image-match/python/py/Lib/site-packages/transformers/models/minimax/modeling_minimax.py

#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
#           This file was automatically generated from src/transformers/models/minimax/modular_minimax.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_minimax.py file directly. One of our CI enforces this.
#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
# Copyright 2025 MiniMaxAI and HuggingFace Inc. teams. 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
from typing import Optional

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

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...integrations import (
    use_experts_implementation,
    use_kernel_forward_from_hub,
    use_kernel_func_from_hub,
    use_kernelized_func,
)
from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask
from ...modeling_flash_attention_utils import FlashAttentionKwargs
from ...modeling_layers import (
    GenericForQuestionAnswering,
    GenericForSequenceClassification,
    GenericForTokenClassification,
    GradientCheckpointingLayer,
)
from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
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
from ...utils.generic import maybe_autocast, merge_with_config_defaults
from ...utils.output_capturing import OutputRecorder, capture_outputs
from .configuration_minimax import MiniMaxConfig


@use_kernel_forward_from_hub("RMSNorm")
class MiniMaxRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps: float = 1e-6) -> None:
        """
        MiniMaxRMSNorm 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 MiniMaxCache(DynamicCache):
    def __init__(self):
        super().__init__()
        self.linear_cache: list[torch.Tensor] = []

    def set_linear_cache(self, layer_idx, linear_cache):
        # There may be skipped layers, fill them with empty lists
        for _ in range(len(self.linear_cache), layer_idx + 1):
            self.linear_cache.append([])
        self.linear_cache[layer_idx] = linear_cache

    def get_linear_cache(self, layer_idx: int):
        if layer_idx < len(self):
            return self.linear_cache[layer_idx]
        return None

    def __len__(self):
        return max(super().__len__(), len(self.linear_cache))

    def batch_repeat_interleave(self, repeats: int):
        for layer_idx in range(len(self)):
            if self.linear_cache[layer_idx] != []:
                self.linear_cache[layer_idx] = self.linear_cache[layer_idx].repeat_interleave(repeats, dim=0)
            else:
                self.layers[layer_idx].batch_repeat_interleave(repeats)

    def batch_select_indices(self, indices: torch.Tensor):
        for layer_idx in range(len(self)):
            if self.linear_cache[layer_idx] != []:
                self.linear_cache[layer_idx] = self.linear_cache[layer_idx][indices, ...]
            else:
                self.layers[layer_idx].batch_select_indices(indices)

    def crop(self, max_length: int):
        raise RuntimeError("MiniMaxCache doesnot support `crop` method")


class MiniMaxLightningAttention(nn.Module):
    def __init__(self, config: MiniMaxConfig, layer_idx: int):
        super().__init__()
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
        self.num_attention_heads = config.num_attention_heads
        self.num_hidden_layers = config.num_hidden_layers
        self.block_size = config.block_size

        self.act_fn = ACT2FN[config.hidden_act]
        self.norm = MiniMaxRMSNorm(self.head_dim * self.num_attention_heads)
        self.qkv_proj = nn.Linear(config.hidden_size, self.num_attention_heads * self.head_dim * 3, bias=False)
        self.out_proj = nn.Linear(self.num_attention_heads * self.head_dim, config.hidden_size, bias=False)
        self.output_gate = nn.Linear(config.hidden_size, self.num_attention_heads * self.head_dim, bias=False)

        slope_rate = self.get_slope_rate()
        query_decay, key_decay, diagonal_decay = self.decay_factors(slope_rate)

        self.register_buffer("slope_rate", slope_rate)
        self.register_buffer("query_decay", query_decay)
        self.register_buffer("key_decay", key_decay)
        self.register_buffer("diagonal_decay", diagonal_decay)

    def get_slope_rate(self):
        base = 1 / (2 ** (8 / self.num_attention_heads))
        exponent = torch.arange(self.num_attention_heads) + 1
        factor = 1 - self.layer_idx / (self.num_hidden_layers - 1 + 1e-5) + 1e-5

        rate = base**exponent
        rate = rate * factor
        rate = rate[:, None, None]

        return rate

    def decay_factors(self, slope_rate):
        block_size_range = torch.arange(self.block_size) + 1

        query_decay = torch.exp(-slope_rate * block_size_range[:, None])
        key_decay = torch.exp(-slope_rate * (self.block_size - block_size_range[:, None]))

        diagonal_decay = block_size_range[:, None] - block_size_range[None, :]
        diagonal_decay = diagonal_decay[None, None, :, :]
        diagonal_decay = slope_rate * diagonal_decay
        diagonal_decay = torch.where(diagonal_decay >= 0, -diagonal_decay, float("-inf"))
        diagonal_decay = torch.exp(diagonal_decay)

        return query_decay, key_decay, diagonal_decay

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor | None,
        past_key_values: Cache | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None, tuple[torch.Tensor] | None]:
        batch_size, seq_len, hidden_size = hidden_states.shape
        num_blocks = (seq_len + self.block_size - 1) // self.block_size

        qkv_states = self.act_fn(self.qkv_proj(hidden_states))
        qkv_states = qkv_states.reshape(batch_size, seq_len, self.num_attention_heads, 3 * self.head_dim)

        query_states, key_states, value_states = torch.split(qkv_states, self.head_dim, dim=3)

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

        # calculated (K.T @ V) and saved as cache
        attn_weights_inter = None
        if past_key_values is not None:
            attn_weights_inter = past_key_values.get_linear_cache(self.layer_idx)

        if attn_weights_inter is None:
            attn_weights_inter = torch.zeros(batch_size, self.num_attention_heads, self.head_dim, self.head_dim).to(
                value_states
            )

            # apply attention_mask
            if attention_mask is not None:
                attention_mask = attention_mask.to(dtype=torch.bool)  # Ensure it's a boolean tensor
                value_states = value_states.masked_fill(~attention_mask.unsqueeze(1).unsqueeze(-1), 0)

            attn_output = []
            for i in range(num_blocks):
                start_idx = i * self.block_size
                end_idx = min(start_idx + self.block_size, seq_len)
                current_block_size = end_idx - start_idx

                current_query_states = query_states[:, :, start_idx:end_idx]
                current_key_states = key_states[:, :, start_idx:end_idx]
                current_value_states = value_states[:, :, start_idx:end_idx]

                current_query_decay = self.query_decay[:, :current_block_size]
                current_key_decay = self.key_decay[:, -current_block_size:]
                current_diagonal_decay = self.diagonal_decay[:, :, :current_block_size, :current_block_size]
                block_decay = torch.exp(-self.slope_rate * current_block_size)

                # intra: ( Q @ K.T ) @ V -> QK * V
                attn_weights_intra = torch.matmul(current_query_states, current_key_states.transpose(-1, -2))
                attn_output_intra = torch.matmul(attn_weights_intra * current_diagonal_decay, current_value_states)

                # inter: Q @ ( K.T @ V ) -> Q * KV
                attn_output_inter = torch.matmul(current_query_states * current_query_decay, attn_weights_inter)

                # final attention output
                current_attn_output = attn_output_inter + attn_output_intra
                attn_output.append(current_attn_output)

                # calculate attn_weights_inter for next block or cache
                next_attn_weights_inter = torch.matmul(
                    (current_key_states * current_key_decay).transpose(-1, -2), current_value_states
                )
                attn_weights_inter = attn_weights_inter * block_decay + next_attn_weights_inter

        else:
            ratio = torch.exp(-self.slope_rate)
            attn_output = []
            for i in range(seq_len):
                current_query_states = query_states[:, :, i : i + 1]
                current_key_states = key_states[:, :, i : i + 1]
                current_value_states = value_states[:, :, i : i + 1]

                current_attn_weights_inter = torch.matmul(current_key_states.transpose(-1, -2), current_value_states)
                attn_weights_inter = ratio * attn_weights_inter + current_attn_weights_inter
                current_attn_output = torch.matmul(current_query_states, attn_weights_inter)

                attn_output.append(current_attn_output)

        # concatenate attention outputs over all blocks
        attn_output = torch.cat(attn_output, dim=-2)

        # final output projection
        attn_output = attn_output.transpose(1, 2)
        attn_output = attn_output.reshape(batch_size, seq_len, self.num_attention_heads * self.head_dim)
        attn_output = self.norm(attn_output)
        attn_output = F.sigmoid(self.output_gate(hidden_states)) * attn_output
        attn_output = self.out_proj(attn_output)

        # update cache
        if past_key_values is not None:
            past_key_values.set_linear_cache(self.layer_idx, attn_weights_inter)

        return attn_output, attn_weights_inter


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

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


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

    def __init__(self, config: MiniMaxConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", None) or 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=False)
        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor | None,
        past_key_values: Cache | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        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,
            sliding_window=getattr(self.config, "sliding_window", None),  # main diff with Llama
            **kwargs,
        )

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


class MiniMaxTopKRouter(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.top_k = config.num_experts_per_tok
        self.num_experts = config.num_local_experts
        self.hidden_dim = config.hidden_size
        self.weight = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim))

    def forward(self, hidden_states):
        hidden_states = hidden_states.reshape(-1, self.hidden_dim)
        router_logits = F.linear(hidden_states, self.weight)  # (seq_len, num_experts)
        router_logits = torch.nn.functional.softmax(router_logits.float(), dim=-1)
        router_top_value, router_indices = torch.topk(router_logits, self.top_k, dim=-1)  # (seq_len, top_k)
        router_top_value /= router_top_value.sum(dim=-1, keepdim=True)
        router_scores = router_top_value
        return router_logits, router_scores, router_indices


@use_experts_implementation
class MiniMaxExperts(nn.Module):
    """Collection of expert weights stored as 3D tensors."""

    def __init__(self, config: MiniMaxConfig):
        super().__init__()
        self.num_experts = config.num_local_experts
        self.hidden_dim = config.hidden_size
        self.intermediate_dim = config.intermediate_size
        self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, 2 * self.intermediate_dim, self.hidden_dim))
        self.down_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_dim, self.intermediate_dim))
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(
        self,
        hidden_states: torch.Tensor,
        top_k_index: torch.Tensor,
        top_k_weights: torch.Tensor,
    ) -> torch.Tensor:
        final_hidden_states = torch.zeros_like(hidden_states)
        with torch.no_grad():
            expert_mask = torch.nn.functional.one_hot(top_k_index, num_classes=self.num_experts)
            expert_mask = expert_mask.permute(2, 1, 0)
            expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()

        for expert_idx in expert_hit:
            expert_idx = expert_idx[0]
            if expert_idx == self.num_experts:
                continue
            top_k_pos, token_idx = torch.where(expert_mask[expert_idx])
            current_state = hidden_states[token_idx]
            gate, up = nn.functional.linear(current_state, self.gate_up_proj[expert_idx]).chunk(2, dim=-1)
            current_hidden_states = self.act_fn(gate) * up
            current_hidden_states = nn.functional.linear(current_hidden_states, self.down_proj[expert_idx])
            current_hidden_states = current_hidden_states * top_k_weights[token_idx, top_k_pos, None]
            final_hidden_states.index_add_(0, token_idx, current_hidden_states.to(final_hidden_states.dtype))

        return final_hidden_states


class MiniMaxSparseMoeBlock(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.top_k = config.num_experts_per_tok
        self.jitter_noise = config.router_jitter_noise
        self.gate = MiniMaxTopKRouter(config)
        self.experts = MiniMaxExperts(config)

    def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        if self.training and self.jitter_noise > 0:
            hidden_states *= torch.empty_like(hidden_states).uniform_(1.0 - self.jitter_noise, 1.0 + self.jitter_noise)
        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
        _, top_k_weights, top_k_index = self.gate(hidden_states)
        hidden_states = self.experts(hidden_states, top_k_index, top_k_weights)
        hidden_states = hidden_states.reshape(batch_size, sequence_length, hidden_dim)
        return hidden_states


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

        self.self_attn = MiniMaxAttention(config, layer_idx)
        self.input_layernorm = MiniMaxRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = MiniMaxRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        self.layer_idx = layer_idx
        self.layer_type = config.layer_types[layer_idx] if hasattr(config, "layer_types") else None
        self.mlp_alpha_factor = config.mlp_alpha_factor
        self.mlp_beta_factor = config.mlp_beta_factor
        self.mlp = MiniMaxSparseMoeBlock(config)
        if self.layer_type == "linear_attention":
            self.self_attn = MiniMaxLightningAttention(config, layer_idx)
            self.attn_alpha_factor = config.linear_attn_alpha_factor
            self.attn_beta_factor = config.linear_attn_beta_factor
        else:
            self.self_attn = MiniMaxAttention(config, layer_idx)
            self.attn_alpha_factor = config.full_attn_alpha_factor
            self.attn_beta_factor = config.full_attn_beta_factor

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
        hidden_states = self.input_layernorm(hidden_states)
        residual = hidden_states
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            **kwargs,
        )
        hidden_states = residual * self.attn_alpha_factor + hidden_states * self.attn_beta_factor
        hidden_states = self.post_attention_layernorm(hidden_states)
        residual = hidden_states
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual * self.mlp_alpha_factor + hidden_states * self.mlp_beta_factor

        return hidden_states


@auto_docstring
class MiniMaxPreTrainedModel(PreTrainedModel):
    config: MiniMaxConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["MiniMaxDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _can_compile_fullgraph = False  # uses a non-compilable custom cache class MiniMaxCache
    _supports_attention_backend = True
    _can_record_outputs = {
        "router_logits": OutputRecorder(MiniMaxTopKRouter, layer_name="mlp.gate", index=0),
        "hidden_states": MiniMaxDecoderLayer,
        "attentions": [MiniMaxAttention, MiniMaxLightningAttention],
    }

    @torch.no_grad()
    def _init_weights(self, module):
        super()._init_weights(module)
        std = self.config.initializer_range
        if isinstance(module, MiniMaxExperts):
            init.normal_(module.gate_up_proj, mean=0.0, std=std)
            init.normal_(module.down_proj, mean=0.0, std=std)
        elif isinstance(module, MiniMaxTopKRouter):
            init.normal_(module.weight, mean=0.0, std=std)
        if isinstance(module, MiniMaxLightningAttention):
            slope_rate = module.get_slope_rate()
            query_decay, key_decay, diagonal_decay = module.decay_factors(slope_rate)
            init.copy_(module.slope_rate, slope_rate)
            init.copy_(module.query_decay, query_decay)
            init.copy_(module.key_decay, key_decay)
            init.copy_(module.diagonal_decay, diagonal_decay)


@auto_docstring
class MiniMaxModel(MiniMaxPreTrainedModel):
    def __init__(self, config: MiniMaxConfig):
        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(
            [MiniMaxDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = MiniMaxRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = MiniMaxRotaryEmbedding(config=config)
        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: MiniMaxCache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple | MoeModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if use_cache and past_key_values is None:
            past_key_values = MiniMaxCache()
        elif use_cache and not isinstance(past_key_values, MiniMaxCache):
            raise ValueError(
                f"MiniMax uses cache of its own and is not compatible with `past_key_values` of type {type(past_key_values)}."
            )

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

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

        mask_function = create_causal_mask if self.config.sliding_window is None else create_sliding_window_causal_mask
        causal_mask = mask_function(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            position_ids=position_ids,
        )

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

        for i, decoder_layer in enumerate(self.layers):
            if self.config.layer_types[i] == "full_attention":
                input_attention_mask = causal_mask
            else:
                # lightning attention uses original attention_mask, and uses it only for the first step
                input_attention_mask = attention_mask

            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=input_attention_mask,
                position_embeddings=position_embeddings,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)

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


def load_balancing_loss_func(
    gate_logits: torch.Tensor | tuple[torch.Tensor] | None,
    num_experts: int | None = None,
    top_k=2,
    attention_mask: torch.Tensor | None = None,
) -> torch.Tensor | int:
    r"""
    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.

    See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss
    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
    experts is too unbalanced.

    Args:
        gate_logits:
            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
            shape [batch_size X sequence_length, num_experts].
        num_experts:
            Number of experts
        top_k:
            The number of experts to route per-token, can be also interpreted as the `top-k` routing
            parameter.
        attention_mask (`torch.Tensor`, *optional*):
            The attention_mask used in forward function
            shape [batch_size X sequence_length] if not None.

    Returns:
        The auxiliary loss.
    """
    if gate_logits is None or not isinstance(gate_logits, tuple):
        return 0

    if isinstance(gate_logits, tuple):
        compute_device = gate_logits[0].device
        concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)

    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)

    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)

    expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)

    if attention_mask is None:
        # Compute the percentage of tokens routed to each experts
        tokens_per_expert = torch.mean(expert_mask.float(), dim=0)

        # Compute the average probability of routing to these experts
        router_prob_per_expert = torch.mean(routing_weights, dim=0)
    else:
        batch_size, sequence_length = attention_mask.shape
        num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)

        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
        expert_attention_mask = (
            attention_mask[None, :, :, None, None]
            .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
            .reshape(-1, top_k, num_experts)
            .to(compute_device)
        )

        # Compute the percentage of tokens routed to each experts
        tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
            expert_attention_mask, dim=0
        )

        # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
        router_per_expert_attention_mask = (
            attention_mask[None, :, :, None]
            .expand((num_hidden_layers, batch_size, sequence_length, num_experts))
            .reshape(-1, num_experts)
            .to(compute_device)
        )

        # Compute the average probability of routing to these experts
        router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
            router_per_expert_attention_mask, dim=0
        )

    overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
    return overall_loss * num_experts


@auto_docstring
class MiniMaxForCausalLM(MiniMaxPreTrainedModel, 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 = MiniMaxModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.router_aux_loss_coef = config.router_aux_loss_coef
        self.num_experts = config.num_local_experts
        self.num_experts_per_tok = config.num_experts_per_tok

        # 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,
        output_router_logits: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs: Unpack[TransformersKwargs],
    ) -> MoeCausalLMOutputWithPast:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Example:

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

        >>> model = MiniMaxForCausalLM.from_pretrained("MiniMaxAI/MiniMax-Text-01-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("MiniMaxAI/MiniMax-Text-01-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."
        ```"""

        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs: MoeModelOutputWithPast = 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,
            output_router_logits=output_router_logits,
            **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, labels, self.vocab_size, **kwargs)

        aux_loss = None
        if output_router_logits:
            aux_loss = load_balancing_loss_func(
                outputs.router_logits,
                self.num_experts,
                self.num_experts_per_tok,
                attention_mask,
            )
            if labels is not None:
                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device

        return MoeCausalLMOutputWithPast(
            loss=loss,
            aux_loss=aux_loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            router_logits=outputs.router_logits,
        )


class MiniMaxForSequenceClassification(GenericForSequenceClassification, MiniMaxPreTrainedModel):
    pass


class MiniMaxForTokenClassification(GenericForTokenClassification, MiniMaxPreTrainedModel):
    pass


class MiniMaxForQuestionAnswering(GenericForQuestionAnswering, MiniMaxPreTrainedModel):
    pass


__all__ = [
    "MiniMaxPreTrainedModel",
    "MiniMaxModel",
    "MiniMaxForCausalLM",
    "MiniMaxForSequenceClassification",
    "MiniMaxForTokenClassification",
    "MiniMaxForQuestionAnswering",
]