modeling_doge.py

# coding=utf-8
# Copyright 2024 Jingze Shi and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on the Wonderful Matrices paper implementation.
#
#     https://arxiv.org/abs/2412.11834
#
# 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.
"""PyTorch Doge model."""

import math
from typing import List, Optional, Tuple, Union

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

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache, StaticCache
from transformers.generation import GenerationMixin
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
    SequenceClassifierOutputWithPast,
)
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)
from .configuration_doge import DogeConfig

try:
    from einx import add as einx_add
except ImportError:
    einx_add = None


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "DogeConfig"


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

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

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


class Residual(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
    
    def forward(self, residual_states, hidden_states):
        return self.weight * residual_states + hidden_states

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}"


class RotaryEmbedding(nn.Module):
    def __init__(self, config: Optional[DogeConfig] = None):
        super().__init__()
        self.rope_kwargs = {}

        if config.rope_scaling is None:
            self.rope_type = "default"
        else:
            self.rope_type = config.rope_scaling
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings
        self.base = config.rope_theta

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, **self.rope_kwargs)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    def _dynamic_frequency_update(self, position_ids, device):
        """
        dynamic RoPE layers should recompute `inv_freq` in the following situations:
        1 - growing beyond the cached sequence length (allow scaling)
        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
        """
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.max_seq_len_cached:  # growth
            inv_freq, self.attention_scaling = self.rope_init_fn(
                self.config, device, seq_len=seq_len, **self.rope_kwargs
            )
            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
            self.max_seq_len_cached = seq_len

        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
            self.max_seq_len_cached = self.original_max_seq_len

    @torch.no_grad()
    def forward(self, x, position_ids):
        if "dynamic" in self.rope_type:
            self._dynamic_frequency_update(position_ids, device=x.device)

        # core RoPE block
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        position_ids_expanded = position_ids[:, None, :].float()
        device_type = x.device.type
        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()

        cos = cos * self.attention_scaling
        sin = 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)


def apply_QK_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

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


class DogeDynamicMaskAttention(nn.Module):
    """Dynamic Mask Attention from 'Wonderful Matrices' paper."""

    def __init__(self, config: DogeConfig, layer_idx: Optional[int] = None):
        super().__init__()

        self.config = config
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.hidden_dim = config.hidden_size
        self.num_attention_heads = config.num_attention_heads
        self.attention_dropout = config.attention_dropout
        self.attention_head_dim = self.hidden_dim // self.num_attention_heads

        # Q K V O projections
        self.q_proj = nn.Linear(
            self.hidden_dim,
            self.num_attention_heads * self.attention_head_dim,
            bias=config.hidden_bias,
        )
        self.k_proj = nn.Linear(
            self.hidden_dim,
            self.num_attention_heads * self.attention_head_dim,
            bias=config.hidden_bias,
        )
        # dynamic mask for the QK^T attention score matrix
        self.A = nn.Parameter(
            torch.ones(self.num_attention_heads)
        )
        self.dt_proj = nn.Linear(
            self.hidden_dim,
            self.num_attention_heads,
            bias=config.hidden_bias,
        )
        self.v_proj = nn.Linear(
            self.hidden_dim,
            self.num_attention_heads * self.attention_head_dim,
            bias=config.hidden_bias,
        )
        self.o_proj = nn.Linear(
            self.hidden_dim,
            self.hidden_dim,
            bias=config.hidden_bias,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[Cache]]:
        bsz, q_len, _ = hidden_states.shape

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

        query_states = query_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_dim).transpose(
            1, 2
        )
        key_states = key_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_dim).transpose(
            1, 2
        )
        value_states = value_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_dim).transpose(
            1, 2
        )

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

        if past_key_value is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        # compute attention scores matrix
        attn_weights = torch.matmul(query_states, key_states.transpose(-1, -2)) / math.sqrt(self.attention_head_dim)

        # add mask to attention scores
        if attention_mask is not None:
            dt_states = self.dt_proj(value_states.transpose(1, 2).reshape(bsz, value_states.shape[-2], -1))
            dynamic_mask = torch.exp(self.A * F.softplus(dt_states)).transpose(-1, -2)
            dynamic_mask = dynamic_mask < 1.0
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]].masked_fill(dynamic_mask[:, :, None, :], torch.finfo(hidden_states.dtype).min)
            attn_weights = attn_weights + causal_mask

        # upcast attention scores to fp32
        attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = F.dropout(attn_weights, p=self.attention_dropout, training=self.training)

        # apply attention scores to value states
        attn_output = torch.matmul(attn_weights, value_states)

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, -1)
        attn_output = self.o_proj(attn_output)

        return attn_output, past_key_value


class DogeSdpaDynamicMaskAttn(DogeDynamicMaskAttention):

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[Cache]]:
        bsz, q_len, _ = hidden_states.shape

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

        query_states = query_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_dim).transpose(1, 2)

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

        if past_key_value is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        if attention_mask is not None:
            dt_states = self.dt_proj(value_states.transpose(1, 2).reshape(bsz, value_states.shape[-2], -1))
            dynamic_mask = torch.exp(self.A * F.softplus(dt_states)).transpose(-1, -2)
            dynamic_mask = dynamic_mask < 1.0
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]].masked_fill(dynamic_mask[:, :, None, :], torch.finfo(hidden_states.dtype).min)

        query_states = query_states.contiguous()
        key_states = key_states.contiguous()
        value_states = value_states.contiguous()

        attn_output = F.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=causal_mask,
            dropout_p=self.attention_dropout,
        )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.view(bsz, q_len, -1)
        attn_output = self.o_proj(attn_output)

        return attn_output, past_key_value


DOGE_ATTENTION_CLASSES = {
    "eager": DogeDynamicMaskAttention,
    "sdpa": DogeSdpaDynamicMaskAttn,
}


class DogeMLP(nn.Module):

    def __init__(self, config: DogeConfig):
        super().__init__()
        self.hidden_dim = config.hidden_size
        self.intermediate_dim = config.intermediate_size
        self.act_fn = ACT2FN[config.hidden_act]

        self.gate_proj = nn.Linear(
            self.hidden_dim,
            self.intermediate_dim,
            bias=config.hidden_bias,
        )
        self.up_proj = nn.Linear(
            self.hidden_dim,
            self.intermediate_dim,
            bias=config.hidden_bias,
        )
        self.down_proj = nn.Linear(
            self.intermediate_dim,
            self.hidden_dim,
            bias=config.hidden_bias,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        **kwargs,
    ) -> torch.Tensor:
        hidden_states = self.down_proj(self.act_fn(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))
        return hidden_states


class DogeCDMoE(DogeMLP):
    """Cross Domain Mixture of Experts from 'Wonderful Matrices' paper."""

    def __init__(self, config: DogeConfig):
        super().__init__(config)
        self.hidden_dim = config.hidden_size
        self.act_fn = ACT2FN[config.hidden_act]

        self.expert_retrieval_dim = config.expert_retrieval_size
        self.num_cdmmoe_experts = config.num_cdmmoe_experts
        self.num_cdmmoe_heads = config.num_cdmmoe_heads
        self.num_cdmmoe_experts_per_head = config.num_cdmmoe_experts_per_head
        self.num_keys = int(math.sqrt(self.num_cdmmoe_experts))

        # queries and keys for retrieval experts
        self.queries = nn.Linear(
            self.hidden_dim,
            self.num_cdmmoe_heads * self.expert_retrieval_dim,
            bias=False,
        )
        self.keys = nn.Parameter(
            torch.zeros(
                self.num_cdmmoe_heads,
                self.num_keys,
                2,
                self.expert_retrieval_dim // 2,
            )
        )

        # experts
        self.down_embed  = nn.Embedding(
            self.num_cdmmoe_experts,
            self.hidden_dim,
        )
        self.up_embed = nn.Embedding(
            self.num_cdmmoe_experts,
            self.hidden_dim,
        )
        

    def forward(
        self,
        hidden_states: torch.Tensor,
        **kwargs,
    ) -> torch.Tensor:
        bsz, seq_len, _ = hidden_states.shape

        # get similarity with queries and keys
        queries = self.queries(hidden_states)
        queries = queries.view(bsz, seq_len, 2, self.num_cdmmoe_heads, -1).permute(2, 0, 1, 3, 4)
        sim = torch.einsum("p b t h n, h k p n -> p b t h k", queries, self.keys)

        # get experts with the highest similarity
        (scores_x, scores_y), (indices_x, indices_y) = sim.topk(self.num_cdmmoe_experts_per_head, dim=-1)
        if einx_add is not None:
            all_scores = einx_add("... i, ... j -> ... (i j)", scores_x, scores_y)
            all_indices = einx_add("... i, ... j -> ... (i j)", indices_x * self.num_keys, indices_y)
        else:
            all_scores = scores_x.unsqueeze(-1) + scores_y.unsqueeze(-2)
            all_scores = all_scores.view(*scores_x.shape[:-1], -1)
            all_indices = (indices_x.unsqueeze(-1) * self.num_keys) + indices_y.unsqueeze(-2)
            all_indices = all_indices.view(*indices_x.shape[:-1], -1)
        scores, pk_indices = all_scores.topk(self.num_cdmmoe_experts_per_head, dim=-1)
        indices = all_indices.gather(-1, pk_indices)
        down_embed = self.down_embed(indices)
        up_embed = self.up_embed(indices)

        # mix experts states with cross domain states
        experts_weights = torch.einsum("b t d, b t h k d -> b t h k", hidden_states, down_embed)
        experts_weights = self.act_fn(experts_weights) * scores.softmax(dim=-1)
        experts_states = torch.einsum("b t h k, b t h k d -> b t d", experts_weights, up_embed)
        hidden_states = self.down_proj(self.act_fn(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))
        hidden_states = hidden_states + experts_states
        return hidden_states


class DogeDecoderLayer(nn.Module):
    def __init__(self, config: DogeConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.hidden_dropout = config.hidden_dropout

        self.pre_sequence_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.attn = DOGE_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
        self.post_sequence_residual = Residual(config.hidden_size)

        self.pre_state_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.feed_forward = DogeMLP(config) if config.is_moe == False else DogeCDMoE(config)
        self.post_state_residual = Residual(config.hidden_size)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
        **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        """
        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_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence
            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
        """

        # sequence transformation
        residual = hidden_states
        hidden_states = self.pre_sequence_layernorm(hidden_states)
        hidden_states, present_key_value = self.attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        self_attn_weights = None
        hidden_states = F.dropout(hidden_states, p=self.hidden_dropout, training=self.training)
        hidden_states = self.post_sequence_residual(residual, hidden_states)

        # state transformation
        residual = hidden_states
        hidden_states = self.pre_state_layernorm(hidden_states)
        hidden_states = self.feed_forward(hidden_states)
        hidden_states = F.dropout(hidden_states, p=self.hidden_dropout, training=self.training)
        hidden_states = self.post_state_residual(residual, hidden_states)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


@add_start_docstrings("The bare Doge Model outputting raw hidden-states without any specific head on top.")
class DogePreTrainedModel(PreTrainedModel):
    config_class = DogeConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["DogeDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_sdpa = True
    _supports_cache_class = True
    _supports_quantized_cache = True
    _supports_static_cache = True

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, (nn.Linear)):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()


DOGE_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

            [What are attention masks?](../glossary#attention-mask)

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
            `past_key_values`).

            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
            information on the default strategy.

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.
        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.n_positions - 1]`.

            [What are position IDs?](../glossary#position-ids)
        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.

            Two formats are allowed:
            - a [`~cache_utils.Cache`] instance, see our
            [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
            cache format.

            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
            legacy cache format will be returned.

            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
            of shape `(batch_size, sequence_length)`.
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
        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`).
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
            tensors for more detail.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
            the complete sequence length.
"""


@add_start_docstrings("The bare Doge Model outputting raw hidden-states without any specific head on top.")
class DogeModel(DogePreTrainedModel):
    def __init__(self, config: DogeConfig):
        super().__init__(config)
        self.config = config
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.word_embed = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
        self.rotary_emb = RotaryEmbedding(config)
        self.layers = nn.ModuleList(
            [DogeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.final_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.gradient_checkpointing = False

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

    def get_input_embeddings(self):
        return self.word_embed

    def set_input_embeddings(self, value):
        self.word_embed = value

    @add_start_docstrings_to_model_forward(DOGE_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[Tuple, BaseModelOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You cannot specify both input_ids and inputs_embeds")

        if self.gradient_checkpointing and self.training and use_cache:
            logger.warning_once(
                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
            )
            use_cache = False

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

        # kept for BC (non `Cache` `past_key_values` inputs)
        return_legacy_cache = False
        if use_cache and not isinstance(past_key_values, Cache):
            return_legacy_cache = True
            if past_key_values is None:
                past_key_values = DynamicCache()
            else:
                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
                logger.warning_once(
                    "We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and "
                    "will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class "
                    "(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)"
                )

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

        causal_mask = self._update_causal_mask(
            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
        )
        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        next_decoder_cache = None

        for decoder_layer in self.layers:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    decoder_layer.__call__,
                    hidden_states,
                    causal_mask,
                    position_ids,
                    past_key_values,
                    output_attentions,
                    use_cache,
                    cache_position,
                    position_embeddings,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=causal_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    cache_position=cache_position,
                    position_embeddings=position_embeddings,
                )

            hidden_states = layer_outputs[0]

            if use_cache:
                next_decoder_cache = layer_outputs[2 if output_attentions else 1]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        hidden_states = self.final_layernorm(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        next_cache = next_decoder_cache if use_cache else None
        if return_legacy_cache:
            next_cache = next_cache.to_legacy_cache()

        if not return_dict:
            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )

    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor = None,
        input_tensor: torch.Tensor = None,
        cache_position: torch.Tensor = None,
        past_key_values: Cache = None,
        output_attentions: bool = False,
    ):
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        using_static_cache = isinstance(past_key_values, StaticCache)

        dtype, device = input_tensor.dtype, input_tensor.device
        sequence_length = input_tensor.shape[1]
        if using_static_cache:
            target_length = past_key_values.get_max_cache_shape()
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        # in case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask=attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            device=device,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
        )

        return causal_mask
    
    @staticmethod
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor = None,
        sequence_length: int = None,
        target_length: int = None,
        dtype: torch.dtype = None,
        device: torch.device = None,
        cache_position: torch.Tensor = None,
        batch_size: int = None,
        **kwargs,
    ):
        """
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
                `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache,
                to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            device (`torch.device`):
                The device to plcae the 4D attention mask on.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        """
        if attention_mask is not None and attention_mask.dim() == 4:
            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
            causal_mask = attention_mask
        else:
            min_dtype = torch.finfo(dtype).min
            causal_mask = torch.full(
                (sequence_length, target_length),
                fill_value=min_dtype, dtype=dtype, device=device,
            )
            if sequence_length != 1:
                causal_mask = torch.triu(causal_mask, diagonal=1)
            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                mask_length = attention_mask.shape[-1]
                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
                padding_mask = padding_mask == 0
                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                    padding_mask, min_dtype
                )

        return causal_mask


class DogeForCausalLM(DogePreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config: DogeConfig):
        super().__init__(config)
        self.config = config
        self.model = DogeModel(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 get_input_embeddings(self):
        return self.model.word_embed

    def set_input_embeddings(self, value):
        self.model.word_embed = value

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model

    @add_start_docstrings_to_model_forward(DOGE_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        num_logits_to_keep: int = 0,
        **loss_kwargs,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        r"""
        Args:
            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]`.

            num_logits_to_keep (`int`, *optional*):
                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.

        Returns:
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # decoder output consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = 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_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )

        hidden_states = outputs[0]

        # only compute necessary logits, and do not upcast them to float if we are not computing the loss
        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.vocab_size, **loss_kwargs)

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@add_start_docstrings(
    """
    The Doge Model transformer with a sequence classification head on top (linear layer).

    [`DogeForSequenceClassification`] uses the last token in order to do the classification, as other causal models
    (e.g. GPT-2) do.

    Since it does classification on the last token, it requires to know the position of the last token. If a
    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
    each row of the batch).
    """
)
class DogeForSequenceClassification(DogePreTrainedModel):
    def __init__(self, config: DogeConfig):
        super().__init__(config)
        self.config = config
        self.num_labels = config.num_labels

        self.model = DogeModel(config)
        self.classifier = nn.Linear(config.hidden_size, self.num_labels, bias=False)

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

    def get_input_embeddings(self):
        return self.model.word_embed

    def set_input_embeddings(self, value):
        self.model.word_embed = value

    @add_start_docstrings_to_model_forward(DOGE_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = 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_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = outputs[0]
        logits = self.classifier(hidden_states)

        if input_ids is not None:
            batch_size = input_ids.shape[0]
        else:
            batch_size = inputs_embeds.shape[0]

        if self.config.pad_token_id is None and batch_size != 1:
            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
        if self.config.pad_token_id is None:
            sequence_lengths = -1
        else:
            if input_ids is not None:
                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
                sequence_lengths = sequence_lengths % input_ids.shape[-1]
                sequence_lengths = sequence_lengths.to(logits.device)
            else:
                sequence_lengths = -1

        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]

        loss = None
        if labels is not None:
            loss = self.loss_function(
                logits=logits,
                labels=labels,
                pooled_logits=pooled_logits,
                config=self.config,
            )

        if not return_dict:
            output = (pooled_logits,) + outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return SequenceClassifierOutputWithPast(
            loss=loss,
            logits=pooled_logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )