#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/olmo_hybrid/modular_olmo_hybrid.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_olmo_hybrid.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2026 the HuggingFace 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
from typing import Any, Optional

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
from ...generation import GenerationMixin
from ...integrations import use_kernel_forward_from_hub, use_kernelized_func
from ...masking_utils import create_causal_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
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, logging
from ...utils.generic import maybe_autocast, merge_with_config_defaults
from ...utils.import_utils import is_flash_linear_attention_available
from ...utils.output_capturing import capture_outputs
from .configuration_olmo_hybrid import OlmoHybridConfig


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

logger = logging.get_logger(__name__)


class OlmoHybridDynamicCache:
    """
    Cache for hybrid model supporting both attention KV cache and linear attention state.

    Inherits from Qwen3NextDynamicCache. The main difference is that this cache
    stores separate conv states for q, k, v (instead of a single conv_states list).
    """

    is_compileable = False

    def __init__(self, config: OlmoHybridConfig):
        super().__init__()
        self.layer_types = config.layer_types
        self.transformer_layers = [
            i for i in range(config.num_hidden_layers) if self.layer_types[i] == "full_attention"
        ]
        self.last_linear_layer = len(self.layer_types) - 1 - self.layer_types[::-1].index("linear_attention")
        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)]
        # Replace single conv_states with separate q, k, v conv states
        self.conv_states_q = [None for _ in range(config.num_hidden_layers)]
        self.conv_states_k = [None for _ in range(config.num_hidden_layers)]
        self.conv_states_v = [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."""
        batch_size = beam_idx.shape[0]
        for layer_idx in range(len(self.key_cache)):
            if self.key_cache[layer_idx] is not None:
                if self.key_cache[layer_idx].shape[0] < batch_size:
                    expand_ratio = batch_size // self.key_cache[layer_idx].shape[0]
                    self.key_cache[layer_idx] = self.key_cache[layer_idx].repeat_interleave(expand_ratio, dim=0)
                    self.value_cache[layer_idx] = self.value_cache[layer_idx].repeat_interleave(expand_ratio, dim=0)
                device = self.key_cache[layer_idx].device
                self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx.to(device))
                self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx.to(device))
            if self.conv_states_q[layer_idx] is not None:
                if self.conv_states_q[layer_idx].shape[0] < batch_size:
                    expand_ratio = batch_size // self.conv_states_q[layer_idx].shape[0]
                    self.conv_states_q[layer_idx] = self.conv_states_q[layer_idx].repeat_interleave(
                        expand_ratio, dim=0
                    )
                    self.conv_states_k[layer_idx] = self.conv_states_k[layer_idx].repeat_interleave(
                        expand_ratio, dim=0
                    )
                    self.conv_states_v[layer_idx] = self.conv_states_v[layer_idx].repeat_interleave(
                        expand_ratio, dim=0
                    )
                    self.recurrent_states[layer_idx] = self.recurrent_states[layer_idx].repeat_interleave(
                        expand_ratio, dim=0
                    )
                device = self.conv_states_q[layer_idx].device
                self.conv_states_q[layer_idx] = self.conv_states_q[layer_idx].index_select(0, beam_idx.to(device))
                self.conv_states_k[layer_idx] = self.conv_states_k[layer_idx].index_select(0, beam_idx.to(device))
                self.conv_states_v[layer_idx] = self.conv_states_v[layer_idx].index_select(0, beam_idx.to(device))
                self.recurrent_states[layer_idx] = self.recurrent_states[layer_idx].index_select(
                    0, beam_idx.to(device)
                )

    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."""
        return self.conv_states_q[self.last_linear_layer] is not None


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

    def forward(self, hidden_states, gate=None):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        # Norm before gate
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        hidden_states = self.weight * hidden_states.to(input_dtype)
        hidden_states = hidden_states * F.silu(gate.to(torch.float32))

        return hidden_states.to(input_dtype)


@use_kernel_forward_from_hub("RMSNorm")
class OlmoHybridRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        OlmoHybridRMSNorm 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:
        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 OlmoHybridShortConvolution(nn.Conv1d):
    def __init__(
        self,
        hidden_size: int,
        kernel_size: int,
        bias: bool = False,
        activation: str | None = "silu",
    ):
        super().__init__(
            in_channels=hidden_size,
            out_channels=hidden_size,
            kernel_size=kernel_size,
            groups=hidden_size,
            padding=kernel_size - 1,
            bias=bias,
        )
        self.hidden_size = hidden_size
        self.conv_kernel_size = kernel_size
        self.act_fn = ACT2FN[activation]

    def forward(
        self,
        hidden_states: torch.Tensor,
        cache: torch.Tensor | None = None,
        use_precomputed: bool = False,
        **kwargs,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        seq_len, dim = hidden_states.shape[-2:]

        hidden_states = hidden_states.transpose(1, 2)

        if use_precomputed:
            # Single token update (decoding mode)
            x_with_state = torch.cat([cache, hidden_states], dim=-1)
            out = F.conv1d(
                x_with_state,
                self.weight,
                self.bias,
                padding=0,
                groups=dim,
            )
            conv_state = x_with_state[:, :, 1:]
        else:
            # Multi-token forward (prefill mode)
            out = F.conv1d(hidden_states, self.weight, self.bias, padding=self.conv_kernel_size - 1, groups=dim)
            out = out[:, :, :seq_len]
            conv_state = F.pad(hidden_states, (self.conv_kernel_size - 1 - hidden_states.shape[-1], 0))

        out = self.act_fn(out)

        return out.transpose(1, 2), conv_state


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(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.
    """
    q_type, k_type = q.dtype, k.dtype
    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(q_type), k_embed.to(k_type)


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_kernelized_func(apply_rotary_pos_emb)
class OlmoHybridAttention(nn.Module):
    """
    Multi-headed attention for OLMo Hybrid that supports optional RoPE (NoPE mode).

    Inherits from Olmo3Attention. The only behavioral difference is that when
    position_embeddings is None, rotary position embeddings are skipped entirely,
    enabling NoPE mode for long context extension.
    """

    def __init__(self, config: OlmoHybridConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )
        self.q_norm = OlmoHybridRMSNorm(config.num_attention_heads * self.head_dim, config.rms_norm_eps)
        self.k_norm = OlmoHybridRMSNorm(config.num_key_value_heads * self.head_dim, config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None,
        attention_mask: torch.Tensor | None,
        past_key_values: Cache | None = None,
        cache_position: torch.LongTensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_norm(self.q_proj(hidden_states))
        key_states = self.k_norm(self.k_proj(hidden_states))
        value_states = self.v_proj(hidden_states)

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

        # NoPE mode: skip RoPE when position_embeddings is None
        cos, sin = None, None
        if position_embeddings is not None:
            cos, sin = position_embeddings
            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_values is not None:
            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 = 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,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class OlmoHybridRotaryEmbedding(nn.Module):
    """
    RoPE for OLMo Hybrid that returns float32 cos/sin to match OLMo-core.
    """

    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, config: OlmoHybridConfig, 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: OlmoHybridConfig | 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
    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):
            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

        # KEY difference from parent: return float32, don't cast to x.dtype
        return cos, sin


def apply_mask_to_padding_states(hidden_states, attention_mask):
    """
    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
    """
    # NOTE: attention mask is a 2D boolean tensor
    if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
        dtype = hidden_states.dtype
        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)

    return hidden_states


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


def torch_chunk_gated_delta_rule(
    query,
    key,
    value,
    g,
    beta,
    chunk_size=64,
    initial_state=None,
    output_final_state=False,
    use_qk_l2norm_in_kernel=False,
):
    initial_dtype = query.dtype
    if use_qk_l2norm_in_kernel:
        query = l2norm(query, dim=-1, eps=1e-6)
        key = l2norm(key, dim=-1, eps=1e-6)
    query, key, value, beta, g = [
        x.transpose(1, 2).contiguous().to(torch.float32) for x in (query, key, value, beta, g)
    ]

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

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

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

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

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


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

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

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

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

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

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


is_fast_path_available = all(
    (ShortConvolution, chunk_gated_delta_rule, fused_recurrent_gated_delta_rule, FusedRMSNormGated)
)


class OlmoHybridGatedDeltaNet(nn.Module):
    """
    GatedDeltaNet linear attention for OLMo Hybrid.

    Key differences from Qwen3NextGatedDeltaNet:
    - Fully separate q/k/v/a/b projections (vs. fused qkvz + partially split ba)
    - Per-projection conv1d for q, k, v (vs. single conv1d over concatenated qkv)
    - Dedicated g_proj gate (vs. z derived from the fused qkvz projection)
    - Supports allow_neg_eigval: scales beta by 2.0 to allow range [0, 2]
    """

    def __init__(self, config: OlmoHybridConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.num_v_heads = config.linear_num_value_heads
        self.num_k_heads = config.linear_num_key_heads
        self.head_k_dim = config.linear_key_head_dim
        self.head_v_dim = config.linear_value_head_dim
        self.key_dim = self.head_k_dim * self.num_k_heads
        self.value_dim = self.head_v_dim * self.num_v_heads
        self.layer_idx = layer_idx
        self.conv_kernel_size = config.linear_conv_kernel_dim
        self.allow_neg_eigval = config.linear_allow_neg_eigval
        self.eps = config.rms_norm_eps

        self.q_proj = nn.Linear(self.hidden_size, self.key_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.key_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, self.value_dim, bias=False)
        self.a_proj = nn.Linear(self.hidden_size, self.num_v_heads, bias=False)
        self.b_proj = nn.Linear(self.hidden_size, self.num_v_heads, bias=False)

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

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

        Conv1dClass = ShortConvolution if ShortConvolution is not None else OlmoHybridShortConvolution

        self.q_conv1d = Conv1dClass(
            hidden_size=self.key_dim,
            kernel_size=self.conv_kernel_size,
            bias=False,
            activation="silu",
        )
        self.k_conv1d = Conv1dClass(
            hidden_size=self.key_dim,
            kernel_size=self.conv_kernel_size,
            bias=False,
            activation="silu",
        )
        self.v_conv1d = Conv1dClass(
            hidden_size=self.value_dim,
            kernel_size=self.conv_kernel_size,
            bias=False,
            activation="silu",
        )

        A = torch.empty(self.num_v_heads, dtype=torch.float32).uniform_(
            config.linear_a_log_min, config.linear_a_log_max
        )
        self.A_log = nn.Parameter(torch.log(A))

        dt = torch.exp(
            torch.rand(self.num_v_heads) * (math.log(config.linear_dt_max) - math.log(config.linear_dt_min))
            + math.log(config.linear_dt_min)
        )
        dt = torch.clamp(dt, min=config.linear_dt_init_floor)
        inv_dt = dt + torch.log(-torch.expm1(-dt))
        self.dt_bias = nn.Parameter(inv_dt)

        # Output norm - NOTE: FLA's FusedRMSNormGated uses eps=1e-5 by default
        self.o_norm = (
            OlmoHybridRMSNormGated(self.head_v_dim, eps=1e-5)
            if FusedRMSNormGated is None
            else FusedRMSNormGated(
                self.head_v_dim,
                eps=1e-5,
                device=torch.cuda.current_device(),
                dtype=config.dtype if config.dtype is not None else torch.get_default_dtype(),
            )
        )

        self.chunk_gated_delta_rule = chunk_gated_delta_rule or torch_chunk_gated_delta_rule
        self.recurrent_gated_delta_rule = fused_recurrent_gated_delta_rule or torch_recurrent_gated_delta_rule

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        cache_params: OlmoHybridDynamicCache | None = None,
        cache_position: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        # Requires LEFT padding to work correctly
        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)

        batch_size, seq_len, _ = hidden_states.shape

        use_cache = cache_params is not None
        use_precomputed = use_cache and getattr(cache_params, "has_previous_state", False) and seq_len == 1

        conv_state_q = cache_params.conv_states_q[self.layer_idx] if cache_params else None
        conv_state_k = cache_params.conv_states_k[self.layer_idx] if cache_params else None
        conv_state_v = cache_params.conv_states_v[self.layer_idx] if cache_params else None
        recurrent_state = cache_params.recurrent_states[self.layer_idx] if cache_params else None

        q = self.q_proj(hidden_states)
        k = self.k_proj(hidden_states)
        v = self.v_proj(hidden_states)

        q, new_conv_state_q = self.q_conv1d(
            q, cache=conv_state_q, use_precomputed=use_precomputed, output_final_state=use_cache
        )
        k, new_conv_state_k = self.k_conv1d(
            k, cache=conv_state_k, use_precomputed=use_precomputed, output_final_state=use_cache
        )
        v, new_conv_state_v = self.v_conv1d(
            v, cache=conv_state_v, use_precomputed=use_precomputed, output_final_state=use_cache
        )

        if cache_params is not None:
            cache_params.conv_states_q[self.layer_idx] = new_conv_state_q
            cache_params.conv_states_k[self.layer_idx] = new_conv_state_k
            cache_params.conv_states_v[self.layer_idx] = new_conv_state_v

        q = q.view(batch_size, seq_len, -1, self.head_k_dim)
        k = k.view(batch_size, seq_len, -1, self.head_k_dim)
        v = v.view(batch_size, seq_len, -1, self.head_v_dim)

        if self.num_v_heads > self.num_k_heads:
            expand_ratio = self.num_v_heads // self.num_k_heads
            q = q.repeat_interleave(expand_ratio, dim=2)
            k = k.repeat_interleave(expand_ratio, dim=2)

        beta = self.b_proj(hidden_states).sigmoid()
        if self.allow_neg_eigval:
            beta = beta * 2.0

        g = -self.A_log.float().exp() * F.softplus(self.a_proj(hidden_states).float() + self.dt_bias)

        if use_precomputed:
            output, new_recurrent_state = self.recurrent_gated_delta_rule(
                q,
                k,
                v,
                g=g,
                beta=beta,
                initial_state=recurrent_state,
                output_final_state=use_cache,
                use_qk_l2norm_in_kernel=True,
            )
        else:
            output, new_recurrent_state = self.chunk_gated_delta_rule(
                q,
                k,
                v,
                g=g,
                beta=beta,
                initial_state=recurrent_state,
                output_final_state=use_cache,
                use_qk_l2norm_in_kernel=True,
            )

        if cache_params is not None:
            cache_params.recurrent_states[self.layer_idx] = new_recurrent_state

        gate = self.g_proj(hidden_states)
        output = output.reshape(-1, self.head_v_dim)
        gate = gate.reshape(-1, self.head_v_dim)
        output = self.o_norm(output, gate)
        output = output.reshape(batch_size, seq_len, -1)

        output = self.o_proj(output)

        return output


class OlmoHybridMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

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


class OlmoHybridAttentionDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: OlmoHybridConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = OlmoHybridAttention(config=config, layer_idx=layer_idx)

        self.mlp = OlmoHybridMLP(config)
        self.post_attention_layernorm = OlmoHybridRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_feedforward_layernorm = OlmoHybridRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.layer_type = "full_attention"

    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,
        cache_position: torch.LongTensor | None = None,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        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 = self.post_attention_layernorm(hidden_states)
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.mlp(hidden_states)
        hidden_states = self.post_feedforward_layernorm(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states


class OlmoHybridLinearAttentionDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: OlmoHybridConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.mlp = OlmoHybridMLP(config)
        self.input_layernorm = OlmoHybridRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = OlmoHybridRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.layer_type = "linear_attention"
        self.linear_attn = OlmoHybridGatedDeltaNet(config, layer_idx=layer_idx)

    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,
        cache_position: torch.LongTensor | None = None,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        output_attentions: bool | None = False,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        # Main difference to llama - signature (`cache_params`) and linear attention
        hidden_states = self.linear_attn(
            hidden_states=hidden_states,
            cache_params=past_key_values,
            cache_position=cache_position,
            attention_mask=attention_mask,
        )
        hidden_states = residual + hidden_states

        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


class OlmoHybridPreTrainedModel(PreTrainedModel):
    config: OlmoHybridConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["OlmoHybridAttentionDecoderLayer", "OlmoHybridLinearAttentionDecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn = True
    _supports_sdpa = True
    _keys_to_ignore_on_load_unexpected = [r"^mtp.*"]
    _can_record_outputs = {
        "hidden_states": (OlmoHybridAttentionDecoderLayer, OlmoHybridLinearAttentionDecoderLayer),
        "attentions": OlmoHybridAttention,
    }
    _is_stateful = True

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, OlmoHybridGatedDeltaNet):
            cfg = self.config
            init.copy_(
                module.A_log,
                torch.empty_like(module.A_log).uniform_(cfg.linear_a_log_min, cfg.linear_a_log_max).log_(),
            )
            dt = torch.exp(
                torch.rand_like(module.dt_bias) * (math.log(cfg.linear_dt_max) - math.log(cfg.linear_dt_min))
                + math.log(cfg.linear_dt_min)
            )
            dt = torch.clamp(dt, min=cfg.linear_dt_init_floor)
            inv_dt = dt + torch.log(-torch.expm1(-dt))
            init.copy_(module.dt_bias, inv_dt)


class OlmoHybridModel(OlmoHybridPreTrainedModel):
    def __init__(self, config: OlmoHybridConfig):
        super().__init__(config)
        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, config.pad_token_id)
        self.layers = nn.ModuleList(
            [
                OlmoHybridLinearAttentionDecoderLayer(config, layer_idx)
                if config.layer_types[layer_idx] == "linear_attention"
                else OlmoHybridAttentionDecoderLayer(config, layer_idx)
                for layer_idx in range(config.num_hidden_layers)
            ]
        )
        self.norm = OlmoHybridRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = (
            OlmoHybridRotaryEmbedding(config=config)
            if getattr(config, "rope_parameters", None) is not None
            and config.rope_parameters.get("rope_theta") is not None
            else None
        )
        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()

    @merge_with_config_defaults
    @capture_outputs
    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,
        cache_position: torch.LongTensor | 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 = self.embed_tokens(input_ids)

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

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

        causal_mask = 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
        # RoPE or NoPE
        position_embeddings = self.rotary_emb(hidden_states, position_ids) if self.rotary_emb is not None else None

        for decoder_layer in self.layers:
            layer_mask = linear_attn_mask if decoder_layer.layer_type == "linear_attention" else causal_mask
            layer_position_embeddings = position_embeddings if decoder_layer.layer_type == "full_attention" else None

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

        hidden_states = self.norm(hidden_states)

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

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


@auto_docstring
class OlmoHybridForCausalLM(OlmoHybridPreTrainedModel, 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"])}

    def __init__(self, config):
        super().__init__(config)
        self.model = OlmoHybridModel(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,
        cache_position: torch.LongTensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> CausalLMOutputWithPast:
        r"""
        Example:

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

        >>> model = OlmoHybridForCausalLM.from_pretrained("meta-olmo_hybrid/OlmoHybrid-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-olmo_hybrid/OlmoHybrid-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: 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__ = ["OlmoHybridForCausalLM", "OlmoHybridModel", "OlmoHybridPreTrainedModel"]