# coding=utf-8
# Copyright 2024 JetMoe AI and the HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch JetMoe model."""

import math
from typing import List, Optional, Tuple, Union

import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import functional as F

from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache, StaticCache
from ...generation import GenerationMixin
from ...modeling_attn_mask_utils import AttentionMaskConverter
from ...modeling_flash_attention_utils import flash_attn_supports_top_left_mask, is_flash_attn_available
from ...modeling_outputs import (
    MoeCausalLMOutputWithPast,
    MoeModelOutputWithPast,
    SequenceClassifierOutputWithPast,
)
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import PreTrainedModel
from ...utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    can_return_tuple,
    is_torch_flex_attn_available,
    logging,
    replace_return_docstrings,
)
from ...utils.deprecation import deprecate_kwarg
from .configuration_jetmoe import JetMoeConfig


if is_torch_flex_attn_available():
    from torch.nn.attention.flex_attention import BlockMask

    from ...integrations.flex_attention import make_flex_block_causal_mask


if is_flash_attn_available():
    from ...modeling_flash_attention_utils import _flash_attention_forward

logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "jetmoe"
_CONFIG_FOR_DOC = "JetMoeConfig"


# Copied from transformers.models.mixtral.modeling_mixtral.load_balancing_loss_func
def load_balancing_loss_func(
    gate_logits: Union[torch.Tensor, Tuple[torch.Tensor], None],
    num_experts: Optional[int] = None,
    top_k=2,
    attention_mask: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, int]:
    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:
            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
            shape [batch_size X sequence_length, num_experts].
        num_experts:
            Number of experts
        top_k:
            The number of experts to route per-token, can be also interpreted as the `top-k` routing
            parameter.
        attention_mask (`torch.Tensor`, *optional*):
            The attention_mask used in forward function
            shape [batch_size X sequence_length] if not None.

    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], 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


class JetMoeParallelExperts(nn.Module):
    def __init__(self, num_experts: int, input_size: int, output_size: int) -> None:
        """
        Initialize the JetMoeParallelExperts module.
        The experts weights are stored in [num_experts, output_size, input_size] format. Such that it's comptible with
        many MoE libraries, such as [Megablock](https://github.com/databricks/megablocks) and
        [ScatterMoE](https://github.com/shawntan/scattermoe), as well as the
        [MoE kernel](https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/fused_moe/fused_moe.py)
        used in vllm.

        Args:
            num_experts (int):
                Number of experts.
            input_size (int):
                Size of the input.
            output_size (int):
                Size of the output.
        """
        super().__init__()
        self.weight = nn.Parameter(torch.empty(num_experts, output_size, input_size))
        self.num_experts = num_experts
        self.input_size = input_size
        self.output_size = output_size

    def forward(self, inputs, expert_size):
        """
        Forward pass of the JetMoeParallelExperts module.

        Args:
            inputs (Tensor):
                Input tensor.
            expert_size:
                Expert size information.

        Returns:
            Tensor: Output tensor.
        """
        input_list = inputs.split(expert_size, dim=0)
        output_list = []
        for i in range(self.num_experts):
            output_list.append(F.linear(input_list[i], self.weight[i]))
        results = torch.cat(output_list, dim=0)
        return results


class JetMoeTopKGating(nn.Module):
    def __init__(self, input_size: int, num_experts: int, top_k: int):
        """
        Initialize the top-k gating mechanism.

        Args:
            input_size (`int`):
                Size of the input.
            num_experts (`int`):
                Number of experts.
            top_k (`int`):
                Number of top experts to select.
        """
        super().__init__()

        self.num_experts = num_experts
        self.input_size = input_size
        self.top_k = top_k

        self.layer = nn.Linear(input_size, num_experts, bias=False)

    def forward(self, hidden_states):
        # compute the top_k routing decision
        logits = self.layer(hidden_states).float()  # [batch_size x seq_len, num_experts]
        top_k_logits, top_k_indices = logits.topk(self.top_k, dim=1)  # [num_tokens, top_k]
        top_k_gates = torch.softmax(top_k_logits, dim=1).type_as(hidden_states)  # [num_tokens, top_k]

        # compute number of input given to each expert
        zeros = torch.zeros(
            [top_k_gates.size(0), self.num_experts], dtype=top_k_gates.dtype, device=top_k_gates.device
        )  # [num_tokens, num_experts]
        gates = zeros.scatter(1, top_k_indices, 1)  # [num_tokens, num_experts]
        expert_size = gates.long().sum(0)  # [num_experts,]
        # (This cause torch.compile to fail with `torch._dynamo.exc.Unsupported: Backend compiler failed with a fake tensor exception at`)
        # (and `DataDependentOutputException`)
        expert_size = expert_size.tolist()

        # sort and group input tokens according to expert assignment
        top_k_experts = top_k_indices.flatten()  # [num_tokens * top_k]
        _, index_sorted_experts = top_k_experts.sort(0)  # [num_tokens * top_k]
        batch_index = index_sorted_experts.div(self.top_k, rounding_mode="trunc")  # [num_tokens * top_k]

        # gather the gate values for grouped input tokens
        top_k_gates = top_k_gates.flatten()  # [num_tokens * top_k]
        batch_gates = top_k_gates[index_sorted_experts]  # [num_tokens * top_k]

        return index_sorted_experts, batch_index, batch_gates, expert_size, logits


class JetMoeMoE(nn.Module):
    """
    A Sparsely gated mixture of experts layer with 1-layer Feed-Forward networks as experts.

    Args:
        config:
            Configuration object with model hyperparameters.
    """

    def __init__(self, config: JetMoeConfig):
        super(JetMoeMoE, self).__init__()

        self.input_size = config.hidden_size
        self.hidden_size = config.intermediate_size
        self.activation = ACT2FN[config.activation_function]
        self.bias = torch.nn.Parameter(torch.empty(self.input_size))
        self.input_linear = JetMoeParallelExperts(config.num_local_experts, self.input_size, self.hidden_size * 2)
        self.output_linear = JetMoeParallelExperts(config.num_local_experts, self.hidden_size, self.input_size)

        self.router = JetMoeTopKGating(
            input_size=self.input_size,
            num_experts=config.num_local_experts,
            top_k=config.num_experts_per_tok,
        )

    def forward(self, layer_input):
        """
        Forward pass of the mixture of experts layer.

        Args:
            layer_input (Tensor):
                Input tensor.

        Returns:
            Tensor:
                Output tensor.
            Tensor:
                Router logits.
        """
        bsz, length, emb_size = layer_input.size()
        layer_input = layer_input.reshape(-1, emb_size)
        _, batch_index, batch_gates, expert_size, router_logits = self.router(layer_input)

        expert_inputs = layer_input[batch_index]
        hidden_states = self.input_linear(expert_inputs, expert_size)
        chunked_hidden_states = hidden_states.chunk(2, dim=-1)
        hidden_states = self.activation(chunked_hidden_states[0]) * chunked_hidden_states[1]
        expert_outputs = self.output_linear(hidden_states, expert_size)

        expert_outputs = expert_outputs * batch_gates[:, None]

        zeros = torch.zeros((bsz * length, self.input_size), dtype=expert_outputs.dtype, device=expert_outputs.device)
        layer_output = zeros.index_add(0, batch_index, expert_outputs)
        layer_output = layer_output.view(bsz, length, self.input_size)
        layer_output = layer_output + self.bias
        return layer_output, router_logits


class JetMoeMoA(nn.Module):
    """
    A Sparsely gated mixture of attention layer with pairs of query- and output-projections as experts.

    Args:
        config:
            Configuration object with model hyperparameters.
    """

    def __init__(self, config: JetMoeConfig):
        super(JetMoeMoA, self).__init__()

        self.num_experts = config.num_local_experts
        self.input_size = config.hidden_size
        self.hidden_size = config.kv_channels * config.num_key_value_heads
        self.top_k = config.num_experts_per_tok
        self.bias = torch.nn.Parameter(torch.empty(self.input_size))

        self.input_linear = JetMoeParallelExperts(self.num_experts, self.input_size, self.hidden_size)
        self.output_linear = JetMoeParallelExperts(self.num_experts, self.hidden_size, self.input_size)

        self.router = JetMoeTopKGating(
            input_size=self.input_size,
            num_experts=self.num_experts,
            top_k=self.top_k,
        )

    def map(self, layer_input):
        """
        Map inputs to attention experts according to routing decision and compute query projection inside each experts.
        """

        # Compute gating topology
        bsz, length, emb_size = layer_input.size()
        layer_input = layer_input.reshape(-1, emb_size)  # [bsz * length, emb_size]
        index_sorted_experts, batch_index, batch_gates, expert_size, router_logits = self.router(layer_input)
        topo_info = (index_sorted_experts, batch_index, batch_gates, expert_size)

        # Group inputs according to topology and compute query projection
        expert_inputs = layer_input[batch_index]  # [bsz * length * top_k, emb_size]
        expert_outputs = self.input_linear(expert_inputs, expert_size)  # [bsz * length * top_k, hidden_size]

        # Ungroup queries back to original order
        zeros = torch.zeros(
            (bsz * length * self.top_k, self.hidden_size), dtype=expert_outputs.dtype, device=expert_outputs.device
        )
        layer_output = zeros.index_add(0, index_sorted_experts, expert_outputs)
        layer_output = layer_output.view(bsz, length, self.top_k, -1)  # [bsz, length, top_k, hidden_size]
        return layer_output, router_logits, topo_info

    def reduce(self, layer_input, topo_info):
        """
        Compute output projection inside each attention experts and merge the outputs of different experts.
        """
        bsz, length, k, hidden_size = layer_input.size()
        layer_input = layer_input.reshape(-1, hidden_size)  # [bsz * length * k, hidden_size]
        index_sorted_experts, batch_index, batch_gates, expert_size = topo_info

        # Group inputs according to topology and compute output projection
        expert_inputs = layer_input[index_sorted_experts]  # [bsz * length * top_k, hidden_size]
        expert_outputs = self.output_linear(expert_inputs, expert_size)  # [bsz * length * top_k, emb_size]

        # Apply gates to attention expert outputs
        expert_outputs = expert_outputs * batch_gates[:, None]

        # Ungroup and merge outputs to original order
        zeros = torch.zeros((bsz * length, self.input_size), dtype=expert_outputs.dtype, device=expert_outputs.device)
        layer_output = zeros.index_add(0, batch_index, expert_outputs)
        layer_output = layer_output.view(bsz, length, self.input_size)
        layer_output = layer_output + self.bias
        return layer_output

    def forward(self, layer_input):
        raise NotImplementedError("This module doesn't support call and forward.")


# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->JetMoe
class JetMoeRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        JetMoeRMSNorm 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)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"


# Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with Gemma->JetMoe
class JetMoeRotaryEmbedding(nn.Module):
    def __init__(self, config: JetMoeConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()

        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


# Copied from transformers.models.llama.modeling_llama.rotate_half
def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


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

    def __init__(self, config: JetMoeConfig, layer_idx: Optional[int] = None):
        """
        Initialize the JetMoeAttention module.

        Args:
            config:
                Configuration object with model hyperparameters.
            layer_idx:
                Index of the layer in the model.
        """
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.is_causal = True
        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.top_k = config.num_experts_per_tok
        self.attention_dropout = config.attention_dropout
        self.kv_projection_size = config.kv_channels * config.num_key_value_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_heads = config.num_attention_heads
        self.head_dim = config.kv_channels

        self.experts = JetMoeMoA(config)

        self.kv_proj = torch.nn.Linear(config.hidden_size, self.kv_projection_size * 2, bias=False)

        self.rotary_emb = JetMoeRotaryEmbedding(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        bsz, q_len, _ = hidden_states.size()

        query_states, router_logits, topo_info = self.experts.map(hidden_states)
        key_states, value_states = self.kv_proj(hidden_states).chunk(2, dim=-1)

        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)

        cos, sin = self.rotary_emb(value_states, position_ids)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        # repeat k/v heads for top-k attention experts
        key_states = key_states.repeat(1, self.top_k, 1, 1)
        value_states = value_states.repeat(1, self.top_k, 1, 1)

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

        if attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
            attn_weights = attn_weights + causal_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.top_k, self.kv_projection_size)

        attn_output = self.experts.reduce(attn_output, topo_info)
        attn_output = attn_output.view(bsz, q_len, -1)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value, router_logits


class JetMoeSdpaAttention(JetMoeAttention):
    """
    JetMoe attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
    `JetMoeAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
    SDPA API.
    """

    # Adapted from JetMoeAttention.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,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]], Optional[torch.Tensor]]:
        if output_attentions:
            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
            logger.warning_once(
                "JetMoeModel is using JetMoeSdpaAttention, 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,
                cache_position=cache_position,
            )

        bsz, q_len, _ = hidden_states.size()

        query_states, router_logits, topo_info = self.experts.map(hidden_states)
        key_states, value_states = self.kv_proj(hidden_states).chunk(2, dim=-1)

        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)

        cos, sin = self.rotary_emb(value_states, position_ids)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        # repeat k/v heads for top-k attention experts
        key_states = key_states.repeat(1, self.top_k, 1, 1)
        value_states = value_states.repeat(1, self.top_k, 1, 1)

        causal_mask = attention_mask
        if attention_mask is not None:
            causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]

        # 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 causal_mask is not None:
            query_states = query_states.contiguous()
            key_states = key_states.contiguous()
            value_states = value_states.contiguous()

        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
        is_causal = True if causal_mask is None and q_len > 1 else False

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=causal_mask,
            dropout_p=self.attention_dropout if self.training else 0.0,
            is_causal=is_causal,
        )

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

        attn_output = self.experts.reduce(attn_output, topo_info)
        attn_output = attn_output.view(bsz, q_len, -1)

        return attn_output, None, past_key_value, router_logits


class JetMoeFlashAttention2(JetMoeAttention):
    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 alignment, 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 = flash_attn_supports_top_left_mask()

    def forward(
        self,
        hidden_states: Optional[torch.FloatTensor],
        attention_mask: Optional[torch.FloatTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        use_cache: Optional[bool] = False,
        output_attentions: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[
        Tuple[torch.Tensor, Tuple[torch.Tensor]],
        Optional[Tuple[torch.Tensor, Tuple[torch.Tensor], Tuple[torch.Tensor, ...]]],
    ]:
        """
        Forward pass of the JetMoeAttention module.

        Args:
            hidden_states (Optional[torch.FloatTensor]): Input hidden states.
            attention_mask (Optional[torch.FloatTensor]): Attention mask.
            layer_past (Optional[Tuple[torch.Tensor]]): Past layer state.
            use_cache (Optional[bool]): Whether to use cached states.
            output_attentions (Optional[bool]): Whether to output attention weights.
            cache_position (Optional[torch.LongTensor]): Position of the cache.

        Returns:
            Union[Tuple[torch.Tensor, Tuple[torch.Tensor]], Optional[Tuple[...]]]: Tuple containing outputs.
        """
        output_attentions = False
        bsz, q_len, hidden_size = hidden_states.size()

        # calculate query, key, values
        query_states, router_logits, topo_info = self.experts.map(hidden_states)
        key_states, value_states = self.kv_proj(hidden_states).chunk(2, dim=-1)

        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)

        cos, sin = self.rotary_emb(value_states, position_ids)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        # repeat k/v heads for top-k attention experts
        key_states = key_states.repeat(1, self.top_k, 1, 1)
        value_states = value_states.repeat(1, self.top_k, 1, 1)

        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
        # to be able to avoid many of these transpose/reshape/view.
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

        dropout_rate = self.attention_dropout if self.training else 0.0

        # 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 the correct dtype just to be sure everything works as expected.
        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
        # in fp32. (LlamaRMSNorm handles it correctly)

        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.kv_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)

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

        # output projection
        attn_output = attn_output.reshape(bsz, q_len, self.top_k, self.kv_projection_size)
        attn_output = self.experts.reduce(attn_output, topo_info)
        attn_output = attn_output.view(bsz, q_len, hidden_size)  # re-assemble all head outputs side by side

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value, router_logits


JETMOE_ATTENTION_CLASSES = {
    "eager": JetMoeAttention,
    "flash_attention_2": JetMoeFlashAttention2,
    "sdpa": JetMoeSdpaAttention,
}


class JetMoeBlock(nn.Module):
    def __init__(self, config: JetMoeConfig, layer_idx: Optional[int] = None):
        """
        Initialize the JetMoeBlock module.

        Args:
            config:
                Configuration object with model hyperparameters.
        """
        super().__init__()
        self.input_layernorm = JetMoeRMSNorm(config.hidden_size)
        self.self_attention = JETMOE_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
        self.post_attention_layernorm = JetMoeRMSNorm(config.hidden_size)

        self.mlp = JetMoeMoE(config)

    def forward(
        self,
        hidden_states: Optional[torch.FloatTensor],
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = False,
        output_router_logits: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
        # Self Attention
        attn_output, self_attn_weights, present_key_value, attn_router_logits = self.self_attention(
            hidden_states=self.input_layernorm(hidden_states),
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
            cache_position=cache_position,
        )

        hidden_states = hidden_states + attn_output
        x_mlp, mlp_router_logits = self.mlp(self.post_attention_layernorm(hidden_states))
        hidden_states = hidden_states + x_mlp

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        if output_router_logits:
            outputs += attn_router_logits, mlp_router_logits

        return outputs


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

    config_class = JetMoeConfig
    base_model_prefix = "transformer"
    supports_gradient_checkpointing = False
    _no_split_modules = ["JetMoeBlock"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True

    def _init_weights(self, module):
        """Initialize the weights."""
        if isinstance(module, (nn.Linear,)):
            # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)
        elif isinstance(module, JetMoeParallelExperts):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
        elif isinstance(module, JetMoeMoA):
            module.bias.data.zero_()
        elif isinstance(module, JetMoeMoE):
            module.bias.data.zero_()


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

    Parameters:
        config ([`JetMoeConfig`]): 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.
"""

JETMOE_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `({0})`):
            Indices of input sequence tokens in the vocabulary.

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

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`torch.FloatTensor` of shape `({0})`, *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)
        position_ids (`torch.LongTensor` of shape `({0})`, *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)
        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_dim)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
            model's internal embedding lookup matrix.
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        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.
        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
            the complete sequence length.
"""


@add_start_docstrings(
    "The bare JetMoe Model outputting raw hidden-states without any specific head on top.",
    JETMOE_START_DOCSTRING,
)
class JetMoeModel(JetMoePreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`JetMoeBlock`]

    Args:
        config:
            JetMoeConfig
    """

    def __init__(self, config: JetMoeConfig):
        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)
        self.layers = nn.ModuleList([JetMoeBlock(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
        self._attn_implementation = config._attn_implementation
        self.norm = JetMoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

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

    # Copied from transformers.models.llama.modeling_llama.LlamaModel.get_input_embeddings
    def get_input_embeddings(self):
        return self.embed_tokens

    # Copied from transformers.models.llama.modeling_llama.LlamaModel.set_input_embeddings
    def set_input_embeddings(self, value):
        self.embed_tokens = value

    @can_return_tuple
    @add_start_docstrings_to_model_forward(JETMOE_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = 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,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> MoeModelOutputWithPast:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

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

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

        # kept for BC (non `Cache` `past_key_values` inputs)
        return_legacy_cache = False
        if use_cache and not isinstance(past_key_values, Cache):
            return_legacy_cache = True
            if past_key_values is None:
                past_key_values = DynamicCache()
            else:
                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
                logger.warning_once(
                    "We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and "
                    "will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class "
                    "(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)"
                )

        if cache_position is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            cache_position = torch.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )
        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
            batch_size = inputs_embeds.shape[0]
            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 JetMoe. Make sure to "
                    " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
                )
        causal_mask = self._update_causal_mask(
            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
        )

        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,
                    position_ids,
                    past_key_values,
                    causal_mask,
                    output_attentions,
                    output_router_logits,
                    use_cache,
                    use_reentrant=False,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=causal_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[-2], layer_outputs[-1])

        hidden_states = self.norm(hidden_states)

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

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

        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,
        )

    # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask
    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool = False,
    ):
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and (attention_mask == 0.0).any():
                return attention_mask
            return None
        if self.config._attn_implementation == "flex_attention":
            if isinstance(attention_mask, torch.Tensor):
                attention_mask = make_flex_block_causal_mask(attention_mask)
            if isinstance(attention_mask, BlockMask):
                return attention_mask

        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
        # to infer the attention mask.
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        using_static_cache = isinstance(past_key_values, StaticCache)

        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                is_training=self.training,
            ):
                return None

        dtype, device = input_tensor.dtype, input_tensor.device
        sequence_length = input_tensor.shape[1]
        if using_static_cache:
            target_length = past_key_values.get_max_cache_shape()
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            device=device,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
        )

        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type in ["cuda", "xpu"]
            and not output_attentions
        ):
            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
            # Details: https://github.com/pytorch/pytorch/issues/110213
            min_dtype = torch.finfo(dtype).min
            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

        return causal_mask

    @staticmethod
    # Copied from transformers.models.llama.modeling_llama.LlamaModel._prepare_4d_causal_attention_mask_with_cache_position
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        device: torch.device,
        cache_position: torch.Tensor,
        batch_size: int,
        **kwargs,
    ):
        """
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
                `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache,
                to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            device (`torch.device`):
                The device to place the 4D attention mask on.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        """
        if attention_mask is not None and attention_mask.dim() == 4:
            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
            causal_mask = attention_mask
        else:
            min_dtype = torch.finfo(dtype).min
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
            )
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
                    causal_mask.device
                )
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

        return causal_mask


class JetMoeForCausalLM(JetMoePreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = JetMoeModel(config)
        self.vocab_size = config.vocab_size
        self.aux_loss_coef = config.aux_loss_coef
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.tie_word_embeddings = config.tie_word_embeddings

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

    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_input_embeddings
    def get_input_embeddings(self):
        return self.model.embed_tokens

    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_input_embeddings
    def set_input_embeddings(self, value):
        self.model.embed_tokens = value

    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_output_embeddings
    def get_output_embeddings(self):
        return self.lm_head

    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_output_embeddings
    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_decoder
    def set_decoder(self, decoder):
        self.model = decoder

    # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_decoder
    def get_decoder(self):
        return self.model

    @can_return_tuple
    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
    @add_start_docstrings_to_model_forward(JETMOE_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: Optional[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,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **kwargs,
    ) -> MoeCausalLMOutputWithPast:
        r"""
            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

            logits_to_keep (`int` or `torch.Tensor`, *optional*):
                If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
                `input_ids` (special case). 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 or large vocabulary size.
                If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
                This is useful when using packed tensor format (single dimension for batch and sequence length).

        Returns:
        """

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

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs: MoeModelOutputWithPast = 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,
            cache_position=cache_position,
        )

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

        loss = None
        if labels is not None:
            # Upcast to float if we need to compute the loss to avoid potential precision issues
            logits = logits.float()
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Ensure tensors are on the same device
            shift_labels = shift_labels.to(shift_logits.device)
            loss = self.loss_function(
                shift_logits,
                shift_labels,
                vocab_size=self.config.vocab_size,
                **kwargs,
            )

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

        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,
        )


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

    [`JetMoeForSequenceClassification`] 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).
    """,
    JETMOE_START_DOCSTRING,
)
# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->JetMoe, LLAMA->JETMOE, BaseModelOutputWithPast->MoeModelOutputWithPast
class JetMoeForSequenceClassification(JetMoePreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.model = JetMoeModel(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

    @can_return_tuple
    @add_start_docstrings_to_model_forward(JETMOE_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = 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,
    ) -> 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).
        """

        transformer_outputs: MoeModelOutputWithPast = 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,
        )
        hidden_states = transformer_outputs.last_hidden_state
        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:
            last_non_pad_token = -1
        elif input_ids is not None:
            # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id
            non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32)
            token_indices = torch.arange(input_ids.shape[-1], device=logits.device, dtype=torch.int32)
            last_non_pad_token = (token_indices * non_pad_mask).argmax(-1)
        else:
            last_non_pad_token = -1
            logger.warning_once(
                f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
                "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
            )

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

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

        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,
        )


__all__ = ["JetMoeForCausalLM", "JetMoeModel", "JetMoePreTrainedModel", "JetMoeForSequenceClassification"]