# Copyright 2026 IBM and 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 dataclasses import dataclass
from typing import Optional

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

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache
from transformers.generation import GenerationMixin
from transformers.integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub, use_kernelized_func
from transformers.masking_utils import create_bidirectional_mask, find_packed_sequence_indices, packed_sequence_mask_function
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 ModelOutput, TransformersKwargs
from transformers.utils.generic import maybe_autocast
from .configuration_granite_speech_nar import (
    GraniteSpeechNarConfig,
    GraniteSpeechNarEncoderConfig,
    GraniteSpeechNarProjectorConfig,
)


@dataclass
class GraniteSpeechNarEncoderOutput(ModelOutput):
    """Output of the GraniteSpeechNar encoder."""

    loss: torch.Tensor | None = None
    logits: torch.FloatTensor | None = None
    last_hidden_state: torch.FloatTensor | None = None
    all_hidden_states: tuple[torch.FloatTensor, ...] | None = None


@dataclass
class GraniteSpeechNarOutput(ModelOutput):
    """Output of the GraniteSpeechNarForASR model.

    Attributes:
        loss: Combined CTC + auxiliary losses (only when labels provided).
        preds: List of predicted token ID tensors per sample (after CTC collapse, inference only).
        logits: List of per-sample logit tensors from the LLM head.
        encoder_logits: Flat BPE CTC logits from the encoder.
        encoder_preds: List of CTC-collapsed encoder predictions per sample.
    """

    loss: torch.Tensor | None = None
    preds: list[torch.Tensor] | None = None
    logits: list[torch.Tensor] | None = None
    encoder_logits: torch.Tensor | None = None
    encoder_preds: list[torch.Tensor] | None = None


### Encoder - conformer is adapted from: https://github.com/lucidrains/conformer.git
class GraniteSpeechNarConformerFeedForward(nn.Module):
    """Feedforward module for conformer encoder blocks."""

    def __init__(self, config: GraniteSpeechNarEncoderConfig):
        super().__init__()
        self.pre_norm = nn.LayerNorm(config.hidden_dim)
        self.up_proj = nn.Linear(config.hidden_dim, config.hidden_dim * config.feedforward_mult)
        self.silu = nn.SiLU()
        self.dropout = nn.Dropout(config.dropout)
        self.down_proj = nn.Linear(config.hidden_dim * config.feedforward_mult, config.hidden_dim)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.pre_norm(hidden_states)
        hidden_states = self.up_proj(hidden_states)
        hidden_states = self.dropout(self.silu(hidden_states))
        hidden_states = self.down_proj(hidden_states)
        hidden_states = self.dropout(hidden_states)
        return hidden_states


class GraniteSpeechNarConformerAttention(nn.Module):
    """Attention for conformer blocks using Shaw's relative positional embeddings.
    See the following [paper](https://huggingface.co/papers/1803.02155) for more details.
    """

    def __init__(self, config: GraniteSpeechNarEncoderConfig):
        super().__init__()

        inner_dim = config.dim_head * config.num_heads
        self.max_pos_emb = config.max_pos_emb
        self.context_size = config.context_size
        self.num_heads = config.num_heads
        self.dim_head = config.dim_head
        self.scale = self.dim_head**-0.5
        self.pre_norm = nn.LayerNorm(config.hidden_dim)
        self.to_q = nn.Linear(config.hidden_dim, inner_dim, bias=False)
        self.to_kv = nn.Linear(config.hidden_dim, inner_dim * 2, bias=False)
        self.to_out = nn.Linear(inner_dim, config.hidden_dim)
        self.rel_pos_emb = nn.Embedding(2 * self.max_pos_emb + 1, self.dim_head)
        self.dropout = nn.Dropout(config.dropout)

        if self.context_size <= 0 or self.context_size > self.max_pos_emb:
            raise ValueError("Context size is either less than 0 or exceeds the max_pos_emb")

    def forward(self, hidden_states: torch.Tensor, attention_dists: torch.Tensor) -> torch.Tensor:
        hidden_states = self.pre_norm(hidden_states)
        bsz, num_features, _ = hidden_states.shape

        num_blocks = math.ceil(num_features / self.context_size)
        remainder = num_features % self.context_size
        if remainder > 0:
            # right padding to reach block size
            hidden_states = torch.nn.functional.pad(hidden_states, (0, 0, 0, self.context_size - remainder))

        query_states = self.to_q(hidden_states)
        key_states, value_states = self.to_kv(hidden_states).chunk(2, dim=-1)

        query_states = query_states.reshape(bsz, num_blocks, self.context_size, self.num_heads, -1).transpose(2, 3)
        key_states = key_states.reshape(bsz, num_blocks, self.context_size, self.num_heads, -1).transpose(2, 3)
        value_states = value_states.reshape(bsz, num_blocks, self.context_size, self.num_heads, -1).transpose(2, 3)

        # shaw's relative positional embedding
        rel_pos_emb = self.rel_pos_emb(attention_dists)
        # alternative computation of `pos_attn` - for readability
        # rel_pos_emb_expanded = rel_pos_emb.view([1, 1, 1] + list(rel_pos_emb.shape))
        # pos_attn = torch.sum(query_states.unsqueeze(-2) * rel_pos_emb_expanded, dim=-1) * self.scale
        # einsum implementation of pos_attn - gives x30 speedup over the alternative
        # TODO (@avihu111) find a fast alternative to einsum
        pos_attn = torch.einsum("b m h c d, c r d -> b m h c r", query_states, rel_pos_emb) * self.scale

        if remainder > 0:
            # masked attention in the extended block
            mask = torch.ones(self.context_size, self.context_size, dtype=bool, device=hidden_states.device)
            mask[:remainder, :remainder] = 0
            mask_value = -torch.finfo(pos_attn.dtype).max
            pos_attn[:, -1, :].masked_fill_(mask, mask_value)

        with torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.MATH):
            out = F.scaled_dot_product_attention(
                query_states, key_states, value_states, attn_mask=pos_attn, scale=self.scale
            )
        out = out.transpose(2, 3).reshape(bsz, hidden_states.shape[1], -1)
        out = self.to_out(out[:, :num_features, :])
        return self.dropout(out)


class GraniteSpeechNarConformerDepthWiseConv1d(nn.Module):
    """Wrapper for padded 1D pointwise convolution."""

    def __init__(self, chan_in: int, chan_out: int, kernel_size: int):
        super().__init__()
        # Padding for the 1D conv is symmetric or close (i.e., offset by one).
        pad = kernel_size // 2
        pad_offset = (kernel_size + 1) % 2
        self.padding = (pad, pad - pad_offset)

        self.conv = nn.Conv1d(chan_in, chan_out, kernel_size, groups=chan_in, bias=False)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = F.pad(hidden_states, self.padding)
        return self.conv(hidden_states)


class GraniteSpeechNarConformerConvModule(nn.Module):
    """Conformer conv module consisting of several 1D/depthwise 1D convolutional layers."""

    def __init__(self, config: GraniteSpeechNarEncoderConfig):
        super().__init__()
        inner_dim = config.hidden_dim * config.conv_expansion_factor

        self.norm = nn.LayerNorm(config.hidden_dim)
        self.up_conv = nn.Conv1d(config.hidden_dim, inner_dim * 2, 1)
        self.glu = nn.GLU(dim=1)
        self.depth_conv = GraniteSpeechNarConformerDepthWiseConv1d(
            inner_dim,
            inner_dim,
            kernel_size=config.conv_kernel_size,
        )
        self.silu = nn.SiLU()
        self.batch_norm = nn.BatchNorm1d(inner_dim)
        self.down_conv = nn.Conv1d(inner_dim, config.hidden_dim, 1)
        self.dropout = nn.Dropout(config.dropout)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.norm(hidden_states)
        hidden_states = self.up_conv(hidden_states.permute(0, 2, 1))
        hidden_states = self.glu(hidden_states)
        hidden_states = self.depth_conv(hidden_states)
        hidden_states = self.silu(self.batch_norm(hidden_states))
        hidden_states = self.down_conv(hidden_states).permute(0, 2, 1)
        hidden_states = self.dropout(hidden_states)
        return hidden_states


class GraniteSpeechNarConformerBlock(nn.Module):
    """Conformer block, consisting largely of linear layers, attention, and convolutional layers."""

    def __init__(self, config: GraniteSpeechNarEncoderConfig):
        super().__init__()
        self.ff1 = GraniteSpeechNarConformerFeedForward(config)
        self.attn = GraniteSpeechNarConformerAttention(config)
        self.conv = GraniteSpeechNarConformerConvModule(config)
        self.ff2 = GraniteSpeechNarConformerFeedForward(config)
        self.post_norm = nn.LayerNorm(config.hidden_dim)

    def forward(self, hidden_states: torch.Tensor, attention_dists: torch.Tensor) -> torch.Tensor:
        hidden_states = 0.5 * self.ff1(hidden_states) + hidden_states
        hidden_states = self.attn(hidden_states, attention_dists=attention_dists) + hidden_states
        hidden_states = self.conv(hidden_states) + hidden_states
        hidden_states = 0.5 * self.ff2(hidden_states) + hidden_states
        hidden_states = self.post_norm(hidden_states)
        return hidden_states


class GraniteSpeechNarQFormerCrossAttention(nn.Module):
    def __init__(self, config: GraniteSpeechNarProjectorConfig):
        super().__init__()
        self.num_heads = config.num_heads
        self.head_dim = config.hidden_size // config.num_heads
        self.hidden_size = config.hidden_size
        self.q_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=config.attn_bias)
        self.k_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=config.attn_bias)
        self.v_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=config.attn_bias)
        self.o_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=config.attn_bias)

    def forward(self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, query_len, _ = hidden_states.shape
        encoder_len = encoder_hidden_states.shape[1]

        query_states = (
            self.q_proj(hidden_states).view(batch_size, query_len, self.num_heads, self.head_dim).transpose(1, 2)
        )
        key_states = (
            self.k_proj(encoder_hidden_states)
            .view(batch_size, encoder_len, self.num_heads, self.head_dim)
            .transpose(1, 2)
        )
        value_states = (
            self.v_proj(encoder_hidden_states)
            .view(batch_size, encoder_len, self.num_heads, self.head_dim)
            .transpose(1, 2)
        )

        attn_output = F.scaled_dot_product_attention(query_states, key_states, value_states, is_causal=False)
        attn_output = attn_output.transpose(1, 2).contiguous().view(batch_size, query_len, self.hidden_size)
        return self.o_proj(attn_output)


class GraniteSpeechNarQFormerMLP(nn.Module):
    def __init__(self, config: GraniteSpeechNarProjectorConfig):
        super().__init__()
        mlp_hidden_size = int(config.hidden_size * config.mlp_ratio)
        self.fc1 = nn.Linear(config.hidden_size, mlp_hidden_size, bias=config.mlp_bias)
        self.act = nn.SiLU()
        self.fc2 = nn.Linear(mlp_hidden_size, config.hidden_size, bias=config.mlp_bias)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        return self.fc2(self.act(self.fc1(hidden_states)))


class GraniteSpeechNarQFormerLayer(nn.Module):
    def __init__(self, config: GraniteSpeechNarProjectorConfig):
        super().__init__()
        self.attn_norm = nn.LayerNorm(config.hidden_size, eps=config.layernorm_eps)
        self.cross_attention = GraniteSpeechNarQFormerCrossAttention(config)
        self.mlp_norm = nn.LayerNorm(config.hidden_size, eps=config.layernorm_eps)
        self.mlp = GraniteSpeechNarQFormerMLP(config)

    def forward(self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = hidden_states + self.cross_attention(self.attn_norm(hidden_states), encoder_hidden_states)
        hidden_states = hidden_states + self.mlp(self.mlp_norm(hidden_states))
        return hidden_states


class GraniteSpeechNarQFormer(nn.Module):
    def __init__(self, config: GraniteSpeechNarProjectorConfig):
        super().__init__()
        self.layers = nn.ModuleList([GraniteSpeechNarQFormerLayer(config) for _ in range(config.num_layers)])

    def forward(self, query_embeds: torch.Tensor, encoder_hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = query_embeds
        for layer in self.layers:
            hidden_states = layer(hidden_states, encoder_hidden_states)
        return hidden_states


class GraniteSpeechNarProjector(nn.Module):
    """Windowed QFormer projector that maps multi-layer encoder features to LLM embedding space."""

    def __init__(self, config: GraniteSpeechNarProjectorConfig):
        super().__init__()
        self.config = config
        self.layer_norms = nn.ModuleList(
            [nn.LayerNorm(config.encoder_dim, eps=config.layernorm_eps) for _ in range(config.num_encoder_layers)]
        )
        self.layer_projector = nn.Linear(config.encoder_dim * config.num_encoder_layers, config.hidden_size)
        self.dropout = nn.Dropout(config.dropout_prob)
        self.projector_act = nn.GELU()
        self.qformer = GraniteSpeechNarQFormer(config)

        query_length = config.block_size // config.downsample_rate
        embed_std = config.hidden_size**-0.5
        self.query = nn.Parameter(torch.randn(1, query_length, config.hidden_size) * embed_std)
        self.window_positions = nn.Parameter(torch.randn(1, config.block_size, config.hidden_size) * embed_std)
        self.out_norm = nn.LayerNorm(config.hidden_size, eps=config.layernorm_eps)
        self.out_linear = nn.Linear(config.hidden_size, config.llm_dim)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, seq_len, dim = hidden_states.size()

        hidden_states = hidden_states.view(
            batch_size, seq_len, self.config.num_encoder_layers, self.config.encoder_dim
        )
        normalized_layers = []
        for i, layer_norm in enumerate(self.layer_norms):
            normalized_layers.append(layer_norm(hidden_states[:, :, i]))
        hidden_states = torch.cat(normalized_layers, dim=-1)

        hidden_states = self.projector_act(self.layer_projector(hidden_states))

        block_size = self.config.block_size
        nblocks = seq_len // block_size
        rest = seq_len % block_size
        if rest > 0:
            hidden_states = F.pad(hidden_states, (0, 0, 0, block_size - rest), "constant", 0)
            nblocks += 1

        hidden_states = hidden_states.view(batch_size * nblocks, block_size, self.config.hidden_size)
        query_length = self.query.shape[1]
        mean_pool = hidden_states.view(
            batch_size * nblocks, query_length, self.config.downsample_rate, self.config.hidden_size
        ).mean(dim=-2)

        hidden_states = self.qformer(
            query_embeds=self.dropout(self.query + mean_pool),
            encoder_hidden_states=self.dropout(hidden_states + self.window_positions),
        )

        hidden_states = hidden_states.view(batch_size, nblocks * query_length, -1)
        hidden_states = self.dropout(self.out_norm(hidden_states))
        return self.out_linear(hidden_states)


class GraniteSpeechNarPreTrainedModel(PreTrainedModel):
    config_class = GraniteSpeechNarConfig
    base_model_prefix = "encoder"
    supports_gradient_checkpointing = True
    _supports_flash_attn = True
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _no_split_modules = ["GraniteSpeechNarConformerBlock", "GraniteDecoderLayer"]
    input_modalities = ("audio",)


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


@use_kernel_func_from_hub("rotary_pos_emb")
def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

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


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


@use_kernelized_func(apply_rotary_pos_emb)
class GraniteSpeechNarAttention(nn.Module):
    """GraniteAttention with is_causal=False for bidirectional attention."""

    is_causal = False

    def __init__(self, config, layer_idx=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 = config.attention_multiplier
        self.attention_dropout = config.attention_dropout
        self.is_causal = False

        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
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        attention_mask: torch.Tensor | None = None,
        past_key_values: Cache | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, 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).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

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

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

        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


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

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

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


class GraniteSpeechNarMLP(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=config.mlp_bias)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
        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 GraniteSpeechNarDecoderLayer(GradientCheckpointingLayer):
    """GraniteDecoderLayer using bidirectional attention."""

    def __init__(self, config, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = GraniteSpeechNarAttention(config=config, layer_idx=layer_idx)

        self.mlp = GraniteSpeechNarMLP(config)
        self.input_layernorm = GraniteSpeechNarRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = GraniteSpeechNarRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.residual_multiplier = config.residual_multiplier

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*):
                attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
                query_sequence_length, key_sequence_length)` if default attention is used.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
            past_key_values (`Cache`, *optional*): cached past key and value projection states
            position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                with `head_dim` being the embedding dimension of each attention head.
            kwargs (`dict`, *optional*):
                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
                into the model
        """
        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        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,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = residual + hidden_states * self.residual_multiplier

        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states * self.residual_multiplier

        return hidden_states


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

    def __init__(self, config: GraniteSpeechNarConfig, 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: GraniteSpeechNarConfig | None = None,
        device: Optional["torch.device"] = None,
        seq_len: int | None = None,
    ) -> tuple["torch.Tensor", float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            device (`torch.device`):
                The device to use for initialization of the inverse frequencies.
            seq_len (`int`, *optional*):
                The current sequence length. Unused for this type of RoPE.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        base = config.rope_parameters["rope_theta"]
        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads

        attention_factor = 1.0  # Unused in this type of RoPE

        # Compute the inverse frequencies
        inv_freq = 1.0 / (
            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
        )
        return inv_freq, attention_factor

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()

        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


class GraniteSpeechNarModel(GraniteSpeechNarPreTrainedModel):
    """GraniteModel with bidirectional (non-causal) attention.

    Uses GraniteSpeechNarDecoderLayer which sets is_causal=False,
    and replaces create_causal_mask() with create_bidirectional_mask() so all
    attention backends (SDPA, FA2, eager, flex) get a proper non-causal mask.
    """

    def __init__(self, config):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

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

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

    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        **kwargs,
    ) -> 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)

        inputs_embeds = inputs_embeds * self.embedding_multiplier

        if position_ids is None:
            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device).unsqueeze(0)

        packed_seq_mask = find_packed_sequence_indices(position_ids)
        and_mask_fn = packed_sequence_mask_function(packed_seq_mask) if packed_seq_mask is not None else None
        bidirectional_mask = create_bidirectional_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            and_mask_function=and_mask_fn,
        )

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

        # KV cache is not needed in a non-autoregressive model
        kwargs["use_cache"] = False
        for decoder_layer in self.layers:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=bidirectional_mask,
                position_ids=position_ids,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)

        return BaseModelOutputWithPast(last_hidden_state=hidden_states)


def _posterior_weighted_pool(hidden: torch.Tensor, importance: torch.Tensor, window_size: int = 4) -> torch.Tensor:
    batch_size, seq_len, hidden_dim = hidden.shape
    pad_len = (window_size - seq_len % window_size) % window_size
    if pad_len > 0:
        hidden = F.pad(hidden, (0, 0, 0, pad_len))
        importance = F.pad(importance, (0, pad_len))
    num_windows = hidden.shape[1] // window_size
    hidden = hidden.view(batch_size, num_windows, window_size, hidden_dim)
    importance = importance.view(batch_size, num_windows, window_size)
    weights = importance / (importance.sum(dim=-1, keepdim=True) + 1e-8)
    pooled = (hidden * weights.unsqueeze(-1)).sum(dim=2)
    return pooled


class GraniteSpeechNarCTCEncoder(GraniteSpeechNarPreTrainedModel):
    """Conformer encoder with BPE CTC head and multi-layer output."""

    config_class = GraniteSpeechNarEncoderConfig

    def __init__(self, config: GraniteSpeechNarEncoderConfig):
        super().__init__(config)
        self.input_linear = nn.Linear(config.input_dim, config.hidden_dim, bias=True)
        self.layers = nn.ModuleList([GraniteSpeechNarConformerBlock(config) for _ in range(config.num_layers)])
        self.out = nn.Linear(config.hidden_dim, config.output_dim, bias=True)
        self.out_mid = nn.Linear(config.output_dim, config.hidden_dim, bias=True)
        self.out_bpe = None
        if config.bpe_output_dim is not None:
            self.out_bpe = nn.Linear(config.hidden_dim, config.bpe_output_dim, bias=True)
        self.dropout = nn.Dropout(config.pred_dropout)
        self.post_init()

    def forward(
        self,
        input_features: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        output_hidden_states: bool | None = None,
        labels: torch.Tensor | None = None,
        label_lengths: torch.Tensor | None = None,
        **kwargs,
    ) -> GraniteSpeechNarEncoderOutput:
        if attention_mask is None:
            attention_mask = torch.ones(input_features.shape[:-1], dtype=torch.bool, device=input_features.device)

        hidden_states = self.input_linear(input_features.to(self.dtype))
        all_hidden_states = (hidden_states,) if output_hidden_states else None
        blank_probs = None

        context_size = self.config.context_size
        seq = torch.arange(context_size, device=hidden_states.device)
        relpos_dist = seq.view(-1, 1) - seq.view(1, -1)
        attention_dists = torch.clamp(relpos_dist, -context_size, context_size) + self.config.max_pos_emb

        for layer_idx, layer in enumerate(self.layers, start=1):
            hidden_states = layer(hidden_states, attention_dists=attention_dists)

            if layer_idx == self.config.self_conditioning_layer:
                mid_logits = self.out(self.dropout(hidden_states))
                mid_probs = torch.softmax(mid_logits.float(), dim=-1)
                blank_probs = mid_probs[:, :, 0]
                hidden_states = hidden_states + self.out_mid(mid_probs.to(hidden_states.dtype))

            if output_hidden_states:
                all_hidden_states += (hidden_states,)

        hidden_states = self.dropout(hidden_states)

        logits = None
        loss = None
        if self.out_bpe is not None and blank_probs is not None:
            pool_window = self.config.bpe_pooling_window
            importance = 1.0 - blank_probs
            pooled = _posterior_weighted_pool(hidden_states.float(), importance, window_size=pool_window).to(
                hidden_states.dtype
            )
            encoder_lengths = attention_mask.sum(dim=1)
            lengths = -(encoder_lengths // -pool_window)
            lengths_list = lengths.tolist()
            logits = self.out_bpe(torch.cat([pooled[i, :length] for i, length in enumerate(lengths_list)]))

            if labels is not None:
                logits_padded = logits.new_zeros(len(lengths_list), max(lengths_list), logits.shape[-1])
                offset = 0
                for i, length in enumerate(lengths_list):
                    logits_padded[i, :length] = logits[offset : offset + length]
                    offset += length

                log_probs = torch.log_softmax(logits_padded.float(), dim=-1)
                loss = (
                    F.ctc_loss(
                        log_probs.transpose(0, 1),
                        labels,
                        lengths,
                        label_lengths,
                        blank=self.config.blank_token_id,
                        reduction="sum",
                        zero_infinity=True,
                    )
                    / lengths.sum()
                )

        return GraniteSpeechNarEncoderOutput(
            loss=loss,
            logits=logits,
            last_hidden_state=hidden_states,
            all_hidden_states=all_hidden_states,
        )


class GraniteSpeechNarLM(GraniteSpeechNarPreTrainedModel, GenerationMixin):
    """GraniteForCausalLM with a bidirectional (non-causal) backbone."""

    _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 = GraniteSpeechNarModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

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

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

        >>> model = GraniteSpeechNarLM.from_pretrained("meta-granite_speech_nar/GraniteSpeechNar-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-granite_speech_nar/GraniteSpeechNar-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,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        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, :])
        logits = logits / self.config.logits_scaling  # main diff with Llama

        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,
        )


class GraniteSpeechNarForASR(GraniteSpeechNarPreTrainedModel):
    def __init__(self, config: GraniteSpeechNarConfig):
        super().__init__(config)

        self.encoder = GraniteSpeechNarCTCEncoder(config.encoder_config)
        self.projector = GraniteSpeechNarProjector(config.projector_config)

        text_config = config.text_config
        if hasattr(config, "_attn_implementation"):
            text_config._attn_implementation = config._attn_implementation
        self.language_model = GraniteSpeechNarLM._from_config(text_config)

        self.post_init()

    def _ctc_collapse_decode(
        self,
        bpe_logits_flat: torch.Tensor,
        bpe_lengths: list[int],
    ) -> list[torch.Tensor]:
        """GPU CTC greedy decode: argmax -> unique_consecutive -> remove blank."""
        blank_id = self.config.blank_token_id
        preds_flat = bpe_logits_flat.argmax(dim=-1)
        per_sample = preds_flat.split(bpe_lengths)
        return [(collapsed := torch.unique_consecutive(seq))[collapsed != blank_id] for seq in per_sample]

    def _add_insertion_slots(self, token_ids: torch.Tensor) -> torch.Tensor:
        """Insert blank tokens between each CTC token as editing slots for the LLM."""
        blank_id = self.config.blank_token_id
        n = token_ids.numel()
        total_len = max(2 * n + 1, self.config.min_edit_sequence_length)
        idx = torch.arange(n, device=token_ids.device)
        out_idx = 2 * idx + 1
        out = torch.full((total_len,), fill_value=blank_id, dtype=token_ids.dtype, device=token_ids.device)
        out[out_idx] = token_ids
        return out

    def _build_flat_inputs(
        self,
        ctc_token_ids: list[torch.Tensor],
        audio_embeds: torch.Tensor,
        audio_lengths: list[int],
    ) -> tuple[torch.Tensor, torch.Tensor, list[int]]:
        """Build flat (pad-free) LLM input: [audio_0, text_0, audio_1, text_1, ...]"""
        embed_tokens = self.language_model.model.embed_tokens

        embeds_list = []
        position_ids_list = []
        text_lengths = []

        for i, audio_len in enumerate(audio_lengths):
            audio_emb = audio_embeds[i, :audio_len]
            text_ids_with_slots = self._add_insertion_slots(ctc_token_ids[i])
            text_emb = embed_tokens(text_ids_with_slots)
            sample_embeds = torch.cat([audio_emb, text_emb], dim=0)
            embeds_list.append(sample_embeds)
            position_ids_list.append(torch.arange(sample_embeds.shape[0], device=audio_embeds.device))
            text_lengths.append(text_ids_with_slots.shape[0])

        flat_embeds = torch.cat(embeds_list, dim=0).unsqueeze(0)
        flat_position_ids = torch.cat(position_ids_list, dim=0).unsqueeze(0)
        return flat_embeds, flat_position_ids, text_lengths

    def forward(
        self,
        *,
        input_features: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        labels: torch.Tensor | None = None,
        label_lengths: torch.Tensor | None = None,
        output_encoder_logits: bool = False,
        **kwargs,
    ) -> GraniteSpeechNarOutput:
        r"""
        Args:
            input_features (`torch.Tensor` of shape `(batch_size, seq_len, input_dim)`):
                Mel spectrogram features.
            attention_mask (`torch.Tensor` of shape `(batch_size, seq_len)`, *optional*):
                Encoder attention mask (1 for valid frames, 0 for padding).
            labels (`torch.Tensor` of shape `(batch_size, max_label_len)`, *optional*):
                Ground truth LLM token IDs for training.
            label_lengths (`torch.Tensor` of shape `(batch_size,)`, *optional*):
                Number of valid tokens per sample in `labels`.
            output_encoder_logits (`bool`, *optional*, defaults to `False`):
                Whether to return encoder BPE logits. When False, the large logits
                tensor is freed early to reduce peak memory.

        Returns:
            [`GraniteSpeechNarOutput`]
        """
        encoder_labels = labels if self.config.encoder_ctc_loss_lambda else None
        enc_out = self.encoder(
            input_features=input_features,
            attention_mask=attention_mask,
            output_hidden_states=True,
            labels=encoder_labels,
            label_lengths=label_lengths if encoder_labels is not None else None,
        )

        if attention_mask is None:
            attention_mask = torch.ones(input_features.shape[:-1], dtype=torch.bool, device=input_features.device)

        encoder_lengths = attention_mask.sum(dim=1)

        pool_window = self.encoder.config.bpe_pooling_window
        bpe_lengths = (-(encoder_lengths // -pool_window)).tolist()
        ctc_token_ids = self._ctc_collapse_decode(enc_out.logits, bpe_lengths)

        multilayer_features = torch.cat(
            [enc_out.all_hidden_states[idx] for idx in self.config.encoder_layer_indices], dim=-1
        )

        encoder_loss = enc_out.loss
        encoder_logits = enc_out.logits if output_encoder_logits else None
        del enc_out

        audio_embeds = self.projector(multilayer_features)
        del multilayer_features
        if self.config.scale_projected_embeddings:
            embedding_multiplier = getattr(self.config.text_config, "embedding_multiplier", 1.0)
            audio_embeds = audio_embeds / embedding_multiplier
        audio_embeds = audio_embeds.to(self.language_model.model.embed_tokens.weight.dtype)

        audio_lengths = (encoder_lengths // self.projector.config.downsample_rate).cpu().tolist()

        flat_embeds, flat_position_ids, text_lengths = self._build_flat_inputs(
            ctc_token_ids, audio_embeds, audio_lengths
        )

        llm_out = self.language_model(
            inputs_embeds=flat_embeds,
            position_ids=flat_position_ids,
        )
        all_logits = llm_out.logits.squeeze(0)

        segment_lengths = [l for a, t in zip(audio_lengths, text_lengths) for l in (a, t)]
        text_logits = torch.cat(list(all_logits.split(segment_lengths)[1::2]))
        logits_per_sample = list(text_logits.split(text_lengths))

        loss = None
        if labels is not None:
            log_probs = torch.log_softmax(text_logits.float(), dim=-1)

            log_probs_padded = log_probs.new_zeros(len(text_lengths), max(text_lengths), log_probs.shape[-1])
            offset = 0
            for i, tl in enumerate(text_lengths):
                log_probs_padded[i, :tl] = log_probs[offset : offset + tl]
                offset += tl

            input_lengths = torch.tensor(text_lengths, device=text_logits.device)

            loss = (
                F.ctc_loss(
                    log_probs_padded.transpose(0, 1),
                    labels,
                    input_lengths,
                    label_lengths,
                    blank=self.config.blank_token_id,
                    reduction="sum",
                    zero_infinity=True,
                )
                / input_lengths.sum()
            )

            if self.config.ce_loss_lambda > 0.0:
                ce_targets = torch.cat([self._add_insertion_slots(ids) for ids in ctc_token_ids])
                ce_loss = F.cross_entropy(
                    text_logits,
                    ce_targets.long(),
                    reduction="mean",
                    ignore_index=-100,
                )
                loss = loss + self.config.ce_loss_lambda * ce_loss

            if encoder_loss is not None:
                loss = loss + self.config.encoder_ctc_loss_lambda * encoder_loss

        return GraniteSpeechNarOutput(
            loss=loss,
            logits=logits_per_sample,
            encoder_logits=encoder_logits,
            encoder_preds=ctc_token_ids,
        )

    @torch.inference_mode()
    def transcribe(
        self,
        input_features: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        output_encoder_logits: bool = False,
    ) -> GraniteSpeechNarOutput:
        """Single-pass non-autoregressive inference: forward + CTC collapse on LLM output.

        Returns token ID tensors in `preds`. Use `GraniteSpeechNarProcessor.batch_decode()`
        to convert to strings.
        """
        output = self.forward(
            input_features=input_features,
            attention_mask=attention_mask,
            output_encoder_logits=output_encoder_logits,
        )

        blank_id = self.config.blank_token_id
        preds = []
        for sample_logits in output.logits:
            pred = torch.unique_consecutive(sample_logits.argmax(-1))
            pred = pred[pred != blank_id]
            preds.append(pred)

        return GraniteSpeechNarOutput(
            preds=preds,
            logits=output.logits,
            encoder_logits=output.encoder_logits,
            encoder_preds=output.encoder_preds,
        )


__all__ = [
    "GraniteSpeechNarModel",
    "GraniteSpeechNarCTCEncoder",
    "GraniteSpeechNarForASR",
    "GraniteSpeechNarLM",
    "GraniteSpeechNarPreTrainedModel",
]