modeling_solar_open2.py

# coding=utf-8
# Copyright 2025 Upstage AI.
# Copyright 2025 The GLM4 & ZhipuAI team and HuggingFace Inc. team.
#
# 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.
#
# This file has been modified by Upstage AI including:
# - Hybrid MoE Architecture: Replaced the standard dense structure with a depth-dependent Hybrid MoE, adding `SolarOpen2MoE` and `SolarOpen2TopkRouter` classes.
# - RoPE Strategy: Changed the rotary position embedding strategy from GLM4's interleaved rotation to Llama-style block rotation (via modified `rotate_half`).
# - Normalization Logic: Simplified the layer normalization structure by removing GLM4's extra post-operation norms and adding optional Query-Key Normalization (`use_qk_norm`).
#
# Based on code from: https://github.com/huggingface/transformers/blob/main/src/transformers/models/glm4/modeling_glm4.py

from typing import Callable, Optional, Union, Any

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

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.generation import GenerationMixin
from transformers.integrations import 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 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.deprecation import deprecate_kwarg
from transformers.utils.generic import check_model_inputs
from .configuration_solar_open2 import SolarOpen2Config

try:
    from fla.modules import FusedRMSNormGated, ShortConvolution
    from fla.ops.kda import chunk_kda, fused_recurrent_kda
    from fla.ops.kda.gate import fused_kda_gate
    from fla.layers.utils import get_unpad_data, index_first_axis, pad_input
except ImportError:
    raise ImportError("Plese run `pip install -U fla-core`")

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 repeat_kv_linear(x: torch.Tensor, n_rep: int) -> torch.Tensor:
    """torch.repeat_interleave(x, dim=-2, repeats=n_rep) for (..., n_kv_heads, head_dim) layout."""
    if n_rep == 1:
        return x
    shape = x.shape
    return (
        x.unsqueeze(-2)
        .expand(*shape[:-1], n_rep, shape[-1])
        .reshape(*shape[:-2], shape[-2] * n_rep, shape[-1])
    )


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    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:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_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 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, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

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

    # 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:]

    # 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


class SolarOpen2DynamicCache:
    """
    Dynamic cache for SolarOpen2 model.
    Inspired by Kimi-Linear
    """
    is_compileable = False

    def __init__(self, config: SolarOpen2Config):
        super().__init__()
        self.config = config

        if config.linear_attn_config is not None:
            self.layer_types = []
            for i in range(config.num_hidden_layers):
                # Determine if this layer uses GQA: gqa_layers takes priority over gqa_interval
                if config.gqa_layers is not None:
                    use_gqa = i in config.gqa_layers
                else:
                    use_gqa = (i + 1) % config.gqa_interval == 0
                if use_gqa:
                    self.layer_types.append("full_attention")
                else:
                    self.layer_types.append("linear_attention")
        else:
            self.layer_types = ["full_attention"] * config.num_hidden_layers

        self.transformer_layers = [
            i for i in range(config.num_hidden_layers) if self.layer_types[i] == "full_attention"
        ]

        linear_layers = [i for i in range(
            config.num_hidden_layers) if self.layer_types[i] == "linear_attention"]
        self.last_linear_layer = linear_layers[-1] if linear_layers else -1

        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 update(
        self,
        key_states: torch.Tensor,
        value_states: torch.Tensor,
        layer_idx: int,
        cache_kwargs: dict[str, Any] | None = 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][0].device
                beam_idx = beam_idx.to(device)
                q_conv, k_conv, v_conv = self.conv_states[layer_idx]
                self.conv_states[layer_idx] = (
                    q_conv.index_select(0, beam_idx),
                    k_conv.index_select(0, beam_idx),
                    v_conv.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: int | None = 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."""
        if self.last_linear_layer == -1:
            return False
        return self.conv_states[self.last_linear_layer] is not None


class SolarOpen2LinearAttention(nn.Module):
    def __init__(self, config: SolarOpen2Config, layer_idx: int):
        super().__init__()
        self.config = config
        self.mode = "chunk"

        self.hidden_size = config.hidden_size
        self.conv_size = config.linear_attn_config["short_conv_kernel_size"]
        self.head_dim = config.linear_attn_config["head_dim"]
        self.num_heads = config.linear_attn_config["num_heads"]
        self.num_kv_heads = config.linear_attn_config.get("num_kv_heads", None) or self.num_heads
        self.n_rep = self.num_heads // self.num_kv_heads
        self.head_k_dim = self.head_dim
        self.num_k_heads = self.num_heads

        self.layer_idx = layer_idx
        self.use_full_proj = getattr(config, 'kda_use_full_proj', False)
        self.gate_lower_bound = getattr(config, 'kda_gate_lower_bound', -5.0)
        self.allow_neg_eigval = getattr(config, 'kda_allow_neg_eigval', False)

        assert self.mode in [
            'chunk', 'fused_recurrent'], f"Not suppoerted mode `{self.mode}`."

        projection_k_size = self.head_k_dim * self.num_k_heads
        projection_size = self.head_dim * self.num_heads
        kv_projection_size = self.head_dim * self.num_kv_heads

        self.q_proj = nn.Linear(
            self.hidden_size, projection_k_size, bias=False)
        self.k_proj = nn.Linear(
            self.hidden_size, kv_projection_size, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, kv_projection_size, bias=False)

        self.q_conv1d = ShortConvolution(
            hidden_size=projection_k_size,
            kernel_size=self.conv_size,
            activation='silu',
        )
        self.k_conv1d = ShortConvolution(
            hidden_size=kv_projection_size,
            kernel_size=self.conv_size,
            activation='silu',
        )
        self.v_conv1d = ShortConvolution(
            hidden_size=kv_projection_size,
            kernel_size=self.conv_size,
            activation='silu',
        )

        self.A_log = torch.nn.Parameter(torch.log(torch.empty(
            self.num_heads, dtype=torch.float32).uniform_(1, 16)).view(1, 1, -1, 1))

        if self.use_full_proj:
            self.f_proj = nn.Linear(self.hidden_size, projection_size, bias=False)
        else:
            self.f_a_proj = nn.Linear(self.hidden_size, self.head_dim, bias=False)
            self.f_b_proj = nn.Linear(self.head_dim, projection_size, bias=False)

        self.dt_bias = nn.Parameter(
            torch.empty(projection_size, dtype=torch.float32))

        self.b_proj = nn.Linear(self.hidden_size, self.num_heads, bias=False)

        if self.use_full_proj:
            self.g_proj = nn.Linear(self.hidden_size, projection_size, bias=False)
        else:
            self.g_a_proj = nn.Linear(self.hidden_size, self.head_dim, bias=False)
            self.g_b_proj = nn.Linear(self.head_dim, projection_size, bias=False)

        self.o_norm = FusedRMSNormGated(
            self.head_dim, eps=config.rms_norm_eps, activation='sigmoid')
        self.o_proj = nn.Linear(projection_size, self.hidden_size, bias=False)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        past_key_values: SolarOpen2DynamicCache | None = None,
        **kwargs: Unpack[dict],
    ) -> tuple[torch.Tensor, torch.Tensor | None, Cache | None]:
        if attention_mask is not None:
            if attention_mask.dim() != 2:
                attention_mask = kwargs.get("padding_mask")

            if attention_mask is not None and attention_mask.dim() != 2:
                raise ValueError(
                    "attention_mask must be a 0-1 matrix of shape [batch_size, seq_len] "
                    "(0 = padding). 3D masks are not supported here.",
                )
        use_cache = past_key_values is not None
        batch_size, q_len, _ = hidden_states.shape
        mode = 'fused_recurrent' if q_len <= 64 else self.mode
        if self.training:
            assert mode == 'chunk', "Only chunk mode is supported in training."

        cu_seqlens = kwargs.get('cu_seqlens')
        indices = None
        if attention_mask is not None:
            indices, cu_seqlens, _ = get_unpad_data(attention_mask[:, -q_len:])
            hidden_states = index_first_axis(
                rearrange(hidden_states, "b s ... -> (b s) ..."), indices).unsqueeze(0)

        conv_state_q, conv_state_k, conv_state_v = None, None, None
        recurrent_state = None
        if past_key_values is not None:
            if past_key_values.conv_states[self.layer_idx] is not None:
                conv_state_q, conv_state_k, conv_state_v = past_key_values.conv_states[
                    self.layer_idx]
            recurrent_state = past_key_values.recurrent_states[self.layer_idx]
        q, conv_state_q = self.q_conv1d(
            x=self.q_proj(hidden_states),
            cache=conv_state_q,
            output_final_state=use_cache,
            cu_seqlens=cu_seqlens,
        )
        k, conv_state_k = self.k_conv1d(
            x=self.k_proj(hidden_states),
            cache=conv_state_k,
            output_final_state=use_cache,
            cu_seqlens=cu_seqlens,
        )
        v, conv_state_v = self.v_conv1d(
            x=self.v_proj(hidden_states),
            cache=conv_state_v,
            output_final_state=use_cache,
            cu_seqlens=cu_seqlens,
        )
        if self.use_full_proj:
            g = self.f_proj(hidden_states)
        else:
            g = self.f_b_proj(self.f_a_proj(hidden_states))
        q, k, g = (rearrange(x, "... (h d) -> ... h d", d=self.head_k_dim) for x in (q, k, g))
        v = rearrange(v, '... (h d) -> ... h d', d=self.head_dim)

        # repeat k/v heads if num_kv_heads < num_heads (GQA)
        k = repeat_kv_linear(k, self.n_rep)
        v = repeat_kv_linear(v, self.n_rep)

        g = fused_kda_gate(g, self.A_log, dt_bias=self.dt_bias, lower_bound=self.gate_lower_bound)
        beta = self.b_proj(hidden_states).sigmoid()
        if self.allow_neg_eigval:
            beta = beta * 2.0

        if mode == 'chunk':
            o, recurrent_state = chunk_kda(
                q=q,
                k=k,
                v=v,
                g=g,
                beta=beta,
                initial_state=recurrent_state,
                output_final_state=True,
                use_qk_l2norm_in_kernel=True,
                cu_seqlens=cu_seqlens,
            )
        else:
            o, recurrent_state = fused_recurrent_kda(
                q=q,
                k=k,
                v=v,
                g=g,
                beta=beta,
                initial_state=recurrent_state,
                output_final_state=True,
                use_qk_l2norm_in_kernel=True,
                cu_seqlens=cu_seqlens,
            )
        if past_key_values is not None:
            past_key_values.recurrent_states[self.layer_idx] = recurrent_state
            past_key_values.conv_states[self.layer_idx] = (
                conv_state_q, conv_state_k, conv_state_v)

        if self.use_full_proj:
            g = self.g_proj(hidden_states)
        else:
            g = self.g_b_proj(self.g_a_proj(hidden_states))
        g = rearrange(g, '... (h d) -> ... h d', d=self.head_dim)
        o = self.o_norm(o, g)

        o = rearrange(o, 'b t h d -> b t (h d)')
        o = self.o_proj(o)
        if attention_mask is not None:
            o = pad_input(o.squeeze(0), indices, batch_size, q_len)

        return o, _


class SolarOpen2Attention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: SolarOpen2Config, layer_idx: Optional[int] = None):
        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.rope_scaling = config.rope_scaling
        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, 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=False)
        self.use_qk_norm = config.use_qk_norm
        if self.use_qk_norm:
            self.q_norm = SolarOpen2RMSNorm(self.head_dim, eps=config.rms_norm_eps)
            self.k_norm = SolarOpen2RMSNorm(self.head_dim, eps=config.rms_norm_eps)

        self.use_gqa_gate = config.use_gqa_gate
        if self.use_gqa_gate:
            self.g_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.use_gqa_gate_bias)

    @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[SolarOpen2DynamicCache] = 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 = self.q_proj(hidden_states).view(hidden_shape)
        key_states = self.k_proj(hidden_states).view(hidden_shape)
        value_states = self.v_proj(hidden_states).view(hidden_shape)

        if self.use_qk_norm:  # main diff from Llama
            query_states = self.q_norm(query_states)
            key_states = self.k_norm(key_states)

        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

        cos, sin = position_embeddings
        if self.config.use_rope:
            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; position_ids 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 = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        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,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        if self.use_gqa_gate:
            gate = self.g_proj(hidden_states)
            attn_output = attn_output * torch.sigmoid(gate)

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


class SolarOpen2MLP(nn.Module):
    def __init__(self, config, hidden_size=None, intermediate_size=None):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size if hidden_size is None else 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 SolarOpen2TopkRouter(nn.Module):
    def __init__(self, config: SolarOpen2Config):
        super().__init__()
        self.config = config
        self.top_k = config.num_experts_per_tok
        self.n_routed_experts = config.n_routed_experts
        self.routed_scaling_factor = config.routed_scaling_factor
        self.n_group = config.n_group
        self.topk_group = config.topk_group
        self.norm_topk_prob = config.norm_topk_prob

        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, config.hidden_size)))
        self.e_score_correction_bias = nn.Parameter(
            torch.zeros((self.n_routed_experts), dtype=torch.float32))

    @torch.no_grad()
    def get_topk_indices(self, scores):
        scores_for_choice = scores.view(-1, self.n_routed_experts) + self.e_score_correction_bias.unsqueeze(0)
        group_scores = (
            scores_for_choice.view(-1, self.n_group, self.n_routed_experts // self.n_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.n_group, self.n_routed_experts // self.n_group)
            .reshape(-1, self.n_routed_experts)
        )
        scores_for_choice = scores_for_choice.masked_fill(~score_mask.bool(), 0.0)
        topk_indices = torch.topk(scores_for_choice, k=self.top_k, dim=-1, sorted=False)[1]
        return topk_indices

    def forward(self, hidden_states):
        hidden_states = hidden_states.view(-1, self.config.hidden_size)
        router_logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
        scores = router_logits.sigmoid()
        topk_indices = self.get_topk_indices(scores)
        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 topk_indices, topk_weights


@use_kernel_forward_from_hub("RMSNorm")
class SolarOpen2RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        SolarOpen2RMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

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

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


class SolarOpen2MoE(nn.Module):
    """
    A mixed expert module containing shared experts.
    """

    def __init__(self, config):
        super().__init__()
        self.config = config
        self.experts = nn.ModuleList(
            [
                SolarOpen2MLP(config, intermediate_size=config.moe_intermediate_size)
                for _ in range(config.n_routed_experts)
            ]
        )
        self.gate = SolarOpen2TopkRouter(config)
        self.shared_experts = SolarOpen2MLP(
            config=config, intermediate_size=config.moe_intermediate_size * config.n_shared_experts
        )

    @torch.compiler.disable()
    def moe(self, hidden_states: torch.Tensor, topk_indices: torch.Tensor, topk_weights: torch.Tensor):
        r"""
        MoE forward pass that only executes selected experts.
        Uses @torch.compiler.disable() to allow dynamic shape operations.
        Requires --enforce-eager flag when serving with vLLM.
        """
        final_hidden_states = torch.zeros_like(hidden_states)

        for expert_idx in range(len(self.experts)):
            expert = self.experts[expert_idx]

            # Find positions where this expert was selected
            batch_idx, topk_pos = torch.where(topk_indices == expert_idx)

            if batch_idx.numel() == 0:
                continue

            # Extract only the tokens routed to this expert
            expert_input = hidden_states[batch_idx]
            expert_output = expert(expert_input)

            # Apply weights and accumulate results
            weights = topk_weights[batch_idx, topk_pos].unsqueeze(-1)
            final_hidden_states.index_add_(0, batch_idx, (expert_output * weights).to(hidden_states.dtype))

        return final_hidden_states

    def forward(self, hidden_states):
        residuals = hidden_states
        orig_shape = hidden_states.shape
        topk_indices, topk_weights = 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


class SolarOpen2HyperConnection(nn.Module):
    """Hyper-Connections module for SolarOpen2 (pure PyTorch, no Triton).

    Learnable replacement for residual connections that expands the hidden
    state to N copies and uses learned mixing (width) and combining (depth)
    connections.

    Reference: https://arxiv.org/abs/2409.19606
    """

    def __init__(self, config: SolarOpen2Config, layer_idx: int):
        super().__init__()
        dim = config.hidden_size
        rate = config.hc_rate
        self.rate = rate
        self.dynamic = config.hc_dynamic
        self.use_tanh = config.hc_use_tanh
        self.use_sigmoid = getattr(config, "hc_use_sigmoid", False)

        self.static_alpha = nn.Parameter(torch.zeros(rate, rate + 1))
        self.static_beta = nn.Parameter(torch.ones(rate))

        if self.dynamic:
            self.dynamic_alpha_fn = nn.Parameter(torch.zeros(dim, rate + 1))
            self.dynamic_alpha_scale = nn.Parameter(torch.ones(1) * 0.01)
            self.dynamic_beta_fn = nn.Parameter(torch.zeros(dim))
            self.dynamic_beta_scale = nn.Parameter(torch.ones(1) * 0.01)

            self.norm = SolarOpen2RMSNorm(dim, eps=config.rms_norm_eps)

    def width_connection(self, h: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """Split h into (h_in, h_res, beta) via learned width mixing.

        Args:
            h: (..., N, D) hidden states with N copies.

        Returns:
            h_in: (..., D) input for the sublayer.
            h_res: (..., N, D) residual for depth connection.
            beta: (N,) or (..., N) depth connection weights.
        """
        if self.dynamic:
            norm_h = self.norm(h)
            # norm_h: (..., N, D), dynamic_alpha_fn: (D, rate+1)
            wc_weight = norm_h @ self.dynamic_alpha_fn  # (..., N, rate+1)
            if self.use_tanh:
                wc_weight = torch.tanh(wc_weight)
            alpha = wc_weight * self.dynamic_alpha_scale + self.static_alpha
            # dc_weight: (..., N)
            dc_weight = (norm_h * self.dynamic_beta_fn).sum(-1)
            if self.use_tanh:
                dc_weight = torch.tanh(dc_weight)
            beta = dc_weight * self.dynamic_beta_scale + self.static_beta
        else:
            alpha = self.static_alpha
            beta = self.static_beta

        if self.use_sigmoid:
            alpha = torch.sigmoid(alpha)
            beta = 2.0 * torch.sigmoid(beta)

        # Mix: alpha^T @ h → (..., N+1, D)  — float32 accumulators to match
        # the Triton kernels in TorchTitan which compute in fp32 internally.
        orig_dtype = h.dtype
        alpha_t = alpha.float().transpose(-2, -1)  # (..., N+1, N) or (N+1, N)
        mix_h = torch.matmul(alpha_t, h.float())  # (..., N+1, D)
        h_in = mix_h[..., 0, :].to(orig_dtype)  # (..., D)
        h_res = mix_h[..., 1:, :].to(orig_dtype)  # (..., N, D)
        return h_in, h_res, beta

    def depth_connection(self, h_res: torch.Tensor, h_out: torch.Tensor, beta: torch.Tensor) -> torch.Tensor:
        """Combine sublayer output with residual via learned depth connection.

        Args:
            h_res: (..., N, D) residual from width connection.
            h_out: (..., D) sublayer output.
            beta: (N,) or (..., N) depth connection weights.

        Returns:
            h: (..., N, D) combined hidden states.
        """
        # float32 accumulators to match Triton kernels in TorchTitan.
        return (h_out.float().unsqueeze(-2) * beta.float().unsqueeze(-1)
                + h_res.float()).to(h_out.dtype)


class SolarOpen2DecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: SolarOpen2Config, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        # Hyper-Connections
        self.use_hc = getattr(config, "hc_rate", 0) > 0
        if self.use_hc:
            self.attn_hc = SolarOpen2HyperConnection(config, layer_idx=2 * layer_idx)
            self.ffn_hc = SolarOpen2HyperConnection(config, layer_idx=2 * layer_idx + 1)

        # Determine if this layer uses GQA: gqa_layers takes priority over gqa_interval
        if config.gqa_layers is not None:
            use_gqa = layer_idx in config.gqa_layers
        else:
            use_gqa = (layer_idx + 1) % config.gqa_interval == 0

        if use_gqa:
            self.is_linear_attn = False
            self.self_attn = SolarOpen2Attention(config=config, layer_idx=layer_idx)
        else:
            self.is_linear_attn = True
            self.self_attn = SolarOpen2LinearAttention(config=config, layer_idx=layer_idx)

        if layer_idx >= config.first_k_dense_replace:
            self.mlp = SolarOpen2MoE(config)
        else:
            self.mlp = SolarOpen2MLP(config)

        self.input_layernorm = SolarOpen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = SolarOpen2RMSNorm(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,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[SolarOpen2DynamicCache] = None,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        if self.use_hc:
            # ═══ Attention with Hyper-Connections ═══
            h_in, h_res, attn_beta = self.attn_hc.width_connection(hidden_states)
            h_in = self.input_layernorm(h_in)
            h_in, _ = self.self_attn(
                hidden_states=h_in,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **kwargs,
            )
            hidden_states = self.attn_hc.depth_connection(h_res, h_in, attn_beta)

            # ═══ FFN/MoE with Hyper-Connections ═══
            ffn_h_in, ffn_h_res, ffn_beta = self.ffn_hc.width_connection(hidden_states)
            ffn_h_in = self.post_attention_layernorm(ffn_h_in)
            ffn_h_in = self.mlp(ffn_h_in)
            hidden_states = self.ffn_hc.depth_connection(ffn_h_res, ffn_h_in, ffn_beta)
            return hidden_states
        else:
            residual = hidden_states
            hidden_states = self.input_layernorm(hidden_states)
            # 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,
                use_cache=use_cache,
                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)
            hidden_states = residual + hidden_states
            return hidden_states


@auto_docstring
class SolarOpen2PreTrainedModel(PreTrainedModel):
    config: SolarOpen2Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["SolarOpen2DecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _can_compile_fullgraph = False
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": SolarOpen2DecoderLayer,
        "attentions": SolarOpen2Attention,
    }

    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, SolarOpen2TopkRouter):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)


class SolarOpen2RotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: SolarOpen2Config, 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)


@auto_docstring
class SolarOpen2Model(SolarOpen2PreTrainedModel):
    _keys_to_ignore_on_load_unexpected = [r"model\.layers\.92.*", r"model\.layers\.46.*"]

    def __init__(self, config: SolarOpen2Config):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size
        self.hc_rate = getattr(config, "hc_rate", 0)

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [SolarOpen2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = SolarOpen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = SolarOpen2RotaryEmbedding(config=config)
        self.gradient_checkpointing = False

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

    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

    @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[SolarOpen2DynamicCache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        cache_position: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = 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 not isinstance(past_key_values, SolarOpen2DynamicCache):
            past_key_values = SolarOpen2DynamicCache(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.Tensor = 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 = create_causal_mask(
            config=self.config,
            input_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            past_key_values=past_key_values,
            position_ids=position_ids,
        )
        linear_attn_mask = self._update_linear_attn_mask(attention_mask, cache_position)

        hidden_states = inputs_embeds
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        # Hyper-Connections: expand (B, L, D) → (B, L, N, D) by replicating N times
        if self.hc_rate > 0:
            hidden_states = hidden_states.unsqueeze(-2).expand(
                *hidden_states.shape[:-1], self.hc_rate, hidden_states.shape[-1]
            ).contiguous()

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            layer_mask = linear_attn_mask if decoder_layer.is_linear_attn else causal_mask
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=layer_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        # Hyper-Connections: contract (B, L, N, D) → (B, L, D) by summing N copies
        if self.hc_rate > 0:
            hidden_states = hidden_states.sum(dim=-2)

        hidden_states = self.norm(hidden_states)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
        )


@auto_docstring
class SolarOpen2ForCausalLM(SolarOpen2PreTrainedModel, 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 = SolarOpen2Model(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: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[SolarOpen2DynamicCache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[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,
            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=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__ = ["SolarOpen2PreTrainedModel", "SolarOpen2Model", "SolarOpen2ForCausalLM"]