modeling_granite.py

import math
import numbers
import warnings
from enum import Enum
from typing import Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import DynamicCache, PreTrainedModel
from transformers.activations import get_activation as get_base_activation
from transformers.modeling_outputs import BaseModelOutputWithPastAndCrossAttentions, CausalLMOutputWithCrossAttentions
from transformers.utils import is_flash_attn_2_available

from .configuration_granite import GraniteConfig


class PositionEmbeddingType(Enum):
    learned_absolute = "learned_absolute"
    alibi = "alibi"
    rope = "rope"


class AttentionHeadType(Enum):
    mha = "mha"
    mqa = "mqa"
    gqa = "gqa"


if is_flash_attn_2_available():
    from flash_attn.bert_padding import IndexFirstAxis, pad_input, unpad_input
    from flash_attn.flash_attn_interface import flash_attn_varlen_func


# Copied from transformers.models.llama.modeling_llama._get_unpad_data
def get_unpad_data(attention_mask: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
    max_seqlen_in_batch = seqlens_in_batch.max().item()
    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
    return indices, cu_seqlens, max_seqlen_in_batch


def repeat_key_value(x: torch.Tensor, num_heads: int, num_key_value_heads: int) -> torch.Tensor:
    num_groups = num_heads // num_key_value_heads

    # mha
    if num_groups == 1:
        return x

    # mqa
    if num_key_value_heads == 1:
        return x.expand(-1, num_heads, -1, -1)

    # gqa
    return x.repeat_interleave(num_groups, dim=1)


##################################################
# activation functions


_GLU_BASE_MAPPING = {
    "geglu": "gelu",
    "miglu": "mish",
    "mishglu": "mish",
    "swiglu": "swish",
}


class GLUActivation(nn.Module):
    def __init__(self, base_activation: nn.Module) -> None:
        super().__init__()
        self.base_activation = base_activation

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = x.chunk(2, dim=-1)
        return x[0] * self.base_activation(x[1])


def is_glu(name: str) -> bool:
    return name.endswith("glu")


def get_activation_function(name: str) -> nn.Module:
    if is_glu(name):
        # for glu and sigmoid_glu, we directly return the pytorch's GLU
        if name in ["glu", "sigmoid_glu"]:
            activation_function = nn.modules.GLU()
        else:
            if name in _GLU_BASE_MAPPING:
                name = _GLU_BASE_MAPPING[name]
            elif name.endswith("_glu"):
                name = name.rstrip("_glu")
            else:
                raise ValueError("invalid activation function")

            base_activation = get_base_activation(name)
            activation_function = GLUActivation(base_activation)
    else:
        activation_function = get_base_activation(name)

    return activation_function


##################################################
# normalization functions


class RMSNorm(nn.Module):
    def __init__(self, normalized_shape: int, eps: float = 1e-6) -> None:
        super().__init__()

        self.weight = nn.Parameter(torch.ones(normalized_shape))
        self.eps = eps

        if isinstance(normalized_shape, numbers.Integral):
            normalized_shape = (normalized_shape,)
        self.normalized_shape = normalized_shape

    def forward(self, input: torch.Tensor) -> torch.Tensor:
        input_dtype = input.dtype

        input = input.to(torch.float32)
        variance = input.pow(2).mean(-1, keepdim=True)
        input = input * torch.rsqrt(variance + self.eps)

        return self.weight * input.to(input_dtype)

    def extra_repr(self) -> str:
        return f"{self.normalized_shape}, eps={self.eps}"

    def reset_parameters(self) -> None:
        nn.init.ones_(self.weight)


_NORMALIZATION_FUNCTIONS = {
    "layernorm": nn.LayerNorm,
    "rmsnorm": RMSNorm,
}


def get_normalization_function(name: str, normalized_shape: int, eps: float = 1e-5) -> nn.Module:
    if name in _NORMALIZATION_FUNCTIONS:
        return _NORMALIZATION_FUNCTIONS[name](normalized_shape, eps=eps)

    raise ValueError(f"unexpected `normalization_function` {name}")


##################################################
# attention modules


class GraniteAttention(nn.Module):
    def __init__(self, config: GraniteConfig, causal: bool, layer_idx: Optional[int] = None) -> None:
        super().__init__()

        self.causal = causal
        self.hidden_size = config.n_embd
        self.num_heads = config.n_head
        self.num_key_value_heads = config.num_key_value_heads
        self.add_bias = config.add_bias

        assert (
            self.hidden_size % self.num_heads == 0
        ), f"`hidden_size` ({self.hidden_size}) must be divisible by `num_heads` ({self.num_heads})"

        self.head_dim = self.hidden_size // self.num_heads
        self.attention_head_type = AttentionHeadType(config.attention_head_type)

        self.position_embedding_type = PositionEmbeddingType(config.position_embedding_type)
        self.scale_attn_weights = config.scale_attn_weights
        self.attention_multiplier = config.attention_multiplier

        self.layer_idx = layer_idx
        self.attention_softmax_in_fp32 = config.attention_softmax_in_fp32
        self.scale_attention_softmax_in_fp32 = (
            config.scale_attention_softmax_in_fp32 and config.attention_softmax_in_fp32
        )

        if self.attention_head_type == AttentionHeadType.mha:
            if self.num_key_value_heads is None:
                self.num_key_value_heads = self.num_heads

            assert (
                self.num_heads == self.num_key_value_heads
            ), f"{self.__class__.__name__} should have same number of heads for query, keys and values"
        elif self.attention_head_type == AttentionHeadType.gqa:
            assert (
                self.num_key_value_heads is not None
            ), "`num_key_value_heads` needs to be specified with GroupedQueryAttention"

            assert self.num_heads % self.num_key_value_heads == 0, (
                f"`num_heads` ({self.num_heads}) should be a multiple of `num_key_value_heads` "
                f"({self.num_key_value_heads})"
            )
        elif self.attention_head_type == AttentionHeadType.mqa:
            if self.num_key_value_heads is None:
                self.num_key_value_heads = 1

            assert self.num_key_value_heads == 1, f"{self.__class__.__name__} should have 1 head for keys and values"
        else:
            raise ValueError(f"unexpected attention_head_type ({self.attention_head_type})")

        # note that the actual layout is different for the output and depends on whether we are using MHA, MQA or GQA
        # (self.hidden_size + 2 * self.num_key_value_heads * self.head_dim) is just the actual number output features
        self.c_attn = nn.Linear(
            self.hidden_size, self.hidden_size + 2 * self.num_key_value_heads * self.head_dim, bias=self.add_bias
        )
        self.c_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=self.add_bias)

        self.attn_pdrop = config.attn_pdrop
        self.resid_pdrop = config.resid_pdrop

        self.attn_dropout = nn.Identity() if self.attn_pdrop == 0 else nn.Dropout(self.attn_pdrop)
        self.resid_dropout = nn.Identity() if self.resid_pdrop == 0 else nn.Dropout(self.resid_pdrop)

    def _prepare_qkv_for_forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        # ==========================================================================================
        # hidden_states -> (batch_size, query_length, num_heads * head_dim)
        # ==========================================================================================

        # the output of following is a tuple if using MQA with tensor parallel
        hidden_states = self.c_attn(hidden_states)

        # ==========================================================================================
        # hidden_states -> (batch_size, query_length, [num_heads + num_key_value_heads * 2] * head_dim)
        # ==========================================================================================

        # for MHA, we can get away with doing just 1 transpose which is not true for GQA
        if self.attention_head_type == AttentionHeadType.mha:
            query, key, value = self._prepare_qkv_for_forward_mha(hidden_states)
        elif self.attention_head_type == AttentionHeadType.gqa:
            query, key, value = self._prepare_qkv_for_forward_gqa(hidden_states)
        elif self.attention_head_type == AttentionHeadType.mqa:
            query, key, value = self._prepare_qkv_for_forward_mqa(hidden_states)
        else:
            raise ValueError(f"unexpected attention_head_type ({self.attention_head_type})")

        # ==========================================================================================
        # query -> (batch_size, num_heads, query_length, head_dim)
        # key -> (batch_size, num_key_value_heads, query_length, head_dim)
        # value -> (batch_size, num_key_value_heads, query_length, head_dim)
        # ==========================================================================================

        return query, key, value

    def _prepare_qkv_for_forward_mha(
        self, hidden_states: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        batch_size, query_length = hidden_states.shape[:-1]

        hidden_states = hidden_states.view(batch_size, query_length, self.num_heads, -1)
        hidden_states = hidden_states.transpose(1, 2)

        query, key, value = hidden_states.chunk(3, dim=-1)

        return query, key, value

    def _prepare_qkv_for_forward_gqa(
        self, hidden_states: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        batch_size, query_length = hidden_states.shape[:-1]

        hidden_states = hidden_states.view(batch_size, query_length, self.num_key_value_heads, -1)

        query, key, value = hidden_states.split(
            ((self.num_heads // self.num_key_value_heads) * self.head_dim, self.head_dim, self.head_dim), dim=-1
        )

        # this needs to be a reshape instead of view sadly
        query = query.reshape(batch_size, query_length, -1, self.head_dim)

        query = query.transpose(1, 2)
        key = key.transpose(1, 2)
        value = value.transpose(1, 2)

        return query, key, value

    def _prepare_qkv_for_forward_mqa(
        self, hidden_states: torch.Tensor
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        batch_size, query_length = hidden_states.shape[:-1]

        query, key, value = hidden_states.split((self.hidden_size, self.head_dim, self.head_dim), dim=-1)

        query = query.view(batch_size, query_length, self.num_heads, -1)

        query = query.transpose(1, 2)
        key = key.unsqueeze(1)
        value = value.unsqueeze(1)

        return query, key, value

    def forward(
        self,
        hidden_states: torch.Tensor,
        past_key_values: Optional[DynamicCache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        rope_cos_sin: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
    ) -> torch.Tensor:
        # ==========================================================================================
        # hidden_states -> (batch_size, query_length, num_heads * head_dim)
        # ==========================================================================================

        query, key, value = self._prepare_qkv_for_forward(hidden_states)

        # ==========================================================================================
        # query -> (batch_size, num_heads, query_length, head_dim)
        # key -> (batch_size, num_key_value_heads, query_length, head_dim)
        # value -> (batch_size, num_key_value_heads, query_length, head_dim)
        # ==========================================================================================

        if self.position_embedding_type == PositionEmbeddingType.rope:
            query = apply_rotary_pos_emb(query, rope_cos_sin)
            key = apply_rotary_pos_emb(key, rope_cos_sin)

        if past_key_values is not None:
            key, value = past_key_values.update(key, value, self.layer_idx)

        # ==========================================================================================
        # query -> (batch_size, num_heads, query_length, head_dim)
        # key -> (batch_size, num_key_value_heads, key_length, head_dim)
        # value -> (batch_size, num_key_value_heads, key_length, head_dim)
        # ==========================================================================================

        key = key.transpose(-1, -2)

        dtype = query.dtype
        softmax_dtype = torch.float32 if self.attention_softmax_in_fp32 else dtype

        if self.scale_attn_weights:
            if self.attention_multiplier is None:
                scale_factor = 1 / self.head_dim**0.5
            else:
                scale_factor = self.attention_multiplier
        else:
            scale_factor = 1

        # ==========================================================================================
        # query -> (batch_size, num_heads, query_length, head_dim)
        # key -> (batch_size, num_key_value_heads, head_dim, key_length)
        # value -> (batch_size, num_key_value_heads, key_length, head_dim)
        # ==========================================================================================

        batch_size = query.shape[0]
        query_length = query.shape[2]
        key_length = key.shape[-1]

        key = repeat_key_value(key, self.num_heads, self.num_key_value_heads)
        value = repeat_key_value(value, self.num_heads, self.num_key_value_heads)

        # Always copies
        query = query.reshape(batch_size * self.num_heads, query_length, self.head_dim)
        # No copy when layer_past is provided.
        key = key.reshape(batch_size * self.num_heads, self.head_dim, key_length)

        # ==========================================================================================
        # query -> (batch_size * num_heads, query_length, head_dim)
        # key -> (batch_size * num_heads, head_dim, key_length)
        # value -> (batch_size, num_heads, key_length, head_dim)
        # ==========================================================================================

        attn_weights = torch.empty(
            (batch_size * self.num_heads, query_length, key_length), device=query.device, dtype=query.dtype
        )

        attn_weights = torch.baddbmm(attn_weights, query, key, beta=0, alpha=scale_factor).view(
            batch_size, self.num_heads, query_length, key_length
        )

        # ==========================================================================================
        # attn_weights -> (batch_size, num_heads, query_length, key_length)
        # ==========================================================================================

        attn_weights = attn_weights.to(softmax_dtype)

        if attention_mask is not None:
            attn_weights = attn_weights + attention_mask

        attn_weights = F.softmax(attn_weights, dim=-1).to(dtype)

        attn_weights = self.attn_dropout(attn_weights)

        # ==========================================================================================
        # value -> (batch_size, num_heads, key_length, head_dim)
        # attn_weights -> (batch_size, num_heads, query_length, key_length)
        # ==========================================================================================

        attn_output = torch.matmul(attn_weights, value)

        # ==========================================================================================
        # attn_output -> (batch_size, num_heads, query_length, head_dim)
        # ==========================================================================================

        attn_output = attn_output.transpose(1, 2)
        attn_output = attn_output.reshape(batch_size, -1, self.num_heads * self.head_dim)

        # ==========================================================================================
        # attn_output -> (batch_size, query_length, num_heads * head_dim)
        # ==========================================================================================

        attn_output = self.c_proj(attn_output)
        attn_output = self.resid_dropout(attn_output)

        return attn_output


class GraniteSDPA(GraniteAttention):
    def forward(
        self,
        hidden_states: torch.Tensor,
        past_key_values: Optional[DynamicCache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        rope_cos_sin: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
    ) -> torch.Tensor:
        # ==========================================================================================
        # hidden_states -> (batch_size, query_length, num_heads * head_dim)
        # ==========================================================================================

        query, key, value = self._prepare_qkv_for_forward(hidden_states)

        # ==========================================================================================
        # query -> (batch_size, num_heads, query_length, head_dim)
        # key -> (batch_size, num_key_value_heads, query_length, head_dim)
        # value -> (batch_size, num_key_value_heads, query_length, head_dim)
        # ==========================================================================================

        if self.position_embedding_type == PositionEmbeddingType.rope:
            query = apply_rotary_pos_emb(query, rope_cos_sin)
            key = apply_rotary_pos_emb(key, rope_cos_sin)

        if past_key_values is not None:
            key, value = past_key_values.update(key, value, self.layer_idx)

        # ==========================================================================================
        # query -> (batch_size, num_heads, query_length, head_dim)
        # key -> (batch_size, num_key_value_heads, key_length, head_dim)
        # value -> (batch_size, num_key_value_heads, key_length, head_dim)
        # ==========================================================================================

        key = repeat_key_value(key, self.num_heads, self.num_key_value_heads)
        value = repeat_key_value(value, self.num_heads, self.num_key_value_heads)

        # ==========================================================================================
        # query -> (batch_size, num_heads, query_length, head_dim)
        # key -> (batch_size, num_heads, key_length, head_dim)
        # value -> (batch_size, num_heads, key_length, head_dim)
        # ==========================================================================================

        attn_output = F.scaled_dot_product_attention(
            query,
            key,
            value,
            attn_mask=attention_mask,
            dropout_p=self.attn_pdrop if self.training else 0,
            is_causal=self.causal if attention_mask is None else False,
            scale=self.attention_multiplier if self.scale_attn_weights else 1,
        )

        # ==========================================================================================
        # attn_output -> (batch_size, num_heads, query_length, head_dim)
        # ==========================================================================================

        batch_size = attn_output.shape[0]
        attn_output = attn_output.transpose(1, 2)
        attn_output = attn_output.reshape(batch_size, -1, self.num_heads * self.head_dim)

        # ==========================================================================================
        # attn_output -> (batch_size, query_length, num_heads * head_dim)
        # ==========================================================================================

        attn_output = self.c_proj(attn_output)
        attn_output = self.resid_dropout(attn_output)

        return attn_output


class GraniteFlashAttention2(GraniteAttention):
    def forward(
        self,
        hidden_states: torch.Tensor,
        past_key_values: Optional[DynamicCache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        rope_cos_sin: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
    ) -> torch.Tensor:
        # ==========================================================================================
        # hidden_states -> (batch_size, query_length, num_heads * head_dim)
        # ==========================================================================================

        query, key, value = self._prepare_qkv_for_forward(hidden_states)

        # ==========================================================================================
        # query -> (batch_size, num_heads, query_length, head_dim)
        # key -> (batch_size, num_key_value_heads, query_length, head_dim)
        # value -> (batch_size, num_key_value_heads, query_length, head_dim)
        # ==========================================================================================

        if self.position_embedding_type == PositionEmbeddingType.rope:
            query = apply_rotary_pos_emb(query, rope_cos_sin)
            key = apply_rotary_pos_emb(key, rope_cos_sin)

        if past_key_values is not None:
            key, value = past_key_values.update(key, value, self.layer_idx)

        # ==========================================================================================
        # query -> (batch_size, num_heads, query_length, head_dim)
        # key -> (batch_size, num_key_value_heads, key_length, head_dim)
        # value -> (batch_size, num_key_value_heads, key_length, head_dim)
        # ==========================================================================================

        # TODO avoid this extra transpose
        query = query.transpose(1, 2)
        if self.attention_head_type == AttentionHeadType.mqa:
            key = key.squeeze(1).unsqueeze(2)
            value = value.squeeze(1).unsqueeze(2)
        else:
            key = key.transpose(1, 2)
            value = value.transpose(1, 2)

        # ==========================================================================================
        # query -> (batch_size, query_length, num_heads, head_dim)
        # key -> (batch_size, key_length, num_heads, head_dim)
        # value -> (batch_size, key_length, num_heads, head_dim)
        # ==========================================================================================

        batch_size, query_length = query.shape[:2]
        key_length = key.shape[1]
        indices_k, cu_seqlens_k, max_seqlen_k = get_unpad_data(attention_mask)

        key = IndexFirstAxis.apply(
            key.reshape(batch_size * key_length, self.num_key_value_heads, self.head_dim), indices_k
        )
        value = IndexFirstAxis.apply(
            value.reshape(batch_size * key_length, self.num_key_value_heads, self.head_dim), indices_k
        )

        if query_length == key_length:
            query = IndexFirstAxis.apply(
                query.reshape(batch_size * key_length, self.num_heads, self.head_dim), indices_k
            )
            cu_seqlens_q = cu_seqlens_k
            max_seqlen_q = max_seqlen_k
            indices_q = indices_k
        elif query_length == 1:
            max_seqlen_q = 1
            cu_seqlens_q = torch.arange(
                batch_size + 1, dtype=torch.int32, device=query.device
            )  # There is a memcpy here, that is very bad.
            indices_q = cu_seqlens_q[:-1]
            query = query.squeeze(1)
        else:
            # The -q_len: slice assumes left padding.
            attention_mask = attention_mask[:, -query_length:]
            query, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(query, attention_mask)

        # ==========================================================================================
        # query -> (total_q, num_heads, head_dim)
        # key -> (total_q, num_heads, head_dim)
        # value -> (total_q, num_heads, head_dim)
        # ==========================================================================================

        attn_output = flash_attn_varlen_func(
            query,
            key,
            value,
            cu_seqlens_q=cu_seqlens_q,
            cu_seqlens_k=cu_seqlens_k,
            max_seqlen_q=max_seqlen_q,
            max_seqlen_k=max_seqlen_k,
            dropout_p=self.attn_pdrop if self.training else 0,
            softmax_scale=self.attention_multiplier if self.scale_attn_weights else 1,
            causal=self.causal,
        )

        # ==========================================================================================
        # attn_output -> (total_q, num_heads, head_dim)
        # ==========================================================================================

        attn_output = pad_input(attn_output, indices_q, batch_size, query_length)
        attn_output = attn_output.view(batch_size, query_length, -1)

        # ==========================================================================================
        # attn_output -> (batch_size, query_length, num_heads * head_dim)
        # ==========================================================================================

        attn_output = self.c_proj(attn_output)
        attn_output = self.resid_dropout(attn_output)

        return attn_output


_ATTENTION_MODULES = {
    "eager": GraniteAttention,
    "sdpa": GraniteSDPA,
    "flash_attention_2": GraniteFlashAttention2,
}


def get_attention_module(
    config: GraniteConfig, causal: bool, attention_implementation: str, layer_idx: int
) -> GraniteAttention:
    if attention_implementation in _ATTENTION_MODULES:
        return _ATTENTION_MODULES[attention_implementation](config, causal=causal, layer_idx=layer_idx)
    raise ValueError(f"unexpected `attention_implementation` {attention_implementation}")


##################################################
# position embeddings


class Alibi(nn.Module):
    def __init__(self, num_heads: int) -> None:
        super().__init__()
        self.num_heads = num_heads

        self.reset_parameters()

    def forward(
        self, attention_mask: torch.Tensor, batch_size: int, key_length: int, device: torch.device, dtype: torch.dtype
    ) -> torch.Tensor:
        """
        Link to paper: https://arxiv.org/abs/2108.12409 Alibi tensor is not causal as the original paper mentions, it
        relies on a translation invariance of softmax for quick implementation: with l being a tensor, and a fixed value
        `softmax(l+a) = softmax(l)`. Based on
        https://github.com/ofirpress/attention_with_linear_biases/blob/a35aaca144e0eb6b789dfcb46784c4b8e31b7983/fairseq/models/transformer.py#L742
        TODO @thomasw21 this doesn't work as nicely due to the masking strategy, and so masking varies slightly.

        Args:
            attention_mask (torch.Tensor): attention_mask tensor of shape (`batch_size`, `key_length`)
            num_heads (int): `num_heads` for the model
            batch_size (int): `batch_size`
            key_length (int): `key_length`
            device (torch.device): device for the tensors
            dtype (torch.dtype): dtype to use for the tensors

        Returns:
            torch.Tensor: alibi tensor of shape (`batch_size`, `num_heads`, `key_length`)
        """

        # Note: alibi will added to the attention bias that will be applied to the query, key product of attention
        # => therefore alibi will have to be of shape (batch_size, num_heads, query_length, key_length)
        # => here we set (batch_size=1, num_heads=num_heads, query_length=1, key_length=max_length)
        # => the query_length dimension will then be broadcasted correctly
        # This is more or less identical to T5's relative position bias:
        # https://github.com/huggingface/transformers/blob/f681437203baa7671de3174b0fa583c349d9d5e1/src/transformers/models/t5/modeling_t5.py#L527
        if attention_mask is None:
            arange_tensor = (
                torch.arange(key_length, device=device).unsqueeze(0).unsqueeze(0).expand(batch_size, -1, -1)
            )
        else:
            arange_tensor = (attention_mask.cumsum(dim=-1) - 1).masked_fill_(attention_mask == 0, 0).unsqueeze(1)

        alibi = self.slopes.unsqueeze(1) * arange_tensor
        return alibi.to(dtype)

    def reset_parameters(self) -> None:
        closest_power_of_2 = 2 ** math.floor(math.log2(self.num_heads))
        base = torch.tensor(2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))), dtype=torch.float32)
        powers = torch.arange(1, 1 + closest_power_of_2, dtype=torch.int32)
        slopes = torch.pow(base, powers)

        if closest_power_of_2 != self.num_heads:
            extra_base = torch.tensor(2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))), dtype=torch.float32)
            num_remaining_heads = min(closest_power_of_2, self.num_heads - closest_power_of_2)
            extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, dtype=torch.int32)
            slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)

        self.register_buffer("slopes", slopes, persistent=False)


class RoPE(nn.Module):
    def __init__(
        self,
        head_dim: int,
        max_position_embeddings: int = 2048,
        base: int = 10000,
    ) -> None:
        super().__init__()

        self.head_dim = head_dim
        self.max_position_embeddings = max_position_embeddings
        self.base = base
        self.mscale = 1

        self.reset_parameters()

    def forward(self, seq_len: int, dtype: torch.dtype, device: torch.device) -> Tuple[torch.Tensor, torch.Tensor]:
        if seq_len > self.max_seq_len_cached:
            self._set_cos_sin_cache(seq_len=seq_len, device=device, dtype=dtype)

        cos = self.cos_cached[:seq_len].to(dtype)
        sin = self.sin_cached[:seq_len].to(dtype)

        return cos, sin

    def reset_parameters(self) -> None:
        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.head_dim, 2).float() / self.head_dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

        # Build here to make `torch.jit.trace` work.
        self._set_cos_sin_cache(
            seq_len=self.max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
        )

    @torch.no_grad()
    def _set_cos_sin_cache(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> None:
        self.max_seq_len_cached = seq_len
        t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)

        freqs = torch.outer(t, self.inv_freq)
        # Different from paper, but it uses a different permutation in order to obtain the same calculation
        emb = torch.cat((freqs, freqs), dim=-1)

        self.register_buffer("cos_cached", (emb.cos() * self.mscale).to(dtype), persistent=False)
        self.register_buffer("sin_cached", (emb.sin() * self.mscale).to(dtype), persistent=False)


def apply_rotary_pos_emb(x: torch.Tensor, cos_sin: Tuple[torch.Tensor, torch.Tensor]) -> torch.Tensor:
    cos, sin = cos_sin
    x = (x * cos) + (_rotate_half(x) * sin)
    return x


def _rotate_half(x: torch.Tensor) -> torch.Tensor:
    x1, x2 = torch.chunk(x, 2, dim=-1)
    return torch.cat((-x2, x1), dim=-1)


##################################################
# MLP


class GraniteMLP(nn.Module):
    def __init__(self, config: GraniteConfig) -> None:
        super().__init__()

        hidden_size = config.n_embd
        intermediate_size = config.n_inner
        activation_function = config.activation_function
        add_bias = config.add_bias
        residual_dropout = config.resid_pdrop

        self.c_fc = nn.Linear(
            hidden_size,
            2 * intermediate_size if is_glu(activation_function) else intermediate_size,
            bias=add_bias,
        )
        self.act = get_activation_function(activation_function)
        self.c_proj = nn.Linear(intermediate_size, hidden_size, bias=add_bias)
        self.dropout = nn.Identity() if residual_dropout == 0 else nn.Dropout(residual_dropout)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.c_fc(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.c_proj(hidden_states)
        hidden_states = self.dropout(hidden_states)
        return hidden_states


##################################################
# transformer layer


class GraniteBlock(nn.Module):
    def __init__(
        self,
        config: GraniteConfig,
        attention_implementation: str,
        layer_idx: Optional[int] = None,
    ) -> None:
        super().__init__()

        hidden_size = config.hidden_size
        self.inner_dim = config.n_inner
        self.layer_idx = layer_idx

        self.ln_1 = get_normalization_function(
            config.normalization_function,
            hidden_size,
            eps=config.layer_norm_epsilon,
        )
        self.attn = get_attention_module(config, True, attention_implementation, layer_idx)
        self.ln_2 = get_normalization_function(
            config.normalization_function,
            hidden_size,
            eps=config.layer_norm_epsilon,
        )
        self.mlp = GraniteMLP(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        past_key_values: Optional[DynamicCache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        rope_cos_sin: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.ln_1(hidden_states)

        attn_output = self.attn(
            hidden_states,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            rope_cos_sin=rope_cos_sin,
        )

        # residual connection
        hidden_states = attn_output + residual

        residual = hidden_states
        hidden_states = self.ln_2(hidden_states)

        feed_forward_hidden_states = self.mlp(hidden_states)

        # residual connection
        hidden_states = residual + feed_forward_hidden_states

        return hidden_states


##################################################
# model classes


class GranitePreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = GraniteConfig
    base_model_prefix = "transformer"
    causal = True
    _no_split_modules = ["GraniteBlock"]
    _skip_keys_device_placement = "past_key_values"
    _supports_sdpa = True
    _supports_flash_attn_2 = True

    def __init__(self, config: GraniteConfig, *inputs, **kwargs):
        super().__init__(config, *inputs, **kwargs)

        self.attention_implementation = self.config._attn_implementation
        self._use_eager_attention = self.attention_implementation == "eager"
        self._use_sdpa = self.attention_implementation == "sdpa"
        self._use_flash_attention_2 = self.attention_implementation == "flash_attention_2"

        self.initializer_range = config.initializer_range

    def _init_weights(self, module: nn.Module) -> None:
        if isinstance(module, (nn.LayerNorm, RMSNorm, Alibi, RoPE)):
            module.reset_parameters()
        elif isinstance(module, nn.Linear):
            nn.init.normal_(module.weight, mean=0, std=self.initializer_range)
            if module.bias is not None:
                module.bias.zero_()
        elif isinstance(module, nn.Embedding):
            nn.init.normal_(module.weight, mean=0, std=self.initializer_range)
            if module.padding_idx is not None:
                module.weight[module.padding_idx].zero_()


class GraniteModel(GranitePreTrainedModel):
    _keys_to_ignore_on_load_missing = ["attn.masked_bias"]
    mask_value = None

    def __init__(self, config: GraniteConfig, **kwargs) -> None:
        super().__init__(config, **kwargs)

        self.attention_head_type = AttentionHeadType(config.attention_head_type)
        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.num_key_value_heads = config.num_key_value_heads

        assert (
            self.embed_dim % self.num_heads == 0
        ), f"`embed_dim` ({self.embed_dim}) must be divisible by `num_heads` ({self.num_heads})"

        self.head_dim = self.embed_dim // self.num_heads

        self.wte = nn.Embedding(config.vocab_size, self.embed_dim)

        self.drop = nn.Identity() if config.embd_pdrop == 0 else nn.Dropout(config.embd_pdrop)
        self.h = nn.ModuleList(
            [GraniteBlock(config, self.attention_implementation, layer_idx=i) for i in range(config.num_hidden_layers)]
        )
        self.ln_f = get_normalization_function(
            config.normalization_function,
            self.embed_dim,
            eps=config.layer_norm_epsilon,
        )

        self.position_embedding_type = PositionEmbeddingType(config.position_embedding_type)

        if self.position_embedding_type == PositionEmbeddingType.learned_absolute:
            self.wpe = nn.Embedding(config.n_positions, self.embed_dim)
        elif self.position_embedding_type == PositionEmbeddingType.alibi:
            assert not self._use_flash_attention_2, "alibi is not implemented with FlashAttention"

            self.alibi = Alibi(self.num_heads)
        elif self.position_embedding_type == PositionEmbeddingType.rope:
            self.rope = RoPE(self.head_dim, max_position_embeddings=config.n_positions, base=config.rope_theta)
        else:
            raise NotImplementedError()

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

    def get_input_embeddings(self) -> nn.Embedding:
        return self.wte

    def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None:
        self.wte = new_embeddings

    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        past_key_values: Optional[DynamicCache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        token_type_ids: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        use_cache: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
        (
            output_hidden_states,
            use_cache,
            return_dict,
            input_shape,
            hidden_states,
            attention_mask,
            position_ids,
            rope_cos_sin,
            past_key_values,
        ) = self._prepare_a_bunch_of_stuff(
            input_ids=input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        # ==========================================================================================
        # flash:
        #     attention_mask -> (batch_size, key_length)
        # else:
        #     attention_mask -> (batch_size, 1, query_length, key_length)
        # ==========================================================================================

        output_shape = input_shape + (hidden_states.size(-1),)

        past_key_values = DynamicCache() if use_cache and past_key_values is None else past_key_values
        all_hidden_states = () if output_hidden_states else None
        for block in self.h:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            hidden_states = block(
                hidden_states,
                past_key_values=past_key_values,
                attention_mask=attention_mask,
                rope_cos_sin=rope_cos_sin,
            )

        hidden_states = self.ln_f(hidden_states)

        hidden_states = hidden_states.view(output_shape)
        # Add last hidden state
        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] if v is not None)

        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
        )

    def _get_position_ids(
        self, attention_mask: torch.Tensor, past_length: int, query_length: int, key_length: int, device: torch.device
    ) -> torch.Tensor:
        if attention_mask is not None and len(attention_mask.shape) == 2:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 0)
            if past_length > 0:
                position_ids = position_ids[:, past_length:key_length:]
        else:
            position_ids = torch.arange(past_length, key_length, dtype=torch.long, device=device)
            position_ids = position_ids.unsqueeze(0).view(-1, query_length)

        return position_ids

    def _get_alibi_bias(
        self,
        attention_mask: torch.Tensor,
        batch_size: int,
        query_length: int,
        key_length: int,
        device: torch.device,
        dtype: torch.dtype,
    ) -> torch.Tensor:
        if self.position_embedding_type != PositionEmbeddingType.alibi:
            return None

        alibi_bias = self.alibi(attention_mask, batch_size, key_length, device, dtype)

        # ==========================================================================================
        # alibi_bias -> (batch_size, num_heads, key_length)
        # ==========================================================================================

        alibi_bias = alibi_bias.unsqueeze(2)
        if query_length != 1:
            alibi_bias = alibi_bias.expand(-1, -1, query_length, -1)

        # ==========================================================================================
        # alibi_bias -> (batch_size, num_heads, query_length, key_length)
        # ==========================================================================================

        return alibi_bias

    def _get_rope_cos_sin(
        self, key_length: int, position_ids: torch.Tensor, dtype: torch.dtype, device: torch.device
    ) -> Optional[Tuple[torch.Tensor, torch.Tensor]]:
        if self.position_embedding_type == PositionEmbeddingType.rope:
            cos, sin = self.rope(key_length, dtype=dtype, device=device)
            cos = cos[position_ids].unsqueeze(1)
            sin = sin[position_ids].unsqueeze(1)
            return cos, sin

    def _prepare_causal_attention_mask(
        self, attention_mask: torch.Tensor, batch_size: int, query_length: int, key_length: int, device: torch.device
    ) -> torch.Tensor:
        past_length = key_length - query_length

        # ==========================================================================================
        # attention_mask -> (batch_size, key_length)
        # ==========================================================================================

        if query_length > 1:
            # (query_length, key_length)
            causal_mask = torch.empty((query_length, key_length), dtype=torch.bool, device=device)
            causal_mask[:, past_length:] = torch.tril(
                torch.ones(query_length, query_length, dtype=torch.bool, device=device)
            )

            if past_length > 0:
                causal_mask[:, :past_length] = True

            # (query_length, key_length) -> (1, query_length, key_length)
            causal_mask = causal_mask.unsqueeze(0)

            if attention_mask is None:
                # (1, query_length, key_length) -> (batch_size, query_length, key_length)
                causal_mask = causal_mask.expand(batch_size, -1, -1)
            else:
                # (1, query_length, key_length) & (batch_size, 1, key_length) -> (batch_size, query_length, key_length)
                causal_mask = causal_mask & attention_mask.unsqueeze(1).to(torch.bool)
        else:
            if attention_mask is None:
                # (batch_size, query_length, key_length)
                causal_mask = torch.ones(batch_size, query_length, key_length, dtype=torch.bool, device=device)
            else:
                # (batch_size, query_length, key_length)
                causal_mask = attention_mask.unsqueeze(1).to(dtype=torch.bool, device=device)

        # ==========================================================================================
        # attention_mask -> (batch_size, query_length, key_length)
        # ==========================================================================================

        causal_mask = causal_mask.unsqueeze(1)

        # ==========================================================================================
        # attention_mask -> (batch_size, 1, query_length, key_length)
        # ==========================================================================================

        return causal_mask

    def _get_initial_hidden_state(
        self,
        input_ids: torch.Tensor,
        inputs_embeds: torch.Tensor,
        position_ids: torch.Tensor,
        token_type_ids: torch.Tensor,
    ) -> torch.Tensor:
        if inputs_embeds is None:
            inputs_embeds = self.wte(input_ids)

        if self.position_embedding_type == PositionEmbeddingType.learned_absolute:
            inputs_embeds = inputs_embeds + self.wpe(position_ids)

        if token_type_ids is not None:
            inputs_embeds = inputs_embeds + self.wte(token_type_ids)

        inputs_embeds = self.drop(inputs_embeds)

        return inputs_embeds

    def _prepare_a_bunch_of_stuff(
        self,
        input_ids: torch.Tensor,
        past_key_values: DynamicCache,
        attention_mask: torch.Tensor,
        token_type_ids: torch.Tensor,
        position_ids: torch.Tensor,
        inputs_embeds: torch.Tensor,
        use_cache: bool,
        output_hidden_states: bool,
        return_dict: bool,
    ) -> Tuple[
        bool,
        bool,
        bool,
        torch.Size,
        torch.Tensor,
        torch.Tensor,
        torch.Tensor,
        Optional[Tuple[torch.Tensor, torch.Tensor]],
        DynamicCache,
    ]:
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )

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

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.size()
        elif inputs_embeds is not None:
            # TODO special handling for padding free transformer needed here if we support inputs_embeds argument
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        batch_size = input_shape[0]
        device = input_ids.device if input_ids is not None else inputs_embeds.device

        if self.position_embedding_type == PositionEmbeddingType.alibi:
            if position_ids is not None:
                warnings.warn("`position_ids` have no functionality with Alibi.", FutureWarning)

        if token_type_ids is not None:
            token_type_ids = token_type_ids.view(-1, input_shape[-1])

        # ==========================================================================================
        # input_ids -> (batch_size, query_length)
        # attention_mask -> None or (batch_size, key_length)
        # position_ids -> None or (batch_size, key_length)
        # ==========================================================================================

        past_length = 0 if past_key_values is None else past_key_values.get_seq_length()
        query_length = input_shape[-1]
        key_length = past_length + query_length

        if position_ids is None:
            position_ids = self._get_position_ids(attention_mask, past_length, query_length, key_length, device)

        # ==========================================================================================
        # input_ids -> (batch_size, query_length)
        # attention_mask -> None or (batch_size, key_length)
        # position_ids -> (batch_size, query_length)
        # ==========================================================================================

        hidden_states = self._get_initial_hidden_state(input_ids, inputs_embeds, position_ids, token_type_ids)

        # ==========================================================================================
        # hidden_states -> (batch_size, query_length, num_heads * head_dim)
        # ==========================================================================================

        alibi_bias = self._get_alibi_bias(
            attention_mask, batch_size, query_length, key_length, device, hidden_states.dtype
        )

        # ==========================================================================================
        # alibi_bias -> (batch_size, num_heads, query_length, key_length)
        # ==========================================================================================

        rope_cos_sin = self._get_rope_cos_sin(
            key_length, position_ids, dtype=hidden_states.dtype, device=hidden_states.device
        )

        # ==========================================================================================
        # rope_cos_sin -> 2 * (key_length, head_dim)
        # ==========================================================================================

        # prepare causal mask only if not using flash attention
        if self._use_flash_attention_2:
            if attention_mask is None:
                attention_mask = torch.ones_like(input_ids)
        elif self._use_sdpa:
            # we use the causal/non-causal argument of SDPA for attention in this case
            if attention_mask is not None:
                attention_mask = self._prepare_causal_attention_mask(
                    attention_mask, batch_size, query_length, key_length, device
                )

                attention_mask = torch.where(
                    attention_mask,
                    ~attention_mask if alibi_bias is None else alibi_bias,
                    self._get_mask_value(attention_mask.device, hidden_states.dtype),
                )
        else:
            attention_mask = self._prepare_causal_attention_mask(
                attention_mask, batch_size, query_length, key_length, device
            )

            attention_mask = torch.where(
                attention_mask,
                ~attention_mask if alibi_bias is None else alibi_bias,
                self._get_mask_value(attention_mask.device, hidden_states.dtype),
            )

        return (
            output_hidden_states,
            use_cache,
            return_dict,
            input_shape,
            hidden_states,
            attention_mask,
            position_ids,
            rope_cos_sin,
            past_key_values,
        )

    def _get_mask_value(self, device: torch.device, dtype: torch.dtype) -> torch.Tensor:
        # torch.where expects a tensor. We use a cache to avoid recreating it every time.
        if self.mask_value is None or self.mask_value.dtype != dtype or self.mask_value.device != device:
            self.mask_value = torch.full([], torch.finfo(torch.float16).min, dtype=dtype, device=device)
        return self.mask_value


class GraniteForCausalLM(GranitePreTrainedModel):
    _keys_to_ignore_on_load_missing = ["lm_head.weight"]

    def __init__(self, config: GraniteConfig, **kwargs) -> None:
        super().__init__(config, **kwargs)
        self.transformer = GraniteModel(config, **kwargs)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)

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

    def get_input_embeddings(self) -> nn.Embedding:
        return self.transformer.wte

    def set_input_embeddings(self, value: nn.Embedding) -> None:
        self.transformer.wte = value

    def get_output_embeddings(self) -> nn.Linear:
        return self.lm_head

    def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
        self.lm_head = new_embeddings

    # FIXME typing
    def prepare_inputs_for_generation(
        self,
        input_ids: torch.Tensor,
        past_key_values: Optional[DynamicCache] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> dict:
        token_type_ids = kwargs.get("token_type_ids", None)
        # Omit tokens covered by past_key_values
        if past_key_values:
            past_length = past_key_values.get_seq_length()

            # Some generation methods already pass only the last input ID
            if input_ids.shape[1] > past_length:
                remove_prefix_length = past_length
            else:
                # Default to old behavior: keep only final ID
                remove_prefix_length = input_ids.shape[1] - 1

            input_ids = input_ids[:, remove_prefix_length:]
            if token_type_ids is not None:
                token_type_ids = token_type_ids[:, -input_ids.shape[1] :]

        attention_mask = kwargs.get("attention_mask", None)
        position_ids = kwargs.get("position_ids", None)

        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 0)
            if past_key_values:
                position_ids = position_ids[:, -input_ids.shape[1] :]
        else:
            position_ids = None

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs.update(
            {
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "position_ids": position_ids,
                "attention_mask": attention_mask,
                "token_type_ids": token_type_ids,
            }
        )
        return model_inputs

    def forward(
        self,
        input_ids: Optional[Union[torch.Tensor]] = None,
        past_key_values: Optional[DynamicCache] = None,
        attention_mask: Optional[torch.Tensor] = None,
        token_type_ids: Optional[Union[torch.Tensor]] = None,
        position_ids: Optional[Union[torch.Tensor]] = None,
        inputs_embeds: Optional[Union[torch.Tensor]] = None,
        labels: Optional[Union[torch.Tensor]] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # ==========================================================================================
        # input_ids -> (batch_size, query_length)
        # attention_mask -> None or (batch_size, key_length)
        # position_ids -> None or (batch_size, key_length)
        # ==========================================================================================

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = transformer_outputs[0]

        lm_logits = self.lm_head(hidden_states)

        loss = None
        # Shift so that tokens < n predict n
        if labels is not None:
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous().to(shift_logits.device)

            # Flatten the tokens
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

        if not return_dict:
            output = (lm_logits,) + transformer_outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return CausalLMOutputWithCrossAttentions(
            loss=loss,
            logits=lm_logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )