#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/qwen3_next/modular_qwen3_next.py.
#               Do NOT edit this file manually as any edits will be overwritten by the generation of
#             the file from the modular. If any change should be done, please apply the change to the
#                          modular_qwen3_next.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# coding=utf-8
# Copyright 2025 The Qwen team, Alibaba Group 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.
# Copyright (c) 2025, HUAWEI CORPORATION.  All rights reserved.

from typing import Any, Callable, Optional, Union

import torch
import torch.nn.functional as F
from torch import nn
from torch.distributed.tensor import DTensor
from einops import rearrange

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache
from transformers.generation import GenerationMixin
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import (
    GenericForQuestionAnswering,
    GenericForSequenceClassification,
    GenericForTokenClassification,
    GradientCheckpointingLayer,
)
from transformers.modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from transformers.modeling_utils import PreTrainedModel
from transformers.models.qwen3_next.configuration_qwen3_next import Qwen3NextConfig
from transformers.processing_utils import Unpack
from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
from transformers.utils.deprecation import deprecate_kwarg

try:
    from transformers.utils.generic import OutputRecorder, check_model_inputs
except ImportError:
    # adapt for transformers 5.x
    from transformers.utils.output_capturing import OutputRecorder, capture_outputs

    check_model_inputs = capture_outputs
from transformers.utils.import_utils import (
    is_causal_conv1d_available,
    is_flash_linear_attention_available,
)

from mindspeed.core.fusions.grouped_matmul import Ops

try:
    import torch_npu
    from mindspeed.ops.npu_moe_token_permute import npu_moe_token_permute  # pylint: disable=ungrouped-imports
    from mindspeed.ops.npu_moe_token_unpermute import npu_moe_token_unpermute  # pylint: disable=ungrouped-imports
except ImportError:
    pass


from mindspeed_llm.fsdp2.utils.global_vars import get_args

if is_causal_conv1d_available():
    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
else:
    causal_conv1d_update, causal_conv1d_fn = None, None

if is_flash_linear_attention_available():
    from fla.modules import FusedRMSNormGated
    from fla.ops.gated_delta_rule import chunk_gated_delta_rule, fused_recurrent_gated_delta_rule
else:
    chunk_gated_delta_rule, fused_recurrent_gated_delta_rule = None, None
    FusedRMSNormGated = None

logger = logging.get_logger(__name__)

_GLOBAL_ATTN_MASK = None


class Qwen3NextRMSNormGated(nn.Module):
    def __init__(self, hidden_size, eps=1e-6, **kwargs):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps
        self.args = get_args()

    def forward(self, hidden_states, gate=None):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        if self.args.use_fused_rmsnorm:
            hidden_states = torch_npu.npu_rms_norm(hidden_states, self.weight.float(), epsilon=self.variance_epsilon)[0]
        else:
            variance = hidden_states.pow(2).mean(-1, keepdim=True)
            # Norm before gate
            hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
            hidden_states = self.weight * hidden_states.to(input_dtype)
        hidden_states = hidden_states * F.silu(gate.to(torch.float32))

        return hidden_states.to(input_dtype)


class Qwen3NextDynamicCache:
    """
    A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the linear attention
    cache (which has a constant shape regardless of seq_len).

    This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states`
    and `ssm_states` for gated deltanet cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor
    For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`,
    while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors).
    For linear attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors),
    while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`,
    and `recurrent_states` represents the recurrent state and has a shape of `(batch_size, d_inner, d_state)`.
    """

    is_compileable = False

    def __init__(self, config: Qwen3NextConfig):
        super().__init__()
        self.layer_types = config.layer_types
        self.transformer_layers = [
            i for i in range(config.num_hidden_layers) if self.layer_types[i] == "full_attention"
        ]
        self.last_linear_layer = len(self.layer_types) - 1 - self.layer_types[::-1].index("linear_attention")

        # Initialize everything to None -> will be lazy initialized to allow multi-gpu (device_map) inference
        self.conv_states = [None for _ in range(config.num_hidden_layers)]
        self.recurrent_states = [None for _ in range(config.num_hidden_layers)]
        self.key_cache = [None for _ in range(config.num_hidden_layers)]
        self.value_cache = [None for _ in range(config.num_hidden_layers)]

    def __len__(self):
        return len(self.layer_types)

    def __getitem__(self, layer_idx: int) -> tuple[torch.Tensor, torch.Tensor]:
        return self.key_cache[layer_idx], self.value_cache[layer_idx]

    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]:
        if self.key_cache[layer_idx] is None:
            self.key_cache[layer_idx] = key_states
            self.value_cache[layer_idx] = value_states
        else:
            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)

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

    def reorder_cache(self, beam_idx: torch.LongTensor):
        """Reorders the cache for beam search, given the selected beam indices."""
        for layer_idx in range(len(self.key_cache)):
            if self.key_cache[layer_idx] is not None:
                device = self.key_cache[layer_idx].device
                beam_idx = beam_idx.to(device)
                self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx)
                self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx)

            if self.conv_states[layer_idx] is not None:
                device = self.conv_states[layer_idx].device
                beam_idx = beam_idx.to(device)
                self.conv_states[layer_idx] = self.conv_states[layer_idx].index_select(0, beam_idx)
                self.recurrent_states[layer_idx] = self.recurrent_states[layer_idx].index_select(0, beam_idx)

    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
        # take any layer that contains cache and not empty tensor
        layer_idx = self.transformer_layers[0] if layer_idx not in self.transformer_layers else layer_idx
        if len(self.key_cache) <= layer_idx or self.key_cache[layer_idx] is None:
            return 0
        return self.key_cache[layer_idx].shape[-2]

    def get_mask_sizes(self, cache_position: torch.Tensor, layer_idx: int) -> tuple[int, int]:
        """
        Return a tuple (kv_length, kv_offset) corresponding to the length and offset that will be returned for
        the given layer at `layer_idx`.
        The masks are then prepared according to the given lengths (kv_length, kv_offset) and patterns for each layer.
        """
        kv_offset = 0
        query_length = cache_position.shape[0]
        past_seen_tokens = self.get_seq_length(layer_idx)
        kv_length = query_length + past_seen_tokens
        return kv_length, kv_offset

    @property
    def has_previous_state(self):
        """We have a previous state if the last linear (conv) layer was already updated."""
        return self.conv_states[self.last_linear_layer] is not None


class Qwen3NextRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: Qwen3NextConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
            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)


class Qwen3NextRMSNorm(nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.zeros(dim))
        self.args = get_args()

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        if self.args.use_fused_rmsnorm:
            return torch_npu.npu_rms_norm(x.float(), 1.0 + self.weight.float(), epsilon=self.eps)[0].type_as(x)
        else:
            output = self._norm(x.float())
            # Llama does x.to(float16) * w whilst Qwen3Next is (x * w).to(float16)
            # See https://github.com/huggingface/transformers/pull/29402
            output = output * (1.0 + self.weight.float())
            return output.type_as(x)

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


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)


# Adapted from transformers.models.glm.modular_glm.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.

    Removes the interleaving of cos and sin from GLM

    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)

    # Keep half or full tensor for later concatenation
    rotary_dim = cos.shape[-1]
    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]

    args = get_args()
    if args.use_fused_rotary_pos_emb:
        q_embed = torch_npu.npu_rotary_mul(q_rot, cos, sin).to(q.dtype)
        k_embed = torch_npu.npu_rotary_mul(k_rot, cos, sin).to(k.dtype)
        q_embed = torch.cat([q_embed, q_pass], dim=-1)
        k_embed = torch.cat([k_embed, k_pass], dim=-1)
        return q_embed, k_embed
    else:
        # Apply rotary embeddings on the first half or full tensor
        q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
        k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)
        # Concatenate back to full shape
        q_embed = torch.cat([q_embed, q_pass], dim=-1)
        k_embed = torch.cat([k_embed, k_pass], dim=-1)
    return q_embed, k_embed


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)


def flash_attention_forward(
    module: torch.nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    input_layout: str,
    dropout: float = 0.0,
    scaling: Optional[float] = None,
    is_causal: Optional[bool] = None,
    **kwargs,
) -> tuple[torch.Tensor, None]:
    if kwargs.get("output_attentions", False) or kwargs.get("head_mask") is not None:
        logger.warning_once(
            "`sdpa` attention does not support `output_attentions=True` or `head_mask`."
            " Please set your attention to `eager` if you want any of these features."
        )

    if hasattr(module, "num_key_value_groups"):
        key = repeat_kv(key, module.num_key_value_groups)
        value = repeat_kv(value, module.num_key_value_groups)

    if attention_mask is not None and attention_mask.ndim == 4:
        attention_mask = attention_mask[:, :, :, : key.shape[-2]]

    if is_causal is None:
        # The last condition is for encoder (decoder) models which specify this by passing their own `is_causal` flag
        # This is mainly due to those models having mixed implementations for encoder, decoder, and encoder-decoder attns
        is_causal = query.shape[2] > 1 and attention_mask is None and getattr(module, "is_causal", True)

    # Shapes (e.g. query.shape[2]) are tensors during jit tracing, resulting in `is_causal` being a tensor.
    # We convert it to a bool for the SDPA kernel that only accepts bools.
    if torch.jit.is_tracing() and isinstance(is_causal, torch.Tensor):
        is_causal = is_causal.item()

    # When `is_causal = False` and the `attention_mask` is not of boolean type, the Ascend NPU's SDPA interface cannot utilize the FlashAttentionScore operator，
    # and falls back to small-operator concatenation. To invoke the FlashAttentionScore, the attention_mask must be converted to boolean type.
    # This adaptation ensures the `attention_mask` meets the requirement for using FlashAttentionScore.
    if attention_mask is not None and attention_mask.dtype != torch.bool:
        # Convert to boolean type, making sdpa to force call FlashAttentionScore to improve performance.
        attention_mask = torch.logical_not(attention_mask.bool()).to(query.device)

    cu_seqlens = None
    if "actual_seq_len" in kwargs:
        cu_seqlens = kwargs.get("actual_seq_len", None).tolist()
    seq_len = query.shape[2]
    if cu_seqlens is not None:
        input_layout = "TND"
        query, key, value = [rearrange(x, 'b h s d -> (b s) h d') for x in [query, key, value]]

    attn_output = torch_npu.npu_fusion_attention(
        query,
        key,
        value,
        head_num=query.shape[1],
        input_layout=input_layout,
        atten_mask=attention_mask,
        keep_prob=1 - dropout,
        actual_seq_qlen=cu_seqlens,
        actual_seq_kvlen=cu_seqlens,
        scale=scaling,
        sparse_mode=2,
    )[0]

    if input_layout == "BNSD":
        attn_output = attn_output.transpose(1, 2).contiguous()
    elif input_layout == "TND":
        attn_output = rearrange(attn_output, '(b s) h d -> b s h d', s=seq_len)

    return attn_output, None


class Qwen3NextAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: Qwen3NextConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True
        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim * 2, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )
        self.q_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # unlike olmo, only on the head dim!
        self.k_norm = Qwen3NextRMSNorm(self.head_dim, eps=config.rms_norm_eps)  # thus post q_norm does not need reshape

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        past_key_values: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states, gate = torch.chunk(
            self.q_proj(hidden_states).view(*input_shape, -1, self.head_dim * 2), 2, dim=-1
        )
        gate = gate.reshape(*input_shape, -1)

        query_states = self.q_norm(query_states.view(hidden_shape)).transpose(1, 2)
        key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_values 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_values.update(key_states, value_states, self.layer_idx, cache_kwargs)

        attention_interface: Callable = flash_attention_forward

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            input_layout="BNSD",
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = attn_output * torch.sigmoid(gate)

        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


def apply_mask_to_padding_states(hidden_states, attention_mask):
    """
    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
    """
    if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
        dtype = hidden_states.dtype
        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)

    return hidden_states


is_fast_path_available = all(
    (causal_conv1d_fn, causal_conv1d_update, chunk_gated_delta_rule, fused_recurrent_gated_delta_rule)
)


def torch_causal_conv1d_update(
    hidden_states,
    conv_state,
    weight,
    bias=None,
    activation=None,
):
    _, hidden_size, seq_len = hidden_states.shape
    state_len = conv_state.shape[-1]

    hidden_states_new = torch.cat([conv_state, hidden_states], dim=-1).to(weight.dtype)
    conv_state.copy_(hidden_states_new[:, :, -state_len:])
    out = F.conv1d(hidden_states_new, weight.unsqueeze(1), bias, padding=0, groups=hidden_size)
    out = F.silu(out[:, :, -seq_len:])
    out = out.to(hidden_states.dtype)
    return out


def l2norm(x: torch.FloatTensor, dim: int = -1, eps: float = 1e-6):
    """This function is intended to align with the l2norm implementation in the FLA library."""
    inv_norm = torch.rsqrt((x * x).sum(dim=dim, keepdim=True) + eps)
    return x * inv_norm


def torch_chunk_gated_delta_rule(
    query,
    key,
    value,
    g,
    beta,
    chunk_size=64,
    initial_state=None,
    output_final_state=False,
    use_qk_l2norm_in_kernel=False,
    cu_seqlens=None,
):
    if cu_seqlens is not None:
        raise NotImplementedError("torch_chunk_gated_delta_rule does not support cu_seqlens")
    initial_dtype = query.dtype
    if use_qk_l2norm_in_kernel:
        query = l2norm(query, dim=-1, eps=1e-6)
        key = l2norm(key, dim=-1, eps=1e-6)
    query, key, value, beta, g = [
        x.transpose(1, 2).contiguous().to(torch.float32) for x in (query, key, value, beta, g)
    ]

    batch_size, num_heads, sequence_length, k_head_dim = key.shape
    v_head_dim = value.shape[-1]
    pad_size = (chunk_size - sequence_length % chunk_size) % chunk_size
    query = F.pad(query, (0, 0, 0, pad_size))
    key = F.pad(key, (0, 0, 0, pad_size))
    value = F.pad(value, (0, 0, 0, pad_size))
    beta = F.pad(beta, (0, pad_size))
    g = F.pad(g, (0, pad_size))
    total_sequence_length = sequence_length + pad_size
    scale = 1 / (query.shape[-1] ** 0.5)
    query = query * scale

    v_beta = value * beta.unsqueeze(-1)
    k_beta = key * beta.unsqueeze(-1)
    # reshape to chunks
    query, key, value, k_beta, v_beta = [
        x.reshape(x.shape[0], x.shape[1], -1, chunk_size, x.shape[-1]) for x in (query, key, value, k_beta, v_beta)
    ]
    g = g.reshape(g.shape[0], g.shape[1], -1, chunk_size)
    mask = torch.triu(torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=0)

    # chunk decay
    g = g.cumsum(dim=-1)
    decay_mask = ((g.unsqueeze(-1) - g.unsqueeze(-2)).tril().exp().float()).tril()
    attn = -((k_beta @ key.transpose(-1, -2)) * decay_mask).masked_fill(mask, 0)
    for i in range(1, chunk_size):
        row = attn[..., i, :i].clone()
        sub = attn[..., :i, :i].clone()
        attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2)
    attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device)
    value = attn @ v_beta
    k_cumdecay = attn @ (k_beta * g.exp().unsqueeze(-1))
    last_recurrent_state = (
        torch.zeros(batch_size, num_heads, k_head_dim, v_head_dim).to(value)
        if initial_state is None
        else initial_state.to(value)
    )
    core_attn_out = torch.zeros_like(value)
    mask = torch.triu(torch.ones(chunk_size, chunk_size, dtype=torch.bool, device=query.device), diagonal=1)

    # for each chunk
    for i in range(0, total_sequence_length // chunk_size):
        q_i, k_i, v_i = query[:, :, i], key[:, :, i], value[:, :, i]
        attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
        v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
        v_new = v_i - v_prime
        attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
        core_attn_out[:, :, i] = attn_inter + attn @ v_new
        last_recurrent_state = (
            last_recurrent_state * g[:, :, i, -1, None, None].exp()
            + (k_i * (g[:, :, i, -1, None] - g[:, :, i]).exp()[..., None]).transpose(-1, -2) @ v_new
        )

    if not output_final_state:
        last_recurrent_state = None
    core_attn_out = core_attn_out.reshape(core_attn_out.shape[0], core_attn_out.shape[1], -1, core_attn_out.shape[-1])
    core_attn_out = core_attn_out[:, :, :sequence_length]
    core_attn_out = core_attn_out.transpose(1, 2).contiguous().to(initial_dtype)
    return core_attn_out, last_recurrent_state


def torch_recurrent_gated_delta_rule(
    query, key, value, g, beta, initial_state, output_final_state, use_qk_l2norm_in_kernel=False
):
    initial_dtype = query.dtype
    if use_qk_l2norm_in_kernel:
        query = l2norm(query, dim=-1, eps=1e-6)
        key = l2norm(key, dim=-1, eps=1e-6)
    query, key, value, beta, g = [
        x.transpose(1, 2).contiguous().to(torch.float32) for x in (query, key, value, beta, g)
    ]

    batch_size, num_heads, sequence_length, k_head_dim = key.shape
    v_head_dim = value.shape[-1]
    scale = 1 / (query.shape[-1] ** 0.5)
    query = query * scale

    core_attn_out = torch.zeros(batch_size, num_heads, sequence_length, v_head_dim).to(value)
    last_recurrent_state = (
        torch.zeros(batch_size, num_heads, k_head_dim, v_head_dim).to(value)
        if initial_state is None
        else initial_state.to(value)
    )

    for i in range(sequence_length):
        q_t = query[:, :, i]
        k_t = key[:, :, i]
        v_t = value[:, :, i]
        g_t = g[:, :, i].exp().unsqueeze(-1).unsqueeze(-1)
        beta_t = beta[:, :, i].unsqueeze(-1)

        last_recurrent_state = last_recurrent_state * g_t
        kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
        delta = (v_t - kv_mem) * beta_t
        last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta.unsqueeze(-2)
        core_attn_out[:, :, i] = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)

    if not output_final_state:
        last_recurrent_state = None
    core_attn_out = core_attn_out.transpose(1, 2).contiguous().to(initial_dtype)
    return core_attn_out, last_recurrent_state


class Qwen3NextGatedDeltaNet(nn.Module):
    def __init__(self, config: Qwen3NextConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.num_v_heads = config.linear_num_value_heads
        self.num_k_heads = config.linear_num_key_heads
        self.head_k_dim = config.linear_key_head_dim
        self.head_v_dim = config.linear_value_head_dim
        self.key_dim = self.head_k_dim * self.num_k_heads
        self.value_dim = self.head_v_dim * self.num_v_heads

        self.conv_kernel_size = config.linear_conv_kernel_dim
        self.layer_idx = layer_idx
        self.activation = config.hidden_act
        self.act = ACT2FN[config.hidden_act]
        self.layer_norm_epsilon = config.rms_norm_eps

        # QKV
        self.conv_dim = self.key_dim * 2 + self.value_dim
        self.conv1d = nn.Conv1d(
            in_channels=self.conv_dim,
            out_channels=self.conv_dim,
            bias=False,
            kernel_size=self.conv_kernel_size,
            groups=self.conv_dim,
            padding=self.conv_kernel_size - 1,
        )

        # projection of the input hidden states
        projection_size_qkvz = self.key_dim * 2 + self.value_dim * 2
        projection_size_ba = self.num_v_heads * 2
        self.in_proj_qkvz = nn.Linear(self.hidden_size, projection_size_qkvz, bias=False)
        self.in_proj_ba = nn.Linear(self.hidden_size, projection_size_ba, bias=False)

        # time step projection (discretization)
        # instantiate once and copy inv_dt in init_weights of PretrainedModel
        self.dt_bias = nn.Parameter(torch.ones(self.num_v_heads))

        A = torch.empty(self.num_v_heads).uniform_(0, 16)
        self.A_log = nn.Parameter(torch.log(A))

        self.norm = (
            Qwen3NextRMSNormGated(self.head_v_dim, eps=self.layer_norm_epsilon)
            if FusedRMSNormGated is None
            else FusedRMSNormGated(
                self.head_v_dim,
                eps=self.layer_norm_epsilon,
                activation=self.activation,
                device=torch.cuda.current_device(),
                dtype=config.dtype if config.dtype is not None else torch.get_current_dtype(),
            )
        )

        self.out_proj = nn.Linear(self.value_dim, self.hidden_size, bias=False)

        self.causal_conv1d_fn = causal_conv1d_fn
        self.causal_conv1d_update = causal_conv1d_update or torch_causal_conv1d_update
        args = get_args()
        self.gdn_chunk_size = args.gdn_chunk_size

        if args.use_flash_gdn:
            from mindspeed_llm.tasks.models.transformer.flash_gated_delta_rule import flash_gated_delta_rule

            self.chunk_gated_delta_rule = flash_gated_delta_rule
        elif args.use_triton_gdn:
            from mindspeed_llm.tasks.models.transformer.chunk_gated_delta_rule import chunk_gated_delta_rule  # pylint: disable=redefined-outer-name

            self.chunk_gated_delta_rule = chunk_gated_delta_rule
        else:
            self.chunk_gated_delta_rule = torch_chunk_gated_delta_rule

        self.recurrent_gated_delta_rule = fused_recurrent_gated_delta_rule or torch_recurrent_gated_delta_rule

        if not is_fast_path_available:
            logger.warning_once(
                "The fast path is not available because one of the required library is not installed. Falling back to "
                "torch implementation. To install follow https://github.com/fla-org/flash-linear-attention#installation and"
                " https://github.com/Dao-AILab/causal-conv1d"
            )

    def fix_query_key_value_ordering(self, mixed_qkvz, mixed_ba):
        """
        Derives `query`, `key` and `value` tensors from `mixed_qkvz` and `mixed_ba`.
        """

        new_tensor_shape_qkvz = mixed_qkvz.size()[:-1] + (
            self.num_k_heads,
            2 * self.head_k_dim + 2 * self.head_v_dim * self.num_v_heads // self.num_k_heads,
        )
        new_tensor_shape_ba = mixed_ba.size()[:-1] + (self.num_k_heads, 2 * self.num_v_heads // self.num_k_heads)

        mixed_qkvz = mixed_qkvz.view(*new_tensor_shape_qkvz)
        mixed_ba = mixed_ba.view(*new_tensor_shape_ba)
        split_arg_list_qkvz = [
            self.head_k_dim,
            self.head_k_dim,
            (self.num_v_heads // self.num_k_heads * self.head_v_dim),
            (self.num_v_heads // self.num_k_heads * self.head_v_dim),
        ]
        split_arg_list_ba = [self.num_v_heads // self.num_k_heads, self.num_v_heads // self.num_k_heads]
        query, key, value, z = torch.split(mixed_qkvz, split_arg_list_qkvz, dim=3)
        b, a = torch.split(mixed_ba, split_arg_list_ba, dim=3)
        # [b, sq, ng, np/ng * hn] -> [b, sq, np, hn]
        value = value.reshape(value.size(0), value.size(1), -1, self.head_v_dim)
        z = z.reshape(z.size(0), z.size(1), -1, self.head_v_dim)
        b = b.reshape(b.size(0), b.size(1), self.num_v_heads)
        a = a.reshape(a.size(0), a.size(1), self.num_v_heads)
        return query, key, value, z, b, a

    def forward(
        self,
        hidden_states: torch.Tensor,
        cache_params: Optional[Qwen3NextDynamicCache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ):
        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)

        # Set up dimensions for reshapes later
        batch_size, seq_len, _ = hidden_states.shape

        use_precomputed_states = (
            cache_params is not None and cache_params.has_previous_state and seq_len == 1 and cache_position is not None
        )
        conv_state = None
        recurrent_state = None

        # getting projected states from cache if it exists
        if cache_params is not None:
            conv_state = cache_params.conv_states[self.layer_idx]
            recurrent_state = cache_params.recurrent_states[self.layer_idx]

        projected_states_qkvz = self.in_proj_qkvz(hidden_states)
        projected_states_ba = self.in_proj_ba(hidden_states)
        query, key, value, z, b, a = self.fix_query_key_value_ordering(projected_states_qkvz, projected_states_ba)
        query, key, value = (x.reshape(x.shape[0], x.shape[1], -1) for x in (query, key, value))

        mixed_qkv = torch.cat((query, key, value), dim=-1)
        mixed_qkv = mixed_qkv.transpose(1, 2)

        if use_precomputed_states:
            # 2. Convolution sequence transformation
            # NOTE: the conv state is updated in `causal_conv1d_update`
            mixed_qkv = self.causal_conv1d_update(
                mixed_qkv,
                conv_state,
                self.conv1d.weight.squeeze(1),
                self.conv1d.bias,
                self.activation,
            )
        else:
            if cache_params is not None:
                conv_state = F.pad(mixed_qkv, (self.conv_kernel_size - mixed_qkv.shape[-1], 0))
                cache_params.conv_states[self.layer_idx] = conv_state
            if self.causal_conv1d_fn is not None:
                mixed_qkv = self.causal_conv1d_fn(
                    x=mixed_qkv,
                    weight=self.conv1d.weight.squeeze(1),
                    bias=self.conv1d.bias,
                    activation=self.activation,
                    seq_idx=None,
                )
            else:
                mixed_qkv = F.silu(self.conv1d(mixed_qkv)[:, :, :seq_len])

        mixed_qkv = mixed_qkv.transpose(1, 2)
        query, key, value = torch.split(
            mixed_qkv,
            [
                self.key_dim,
                self.key_dim,
                self.value_dim,
            ],
            dim=-1,
        )
        query = query.reshape(query.shape[0], query.shape[1], -1, self.head_k_dim)
        key = key.reshape(key.shape[0], key.shape[1], -1, self.head_k_dim)
        value = value.reshape(value.shape[0], value.shape[1], -1, self.head_v_dim)

        beta = b.sigmoid()
        # If the model is loaded in fp16, without the .float() here, A might be -inf
        g = -self.A_log.float().exp() * F.softplus(a.float() + self.dt_bias)
        if self.num_v_heads // self.num_k_heads > 1:
            query = query.repeat_interleave(self.num_v_heads // self.num_k_heads, dim=2)
            key = key.repeat_interleave(self.num_v_heads // self.num_k_heads, dim=2)

        if not use_precomputed_states:
            cu_seqlens = None
            input_layout = "BSND"
            if "actual_seq_len" in kwargs:
                cu_seqlens = kwargs.get("actual_seq_len", None)
            if cu_seqlens is not None:
                cu_seqlens = F.pad(cu_seqlens, pad=(1, 0), value=0)
                input_layout = "TND"
                query, key, value = [rearrange(x, 'b s h d -> 1 (b s) h d') for x in [query, key, value]]

            core_attn_out, last_recurrent_state = self.chunk_gated_delta_rule(
                query,
                key,
                value,
                g=g,
                beta=beta,
                initial_state=None,
                output_final_state=cache_params is not None,
                use_qk_l2norm_in_kernel=True,
                chunk_size=self.gdn_chunk_size,
                cu_seqlens=cu_seqlens,
            )
            if input_layout == "TND":
                core_attn_out = rearrange(core_attn_out, '1 (b s) h d -> b s h d', s=seq_len)

        else:
            core_attn_out, last_recurrent_state = self.recurrent_gated_delta_rule(
                query,
                key,
                value,
                g=g,
                beta=beta,
                initial_state=recurrent_state,
                output_final_state=cache_params is not None,
                use_qk_l2norm_in_kernel=True,
            )

        # Update cache
        if cache_params is not None:
            cache_params.recurrent_states[self.layer_idx] = last_recurrent_state

        z_shape_og = z.shape
        # reshape input data into 2D tensor
        core_attn_out = core_attn_out.reshape(-1, core_attn_out.shape[-1])
        z = z.reshape(-1, z.shape[-1])
        core_attn_out = self.norm(core_attn_out, z)
        core_attn_out = core_attn_out.reshape(z_shape_og)
        core_attn_out = core_attn_out.reshape(core_attn_out.shape[0], core_attn_out.shape[1], -1)

        output = self.out_proj(core_attn_out)
        return output


class Qwen3NextMLP(nn.Module):
    def __init__(self, config, intermediate_size=None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = intermediate_size if intermediate_size is not None else config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

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


class Qwen3NextSparseMoeBlock(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.num_experts = config.num_experts
        self.top_k = config.num_experts_per_tok
        self.norm_topk_prob = config.norm_topk_prob

        # gating
        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
        self.experts = nn.ModuleList(
            [Qwen3NextMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)]
        )

        self.shared_expert = Qwen3NextMLP(config, intermediate_size=config.shared_expert_intermediate_size)
        self.shared_expert_gate = torch.nn.Linear(config.hidden_size, 1, bias=False)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
        # router_logits: (batch * sequence_length, n_experts)
        router_logits = self.gate(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)
        if self.norm_topk_prob:
            routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
        # 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
        expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
        for expert_idx in expert_hit:
            expert_layer = self.experts[expert_idx]
            idx, top_x = torch.where(expert_mask[expert_idx].squeeze(0))

            # 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].reshape(-1, hidden_dim)
            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, 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))

        shared_expert_output = self.shared_expert(hidden_states)
        shared_expert_output = F.sigmoid(self.shared_expert_gate(hidden_states)) * shared_expert_output

        final_hidden_states = final_hidden_states + shared_expert_output

        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
        return final_hidden_states, router_logits


class Qwen3NextMoeExperts(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.num_experts = config.num_experts
        self.hidden_dim = config.hidden_size
        self.intermediate_size = config.moe_intermediate_size
        self.gate_up_proj = torch.nn.Parameter(
            torch.empty(self.num_experts * self.hidden_dim, 2 * self.intermediate_size)
        )

        self.down_proj = torch.nn.Parameter(torch.empty(self.num_experts * self.intermediate_size, self.hidden_dim))
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, hidden_states, routing_weights=None, selected_experts=None):
        gate_up_proj, down_proj = self._view_experts_weight()

        # permute
        permuted_hidden_states, row_ids_map = npu_moe_token_permute(hidden_states, selected_experts.to(torch.int32))
        tokens_per_expert = torch.histc(selected_experts, bins=self.num_experts, min=0, max=self.num_experts)

        fc1_output = Ops.gmm(permuted_hidden_states, gate_up_proj, tokens_per_expert.to(torch.int64), trans_b=False)
        fc1_activation = torch_npu.npu_swiglu(fc1_output, dim=-1)
        fc2_out = Ops.gmm(fc1_activation, down_proj, tokens_per_expert.to(torch.int64), trans_b=False)

        # unpermute
        output = npu_moe_token_unpermute(fc2_out, row_ids_map, probs=routing_weights)
        return output

    def _view_experts_weight(self):
        gate_up_proj = self.gate_up_proj.to_local() if isinstance(self.gate_up_proj, DTensor) else self.gate_up_proj
        gate_up_proj = gate_up_proj.view(self.num_experts, self.hidden_dim, -1)

        down_proj = self.down_proj.to_local() if isinstance(self.down_proj, DTensor) else self.down_proj
        down_proj = down_proj.view(self.num_experts, -1, self.hidden_dim)
        return gate_up_proj, down_proj


class Qwen3NextMoeSharedExpert(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.hidden_dim = config.hidden_size
        self.intermediate_size = config.shared_expert_intermediate_size

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

    def forward(self, hidden_states):
        gate_out = self.gate_proj(hidden_states)
        up_out = self.up_proj(hidden_states)
        activated = F.silu(gate_out) * up_out
        output = self.down_proj(activated)
        return output


class Qwen3NextSparseFusedMoeBlock(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.num_experts = config.num_experts
        self.top_k = config.num_experts_per_tok
        self.norm_topk_prob = config.norm_topk_prob
        self.hidden_size = config.hidden_size

        # --- gating ---
        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
        self.experts = Qwen3NextMoeExperts(config)

        # --- Shared expert ---
        self.shared_expert = Qwen3NextMoeSharedExpert(config)
        self.shared_expert_gate = torch.nn.Linear(config.hidden_size, 1, bias=False)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
        router_logits = self.gate(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)
        if self.norm_topk_prob:
            routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
        routing_weights = routing_weights.to(hidden_states.dtype)

        final_hidden_states = self.experts(
            hidden_states, routing_weights=routing_weights, selected_experts=selected_experts
        )

        # --- Shared expert ---
        shared_expert_output = self.shared_expert(hidden_states)
        shared_expert_output = F.sigmoid(self.shared_expert_gate(hidden_states)) * shared_expert_output

        final_hidden_states = final_hidden_states + shared_expert_output
        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
        return final_hidden_states, router_logits


class Qwen3NextDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Qwen3NextConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        # token mixer
        self.layer_type = config.layer_types[layer_idx]
        if self.layer_type == "linear_attention":
            self.linear_attn = Qwen3NextGatedDeltaNet(config, layer_idx)
        elif self.layer_type == "full_attention":
            self.self_attn = Qwen3NextAttention(config, layer_idx)

        if (layer_idx not in config.mlp_only_layers) and (
            config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
        ):
            self.args = get_args()
            if self.args.moe_grouped_gemm:
                self.mlp = Qwen3NextSparseFusedMoeBlock(config)
            else:
                self.mlp = Qwen3NextSparseMoeBlock(config)
        else:
            self.mlp = Qwen3NextMLP(config, intermediate_size=config.intermediate_size)

        self.input_layernorm = Qwen3NextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Qwen3NextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> torch.FloatTensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            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`).
            past_key_values (`Cache`, *optional*): cached past key and value projection states
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence.
            position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                with `head_dim` being the embedding dimension of each attention head.
            kwargs (`dict`, *optional*):
                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
                into the model
        """
        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Token Mixer
        if self.layer_type == "linear_attention":
            hidden_states = self.linear_attn(
                hidden_states=hidden_states,
                cache_params=past_key_values,
                cache_position=cache_position,
                attention_mask=attention_mask,
                **kwargs,
            )
        elif self.layer_type == "full_attention":
            # Self Attention
            hidden_states, _ = self.self_attn(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        # For the MoE layers, we need to unpack
        if isinstance(hidden_states, tuple):
            hidden_states, _ = hidden_states
        hidden_states = residual + hidden_states

        return hidden_states


class Qwen3NextPreTrainedModel(PreTrainedModel):
    config: Qwen3NextConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Qwen3NextDecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _keys_to_ignore_on_load_unexpected = [r"^mtp.*"]
    _can_record_outputs = {
        "router_logits": OutputRecorder(Qwen3NextSparseFusedMoeBlock, index=1),
        "hidden_states": Qwen3NextDecoderLayer,
        "attentions": Qwen3NextAttention,
    }
    _is_stateful = True

    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, Qwen3NextGatedDeltaNet):
            module.dt_bias.data.fill_(1.0)
            module.A_log.data.uniform_(0, 16).log_()
        # We initialize with 0s to be 1 centered as the RMSNorm here does (1 + weight)
        elif isinstance(module, Qwen3NextRMSNorm):
            module.weight.data.zero_()


class Qwen3NextModel(Qwen3NextPreTrainedModel):
    def __init__(self, config: Qwen3NextConfig):
        super().__init__(config)
        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, config.pad_token_id)
        self.layers = nn.ModuleList(
            [Qwen3NextDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = Qwen3NextRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Qwen3NextRotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()

    @check_model_inputs
    @auto_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,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> MoeModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

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

        if use_cache and past_key_values is None:
            past_key_values = Qwen3NextDynamicCache(config=self.config)

        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)

        causal_mask = get_attention_mask_in_transformers()
        linear_attn_mask = None

        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            layer_mask = linear_attn_mask if decoder_layer.layer_type == "linear_attention" else causal_mask

            hidden_states = decoder_layer(
                hidden_states,
                position_embeddings=position_embeddings,
                attention_mask=layer_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                cache_position=cache_position,
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)

        return MoeModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )

    def _update_linear_attn_mask(self, attention_mask, cache_position):
        """
        NOTE: Left-padding is used for linear attention mask.
        No need for zeroing states when
            1. Cached forward
            2. Attending to all inputs
        """
        linear_attn_mask = attention_mask
        if cache_position[0] > 0 or (attention_mask is not None and torch.all(attention_mask == 1)):
            linear_attn_mask = None
        return linear_attn_mask


def get_attention_mask_in_transformers():
    global _GLOBAL_ATTN_MASK
    if _GLOBAL_ATTN_MASK is not None:
        return _GLOBAL_ATTN_MASK

    _GLOBAL_ATTN_MASK = torch.triu(
        torch.ones((2048, 2048), device=torch.accelerator.current_accelerator().type, dtype=torch.bool), diagonal=1
    )

    return _GLOBAL_ATTN_MASK


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://huggingface.co/papers/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

    concatenated_gate_logits = None

    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


@auto_docstring
class Qwen3NextForCausalLM(Qwen3NextPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]
    _tp_plan = {"lm_head": "colwise_rep"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}

    def __init__(self, config):
        super().__init__(config)
        self.model = Qwen3NextModel(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()

    @can_return_tuple
    @auto_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[Qwen3NextDynamicCache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_router_logits: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> 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]`.

        Example:

        ```python
        >>> from transformers import AutoTokenizer, Qwen3NextForCausalLM

        >>> model = Qwen3NextForCausalLM.from_pretrained("Qwen/Qwen3-Next-80B-A3B-Instruct")
        >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-Next-80B-A3B-Instruct")

        >>> prompt = "Hey, are you conscious? Can you talk to me?"
        >>> inputs = tokenizer(prompt, return_tensors="pt")

        >>> # Generate
        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
        ```
        """

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

        # 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_router_logits=output_router_logits,
            cache_position=cache_position,
            **kwargs,
        )

        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:
            loss = self.loss_function(logits, labels, self.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.router_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,
        )


class Qwen3NextForSequenceClassification(GenericForSequenceClassification, Qwen3NextPreTrainedModel):
    pass


class Qwen3NextForTokenClassification(GenericForTokenClassification, Qwen3NextPreTrainedModel):
    pass


class Qwen3NextForQuestionAnswering(GenericForQuestionAnswering, Qwen3NextPreTrainedModel):
    base_model_prefix = "transformer"  # For BC, where `transformer` was used instead of `model`


__all__ = [
    "Qwen3NextForCausalLM",
    "Qwen3NextForQuestionAnswering",
    "Qwen3NextModel",
    "Qwen3NextPreTrainedModel",
    "Qwen3NextForSequenceClassification",
    "Qwen3NextForTokenClassification",
]