modeling_jamba.py

# coding=utf-8
# Copyright 2024 AI21 Labs Ltd. and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# 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.
""" PyTorch Jamba model."""
import inspect
import math
import warnings
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Tuple, Union

import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_attn_mask_utils import (
    _prepare_4d_causal_attention_mask,
    _prepare_4d_causal_attention_mask_for_sdpa,
)
from transformers.modeling_outputs import (
    MoeCausalLMOutputWithPast,
    MoeModelOutputWithPast,
    SequenceClassifierOutputWithPast,
)
from transformers.modeling_utils import PreTrainedModel
from transformers.pytorch_utils import is_torch_greater_or_equal_than_1_13
from transformers.utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_flash_attn_greater_or_equal_2_10,
    logging,
    replace_return_docstrings,
)
from transformers.utils.import_utils import is_torch_fx_available
from .configuration_jamba import JambaConfig


# try except block so it'll work with trust_remote_code. Later we can have `if is_flash_attn_2_available():`
try:
    from flash_attn import flash_attn_func, flash_attn_varlen_func
    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa

    _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
except ImportError:
    pass

# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
# It means that the function will not be traced through and simply appear as a node in the graph.
if is_torch_fx_available():
    if not is_torch_greater_or_equal_than_1_13:
        import torch.fx

    _prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)

# try except block so it'll work with trust_remote_code. Later we can have `if is_mamba_ssm_available():`
try:
    from mamba_ssm.ops.selective_scan_interface import mamba_inner_fn, selective_scan_fn
    from mamba_ssm.ops.triton.selective_state_update import selective_state_update
except ImportError:
    selective_state_update, selective_scan_fn, mamba_inner_fn = None, None, None

# try except block so it'll work with trust_remote_code. Later we can have `if is_causal_conv1d_available():`
try:
    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
except ImportError:
    causal_conv1d_update, causal_conv1d_fn = None, None

is_fast_path_available = all(
    (selective_state_update, selective_scan_fn, causal_conv1d_fn, causal_conv1d_update, mamba_inner_fn)
)


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "JambaConfig"


# Adapted from transformers.models.mixtral.modeling_mixtral.load_balancing_loss_func
def load_balancing_loss_func(
        gate_logits: torch.Tensor, num_experts: torch.Tensor = None, top_k=2, attention_mask: Optional[torch.Tensor] = None
) -> float:
    r"""
    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.

    See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss
    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
    experts is too unbalanced.

    Args:
        gate_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]):
            Logits from the `router`, should be a tuple of model.config.num_hidden_layers tensors of
            shape [batch_size X sequence_length, num_experts].
        attention_mask (`torch.Tensor`, None):
            The attention_mask used in forward function
            shape [batch_size X sequence_length] if not None.
        num_experts (`int`, *optional*):
            Number of experts

    Returns:
        The auxiliary loss.
    """
    if gate_logits is None or not isinstance(gate_logits, tuple):
        return 0

    if isinstance(gate_logits, tuple):
        compute_device = gate_logits[0].device
        concatenated_gate_logits = torch.cat(
            [layer_gate.to(compute_device) for layer_gate in gate_logits if layer_gate.shape[1] > 1], dim=0
        )

    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)

    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)

    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)

    if attention_mask is None:
        # Compute the percentage of tokens routed to each experts
        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)

        # Compute the average probability of routing to these experts
        router_prob_per_expert = torch.mean(routing_weights, dim=0)
    else:
        batch_size, sequence_length = attention_mask.shape
        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)

        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
        expert_attention_mask = (
            attention_mask[None, :, :, None, None]
                .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
                .reshape(-1, top_k, num_experts)
                .to(compute_device)
        )

        # Compute the percentage of tokens routed to each experts
        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
            expert_attention_mask, dim=0
        )

        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
        router_per_expert_attention_mask = (
            attention_mask[None, :, :, None]
                .expand((num_hidden_layers, batch_size, sequence_length, num_experts))
                .reshape(-1, num_experts)
                .to(compute_device)
        )

        # Compute the average probability of routing to these experts
        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
            router_per_expert_attention_mask, dim=0
        )

    overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
    return overall_loss * num_experts


# Copied from transformers.models.llama.modeling_llama._get_unpad_data
def _get_unpad_data(attention_mask):
    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,
    )


# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Jamba
class JambaRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        JambaRMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)


# Copied from transformers.models.llama.modeling_llama.repeat_kv
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


# Adapted from transformers.models.mistral.modeling_mistral.MistralAttention with Mistral->Jamba
class JambaAttention(nn.Module):
    """
    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
    and "Generating Long Sequences with Sparse Transformers".
    """

    def __init__(self, config: JambaConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing a `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.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.is_causal = True
        self.attention_dropout = config.attention_dropout

        if (self.head_dim * self.num_heads) != self.hidden_size:
            raise ValueError(
                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                f" and `num_heads`: {self.num_heads})."
            )
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)

    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()

    def forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[Cache] = None,
            output_attentions: bool = False,
            use_cache: bool = False,
            **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError(
                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                    "with a layer index."
                )
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        if past_key_value is not None:
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)

        # repeat k/v heads if n_kv_heads < n_heads
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)

        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
                f" {attn_weights.size()}"
            )

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                )

            attn_weights = attn_weights + attention_mask

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, value_states)

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

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)

        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


# Adapted from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with Mistral->Jamba
class JambaFlashAttention2(JambaAttention):
    """
    Jamba flash attention module. This module inherits from `JambaAttention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    """

    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()

    def forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[Cache] = None,
            output_attentions: bool = False,
            use_cache: bool = False,
            **kwargs,
    ):
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )

            # overwrite attention_mask with padding_mask
            attention_mask = kwargs.pop("padding_mask")
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError(
                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                    "with a layer index."
                )
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        use_sliding_windows = (
                _flash_supports_window_size
                and getattr(self.config, "sliding_window", None) is not None
                and kv_seq_len > self.config.sliding_window
        )

        if not _flash_supports_window_size:
            logger.warning_once(
                "The current flash attention version does not support sliding window attention, for a more memory efficient implementation"
                " make sure to upgrade flash-attn library."
            )

        if past_key_value is not None:
            # Activate slicing cache only if the config has a value `sliding_windows` attribute
            cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
            if (
                    getattr(self.config, "sliding_window", None) is not None
                    and kv_seq_len > self.config.sliding_window
                    and cache_has_contents
            ):
                slicing_tokens = 1 - self.config.sliding_window

                past_key = past_key_value[self.layer_idx][0]
                past_value = past_key_value[self.layer_idx][1]

                past_key = past_key[:, :, slicing_tokens:, :].contiguous()
                past_value = past_value[:, :, slicing_tokens:, :].contiguous()

                if past_key.shape[-2] != self.config.sliding_window - 1:
                    raise ValueError(
                        f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
                        f" {past_key.shape}"
                    )

                if attention_mask is not None:
                    attention_mask = attention_mask[:, slicing_tokens:]
                    attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)

            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)

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

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

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

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

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

        attn_output = self._flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=dropout_rate,
            use_sliding_windows=use_sliding_windows,
        )

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

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value

    def _flash_attention_forward(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            query_length,
            dropout=0.0,
            softmax_scale=None,
            use_sliding_windows=False,
    ):
        """
        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
        first unpad the input, then computes the attention scores and pad the final attention scores.

        Args:
            query_states (`torch.Tensor`):
                Input query states to be passed to Flash Attention API
            key_states (`torch.Tensor`):
                Input key states to be passed to Flash Attention API
            value_states (`torch.Tensor`):
                Input value states to be passed to Flash Attention API
            attention_mask (`torch.Tensor`):
                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                position of padding tokens and 1 for the position of non-padding tokens.
            dropout (`int`, *optional*):
                Attention dropout
            softmax_scale (`float`, *optional*):
                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
            use_sliding_windows (`bool`, *optional*):
                Whether to activate sliding window attention.
        """
        if not self._flash_attn_uses_top_left_mask:
            causal = self.is_causal
        else:
            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
            causal = self.is_causal and query_length != 1

        # Contains at least one padding token in the sequence
        if attention_mask is not None:
            batch_size = query_states.shape[0]
            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
                query_states, key_states, value_states, attention_mask, query_length
            )

            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens

            if not use_sliding_windows:
                attn_output_unpad = flash_attn_varlen_func(
                    query_states,
                    key_states,
                    value_states,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_k=cu_seqlens_k,
                    max_seqlen_q=max_seqlen_in_batch_q,
                    max_seqlen_k=max_seqlen_in_batch_k,
                    dropout_p=dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                )
            else:
                attn_output_unpad = flash_attn_varlen_func(
                    query_states,
                    key_states,
                    value_states,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_k=cu_seqlens_k,
                    max_seqlen_q=max_seqlen_in_batch_q,
                    max_seqlen_k=max_seqlen_in_batch_k,
                    dropout_p=dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                    window_size=(self.config.sliding_window, self.config.sliding_window),
                )

            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
        else:
            if not use_sliding_windows:
                attn_output = flash_attn_func(
                    query_states,
                    key_states,
                    value_states,
                    dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                )
            else:
                attn_output = flash_attn_func(
                    query_states,
                    key_states,
                    value_states,
                    dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                    window_size=(self.config.sliding_window, self.config.sliding_window),
                )

        return attn_output

    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
        batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape

        # On the first iteration we need to properly re-create the padding mask
        # by slicing it on the proper place
        if kv_seq_len != attention_mask.shape[-1]:
            attention_mask_num_tokens = attention_mask.shape[-1]
            attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]

        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)

        key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
        value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)

        if query_length == kv_seq_len:
            query_layer = index_first_axis(
                query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
            )
            cu_seqlens_q = cu_seqlens_k
            max_seqlen_in_batch_q = max_seqlen_in_batch_k
            indices_q = indices_k
        elif query_length == 1:
            max_seqlen_in_batch_q = 1
            cu_seqlens_q = torch.arange(
                batch_size + 1, dtype=torch.int32, device=query_layer.device
            )  # There is a memcpy here, that is very bad.
            indices_q = cu_seqlens_q[:-1]
            query_layer = query_layer.squeeze(1)
        else:
            # The -q_len: slice assumes left padding.
            attention_mask = attention_mask[:, -query_length:]
            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)

        return (
            query_layer,
            key_layer,
            value_layer,
            indices_q,
            (cu_seqlens_q, cu_seqlens_k),
            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
        )


# Adapted from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Jamba
class JambaSdpaAttention(JambaAttention):
    """
    Jamba attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
    `JambaAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
    SDPA API.
    """

    # Adapted from JambaAttention.forward
    def forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[Cache] = None,
            output_attentions: bool = False,
            use_cache: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if output_attentions:
            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
            logger.warning_once(
                "JambaModel is using JambaSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
            )
            return super().forward(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
            )

        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        if past_key_value is not None:
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)

        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                )

        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
        # Reference: https://github.com/pytorch/pytorch/issues/112577.
        if query_states.device.type == "cuda" and attention_mask is not None:
            query_states = query_states.contiguous()
            key_states = key_states.contiguous()
            value_states = value_states.contiguous()

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=attention_mask,
            dropout_p=self.attention_dropout if self.training else 0.0,
            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
            is_causal=self.is_causal and attention_mask is None and q_len > 1,
        )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.view(bsz, q_len, self.hidden_size)

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value


JAMBA_ATTENTION_CLASSES = {
    "eager": JambaAttention,
    "flash_attention_2": JambaFlashAttention2,
    "sdpa": JambaSdpaAttention,
}


class HybridMambaAttentionDynamicCache(DynamicCache):
    """
    A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache
    (which has a constant shape regardless of seq_len).

    It stores the Key and Value states as a list of tensors, one for each layer.
    The expected shape for each tensor for attention layers is `[batch_size, num_heads, seq_len, head_dim]`.
    For the mamba layers, the `key_cache` represents the convolution state and has a shape of `[batch_size, d_inner, 1, d_conv]`,
    and the `value_cache` represents the ssm state and has a shape of `[batch_size, d_inner, 1, d_state]`. Mamba cache
    shape[2] is a dummy "seqlen" dimension to match the number of attention cache dimensions. For mamba, the cache
    doesn't grow with seqlen so this dimension is always 1.
    """

    def __init__(self) -> None:
        super().__init__()
        self.attention_layer_idx = None  # used to know which layer has data on seqlen in the cache shape

    def update(
            self,
            key_states: torch.Tensor,
            value_states: torch.Tensor,
            layer_idx: int,
            cache_kwargs: Optional[Dict[str, Any]] = None,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Updates the cache with the new `key_states` and `value_states` for the layer `layer_idx`.

        Parameters:
            key_states (`torch.Tensor`):
                The new key states to cache.
            value_states (`torch.Tensor`):
                The new value states to cache.
            layer_idx (`int`):
                The index of the layer to cache the states for.
            cache_kwargs (`Dict[str, Any]`, `optional`):
                Additional arguments for the cache subclass. No additional arguments are used in `HybridMambaAttentionDynamicCache`.

        Return:
            A tuple containing the updated key and value states.
        """
        # Update the number of seen tokens
        if self.attention_layer_idx is None and self._is_attn_layer(key_states, value_states):
            self.attention_layer_idx = layer_idx
        if self.attention_layer_idx is not None and layer_idx == self.attention_layer_idx:
            if hasattr(self, "_seen_tokens"):
                self._seen_tokens += key_states.shape[-2]
            else:
                self.seen_tokens += key_states.shape[-2]

        # Update the cache
        if len(self.key_cache) <= layer_idx:
            self.key_cache.append(key_states)
            self.value_cache.append(value_states)
        else:
            if self._is_attn_layer(self.key_cache[layer_idx], self.value_cache[layer_idx]):
                # attention layer - append the new states to the existing cache on the seqlen dimension
                self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=-2)
                self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=-2)
            else:
                # mamba layer - replace the cache with the new states
                self.key_cache[layer_idx] = key_states
                self.value_cache[layer_idx] = value_states

        return self.key_cache[layer_idx], self.value_cache[layer_idx]

    def get_seq_length(self, layer_idx: Optional[int] = None) -> int:
        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
        if layer_idx is not None:
            if len(self.key_cache) <= layer_idx:
                return 0
            if self._is_attn_layer(self.key_cache[layer_idx], self.value_cache[layer_idx]):
                return self.key_cache[layer_idx].shape[-2]
            else:
                warnings.warn(
                    f"Asked to get the sequence length from cache of layer {layer_idx} which is not an attention layer. "
                    f"Ignoring that and using an attention layer cache"
                )
        if self.attention_layer_idx is None or len(self.key_cache) <= self.attention_layer_idx:
            return 0
        return self.key_cache[self.attention_layer_idx].shape[-2]

    @staticmethod
    def _is_attn_layer(key_states: torch.Tensor, value_states: torch.Tensor):
        return key_states.shape[-1] == value_states.shape[-1]


@dataclass
class MambaCacheParams:
    seqlen_offset: int = 0
    conv_states: Dict[int, torch.Tensor] = field(default_factory=dict)
    ssm_states: Dict[int, torch.Tensor] = field(default_factory=dict)


# Adapted from transformers.models.mamba.modeling_mamba.MambaMixer
class JambaMambaMixer(nn.Module):
    """
    Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`.
    A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective)
    ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4,
    and is why Mamba is called **selective** state spaces)
    """

    def __init__(self, config: JambaConfig, layer_idx):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.hidden_size = config.hidden_size
        self.ssm_state_size = config.mamba_d_state
        self.conv_kernel_size = config.mamba_d_conv
        self.intermediate_size = config.mamba_expand * config.hidden_size
        self.time_step_rank = config.mamba_dt_rank
        self.use_conv_bias = config.mamba_conv_bias
        self.use_bias = config.mamba_proj_bias
        self.conv1d = nn.Conv1d(
            in_channels=self.intermediate_size,
            out_channels=self.intermediate_size,
            bias=self.use_conv_bias,
            kernel_size=self.conv_kernel_size,
            groups=self.intermediate_size,
            padding=self.conv_kernel_size - 1,
        )

        self.activation = config.hidden_act
        self.act = ACT2FN[config.hidden_act]
        self.apply_inner_layernorms = config.mamba_inner_layernorms

        self.use_fast_kernels = config.use_mamba_kernels

        # projection of the input hidden states
        self.in_proj = nn.Linear(self.hidden_size, self.intermediate_size * 2, bias=self.use_bias)
        # selective projection used to make dt, B and C input dependant
        self.x_proj = nn.Linear(self.intermediate_size, self.time_step_rank + self.ssm_state_size * 2, bias=False)
        # time step projection (discretization)
        self.dt_proj = nn.Linear(self.time_step_rank, self.intermediate_size, bias=True)

        # S4D real initialization. These are not discretized!
        # The core is to load them, compute the discrete states, then write the updated state. Keeps the memory bounded
        A = torch.arange(1, self.ssm_state_size + 1, dtype=torch.float32)[None, :]
        A = A.expand(self.intermediate_size, -1).contiguous()

        self.A_log = nn.Parameter(torch.log(A))
        self.D = nn.Parameter(torch.ones(self.intermediate_size))
        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=self.use_bias)

        if self.apply_inner_layernorms:
            self.dt_layernorm = JambaRMSNorm(self.time_step_rank, eps=config.rms_norm_eps)
            self.B_layernorm = JambaRMSNorm(self.ssm_state_size, eps=config.rms_norm_eps)
            self.C_layernorm = JambaRMSNorm(self.ssm_state_size, eps=config.rms_norm_eps)
        else:
            self.dt_layernorm = None
            self.B_layernorm = None
            self.C_layernorm = None

        if not is_fast_path_available:
            logger.warning_once(
                "The fast path is not available because on of `(selective_state_update, selective_scan_fn, causal_conv1d_fn, causal_conv1d_update, mamba_inner_fn)`"
                " is None. To install follow https://github.com/state-spaces/mamba/#installation and"
                " https://github.com/Dao-AILab/causal-conv1d. If you want to use the naive implementation, set `use_mamba_kernels=False` in the model config"
            )

    def _apply_layernorms(self, dt, B, C):
        if self.dt_layernorm is not None:
            dt = self.dt_layernorm(dt)
        if self.B_layernorm is not None:
            B = self.B_layernorm(B)
        if self.C_layernorm is not None:
            C = self.C_layernorm(C)
        return dt, B, C

    def cuda_kernels_forward(self, hidden_states: torch.Tensor, cache_params: MambaCacheParams = None):
        # 1. Gated MLP's linear projection
        projected_states = self.in_proj(hidden_states).transpose(1, 2)

        if (
                self.training and cache_params is None and not self.apply_inner_layernorms
        ):  # Doesn't support outputting the states -> used for training
            contextualized_states = mamba_inner_fn(
                projected_states,
                self.conv1d.weight,
                self.conv1d.bias if self.use_conv_bias else None,
                self.x_proj.weight,
                self.dt_proj.weight,
                self.out_proj.weight,
                self.out_proj.bias.float() if self.use_bias else None,
                -torch.exp(self.A_log.float()),
                None,  # input-dependent B
                None,  # input-dependent C
                self.D.float(),
                delta_bias=self.dt_proj.bias.float(),
                delta_softplus=True,
            )

        else:
            hidden_states, gate = projected_states.chunk(2, dim=1)

            # 2. Convolution sequence transformation
            conv_weights = self.conv1d.weight.view(self.conv1d.weight.size(0), self.conv1d.weight.size(2))
            if cache_params is not None and cache_params.seqlen_offset > 0:
                hidden_states = causal_conv1d_update(
                    hidden_states.squeeze(-1),
                    cache_params.conv_states[self.layer_idx],
                    conv_weights,
                    self.conv1d.bias,
                    self.activation,
                )
                hidden_states = hidden_states.unsqueeze(-1)
            else:
                if cache_params is not None:
                    conv_states = nn.functional.pad(
                        hidden_states, (self.conv_kernel_size - hidden_states.shape[-1], 0)
                    )
                    cache_params.conv_states[self.layer_idx].copy_(conv_states)
                hidden_states = causal_conv1d_fn(
                    hidden_states, conv_weights, self.conv1d.bias, activation=self.activation
                )

            # 3. State Space Model sequence transformation
            # 3.a. input varying initialization of time_step, B and C
            ssm_parameters = self.x_proj(hidden_states.transpose(1, 2))
            time_step, B, C = torch.split(
                ssm_parameters, [self.time_step_rank, self.ssm_state_size, self.ssm_state_size], dim=-1
            )
            time_step, B, C = self._apply_layernorms(time_step, B, C)

            # Here we need to apply dt_proj without the bias, as the bias is added in the selective scan kernel.
            # This is a hack to apply dt_proj while still using the forward pass of `torch.nn.Linear`, which is needed
            # in order to make quantization work. Quantization code replaces `torch.nn.Linear` layers with quantized
            # linear layers, and requires to call the forward pass directly.
            # The original code here was: ```discrete_time_step = self.dt_proj.weight @ time_step.transpose(1, 2)```
            if hasattr(self.dt_proj, "base_layer"):
                # In case of LoRA, we need to access the base layer to get the weight
                time_proj_bias = self.dt_proj.base_layer.bias
                self.dt_proj.base_layer.bias = None
            else:
                time_proj_bias = self.dt_proj.bias
                self.dt_proj.bias = None
            discrete_time_step = self.dt_proj(time_step).transpose(1, 2)
            if hasattr(self.dt_proj, "base_layer"):
                self.dt_proj.base_layer.bias = time_proj_bias
            else:
                self.dt_proj.bias = time_proj_bias

            A = -torch.exp(self.A_log.float())
            # 3.c perform the recurrence y ← SSM(A, B, C)(x)
            time_proj_bias = time_proj_bias.float() if time_proj_bias is not None else None
            if cache_params is not None and cache_params.seqlen_offset > 0:
                scan_outputs = selective_state_update(
                    cache_params.ssm_states[self.layer_idx],
                    hidden_states[..., 0],
                    discrete_time_step[..., 0],
                    A,
                    B[:, 0],
                    C[:, 0],
                    self.D,
                    gate[..., 0],
                    time_proj_bias,
                    dt_softplus=True,
                ).unsqueeze(-1)
            else:
                scan_outputs, ssm_state = selective_scan_fn(
                    hidden_states,
                    discrete_time_step,
                    A,
                    B.transpose(1, 2),
                    C.transpose(1, 2),
                    self.D.float(),
                    gate,
                    time_proj_bias,
                    delta_softplus=True,
                    return_last_state=True,
                )
                if ssm_state is not None and cache_params is not None:
                    cache_params.ssm_states[self.layer_idx].copy_(ssm_state)

            # 4. Final linear projection
            contextualized_states = self.out_proj(scan_outputs.transpose(1, 2))
        return contextualized_states

    # fmt: off
    def slow_forward(self, input_states, cache_params: MambaCacheParams = None):
        batch_size, seq_len, _ = input_states.shape
        dtype = input_states.dtype
        # 1. Gated MLP's linear projection
        projected_states = self.in_proj(input_states).transpose(1, 2)                   # [batch, 2 * intermediate_size, seq_len]
        hidden_states, gate = projected_states.chunk(2, dim=1)

        # 2. Convolution sequence transformation
        if cache_params is not None:
            if self.training:
                # In training mode, we don't want to perform in-place operations on ssm_state so we can compute the backwards pass
                ssm_state = cache_params.ssm_states[self.layer_idx].clone()
            else:
                ssm_state = cache_params.ssm_states[self.layer_idx]

            if cache_params.seqlen_offset > 0:
                conv_state = cache_params.conv_states[self.layer_idx]                   # [batch, intermediate_size, conv_kernel_size]
                conv_state = torch.roll(conv_state, shifts=-1, dims=-1)
                conv_state[:, :, -1] = hidden_states[:, :, 0]
                cache_params.conv_states[self.layer_idx].copy_(conv_state)
                hidden_states = torch.sum(conv_state * self.conv1d.weight[:, 0, :], dim=-1)
                if self.use_conv_bias:
                    hidden_states += self.conv1d.bias
                hidden_states = self.act(hidden_states).to(dtype).unsqueeze(-1)         # [batch, intermediate_size, 1] : decoding
            else:
                conv_state = nn.functional.pad(
                    hidden_states,
                    (self.conv_kernel_size - hidden_states.shape[-1], 0)
                )
                cache_params.conv_states[self.layer_idx].copy_(conv_state)
                hidden_states = self.act(self.conv1d(hidden_states)[..., :seq_len])     # [batch, intermediate_size, seq_len]
        else:
            ssm_state = torch.zeros(
                (batch_size, self.intermediate_size, self.ssm_state_size),
                device=hidden_states.device, dtype=dtype
            )
            hidden_states = self.act(self.conv1d(hidden_states)[..., :seq_len])         # [batch, intermediate_size, seq_len]

        # 3. State Space Model sequence transformation
        # 3.a. Selection:  [batch, seq_len, self.time_step_rank + self.ssm_state_size * 2]
        ssm_parameters = self.x_proj(hidden_states.transpose(1, 2))
        time_step, B, C = torch.split(
            ssm_parameters, [self.time_step_rank, self.ssm_state_size, self.ssm_state_size], dim=-1
        )
        time_step, B, C = self._apply_layernorms(time_step, B, C)
        discrete_time_step = self.dt_proj(time_step)                                    # [batch, seq_len, intermediate_size]
        discrete_time_step = nn.functional.softplus(discrete_time_step).transpose(1, 2) # [batch, intermediate_size, seq_len]

        # 3.b. Discretization: B and C to [batch, seq_len, intermediate_size, ssm_state_size] (SRAM)
        A = -torch.exp(self.A_log.float())                                              # [intermediate_size, ssm_state_size]
        discrete_A = torch.exp(A[None, :, None, :] * discrete_time_step[:, :, :, None]) # [batch, intermediate_size, seq_len, ssm_state_size]
        discrete_B = discrete_time_step[:, :, :, None] * B[:, None, :, :].float()       # [batch, intermediade_size, seq_len, ssm_state_size]
        deltaB_u = discrete_B * hidden_states[:, :, :, None].float()

        # 3.c perform the recurrence y ← SSM(A, B, C)(x)
        scan_outputs = []
        for i in range(seq_len):
            ssm_state = discrete_A[:, :, i, :] * ssm_state + deltaB_u[:, :, i, :]      # [batch, intermediade_size, ssm_state]
            scan_output = torch.matmul(ssm_state.to(dtype), C[:, i, :].unsqueeze(-1))  # [batch, intermediade_size, 1]
            scan_outputs.append(scan_output[:, :, 0])
        scan_output = torch.stack(scan_outputs, dim=-1)                                # [batch, intermediade_size, seq_len]
        scan_output = scan_output + (hidden_states * self.D[None, :, None])
        scan_output = (scan_output * self.act(gate))

        if cache_params is not None:
            cache_params.ssm_states[self.layer_idx].copy_(ssm_state)

        # 4. Final linear projection
        contextualized_states = self.out_proj(scan_output.transpose(1, 2))             # [batch, seq_len, hidden_size]
        return contextualized_states
    # fmt: on

    def mixer_forward(self, hidden_states, cache_params: MambaCacheParams = None):
        if self.use_fast_kernels:
            if not is_fast_path_available or "cuda" not in self.x_proj.weight.device.type:
                raise ValueError(
                    "Fast Mamba kernels are not available. Make sure to they are installed and that the mamba module is on a CUDA device"
                )
            return self.cuda_kernels_forward(hidden_states, cache_params)
        return self.slow_forward(hidden_states, cache_params)

    def forward(
            self,
            hidden_states: torch.Tensor,
            past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
            **kwargs,
    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor]]]:
        if past_key_value is not None:
            cache_params = MambaCacheParams(
                seqlen_offset=0 if hidden_states.shape[1] > 1 else past_key_value.seen_tokens,
            )
            if len(past_key_value.key_cache) > self.layer_idx:
                # we already have cache for this layer, use it
                # remove the dummy seqlen dim (dim=2)
                cache_params.conv_states[self.layer_idx] = past_key_value.key_cache[self.layer_idx].squeeze(2)
                cache_params.ssm_states[self.layer_idx] = past_key_value.value_cache[self.layer_idx].squeeze(2)
            else:
                # we don't have cache for this layer, initialize it with zeros
                batch_size = hidden_states.shape[0]
                cache_params.conv_states[self.layer_idx] = torch.zeros(
                    batch_size,
                    self.intermediate_size,
                    self.conv_kernel_size,
                    device=hidden_states.device,
                    dtype=hidden_states.dtype,
                )
                cache_params.ssm_states[self.layer_idx] = torch.zeros(
                    batch_size,
                    self.intermediate_size,
                    self.ssm_state_size,
                    device=hidden_states.device,
                    dtype=hidden_states.dtype,
                )
        else:
            cache_params = None

        res = self.mixer_forward(hidden_states, cache_params)

        if past_key_value is not None:
            past_key_value.update(
                # add dummy seqlen dim (dim=2) to match the number of dimensions of the attention cache
                cache_params.conv_states[self.layer_idx].unsqueeze(2),
                cache_params.ssm_states[self.layer_idx].unsqueeze(2),
                self.layer_idx,
            )

        return res, past_key_value


class JambaMLP(nn.Module):
    def __init__(self, config: JambaConfig):
        super().__init__()
        self.ffn_dim = config.intermediate_size
        self.hidden_dim = config.hidden_size

        self.gate_proj = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
        self.down_proj = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)
        self.up_proj = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)

        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))


# Adapted from transformers.models.mixtral.modeling_mixtral.MixtralSparseMoeBlock with Mistral->Jamba
class JambaSparseMoeBlock(nn.Module):
    """
    This implementation is
    strictly equivalent to standard MoE with full capacity (no
    dropped tokens). It's faster since it formulates MoE operations
    in terms of block-sparse operations to accomodate imbalanced
    assignments of tokens to experts, whereas standard MoE either
    (1) drop tokens at the cost of reduced performance or (2) set
    capacity factor to number of experts and thus waste computation
    and memory on padding.
    """

    def __init__(self, config: JambaConfig, num_experts: int, num_experts_per_tok: int):
        super().__init__()
        self.hidden_dim = config.hidden_size
        self.ffn_dim = config.intermediate_size

        #   these values are decided on runtime depending on the layer index
        self.num_experts = num_experts
        self.top_k = num_experts_per_tok

        if num_experts > 1:
            # expert routing
            self.router = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
        else:
            self.router = None

        self.experts = nn.ModuleList([JambaMLP(config) for _ in range(self.num_experts)])

    def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        """ """
        batch_size, sequence_length, hidden_dim = hidden_states.shape

        if self.num_experts == 1:
            # in this case we have a single MLP block and don't need to do any routing
            final_hidden_states = self.experts[0](hidden_states)
            router_logits = torch.ones(
                (batch_size * sequence_length, 1),
                device=hidden_states.device,
                dtype=hidden_states.dtype,
                requires_grad=hidden_states.requires_grad,
            )
            return final_hidden_states, router_logits

        # in this case we have multiple experts and need to do routing
        hidden_states = hidden_states.view(-1, hidden_dim)
        # router_logits: (batch * sequence_length, n_experts)
        router_logits = self.router(hidden_states)
        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
        # we cast back to the input dtype
        routing_weights = routing_weights.to(hidden_states.dtype)

        final_hidden_states = torch.zeros(
            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
        )

        # One hot encode the selected experts to create an expert mask
        # this will be used to easily index which expert is going to be sollicitated
        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)

        # Loop over all available experts in the model and perform the computation on each expert
        for expert_idx in range(self.num_experts):
            expert_layer = self.experts[expert_idx]
            idx, top_x = torch.where(expert_mask[expert_idx])

            if top_x.shape[0] == 0:
                continue

            # in torch it is faster to index using lists than torch tensors
            top_x_list = top_x.tolist()
            idx_list = idx.tolist()

            # Index the correct hidden states and compute the expert hidden state for
            # the current expert. We need to make sure to multiply the output hidden
            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
            current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)
            current_hidden_states = expert_layer(current_state) * routing_weights[top_x_list, idx_list, None]

            # However `index_add_` only support torch tensors for indexing so we'll use
            # the `top_x` tensor here.
            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
        return final_hidden_states, router_logits


class JambaAttentionDecoderLayer(nn.Module):
    def __init__(self, config: JambaConfig, num_experts: int, layer_idx: int):
        super().__init__()

        self.self_attn = JAMBA_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)

        num_experts_per_tok = config.num_experts_per_tok if num_experts > 1 else 1
        self.moe = JambaSparseMoeBlock(config, num_experts=num_experts, num_experts_per_tok=num_experts_per_tok)
        self.input_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.pre_moe_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[Tuple[torch.Tensor]] = None,
            output_attentions: Optional[bool] = False,
            output_router_logits: Optional[bool] = False,
            use_cache: Optional[bool] = False,
            **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): 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.
            output_router_logits (`bool`, *optional*):
                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
                should not be returned during inference.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
        """

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
        )

        # residual connection after attention
        hidden_states = residual + hidden_states

        # Experts
        residual = hidden_states
        hidden_states = self.pre_moe_layernorm(hidden_states)
        hidden_states, router_logits = self.moe(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        if output_router_logits:
            outputs += (router_logits,)

        return outputs


class JambaMambaDecoderLayer(nn.Module):
    def __init__(self, config: JambaConfig, num_experts: int, layer_idx: int):
        super().__init__()

        self.mamba = JambaMambaMixer(config=config, layer_idx=layer_idx)

        num_experts_per_tok = config.num_experts_per_tok if num_experts > 1 else 1
        self.moe = JambaSparseMoeBlock(config, num_experts=num_experts, num_experts_per_tok=num_experts_per_tok)
        self.input_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.pre_moe_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
            output_attentions: Optional[bool] = False,
            output_router_logits: Optional[bool] = False,
            use_cache: Optional[bool] = False,
            **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): 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.
            output_router_logits (`bool`, *optional*):
                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
                should not be returned during inference.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
        """

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        hidden_states, present_key_value = self.mamba(
            hidden_states=hidden_states,
            past_key_value=past_key_value,
        )
        bs, seqlen, _ = hidden_states.shape
        past_seqlen = self._get_past_seqlen(past_key_value, seqlen)
        num_attention_heads = self.mamba.config.num_attention_heads
        self_attn_weights = torch.empty(bs, num_attention_heads, seqlen, past_seqlen, device="meta")

        # residual connection after mamba
        hidden_states = residual + hidden_states

        # Experts
        residual = hidden_states
        hidden_states = self.pre_moe_layernorm(hidden_states)
        hidden_states, router_logits = self.moe(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        if output_router_logits:
            outputs += (router_logits,)

        return outputs

    def _get_past_seqlen(self, past_key_value, seqlen):
        if past_key_value is None:
            return seqlen
        past_seqlen = past_key_value.get_seq_length()
        if past_seqlen == 0:
            return seqlen
        if past_key_value.attention_layer_idx is None:
            return seqlen
        if self.mamba.layer_idx < past_key_value.attention_layer_idx:
            return past_seqlen + 1
        return past_seqlen


JAMBA_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`JambaConfig`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""


@add_start_docstrings(
    "The bare Jamba Model outputting raw hidden-states without any specific head on top.",
    JAMBA_START_DOCSTRING,
)
# Adapted from transformers.models.mistral.modeling_mistral.MistralPreTrainedModel with Mistral->Jamba
class JambaPreTrainedModel(PreTrainedModel):
    config_class = JambaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["JambaAttentionDecoderLayer", "JambaMambaDecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, (nn.Linear, nn.Conv1d)):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()

    @staticmethod
    def _convert_to_standard_cache(
            past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]], batch_size: int
    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
        """
        Standardizes the format of the cache so as to match most implementations, i.e. have the seqlen as the third dim
        also for mamba layers
        """
        attn_layer_index = [k.shape == v.shape for k, v in past_key_value].index(True)
        seqlen = past_key_value[attn_layer_index][0].shape[2]
        standard_past_key_value = ()
        for k, v in past_key_value:
            if k.shape != v.shape:
                # mamba layer
                # expand doesn't use more memory, so it's fine to do it here
                standard_past_key_value += ((k.expand(-1, -1, seqlen, -1), v.expand(-1, -1, seqlen, -1)),)
            else:
                standard_past_key_value += ((k, v),)
        return standard_past_key_value

    @staticmethod
    def _convert_to_jamba_cache(
            past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]],
    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
        """
        Converts the cache to the format expected by Jamba, i.e. dummy seqlen dimesion with size 1 for mamba layers
        """
        jamba_past_key_value = ()
        for k, v in past_key_value:
            if k.shape != v.shape:
                # mamba layer
                jamba_past_key_value += ((k[:, :, :1, :], v[:, :, :1, :]),)
            else:
                jamba_past_key_value += ((k, v),)
        return jamba_past_key_value


JAMBA_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

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

            [What are attention masks?](../glossary#attention-mask)

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
            `past_key_values`).

            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
            information on the default strategy.

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.n_positions - 1]`.

            [What are position IDs?](../glossary#position-ids)
        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
            Tuple of `tuple(torch.FloatTensor)` of length `config.num_hidden_layers`, with each tuple having 2 tensors
            corresponding to the cache of the layer.
            For attention layers, both tensors have shape of `(batch_size, num_kv_heads, sequence_length, embed_size_per_head)`
            For mamba layers, the first tensor represents the convolution state and has shape of `(batch_size, d_inner, 1, d_conv)`,
            and the second tensor represents the ssm state and has shape of `(batch_size, d_inner, 1, d_state)`. Mamba
            cache shape[2] is a dummy "seqlen" dimension to match the number of attention cache dimensions. For mamba,
            the cache doesn't grow with seqlen so this dimension is always 1.

            Contains pre-computed hidden-states (key and values in the self-attention blocks and convolution and
            ssm states in the mamba 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 `input_ids` (those that
            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
            `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.
        use_cache (`bool`, *optional*):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).
        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.
        output_router_logits (`bool`, *optional*):
            Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
            should not be returned during inference.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""


@add_start_docstrings(
    "The bare Jamba Model outputting raw hidden-states without any specific head on top.",
    JAMBA_START_DOCSTRING,
)
# Adapted from transformers.models.mistral.modeling_mistral.MistralModel with MISTRAL->JAMBA, Mistral->Jamba
class JambaModel(JambaPreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`JambaDecoderLayer`]

    Args:
        config: JambaConfig
    """

    def __init__(self, config: JambaConfig):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)

        # init each model layer, decide if it's mamba/attention and has experts or not
        decoder_layers = []
        for i in range(config.num_hidden_layers):
            is_attn = True if (i - self.config.attn_layer_offset) % self.config.attn_layer_period == 0 else False
            is_expert = True if (i - self.config.expert_layer_offset) % self.config.expert_layer_period == 0 else False

            num_experts = self.config.num_experts if is_expert else 1
            if is_attn:
                decoder_layers.append(JambaAttentionDecoderLayer(config, num_experts=num_experts, layer_idx=i))
            else:
                decoder_layers.append(JambaMambaDecoderLayer(config, num_experts=num_experts, layer_idx=i))

        if not any(isinstance(layer, JambaAttentionDecoderLayer) for layer in decoder_layers):
            raise ValueError("At least one layer in the decoder must be an attention layer")
        self._attn_layer_index = [isinstance(layer, JambaAttentionDecoderLayer) for layer in decoder_layers].index(
            True
        )

        if not any(isinstance(layer, JambaMambaDecoderLayer) for layer in decoder_layers):
            raise ValueError("At least one layer in the decoder must be a Mamba layer")
        self._mamba_layer_index = [isinstance(layer, JambaMambaDecoderLayer) for layer in decoder_layers].index(True)

        if (
                decoder_layers[self._mamba_layer_index].mamba.ssm_state_size
                == decoder_layers[self._mamba_layer_index].mamba.conv_kernel_size
        ):
            raise ValueError("Mamba state size and convolution size must be different")

        self.layers = nn.ModuleList(decoder_layers)

        self._attn_implementation = config._attn_implementation
        self.final_layernorm = JambaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

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

    def get_input_embeddings(self):
        return self.embed_tokens

    def set_input_embeddings(self, value):
        self.embed_tokens = value

    # Ignore copy
    @add_start_docstrings_to_model_forward(JAMBA_INPUTS_DOCSTRING)
    def forward(
            self,
            input_ids: torch.LongTensor = None,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_values: Optional[Union[List[torch.FloatTensor], HybridMambaAttentionDynamicCache]] = None,
            inputs_embeds: Optional[torch.FloatTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            output_router_logits: Optional[bool] = None,
            return_dict: Optional[bool] = None,
    ) -> Union[Tuple, MoeModelOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )
        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

        # retrieve input_ids and inputs_embeds
        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:
            batch_size, seq_length = input_ids.shape
        elif inputs_embeds is not None:
            batch_size, seq_length, _ = inputs_embeds.shape
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        past_key_values_length = 0

        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

        if use_cache:
            if isinstance(past_key_values, Cache) and not isinstance(
                    past_key_values, HybridMambaAttentionDynamicCache
            ):
                past_key_values = HybridMambaAttentionDynamicCache.from_legacy_cache(past_key_values.to_legacy_cache())
            use_legacy_cache = not isinstance(past_key_values, HybridMambaAttentionDynamicCache)
            if use_legacy_cache:
                past_key_values = HybridMambaAttentionDynamicCache.from_legacy_cache(past_key_values)
            past_key_values_length = past_key_values.get_usable_length(seq_length, self._attn_layer_index)

        if position_ids is None:
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(
                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
            )
            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
        else:
            position_ids = position_ids.view(-1, seq_length).long()

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

        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
            if is_padding_right:
                raise ValueError(
                    "You are attempting to perform batched generation with padding_side='right'"
                    " this may lead to unexpected behaviour for Flash Attention version of Jamba. Make sure to "
                    " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
                )

        if self._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._attn_implementation == "sdpa" and not output_attentions:
            # output_attentions=True 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,
                (batch_size, seq_length),
                inputs_embeds,
                past_key_values_length,
            )
        else:
            # 4d mask is passed through the layers
            attention_mask = _prepare_4d_causal_attention_mask(
                attention_mask,
                (batch_size, seq_length),
                inputs_embeds,
                past_key_values_length,
                sliding_window=self.config.sliding_window,
            )

        hidden_states = inputs_embeds

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        all_router_logits = () if output_router_logits else None
        next_decoder_cache = None

        for decoder_layer in self.layers:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    decoder_layer.__call__,
                    hidden_states,
                    attention_mask,
                    position_ids,
                    past_key_values,
                    output_attentions,
                    output_router_logits,
                    use_cache,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    output_router_logits=output_router_logits,
                    use_cache=use_cache,
                )

            hidden_states = layer_outputs[0]

            if use_cache:
                next_decoder_cache = layer_outputs[2 if output_attentions else 1]

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

            if output_router_logits:
                all_router_logits += (layer_outputs[-1],)

        hidden_states = self.final_layernorm(hidden_states)

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

        next_cache = None
        if use_cache:
            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache

        if not return_dict:
            return tuple(
                v
                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits]
                if v is not None
            )
        return MoeModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            router_logits=all_router_logits,
        )


# Adapted from transformers.models.mixtral.modeling_mixtral.MixtralForCausalLM with MIXTRAL->JAMBA, Mixtral->Jamba
class JambaForCausalLM(JambaPreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config: JambaConfig):
        super().__init__(config)
        self.model = JambaModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.router_aux_loss_coef = config.router_aux_loss_coef
        self.num_experts = config.num_experts
        self.num_experts_per_tok = config.num_experts_per_tok
        # Initialize weights and apply final processing
        self.post_init()

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

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

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model

    @add_start_docstrings_to_model_forward(JAMBA_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    # Ignore copy
    def forward(
            self,
            input_ids: torch.LongTensor = None,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_values: Optional[List[torch.FloatTensor]] = None,
            inputs_embeds: Optional[torch.FloatTensor] = None,
            labels: Optional[torch.LongTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            output_router_logits: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            calc_logits_for_entire_prompt: Optional[bool] = True,
    ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
        r"""
        Args:
            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

            calc_logits_for_entire_prompt (`bool`, *optional*):
                Whether or not to calculate the logits for the entire prompt, or just the last token. Only last token
                logits are needed for generation, and calculating them only for that token can save memory,
                which becomes pretty significant for long sequences.

        Returns:
        ```"""

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )

        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

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            output_router_logits=output_router_logits,
            return_dict=return_dict,
        )

        hidden_states = outputs[0]
        if calc_logits_for_entire_prompt:
            logits = self.lm_head(hidden_states)
        else:
            logits = self.lm_head(hidden_states[..., -1:, :])
        logits = logits.float()

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

        aux_loss = None
        if output_router_logits:
            aux_loss = load_balancing_loss_func(
                outputs.router_logits if return_dict else outputs[-1],
                self.num_experts,
                self.num_experts_per_tok,
                attention_mask,
            )
            if labels is not None:
                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device

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

        return MoeCausalLMOutputWithPast(
            loss=loss,
            aux_loss=aux_loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            router_logits=outputs.router_logits,
        )

    def prepare_inputs_for_generation(
            self,
            input_ids,
            past_key_values=None,
            attention_mask=None,
            inputs_embeds=None,
            output_router_logits=False,
            **kwargs,
    ):
        # Omit tokens covered by past_key_values
        if past_key_values is not None:
            # the cache may be in the stardard format (e.g. in contrastive search), convert to Jamba's format if needed
            if isinstance(past_key_values, Tuple):
                if past_key_values[self.model._mamba_layer_index][0].shape[2] > 1:
                    past_key_values = self._convert_to_jamba_cache(past_key_values)

            if isinstance(past_key_values, Cache):
                if not isinstance(past_key_values, HybridMambaAttentionDynamicCache):
                    past_key_values = HybridMambaAttentionDynamicCache.from_legacy_cache(
                        past_key_values.to_legacy_cache()
                    )
                cache_length = past_key_values.get_seq_length()
                past_length = past_key_values.seen_tokens
                max_cache_length = past_key_values.get_max_length()
            else:
                cache_length = past_length = past_key_values[self.model._attn_layer_index][0].shape[2]
                max_cache_length = None

            # Keep only the unprocessed tokens:
            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
            # input)
            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
            # input_ids based on the past_length.
            elif past_length < input_ids.shape[1]:
                input_ids = input_ids[:, past_length:]
            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.

            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
            if (
                    max_cache_length is not None
                    and attention_mask is not None
                    and cache_length + input_ids.shape[1] > max_cache_length
            ):
                attention_mask = attention_mask[:, -max_cache_length:]

        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, 1)
            if past_key_values:
                position_ids = position_ids[:, -input_ids.shape[1] :]

        # 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(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
                "output_router_logits": output_router_logits,
                "calc_logits_for_entire_prompt": self.config.calc_logits_for_entire_prompt,
            }
        )
        return model_inputs

    @staticmethod
    def _reorder_cache(past_key_values, beam_idx):
        reordered_past = ()
        for layer_past in past_key_values:
            reordered_past += (
                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
            )
        return reordered_past


@add_start_docstrings(
    """
    The Jamba Model with a sequence classification head on top (linear layer).

    [`JambaForSequenceClassification`] uses the last token in order to do the classification, as other causal models
    (e.g. GPT-2) do.

    Since it does classification on the last token, it requires to know the position of the last token. If a
    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
    each row of the batch).
    """,
    JAMBA_START_DOCSTRING,
)
# Copied from transformers.models.mixtral.modeling_mixtral.MixtralForSequenceClassification with Mixtral->Jamba, MIXTRAL->JAMBA
class JambaForSequenceClassification(JambaPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.model = JambaModel(config)
        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)

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

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

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

    @add_start_docstrings_to_model_forward(JAMBA_INPUTS_DOCSTRING)
    def forward(
            self,
            input_ids: torch.LongTensor = None,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_values: Optional[List[torch.FloatTensor]] = None,
            inputs_embeds: Optional[torch.FloatTensor] = None,
            labels: Optional[torch.LongTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        transformer_outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = transformer_outputs[0]
        logits = self.score(hidden_states)

        if input_ids is not None:
            batch_size = input_ids.shape[0]
        else:
            batch_size = inputs_embeds.shape[0]

        if self.config.pad_token_id is None and batch_size != 1:
            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
        if self.config.pad_token_id is None:
            sequence_lengths = -1
        else:
            if input_ids is not None:
                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
                sequence_lengths = sequence_lengths % input_ids.shape[-1]
                sequence_lengths = sequence_lengths.to(logits.device)
            else:
                sequence_lengths = -1

        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]

        loss = None
        if labels is not None:
            labels = labels.to(logits.device)
            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.num_labels == 1:
                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(pooled_logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(pooled_logits, labels)
        if not return_dict:
            output = (pooled_logits,) + transformer_outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return SequenceClassifierOutputWithPast(
            loss=loss,
            logits=pooled_logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )