#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/muse_glimmer_assistant/modular_muse_glimmer_assistant.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_muse_glimmer_assistant.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.


from collections.abc import Callable

import torch
import torch.nn as nn

from ...activations import ACT2FN
from ...cache_utils import DFlashCache
from ...integrations import use_kernel_forward_from_hub
from ...masking_utils import create_bidirectional_mask, create_bidirectional_sliding_window_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutputWithPast
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
from ...utils.deprecation import deprecate_kwarg
from ...utils.generic import maybe_autocast, merge_with_config_defaults
from ...utils.output_capturing import capture_outputs
from .configuration_muse_glimmer_assistant import MuseGlimmerAssistantConfig


@use_kernel_forward_from_hub("RMSNorm")
class MuseGlimmerAssistantRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        MuseGlimmerAssistantRMSNorm 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}"


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


def 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 where keys have concatenated context
    from main model. This means that `q` and `k` do not have the same seq_len, so we slice to get the `cos` and `sin` corresponding
    to latest positions for `q`.

    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)
    # Due to the added context from the main model, k/v and q do not have the same seq_len, so we have to slice here
    q_len = q.size(-2)
    q_embed = (q * cos[..., -q_len:, :]) + (rotate_half(q) * sin[..., -q_len:, :])
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


class MuseGlimmerAssistantAttention(nn.Module):
    def __init__(self, config: MuseGlimmerAssistantConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.num_attention_heads = config.num_attention_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.hidden_size = config.hidden_size
        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.attention_dropout = config.attention_dropout
        self.is_causal = False
        self.scaling = self.head_dim**-0.5
        self.sliding_window = config.sliding_window
        layer_type = config.layer_types[layer_idx] if hasattr(config, "layer_types") else None
        self.is_sliding = layer_type == "sliding_attention"

        self.q_proj = nn.Linear(self.hidden_size, self.num_attention_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(self.num_attention_heads * self.head_dim, self.hidden_size, bias=False)

        self.q_norm = MuseGlimmerAssistantRMSNorm(self.head_dim, eps=config.rms_norm_eps)
        self.k_norm = MuseGlimmerAssistantRMSNorm(self.head_dim, eps=config.rms_norm_eps)

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

        # The total k/v states in Dflash are the concatenation of the previous `context_hidden_states` (same for every layer)
        # and the actual projections on the diffusion window (the actual `hidden_states` input). Everything gets appended to the
        # `past_key_values`, and then the diffusion window will be evicted from it so that the cache is effectively only the context
        # from the main model
        if context_hidden_states is not None:
            kv_hidden_states = torch.cat([context_hidden_states, hidden_states], dim=1)
            kv_hidden_shape = (*kv_hidden_states.shape[:-1], -1, self.head_dim)

        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        key_states = self.k_proj(kv_hidden_states).view(kv_hidden_shape).transpose(1, 2)
        value_states = self.v_proj(kv_hidden_states).view(kv_hidden_shape).transpose(1, 2)

        # We use QK-norm
        query_states = self.q_norm(query_states)
        key_states = self.k_norm(key_states)

        cos, sin = position_embeddings
        # Note that here q_states and k_states do not have the same seq_len (even before the `cache.update()` call), so this
        # will take care of slicing for latest positions for q_states
        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,
            sliding_window=self.sliding_window if self.is_sliding else None,
            **kwargs,
        )

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


class MuseGlimmerAssistantMLP(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 MuseGlimmerAssistantDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: MuseGlimmerAssistantConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = MuseGlimmerAssistantAttention(config=config, layer_idx=layer_idx)

        self.mlp = MuseGlimmerAssistantMLP(config)
        self.post_attention_layernorm = MuseGlimmerAssistantRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.input_layernorm = MuseGlimmerAssistantRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        context_hidden_states: torch.FloatTensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor | None = None,
        past_key_values: DFlashCache | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            context_hidden_states=context_hidden_states,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            **kwargs,
        )
        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


@auto_docstring
class MuseGlimmerAssistantPreTrainedModel(PreTrainedModel):
    config: MuseGlimmerAssistantConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["MuseGlimmerAssistantDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True

    _can_compile_fullgraph = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": MuseGlimmerAssistantDecoderLayer,
        "attentions": MuseGlimmerAssistantAttention,
    }
    config_class = MuseGlimmerAssistantConfig
    main_input_name = "noise_embeds"


class MuseGlimmerAssistantContextProjection(nn.Module):
    def __init__(self, config: MuseGlimmerAssistantConfig):
        super().__init__()
        # Project concatenated context hidden states -> hidden_size
        self.target_layer_ids = config.target_layer_ids
        encoder_input_size = len(config.target_layer_ids) * config.hidden_size
        self.fc = nn.Linear(encoder_input_size, config.hidden_size, bias=False)
        self.output_norm_enc = MuseGlimmerAssistantRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(self, context_hidden_states: torch.Tensor) -> torch.Tensor:
        """
        context_hidden_states (`torch.FloatTensor` of shape `[batch_size, seq_length, dim * len(config.target_layer_ids)]`):
            Context hidden states from target model's selected layer ids concatenated in the last dim.
        """
        context_hidden_states = self.fc(context_hidden_states)
        context_hidden_states = self.output_norm_enc(context_hidden_states)
        return context_hidden_states


class MuseGlimmerAssistantRotaryEmbedding(nn.Module):
    @deprecate_kwarg("device", version="5.18")
    def __init__(self, config: MuseGlimmerAssistantConfig, 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.inv_freq = nn.Buffer(inv_freq, persistent=False)
        self.original_inv_freq = nn.Buffer(inv_freq.clone(), persistent=False)

    @staticmethod
    @deprecate_kwarg("device", version="5.18")
    def compute_default_rope_parameters(
        config: MuseGlimmerAssistantConfig, device=None, **kwargs
    ) -> tuple[torch.Tensor, float]:
        """
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        """
        base = config.rope_parameters["rope_theta"]
        dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads

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


@auto_docstring
class MuseGlimmerAssistantModel(MuseGlimmerAssistantPreTrainedModel):
    def __init__(self, config: MuseGlimmerAssistantConfig):
        super().__init__(config)
        self.layers = nn.ModuleList(
            [MuseGlimmerAssistantDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = MuseGlimmerAssistantRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = MuseGlimmerAssistantRotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        self.encoder = MuseGlimmerAssistantContextProjection(config)

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

    @merge_with_config_defaults
    @capture_outputs
    def forward(
        self,
        noise_embeds: torch.FloatTensor,
        context_hidden_states: torch.FloatTensor,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: DFlashCache | None = None,
        use_cache: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | BaseModelOutputWithPast:
        r"""
        noise_embeds (`torch.FloatTensor` of shape `[batch_size, config.block_size, dim]`):
            Input embedding for the last generated anchor token and mask tokens to be denoised.
        context_hidden_states (`torch.FloatTensor` of shape `[batch_size, number_of_previous_accepted_tokens, dim * len(config.target_layer_ids)]`):
            Context hidden states from target model's selected layer ids concatenated in the last dim.
        attention_mask (`torch.Tensor` of shape `[batch_size, number_of_previous_accepted_tokens + config.block_size]`):
            Similar to the usual attention_mask, but note that it has length `number_of_previous_accepted_tokens + config.block_size`,
            because the Attention will first concatenate `context_hidden_states` and the hidden states derived from `noise_embeds`, so that
            k/v states do not have the same length as q_states, even before the `cache.update()` call. Thus the kv_seq_len dimension of
            the attention mask needs to span the additional positions.
        position_ids (`torch.Tensor` of shape `[batch_size, number_of_previous_accepted_tokens + config.block_size]`):
            Similar to the usual position_ids, but note that it has length `number_of_previous_accepted_tokens + config.block_size`,
            because the Attention will first concatenate `context_hidden_states` and the hidden states derived from `noise_embeds`, so that
            k/v states do not have the same length as q_states, even before the `cache.update()` call. Thus the `position_ids` and the derived
            `position_embeddings` need to span all the additional positions.
        """
        if use_cache and past_key_values is None:
            past_key_values = DFlashCache(config=self.config)

        # project targets to the same hidden dim as the model
        context_hidden_states = self.encoder(context_hidden_states)

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

        if not isinstance(mask_mapping := attention_mask, dict):
            mask_kwargs = {
                "config": self.config,
                "inputs_embeds": noise_embeds,
                "attention_mask": attention_mask,
                "past_key_values": past_key_values,
            }
            # The queries corresponding to `noise_embeds` must attend bi-directionally to each other, and causally to previous k/v states
            # derived from the main model's `context_hidden_states`. However, since they strictly correspond to positions larger than
            # the cache derived to `context_hidden_states`, this corresponds to bi-directional attention everywhere for the queries
            mask_mapping = {
                "full_attention": create_bidirectional_mask(**mask_kwargs),
                "sliding_attention": create_bidirectional_sliding_window_mask(**mask_kwargs),
            }

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

        for i, decoder_layer in enumerate(self.layers):
            layer_type = self.config.layer_types[i]
            hidden_states = decoder_layer(
                hidden_states,
                context_hidden_states=context_hidden_states,
                position_embeddings=position_embeddings,
                attention_mask=mask_mapping[layer_type],
                past_key_values=past_key_values,
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)

        return BaseModelOutputWithPast(last_hidden_state=hidden_states, past_key_values=past_key_values)


__all__ = ["MuseGlimmerAssistantModel", "MuseGlimmerAssistantPreTrainedModel"]