".

                The `sequence_length` here is equal to `config.context_length` + `max(config.lags_sequence)`, which if
                no `lags_sequence` is configured, is equal to `config.context_length` + 7 (as by default, the largest
                look-back index in `config.lags_sequence` is 7). The property `_past_length` returns the actual length
                of the past.

                The `past_values` is what the Transformer encoder gets as input (with optional additional features,
                such as `static_categorical_features`, `static_real_features`, `past_time_features` and lags).

                Optionally, missing values need to be replaced with zeros and indicated via the `past_observed_mask`.

                For multivariate time series, the `input_size` > 1 dimension is required and corresponds to the number
                of variates in the time series per time step.
            past_time_features (`torch.FloatTensor` of shape `(batch_size, sequence_length, num_features)`):
                Required time features, which the model internally will add to `past_values`. These could be things
                like "month of year", "day of the month", etc. encoded as vectors (for instance as Fourier features).
                These could also be so-called "age" features, which basically help the model know "at which point in
                life" a time-series is. Age features have small values for distant past time steps and increase
                monotonically the more we approach the current time step. Holiday features are also a good example of
                time features.

                These features serve as the "positional encodings" of the inputs. So contrary to a model like BERT,
                where the position encodings are learned from scratch internally as parameters of the model, the Time
                Series Transformer requires to provide additional time features. The Time Series Transformer only
                learns additional embeddings for `static_categorical_features`.

                Additional dynamic real covariates can be concatenated to this tensor, with the caveat that these
                features must but known at prediction time.

                The `num_features` here is equal to `config.`num_time_features` + `config.num_dynamic_real_features`.
            future_time_features (`torch.FloatTensor` of shape `(batch_size, prediction_length, num_features)`):
                Required time features for the prediction window, which the model internally will add to sampled
                predictions. These could be things like "month of year", "day of the month", etc. encoded as vectors
                (for instance as Fourier features). These could also be so-called "age" features, which basically help
                the model know "at which point in life" a time-series is. Age features have small values for distant
                past time steps and increase monotonically the more we approach the current time step. Holiday features
                are also a good example of time features.

                These features serve as the "positional encodings" of the inputs. So contrary to a model like BERT,
                where the position encodings are learned from scratch internally as parameters of the model, the Time
                Series Transformer requires to provide additional time features. The Time Series Transformer only
                learns additional embeddings for `static_categorical_features`.

                Additional dynamic real covariates can be concatenated to this tensor, with the caveat that these
                features must but known at prediction time.

                The `num_features` here is equal to `config.`num_time_features` + `config.num_dynamic_real_features`.
            past_observed_mask (`torch.BoolTensor` of shape `(batch_size, sequence_length)` or `(batch_size, sequence_length, input_size)`, *optional*):
                Boolean mask to indicate which `past_values` were observed and which were missing. Mask values selected
                in `[0, 1]`:

                - 1 for values that are **observed**,
                - 0 for values that are **missing** (i.e. NaNs that were replaced by zeros).

            static_categorical_features (`torch.LongTensor` of shape `(batch_size, number of static categorical features)`, *optional*):
                Optional static categorical features for which the model will learn an embedding, which it will add to
                the values of the time series.

                Static categorical features are features which have the same value for all time steps (static over
                time).

                A typical example of a static categorical feature is a time series ID.
            static_real_features (`torch.FloatTensor` of shape `(batch_size, number of static real features)`, *optional*):
                Optional static real features which the model will add to the values of the time series.

                Static real features are features which have the same value for all time steps (static over time).

                A typical example of a static real feature is promotion information.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers.
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers.

        Return:
            [`SampleTSPredictionOutput`] where the outputs `sequences` tensor will have shape `(batch_size, number of
            samples, prediction_length)` or `(batch_size, number of samples, prediction_length, input_size)` for
            multivariate predictions.
        """
        outputs = self(
            static_categorical_features=static_categorical_features,
            static_real_features=static_real_features,
            past_time_features=past_time_features,
            past_values=past_values,
            past_observed_mask=past_observed_mask,
            future_time_features=future_time_features,
            future_values=None,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            use_cache=True,
        )

        decoder = self.model.get_decoder()
        enc_last_hidden = outputs.encoder_last_hidden_state
        loc = outputs.loc
        scale = outputs.scale
        static_feat = outputs.static_features

        num_parallel_samples = self.config.num_parallel_samples
        repeated_loc = loc.repeat_interleave(repeats=num_parallel_samples, dim=0)
        repeated_scale = scale.repeat_interleave(repeats=num_parallel_samples, dim=0)

        repeated_past_values = (
            past_values.repeat_interleave(repeats=num_parallel_samples, dim=0) - repeated_loc
        ) / repeated_scale

        expanded_static_feat = static_feat.unsqueeze(1).expand(-1, future_time_features.shape[1], -1)
        features = torch.cat((expanded_static_feat, future_time_features), dim=-1)
        repeated_features = features.repeat_interleave(repeats=num_parallel_samples, dim=0)

        repeated_enc_last_hidden = enc_last_hidden.repeat_interleave(repeats=num_parallel_samples, dim=0)

        future_samples = []

        # greedy decoding
        for k in range(self.config.prediction_length):
            lagged_sequence = self.model.get_lagged_subsequences(
                sequence=repeated_past_values,
                subsequences_length=1 + k,
                shift=1,
            )

            lags_shape = lagged_sequence.shape
            reshaped_lagged_sequence = lagged_sequence.reshape(lags_shape[0], lags_shape[1], -1)

            decoder_input = torch.cat((reshaped_lagged_sequence, repeated_features[:, : k + 1]), dim=-1)

            dec_output = decoder(inputs_embeds=decoder_input, encoder_hidden_states=repeated_enc_last_hidden)
            dec_last_hidden = dec_output.last_hidden_state

            params = self.parameter_projection(dec_last_hidden[:, -1:])
            distr = self.output_distribution(params, loc=repeated_loc, scale=repeated_scale)
            next_sample = distr.sample()

            repeated_past_values = torch.cat(
                (repeated_past_values, (next_sample - repeated_loc) / repeated_scale), dim=1
            )
            future_samples.append(next_sample)

        concat_future_samples = torch.cat(future_samples, dim=1)

        return SampleTSPredictionOutput(
            sequences=concat_future_samples.reshape(
                (-1, num_parallel_samples, self.config.prediction_length) + self.target_shape,
            )
        )


__all__ = ["InformerForPrediction", "InformerModel", "InformerPreTrainedModel"]