modeling_xing4_0.py

import math
from collections.abc import Callable
from typing import Optional

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

from transformers import initialization as init
from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.generation import GenerationMixin
from transformers.integrations import use_experts_implementation, use_kernel_forward_from_hub
from transformers.masking_utils import create_causal_mask
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import (
    GenericForSequenceClassification,
    GenericForTokenClassification,
    GradientCheckpointingLayer,
)
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import TransformersKwargs, auto_docstring, can_return_tuple
from transformers.utils.generic import is_flash_attention_requested, maybe_autocast, merge_with_config_defaults
from transformers.utils.output_capturing import capture_outputs
from .configuration_xing4_0 import Xing4_0Config


class Xing4_0RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        Xing4_0RMSNorm 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 Xing4_0UnweightedRMSNorm(nn.Module):
    def __init__(self, eps: float = 1.0e-6):
        super().__init__()
        self.eps = eps

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return (x.float() * torch.rsqrt(x.float().square().mean(-1, keepdim=True) + self.eps)).to(x.dtype)


class Xing4_0RotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: Xing4_0Config, device=None):
        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, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.register_buffer("original_inv_freq", inv_freq.clone(), persistent=False)

    @staticmethod
    def compute_default_rope_parameters(
            config: Xing4_0Config | None = None,
            device: Optional["torch.device"] = None,
            seq_len: int | None = None,
    ) -> tuple["torch.Tensor", float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            device (`torch.device`):
                The device to use for initialization of the inverse frequencies.
            seq_len (`int`, *optional*):
                The current sequence length. Unused for this type of RoPE.
        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.int64).to(device=device, 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):
        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 maybe_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.float(), sin.float()


class Xing4_0MLP(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 Xing4_0TopkRouter(nn.Module):
    def __init__(self, config):
        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.register_buffer("e_score_correction_bias", torch.zeros(self.num_experts))

    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()
        topk_indices = torch.topk(scores + self.e_score_correction_bias, self.top_k, dim=-1, sorted=False).indices
        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


class Xing4_0MoE(nn.Module):
    """
    A mixed expert module containing shared experts.
    """

    def __init__(self, config: Xing4_0Config):
        super().__init__()
        self.config = config
        self.experts = nn.ModuleList(
            [
                Xing4_0MLP(
                    config, intermediate_size=config.moe_intermediate_size
                )
                for i in range(config.n_routed_experts)
            ]
        )
        self.gate = Xing4_0TopkRouter(config)
        self.shared_experts = Xing4_0MLP(
            config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
        )

    def moe(self, hidden_states: torch.Tensor, topk_indices: torch.Tensor, topk_weights: torch.Tensor):
        final_hidden_states = torch.zeros_like(hidden_states, dtype=topk_weights.dtype)
        expert_mask = torch.nn.functional.one_hot(topk_indices, num_classes=len(self.experts))
        expert_mask = expert_mask.permute(2, 0, 1)

        for expert_idx in range(len(self.experts)):
            expert = self.experts[expert_idx]
            mask = expert_mask[expert_idx]
            token_indices, weight_indices = torch.where(mask)

            if token_indices.numel() > 0:
                expert_weights = topk_weights[token_indices, weight_indices]
                expert_input = hidden_states[token_indices]
                expert_output = expert(expert_input)
                weighted_output = expert_output * expert_weights.unsqueeze(-1)
                final_hidden_states.index_add_(0, token_indices, weighted_output)
        return final_hidden_states.type(hidden_states.dtype)

    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.moe(hidden_states, topk_indices, topk_weights).view(*orig_shape)
        hidden_states = hidden_states + self.shared_experts(residuals)
        return hidden_states


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)


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.
    """
    ori_dtype = q.dtype
    q = q.float()
    k = k.float()
    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.to(ori_dtype), k_embed.to(ori_dtype)


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 = 0.0,
        **kwargs: Unpack[TransformersKwargs],
):
    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

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights


def apply_rotary_pos_emb_interleave(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    r"""
    Applies interleaved Rotary Position Embedding to the query and key tensors.

    Xing4_0 lays the rotary dimensions out in interleaved pairs `(x0, x1), (x2, x3), ...`, each rotated by a
    single frequency. We compute that rotation directly on the even/odd slices instead of de-interleaving with a
    `view`/`transpose`/`reshape`; the output is bit-identical to the de-interleaved `rotate_half` formulation while
    avoiding the extra contiguous copy.

    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`):
            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
            used to pass offsetted position ids when working with a KV-cache.
        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`/`sin` are `cat(freqs, freqs)`; the first half holds the per-pair angle.
    ori_dtype = q.dtype
    q = q.float()
    k = k.float()
    cos = cos[..., : cos.shape[-1] // 2].unsqueeze(unsqueeze_dim)
    sin = sin[..., : sin.shape[-1] // 2].unsqueeze(unsqueeze_dim)

    q1, q2 = q[..., 0::2], q[..., 1::2]
    k1, k2 = k[..., 0::2], k[..., 1::2]
    q_embed = torch.cat([q1 * cos - q2 * sin, q2 * cos + q1 * sin], dim=-1)
    k_embed = torch.cat([k1 * cos - k2 * sin, k2 * cos + k1 * sin], dim=-1)
    return q_embed.to(ori_dtype), k_embed.to(ori_dtype)


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


class Xing4_0Attention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: Xing4_0Config, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.attention_dropout = config.attention_dropout
        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_head_dim

        self.is_causal = True
        if self.q_lora_rank is None:
            self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.qk_head_dim, bias=False)
        else:
            self.q_a_proj = nn.Linear(config.hidden_size, config.q_lora_rank, bias=config.attention_bias)
            self.q_a_layernorm = Xing4_0RMSNorm(config.q_lora_rank)
            self.q_b_proj = nn.Linear(config.q_lora_rank, self.num_heads * self.qk_head_dim, bias=False)

        self.kv_a_proj_with_mqa = nn.Linear(
            config.hidden_size,
            self.kv_lora_rank + self.qk_rope_head_dim,
            bias=config.attention_bias,
        )
        self.kv_a_layernorm = Xing4_0RMSNorm(self.kv_lora_rank)
        self.kv_b_proj = nn.Linear(
            self.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,
            config.hidden_size,
            bias=config.attention_bias,
        )

        self.scaling = self.qk_head_dim ** (-0.5)
        if self.config.rope_parameters.get("rope_type", "default") != "default":
            mscale_all_dim = self.config.rope_parameters.get("mscale_all_dim", 0)
            scaling_factor = self.config.rope_parameters["factor"]
            if mscale_all_dim:
                mscale = yarn_get_mscale(scaling_factor, mscale_all_dim)
                self.scaling = self.scaling * mscale * mscale

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

        if self.q_lora_rank is None:
            q_states = self.q_proj(hidden_states)
        else:
            q_states = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
        q_states = q_states.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)
        k_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)

        k_pass = self.kv_b_proj(self.kv_a_layernorm(k_pass)).view(key_shape).transpose(1, 2)
        k_pass, value_states = torch.split(k_pass, [self.qk_nope_head_dim, self.v_head_dim], dim=-1)

        k_rot = k_rot.view(batch_size, 1, seq_length, self.qk_rope_head_dim)

        cos, sin = position_embeddings
        if self.config.rope_interleave:  # support using interleaved weights for efficiency
            q_rot, k_rot = apply_rotary_pos_emb_interleave(q_rot, k_rot, cos, sin)
        else:
            q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin)
        k_rot = k_rot.expand(*k_pass.shape[:-1], -1)

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

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

        if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
            value_states = F.pad(value_states, [0, self.qk_head_dim - self.v_head_dim])

        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,
            **kwargs,
        )

        if is_flash_attention_requested(self.config) and self.qk_head_dim != self.v_head_dim:
            attn_output = attn_output[:, :, :, : self.v_head_dim]

        attn_output = attn_output.reshape(batch_size, seq_length, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class Xing4_0HyperConnection(nn.Module):

    def __init__(self, config: Xing4_0Config):
        super().__init__()
        self.hc_mult = config.hc_mult
        self.hc_sinkhorn_iters = config.hc_sinkhorn_iters
        self.hc_eps = config.hc_eps
        self.input_norm = Xing4_0UnweightedRMSNorm(eps=config.rms_norm_eps)
        mix = (2 + self.hc_mult) * self.hc_mult
        self.hc_fn = nn.Parameter(torch.empty(mix, self.hc_mult * config.hidden_size))
        self.hc_base = nn.Parameter(torch.empty(mix))
        self.hc_scale = nn.Parameter(torch.empty(3))
        self.mhc_h_res_clamp_max = config.mhc_h_res_clamp_max
        self.mhc_h_res_clamp_min = config.mhc_h_res_clamp_min

    def forward(self, hidden_streams: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        ori_dtype = hidden_streams.dtype
        hc = self.hc_mult
        flat = self.input_norm(hidden_streams.flatten(start_dim=2).float())
        pre_w, post_w, comb_w = F.linear(flat.to(ori_dtype), self.hc_fn.to(ori_dtype)).float().split([hc, hc, hc * hc], dim=-1)
        pre_b, post_b, comb_b = self.hc_base.split([hc, hc, hc * hc])
        pre_scale, post_scale, comb_scale = self.hc_scale.unbind(0)
        pre = torch.sigmoid(pre_w * pre_scale + pre_b)
        post = 2 * torch.sigmoid(post_w * post_scale + post_b)
        comb_logits = comb_w.view(*comb_w.shape[:-1], hc, hc) * comb_scale + comb_b.view(hc, hc)
        comb_logits = torch.clamp(comb_logits, min=self.mhc_h_res_clamp_min, max=self.mhc_h_res_clamp_max)
        comb_max = comb_logits.amax(dim=-1, keepdim=True)
        comb = torch.exp(comb_logits - comb_max)
        for _ in range(self.hc_sinkhorn_iters):
            comb = comb / (comb.sum(dim=-1, keepdim=True) + self.hc_eps)
            comb = comb / (comb.sum(dim=-2, keepdim=True) + self.hc_eps)
        collapsed = (pre.unsqueeze(-1).to(ori_dtype) * hidden_streams).sum(dim=2)
        return post.to(ori_dtype), comb.to(ori_dtype), collapsed.to(ori_dtype)


class Xing4_0DecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: Xing4_0Config, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.self_attn = Xing4_0Attention(config=config, layer_idx=layer_idx)

        if layer_idx >= config.first_k_dense_replace:
            self.mlp = Xing4_0MoE(config)
        else:
            self.mlp = Xing4_0MLP(config)

        self.input_layernorm = Xing4_0RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Xing4_0RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.attn_hc = Xing4_0HyperConnection(config)
        self.ffn_hc = Xing4_0HyperConnection(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,
            **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        post, comb, collapsed = self.attn_hc(hidden_states)
        # Self Attention
        collapsed = self.input_layernorm(collapsed)
        attn_output, _ = self.self_attn(
            hidden_states=collapsed,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = post.unsqueeze(-1) * attn_output.unsqueeze(-2) + torch.matmul(
            comb, hidden_states
        )
        # Fully Connected
        post, comb, collapsed = self.ffn_hc(hidden_states)
        collapsed = self.post_attention_layernorm(collapsed)
        mlp_output = self.mlp(collapsed)
        return post.unsqueeze(-1) * mlp_output.unsqueeze(-2) + torch.matmul(
            comb, hidden_states
        )


@auto_docstring
class Xing4_0PreTrainedModel(PreTrainedModel):
    config: Xing4_0Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["Xing4_0DecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": Xing4_0DecoderLayer,
        "attentions": Xing4_0Attention,
    }
    _keep_in_fp32_modules_strict = [
        "e_score_correction_bias",
    ]
    _keys_to_ignore_on_load_unexpected = [r"model\.layers\.40.*"]

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, Xing4_0TopkRouter):
            init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
            init.zeros_(module.e_score_correction_bias)
        elif isinstance(module, Xing4_0HyperConnection):
            init.normal_(module.fn, mean=0.0, std=self.config.initializer_range)
            init.zeros_(module.base)
            init.ones_(module.scale)


@auto_docstring
class Xing4_0Model(Xing4_0PreTrainedModel):
    def __init__(self, config: Xing4_0Config):
        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(
            [Xing4_0DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = Xing4_0RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = Xing4_0RotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        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],
    ) -> BaseModelOutputWithPast:
        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)

        causal_mask = create_causal_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            position_ids=position_ids,
        )

        hidden_states = inputs_embeds.unsqueeze(2).expand(-1, -1, self.config.hc_mult, -1).contiguous()
        position_embeddings = self.rotary_emb(hidden_states, position_ids=position_ids)
        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=causal_mask,
                position_embeddings=position_embeddings,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                **kwargs,
            )
        hidden_states = hidden_states.mean(dim=2)
        hidden_states = self.norm(hidden_states)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )


class Xing4_0ForCausalLM(Xing4_0PreTrainedModel, 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"}

    def __init__(self, config):
        super().__init__(config)
        self.model = Xing4_0Model(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()

    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],
    ) -> CausalLMOutputWithPast:
        outputs: BaseModelOutputWithPast = 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
        logits = self.lm_head(hidden_states[:, slice_indices, :])
        loss = None
        if labels is not None:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)

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


__all__ = [
    "Xing4_0PreTrainedModel",
    "Xing4_0Model",
    "Xing4_0ForCausalLM",
]