#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/hy_v4/modular_hy_v4.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_hy_v4.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2026 Tencent HunYuan Team 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.

import math
from collections.abc import Callable

import torch
import torch.nn as nn
import torch.nn.functional as F

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...integrations import use_experts_implementation, use_kernel_forward_from_hub
from ...masking_utils import create_causal_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from ...processing_utils import Unpack
from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
from ...utils.generic import maybe_autocast, merge_with_config_defaults
from ...utils.output_capturing import OutputRecorder, capture_outputs
from .configuration_hy_v4 import HYV4Config


@use_kernel_forward_from_hub("RMSNorm")
class HYV4RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        HYV4RMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        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}"


class HYV4UnweightedRMSNorm(nn.Module):
    def __init__(self, eps: float = 1.0e-6):
        super().__init__()
        self.eps = eps

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # We compute the inverse (i.e. we do not apply `x * ...`)
        return torch.rsqrt(x.float().square().mean(-1, keepdim=True) + self.eps).to(x.dtype)


class HYV4RotaryEmbedding(nn.Module):
    def __init__(self, config: HYV4Config):
        super().__init__()
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config

        self.rope_type = self.config.rope_parameters["rope_type"]
        rope_init_fn: Callable = self.compute_default_rope_parameters
        if self.rope_type != "default":
            rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
        inv_freq, self.attention_scaling = rope_init_fn(self.config)

        self.inv_freq = nn.Buffer(inv_freq, persistent=False)
        self.original_inv_freq = nn.Buffer(inv_freq.clone(), persistent=False)

    @staticmethod
    def compute_default_rope_parameters(config: HYV4Config, **kwargs) -> tuple[torch.Tensor, float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        base = config.rope_parameters["rope_theta"]
        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads

        attention_factor = 1.0  # Unused in this type of RoPE
        # Compute the inverse frequencies
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
        return inv_freq, attention_factor

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        freqs = position_ids[..., None].float() * self.inv_freq.to(device=x.device, dtype=torch.float)
        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)


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)


@use_kernel_forward_from_hub("rotary_pos_emb")
def apply_rotary_pos_emb(q, k, cos, sin, 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.
        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 HYV4Indexer(nn.Module):
    """
    DeepSeek Sparse Attention (DSA) indexer for selecting top-k tokens.

    The Indexer has its own lightweight projections (wq_b, wk) separate from the main MLA attention,
    and returns the additive top-k sparse mask directly (`0` at the selected tokens, `-inf` elsewhere);
    the raw top-k indices are only ever scattered into that mask, so they are not surfaced.

    **Cache strategy**: the indexer key cache lives on the per-layer `DynamicIndexedLayer` (or the
    `StaticIndexedLayer` for static caches) inside the shared cache, accessed via
    `past_key_values.update_indexer()`.
    """

    def __init__(self, config: HYV4Config, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx

        self.hidden_size: int = config.hidden_size
        self.n_heads: int = config.index_n_heads
        self.head_dim: int = config.index_head_dim
        self.qk_rope_head_dim: int = config.qk_rope_head_dim
        self.index_topk: int = config.index_topk
        self.q_lora_rank: int = config.q_lora_rank

        self.wq_b = nn.Linear(self.q_lora_rank, self.n_heads * self.head_dim, bias=False)
        self.wk = nn.Linear(self.hidden_size, self.head_dim, bias=False)
        self.k_norm = nn.LayerNorm(self.head_dim, eps=config.rms_norm_eps)
        self.weights_proj = nn.Linear(self.hidden_size, self.n_heads, bias=False)
        self.softmax_scale = self.head_dim**-0.5

    @torch.no_grad()
    def forward(
        self,
        hidden_states: torch.Tensor,
        q_resid: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor,
        position_ids: torch.Tensor,  # Kept for BC
        past_key_values: Cache | None = None,
    ) -> torch.Tensor:
        """
        Selects the top-k tokens per query for DeepSeek Sparse Attention (DSA).

        This is the bf16 equivalent of the reference Indexer which uses `rotate_activation` (Hadamard transform)
        and `fp8_index` (FP8 quantized scoring kernel). Since the Hadamard transform is orthogonal (dot products
        are preserved: Hq·Hk = q·k), and FP8 quantization is a precision optimization, we skip both and compute
        scores directly in bf16/fp32.

        The scoring logic computes:
            index_score[b,s,t] = Σ_h (weight[b,s,h] · softmax_scale · q[b,s,h,:] · k[b,t,:])

        Args:
            hidden_states: Input hidden states `[B, S, hidden_size]`.
            q_resid: Query residual from `q_a_layernorm(q_a_proj(x))`, shape `[B, S, q_lora_rank]`.
            position_embeddings: `(cos, sin)` from RotaryEmbedding.
            attention_mask: Causal mask, broadcastable to `[B, S, T]`.
            past_key_values: Cache object containing the indexer key cache for this layer.

        Returns:
            `torch.Tensor`: the `int32` top-k token indices of shape `[B, S, topk]`. The eager / SDPA paths
                turn these into an additive sparse mask; the `flash-mla` kernel consumes them directly.
        """
        batch_size, seq_len, _ = hidden_states.shape
        cos, sin = position_embeddings
        q = self.wq_b(q_resid)  # [B, S, H*D]
        q = q.view(batch_size, seq_len, self.n_heads, self.head_dim)  # [B, S, H, D]
        # Flipped RoPE position (later portion instead of the first)
        q_pass, q_rot = torch.split(q, [self.head_dim - self.qk_rope_head_dim, self.qk_rope_head_dim], dim=-1)

        # Norm is kept in fp32
        k = (
            self.k_norm(self.wk(hidden_states).to(self.k_norm.weight.dtype)).to(hidden_states.dtype).unsqueeze(2)
        )  # [B, S, 1, D]
        k_pass, k_rot = torch.split(k, [self.head_dim - self.qk_rope_head_dim, self.qk_rope_head_dim], dim=-1)

        q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin, unsqueeze_dim=2)
        q = torch.cat([q_pass, q_rot], dim=-1)  # [B, S, H, D]
        k = torch.cat([k_pass, k_rot], dim=-1).squeeze(2)  # [B, S, D]

        if past_key_values is not None:
            k = past_key_values.update_indexer(k, self.layer_idx)

        scores = torch.matmul(q.float(), k.transpose(-1, -2).float().unsqueeze(1))
        scores = F.relu(scores)

        # Weight per head and sum across heads: [B, S, 1, H] @ [B, S, H, T] → [B, S, T]
        # Apply softmax scale later
        weights = (
            self.weights_proj(hidden_states.to(self.weights_proj.weight.dtype)).float()
            * (self.n_heads**-0.5)
            * self.softmax_scale
        )
        index_scores = torch.matmul(weights.unsqueeze(-2), scores).squeeze(-2)

        # Causality needs to be taken into account when computing scores so padding tokens don't affect computation
        if attention_mask.dtype == torch.bool:
            index_scores = index_scores.masked_fill(~attention_mask, float("-inf"))
        else:
            index_scores = index_scores + attention_mask

        topk = min(self.index_topk, index_scores.shape[-1])
        return index_scores.topk(topk, dim=-1).indices.to(torch.int32)  # [B, S, topk]


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 eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    scaling: float,
    dropout: float | int = 0.0,
    **kwargs,
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)
    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        attn_weights = attn_weights + attention_mask

    sinks = module.sinks.reshape(1, -1, 1, 1).expand(query.shape[0], -1, query.shape[-2], -1)
    combined_logits = torch.cat([attn_weights, sinks], dim=-1)

    # This was not in the original implementation and slightly affect results; it prevents overflow in BF16/FP16
    # when training with bsz>1 we clamp max values.

    combined_logits = combined_logits - combined_logits.max(dim=-1, keepdim=True).values
    probs = F.softmax(combined_logits, dim=-1, dtype=combined_logits.dtype)
    scores = probs[..., :-1]  # we drop the sink here
    attn_weights = nn.functional.dropout(scores, p=dropout, training=module.training).to(value_states.dtype)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()
    return attn_output, attn_weights


def yarn_get_mscale(scale=1, mscale=1):
    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


def yarn_apply_mscale(rope_parameters, scaling):
    if rope_parameters.get("rope_type", "default") != "default":
        mscale_all_dim = rope_parameters.get("mscale_all_dim", 0)
        scaling_factor = rope_parameters["factor"]
        if mscale_all_dim:
            mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
            scaling = scaling * mscale * mscale
    return scaling


class HYV4Attention(nn.Module):
    """
    DeepSeek-V3.2 DSA extended with **cross-layer top-k sharing**.

    `config.indexer_types[layer_idx]` decides whether this layer runs its own indexer (`"full"`) or reuses the previous
    full layer's top-k selection (`"shared"`). `next_skip_topk` signals that the *next* layer will reuse this layer's
    top-k, so it is propagated upward via `prev_topk_indices`.
    """

    def __init__(self, config: HYV4Config, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.attention_dropout = config.attention_dropout
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads

        self.q_lora_rank = config.q_lora_rank
        self.qk_rope_head_dim = config.qk_rope_head_dim
        self.kv_lora_rank = config.kv_lora_rank
        self.v_head_dim = config.v_head_dim
        self.qk_nope_head_dim = config.qk_nope_head_dim
        self.qk_head_dim = config.qk_nope_head_dim + config.qk_rope_head_dim
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads

        self.is_causal = True
        self.q_a_proj = nn.Linear(self.hidden_size, self.q_lora_rank, bias=config.attention_bias)
        self.q_a_layernorm = HYV4RMSNorm(self.q_lora_rank)
        self.q_b_proj = nn.Linear(self.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)

        self.kv_a_proj_with_mqa = nn.Linear(
            self.hidden_size,
            config.kv_lora_rank + config.qk_rope_head_dim,
            bias=config.attention_bias,
        )
        self.kv_a_layernorm = HYV4RMSNorm(config.kv_lora_rank)
        self.kv_b_proj = nn.Linear(
            config.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
        )

        self.o_proj = nn.Linear(
            self.num_heads * self.v_head_dim,
            self.hidden_size,
            bias=config.attention_bias,
        )

        self.scaling = yarn_apply_mscale(config.rope_parameters, self.qk_head_dim ** (-0.5))
        # Refer: https://arxiv.org/abs/2603.12201 for more details.
        self.skip_topk = config.indexer_types[layer_idx] == "shared"
        self.indexer = None if self.skip_topk else HYV4Indexer(config, layer_idx)
        self.gate_projection_size = self.v_head_dim
        self.gate_proj = nn.Linear(config.hidden_size, self.num_heads * self.gate_projection_size, bias=False)
        self.sinks = nn.Parameter(torch.full((self.num_heads,), config.learnable_sink_init))

    def expand_kv(self, kv_nope: torch.Tensor, k_rot: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        """Expands the compressed latents into key and value states. Args:
            - kv_nope: key + value without positional encoding, shape [batch_size, 1, seqlen, self.kv_lora_rank]
            - k_rot: shared key with positional encoding, shape [batch_size, 1, seqlen, self.qk_rope_head_dim]
        Returns the key and value states, two tensors of shape [batch, num_heads, seq, (k or v)_head_dim].
        """
        batch_size, _, seq_length, _ = kv_nope.shape
        key_shape = (batch_size, seq_length, -1, self.qk_nope_head_dim + self.v_head_dim)

        kv_nope = self.kv_b_proj(kv_nope).view(key_shape).transpose(1, 2)
        k_nope, value_states = torch.split(kv_nope, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)
        k_rot = k_rot.expand(-1, k_nope.shape[1], -1, -1)  # does not affect the underlying storage

        # Concatenate k_nope and k_rot in a new tensor
        key_states = kv_nope.new_empty(*kv_nope.shape[:-1], self.qk_nope_head_dim + self.qk_rope_head_dim)
        key_states[..., : self.qk_nope_head_dim].copy_(k_nope)
        key_states[..., self.qk_nope_head_dim :].copy_(k_rot)
        return key_states, value_states

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor,
        past_key_values: Cache | None = None,
        position_ids: torch.Tensor | None = None,
        prev_topk_indices: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None, torch.Tensor | None]:
        batch_size, seq_length = hidden_states.shape[:-1]
        query_shape = (batch_size, seq_length, -1, self.qk_head_dim)

        # Key difference is the gating mechanism
        gate_states = self.gate_proj(hidden_states).view(batch_size, seq_length, -1, self.gate_projection_size)

        q_resid = self.q_a_layernorm(self.q_a_proj(hidden_states))
        q_states = self.q_b_proj(q_resid).view(query_shape).transpose(1, 2)
        q_pass, q_rot = torch.split(q_states, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)

        compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
        kv_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
        # Both latents are viewed as single-head, 4D tensors, as expected by `expand_kv`
        k_pass = self.kv_a_layernorm(kv_pass).view(batch_size, 1, seq_length, self.kv_lora_rank)

        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)
        cos, sin = position_embeddings
        # Non-interleave RoPE
        q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin)

        # Cache read / write is performed while latent KV is still compressed
        if past_key_values is not None:
            k_pass, k_rot = past_key_values.update(k_pass, k_rot, self.layer_idx)

        query_states = torch.cat((q_pass, q_rot), dim=-1)

        key_states, value_states = self.expand_kv(k_pass, k_rot)

        # DSA: select this layer's top-k tokens, or reuse the previous full layer's on `"shared"` layers.
        if self.indexer is not None:
            topk_indices = self.indexer(
                hidden_states,
                q_resid,
                position_embeddings,
                attention_mask[:, 0, :, :],
                position_ids,  # Kept for BC
                past_key_values=past_key_values,
            )  # [B, S, topk]
        else:
            if prev_topk_indices is None:
                raise ValueError("Shared DSA layers require top-k indices from a previous full indexer layer.")
            topk_indices = prev_topk_indices

        sparse_indices = None
        if self.config._attn_implementation in ("eager", "sdpa"):
            index_mask = (
                topk_indices.new_ones((batch_size, seq_length, key_states.shape[2]), dtype=torch.bool)
                .scatter(-1, topk_indices.long(), False)
                .unsqueeze(1)
            )

            if attention_mask.dtype == torch.bool:
                attention_mask = attention_mask & ~index_mask
            else:
                attention_mask = attention_mask.masked_fill(index_mask, torch.finfo(hidden_states.dtype).min)
        else:
            sparse_indices = topk_indices

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )
        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            indices=sparse_indices,  # consumed by flash_mla_with_kvcache; ignored by eager / SDPA
            s_aux=self.sinks,
            **kwargs,
        )

        attn_output = attn_output * torch.sigmoid(gate_states)
        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
        return self.o_proj(attn_output), attn_weights, topk_indices


class HYV4MLP(nn.Module):
    def __init__(self, config, intermediate_size=None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size if intermediate_size is None else 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 HYV4TopkRouter(nn.Module):
    def __init__(self, config: HYV4Config):
        super().__init__()
        self.top_k = config.num_experts_per_tok
        self.num_experts = config.num_local_experts
        self.hidden_dim = config.hidden_size
        self.weight = nn.Parameter(torch.zeros(self.num_experts, self.hidden_dim))
        self.routed_scaling_factor = config.routed_scaling_factor
        self.num_group = config.n_group
        self.topk_group = config.topk_group
        self.norm_topk_prob = config.norm_topk_prob
        self.e_score_correction_bias = nn.Buffer(torch.zeros((self.num_experts), dtype=torch.float32))

    def forward(self, hidden_states):
        hidden_states = hidden_states.view(-1, self.hidden_dim)
        router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
        scores = router_logits.sigmoid()
        scores_for_choice = scores + self.e_score_correction_bias
        group_scores = (
            scores_for_choice.view(-1, self.num_group, self.num_experts // self.num_group)
            .topk(2, dim=-1)[0]
            .sum(dim=-1)
        )
        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
        group_mask = torch.zeros_like(group_scores)
        group_mask.scatter_(1, group_idx, 1)
        score_mask = (
            group_mask.unsqueeze(-1)
            .expand(-1, self.num_group, self.num_experts // self.num_group)
            .reshape(-1, self.num_experts)
        )
        scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), float("-inf"))
        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
        topk_weights = scores.gather(1, topk_indices)
        if self.norm_topk_prob:
            denominator = topk_weights.sum(dim=-1, keepdim=True) + 1e-20
            topk_weights /= denominator
        topk_weights = topk_weights * self.routed_scaling_factor
        return router_logits, topk_weights, topk_indices


@use_experts_implementation
class HYV4Experts(nn.Module):
    """Collection of expert weights stored as 3D tensors."""

    def __init__(self, config):
        super().__init__()
        self.num_experts = config.num_local_experts
        self.hidden_dim = config.hidden_size
        self.intermediate_dim = config.moe_intermediate_size
        self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
        self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
        self.swiglu_limit = config.swiglu_limit

    def forward(
        self, hidden_states: torch.Tensor, top_k_index: torch.Tensor, top_k_weights: torch.Tensor
    ) -> torch.Tensor:
        final = torch.zeros_like(hidden_states)
        with torch.no_grad():
            mask = F.one_hot(top_k_index, num_classes=self.num_experts + 1).permute(2, 1, 0)
            hit = torch.greater(mask.sum(dim=(-1, -2)), 0).nonzero()
        for expert_idx in hit:
            expert_idx = expert_idx[0]
            if expert_idx == self.num_experts:
                continue
            top_k_pos, token_idx = torch.where(mask[expert_idx])
            current = self._apply_gate(F.linear(hidden_states[token_idx], self.gate_up_proj[expert_idx]))
            current = F.linear(current, self.down_proj[expert_idx]) * top_k_weights[token_idx, top_k_pos, None]
            final.index_add_(0, token_idx, current.to(final.dtype))
        return final

    def _apply_gate(self, gate_up: torch.Tensor) -> torch.Tensor:
        gate, up = gate_up.chunk(2, dim=-1)
        gate = gate.clamp(min=None, max=self.swiglu_limit)
        up = up.clamp(min=-self.swiglu_limit, max=self.swiglu_limit)
        # Simple swiglu instead of alpha
        return F.silu(gate) * up


class HYV4MoE(nn.Module):
    """
    A mixed expert module containing shared experts.
    """

    def __init__(self, config: HYV4Config):
        super().__init__()
        self.config = config
        self.experts = HYV4Experts(config)
        self.gate = HYV4TopkRouter(config)
        self.shared_experts = HYV4MLP(
            config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        residuals = hidden_states
        orig_shape = hidden_states.shape
        _, topk_weights, topk_indices = self.gate(hidden_states)
        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
        hidden_states = self.experts(hidden_states, topk_indices, topk_weights).view(*orig_shape)
        hidden_states = hidden_states + self.shared_experts(residuals)
        return hidden_states


class HYV4HyperConnection(nn.Module):
    """
    Independent Hyper-Connection following the DeepSeek-V4 form (parameters on the module).

    The overall difference lies in the dropping of the `comb` output which skips the sinkhorn part of the algorithm.
    """

    def __init__(self, config: HYV4Config):
        super().__init__()
        self.hc_mult = config.hc_mult  # number of streams, referred as N below
        self.hc_eps = config.hc_eps
        self.input_norm = HYV4UnweightedRMSNorm(eps=config.rms_norm_eps)
        concatenated_weights_size = 2 * self.hc_mult  # noqa: F841
        self.fn = nn.Parameter(torch.empty(concatenated_weights_size, self.hc_mult * config.hidden_size))
        self.base = nn.Parameter(torch.empty(concatenated_weights_size))
        self.scale = nn.Parameter(torch.empty(2))
        self.hc_post_magnitude = config.hc_magnitude

    def forward(self, hidden_streams: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        """Independent HC implementation with forced fp32 application"""
        # Key difference is to force fp32 in any case
        device_type = hidden_streams.device.type
        with maybe_autocast(device_type=device_type, enabled=False):
            flat = hidden_streams.flatten(2).float()
            # Norm as residual
            mixes = F.linear(flat, self.fn.float()) * self.input_norm(flat)

            # Can be seen as non-combined variation of the DSv4 implementation
            pre_b, post_b = self.base.split(self.hc_mult, dim=-1)
            pre_scale, post_scale = self.scale.unbind(0)
            pre_logits, post_logits = mixes.split(self.hc_mult, dim=-1)

            pre = torch.sigmoid(pre_logits * pre_scale + pre_b) + self.hc_eps
            # Magnitude can be set within the config (difference to the constant to 2 in dsv4)
            post = self.hc_post_magnitude * torch.sigmoid(post_logits * post_scale + post_b) + self.hc_eps
            out = torch.sum(pre.unsqueeze(-1) * hidden_streams, dim=2)

        return post, out.to(hidden_streams.dtype)


class HYV4HyperHead(nn.Module):
    """Final HC-stream collapse; used by `HYV4Model` before the shared RMSNorm."""

    def __init__(self, config: HYV4Config):
        super().__init__()
        self.hc_mult = config.hc_mult
        self.input_norm = HYV4UnweightedRMSNorm(eps=config.rms_norm_eps)
        self.eps = config.hc_eps
        self.hc_fn = nn.Parameter(torch.empty(self.hc_mult, self.hc_mult * config.hidden_size))
        self.hc_base = nn.Parameter(torch.empty(self.hc_mult))
        self.hc_scale = nn.Parameter(torch.empty(1))

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # Key difference is to force fp32 in any case
        device_type = x.device.type
        with maybe_autocast(device_type=device_type, enabled=False):
            flat = x.flatten(2).float()
            # Norm as residual
            mixes = F.linear(flat, self.hc_fn.float()) * self.input_norm(flat)
            pre = torch.sigmoid(mixes * self.hc_scale.float() + self.hc_base.float()) + self.eps
            out = (pre.unsqueeze(-1) * x).sum(dim=2)
        return out.to(x.dtype)


class HYV4DecoderLayer(GradientCheckpointingLayer):
    """Similar to DSv4 but with a different Hyper Connection and mixed MLP/MoE patterns"""

    def __init__(self, config: HYV4Config, layer_idx: int):
        super().__init__()
        self.layer_idx = layer_idx
        self.self_attn = HYV4Attention(config, layer_idx)
        self.mlp = HYV4MoE(config) if config.mlp_layer_types[layer_idx] == "sparse" else HYV4MLP(config)
        self.input_layernorm = HYV4RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = HYV4RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.attn_hc = HYV4HyperConnection(config)
        self.ffn_hc = HYV4HyperConnection(config)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        prev_topk_indices: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.LongTensor | None]:
        dtype = hidden_states.dtype

        # Key difference is the hyper connection and its residual connection
        residual = hidden_states
        post, hidden_states = self.attn_hc(hidden_states)
        # Self attn
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states, _, topk_indices = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            position_embeddings=position_embeddings,
            prev_topk_indices=prev_topk_indices,
            **kwargs,
        )
        hidden_states = (post.float().unsqueeze(-1) * hidden_states.float().unsqueeze(-2) + residual.float()).to(dtype)

        residual = hidden_states
        post, hidden_states = self.ffn_hc(hidden_states)
        # Feed forward
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = (post.float().unsqueeze(-1) * hidden_states.float().unsqueeze(-2) + residual.float()).to(dtype)

        return hidden_states, topk_indices


@auto_docstring
class HYV4PreTrainedModel(PreTrainedModel):
    config: HYV4Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["HYV4DecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    # Combination of sinks and DSA disable anything but eager atm
    _supports_flash_attn = False
    _supports_sdpa = False
    _supports_flex_attn = False

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": HYV4DecoderLayer,
        "attentions": HYV4Attention,
        "router_logits": OutputRecorder(HYV4TopkRouter, index=0),
    }
    _keep_in_fp32_modules_strict = [
        "e_score_correction_bias",
        "fn",
        "scale",
        "base",
        "hc_fn",
        "hc_scale",
        "hc_base",
        "weights_proj",
        "k_norm",
        "sinks",
    ]
    _keys_to_ignore_on_load_unexpected = [r"model\.mtp_layers\..*"]

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, HYV4TopkRouter):
            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
            init.zeros_(module.e_score_correction_bias)
        elif isinstance(module, HYV4Experts):
            init.normal_(module.gate_up_proj, mean=0.0, std=self.config.initializer_range)
            init.normal_(module.down_proj, mean=0.0, std=self.config.initializer_range)
        elif isinstance(module, HYV4HyperConnection):
            init.normal_(module.fn, mean=0.0, std=self.config.initializer_range)
            init.constant_(module.scale, 0.01)

            base_value = -math.log(max(module.hc_mult - 1, 1))
            base = torch.zeros_like(module.base)
            base[: module.hc_mult] = base_value
            init.copy_(module.base, base)
        elif isinstance(module, HYV4HyperHead):
            init.normal_(module.hc_fn, mean=0.0, std=self.config.initializer_range)
            init.constant_(module.hc_scale, 0.01)

            base_value = -math.log(max(module.hc_mult - 1, 1))
            init.constant_(module.hc_base, base_value)
        elif isinstance(module, HYV4Attention):
            init.constant_(module.sinks, self.config.learnable_sink_init)


@auto_docstring
class HYV4Model(HYV4PreTrainedModel):
    def __init__(self, config: HYV4Config):
        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(
            [HYV4DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = HYV4RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = HYV4RotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        self.hc_head = HYV4HyperHead(config)

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

    @merge_with_config_defaults
    @capture_outputs
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = 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: torch.Tensor = self.embed_tokens(input_ids)

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

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

        # It may already have been prepared by e.g. `generate`
        if not isinstance(causal_mask_mapping := attention_mask, dict):
            mask_kwargs = {
                "config": self.config,
                "inputs_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "past_key_values": past_key_values,
                "position_ids": position_ids,
                "allow_is_causal_skip": False,  # Always force creation to account for causality in the indexer
            }
            causal_mask_mapping = {"indexed_attention": create_causal_mask(**mask_kwargs)}

        hidden_states = inputs_embeds
        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)
        # Prepare HC connections
        hidden_states = inputs_embeds.unsqueeze(2).expand(-1, -1, self.config.hc_mult, -1).contiguous()

        topk_indices = None
        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states, topk_indices = decoder_layer(
                hidden_states,
                attention_mask=causal_mask_mapping["indexed_attention"],
                position_embeddings=position_embeddings,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                prev_topk_indices=topk_indices,
                **kwargs,
            )

        # Difference with the HC head at the end
        hidden_states = self.norm(self.hc_head(hidden_states))

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


@auto_docstring
class HYV4ForCausalLM(HYV4PreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
    _tp_plan = {"lm_head": "colwise_gather_output"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
    _fsdp_plan = {"lm_head": "keep_full_weight"}
    # Same as base but with the additional lm head
    _keep_in_fp32_modules_strict = [
        "e_score_correction_bias",
        "fn",
        "scale",
        "base",
        "hc_fn",
        "hc_scale",
        "hc_base",
        "weights_proj",
        "k_norm",
        "sinks",
        "lm_head",
    ]

    def __init__(self, config):
        super().__init__(config)
        self.model = HYV4Model(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

    @can_return_tuple
    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> MoeCausalLMOutputWithPast:
        r"""
        Example:

        ```python
        >>> from transformers import AutoTokenizer, HYV4ForCausalLM

        >>> model = HYV4ForCausalLM.from_pretrained("meta-hy_v4/HYV4-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-hy_v4/HYV4-2-7b-hf")

        >>> 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."
        ```"""
        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,
            **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
        # Key difference of the lm_head being kept in float
        logits = self.lm_head(hidden_states[:, slice_indices, :].to(dtype=self.lm_head.weight.dtype))

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

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


__all__ = ["HYV4PreTrainedModel", "HYV4Model", "HYV4ForCausalLM"]