modeling_dense_gau_retnet.py

# coding=utf-8
# Copyright 2022 EleutherAI and the Huggingface Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# 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 DenseGAU RetNet model."""
from typing import List, Optional, Tuple, Union
import math
import torch
import torch.utils.checkpoint
import torch.nn.functional as F
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, \
    SequenceClassifierOutputWithPast
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, \
    replace_return_docstrings
from transformers import top_k_top_p_filtering
from transformers.generation.configuration_utils import GenerationConfig
from .configuration_dense_gau_retnet import DenseGauRetNetConfig

logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "DenseGauRetNetConfig"


# Copied from transformers.models.bart.modeling_bart._make_causal_mask
def _make_causal_mask(
        input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
):
    """
    Make causal mask used for bi-directional self-attention.
    """
    bsz, tgt_len = input_ids_shape
    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
    mask_cond = torch.arange(mask.size(-1), device=device)
    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
    mask = mask.to(dtype)

    if past_key_values_length > 0:
        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype, device=device), mask], dim=-1)
    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)


# Copied from transformers.models.bart.modeling_bart._expand_mask
def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
    """
    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
    """
    bsz, src_len = mask.size()
    tgt_len = tgt_len if tgt_len is not None else src_len

    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)

    inverted_mask = 1.0 - expanded_mask

    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)

# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm
class DenseGauRetNetRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

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

        return (self.weight * hidden_states).to(input_dtype)

# added for retention
# Copied from https://github.com/microsoft/torchscale/blob/main/torchscale/component/multiscale_retention.py
def rotate_every_two(x):
    x1 = x[:, :, :, ::2]
    x2 = x[:, :, :, 1::2]
    x = torch.stack((-x2, x1), dim=-1)
    return x.flatten(-2)  # in einsum notation: rearrange(x, '... d j -> ... (d j)')\
def theta_shift(x, sin, cos):
    return (x * cos) + (rotate_every_two(x) * sin)


#Parameter efficient HiddenProjection
class HiddenProjection(nn.Module):
    def __init__(self, input_dim, mid_reduction_ratio=16, final_reduction_ratio=4):
        super(HiddenProjection, self).__init__()
        self.fc1 = nn.Linear(input_dim, input_dim // mid_reduction_ratio, bias=False)
        self.fc2 = nn.Linear(input_dim // mid_reduction_ratio, int(input_dim // final_reduction_ratio), bias=False)

    def forward(self, x):
        fc1_output = F.silu(self.fc1(x))
        fc2_output = self.fc2(fc1_output)
        return fc2_output

# Copied and modified from transformers.models.bart.modeling_bart._expand_mask

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

    def __init__(self, config: DenseGauRetNetConfig):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.query_key_dim = config.query_key_dim
        self.num_heads = config.num_attention_heads
        self.factor = config.v_factor
        self.head_dim = config.hidden_size * self.factor // self.num_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.q_proj = nn.Linear(self.hidden_size, self.query_key_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.query_key_dim, bias=False)
        self.key_dim = self.query_key_dim // self.num_heads
        self.scaling = self.key_dim ** -0.5
        self.expansion_dim = int(config.hidden_size * self.factor)
        self.group_norm = DenseGauRetNetRMSNorm(self.expansion_dim // config.num_attention_heads, eps=config.rms_norm_eps)
        self.to_hidden = nn.Sequential(
            nn.Linear(config.hidden_size, self.expansion_dim * 2, bias=False),
            nn.SiLU()
        )
        self.to_out = nn.Sequential(
            nn.Linear(self.expansion_dim, config.hidden_size, bias=False),
            nn.Dropout(0)
        )
        self.config = config
        self.k_select = HiddenProjection(self.hidden_size, config.intermediate_k_select_scale, 2)
        self.v_select = HiddenProjection(self.hidden_size, config.intermediate_v_select_scale, 0.5)
        self.k_norm = DenseGauRetNetRMSNorm(self.query_key_dim, eps=config.rms_norm_eps)
        self.v_norm = DenseGauRetNetRMSNorm(self.expansion_dim, eps=config.rms_norm_eps)
        if config.deepnorm:
            self.alpha = math.pow(2.0 * config.num_hidden_layers, 0.25)
        else:
            self.alpha = 1.0
        self.dropout_module = torch.nn.Dropout(config.dropout)
        self.reset_parameters()

    #
    def reset_parameters(self):
        nn.init.xavier_uniform_(self.q_proj.weight, gain=2 ** -2.5)
        nn.init.xavier_uniform_(self.k_proj.weight, gain=2 ** -2.5)
        nn.init.xavier_uniform_(self.k_select.fc1.weight, gain=2 ** -2.5)
        nn.init.xavier_uniform_(self.k_select.fc2.weight, gain=2 ** -2.5)
        nn.init.xavier_uniform_(self.v_select.fc1.weight, gain=2 ** -2.5)
        nn.init.xavier_uniform_(self.v_select.fc2.weight, gain=2 ** -2.5)
        for module in self.to_out.modules():
            if isinstance(module, nn.Linear):
                nn.init.xavier_uniform_(module.weight, gain=2 ** -1)
        for module in self.to_hidden.modules():
            if isinstance(module, nn.Linear):
                nn.init.xavier_uniform_(module.weight, gain=2 ** -2.5)

    def forward(
            self,
            forward_impl: 'parallel',
            hidden_states: torch.Tensor,
            rel_pos,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[Tuple[torch.Tensor]] = None,
            output_attentions: bool = False,
            k_features=None,  # dense
            v_features=None,  # dense
            dense=False,
            dense_layers=0,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        bsz, tgt_len, _ = hidden_states.size()
        (sin, cos), inner_mask = rel_pos
        x = hidden_states
        q = F.silu(self.q_proj(x))
        k = F.silu(self.k_proj(x))
        v, gate = self.to_hidden(x).chunk(2, dim=-1)

        k *= self.scaling
        k_curr = k
        v_curr = v

        if dense:
            k_gate = self.k_select(hidden_states.clone())
            for i, k_past in enumerate(k_features):
                k = k.clone() + F.silu(k_gate) * k_past
            k = self.k_norm(k)

            v_gate = self.v_select(hidden_states.clone())
            for i, v_past in enumerate(v_features):
                v = v.clone() + F.silu(v_gate) * v_past
            v = self.v_norm(v)

        q = q.view(bsz, tgt_len, self.num_heads, self.key_dim).transpose(1, 2)
        k = k.view(bsz, tgt_len, self.num_heads, self.key_dim).transpose(1, 2)
        qr = theta_shift(q, sin, cos)
        kr = theta_shift(k, sin, cos)

        if forward_impl == 'parallel':
            output = self.parallel_forward(qr, kr, v, inner_mask)
        elif forward_impl == 'recurrent':
            output, past_key_value = self.recurrent_forward(qr, kr, v, inner_mask, past_key_value=past_key_value)

        output = self.group_norm(output)
        output = output.reshape(bsz, tgt_len, self.expansion_dim) * gate  # gate
        output = self.to_out(output)
        output = self.dropout_module(output)

        return output, past_key_value, k_curr, v_curr

    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()

    # retntion parallel forward
    def recurrent_forward(
            self,
            qr, kr, v,
            decay,
            past_key_value,
    ):
        bsz = v.size(0)

        v = v.view(bsz, self.num_heads, self.head_dim, 1)
        kv = kr * v
        if "prev_key_value" in past_key_value:
            prev_kv = past_key_value["prev_key_value"]
            prev_scale = past_key_value["scale"]
            scale = prev_scale * decay + 1
            kv = prev_kv * (prev_scale.sqrt() * decay / scale.sqrt()).view(self.num_heads, 1,
                                                                           1) + kv / scale.sqrt().view(self.num_heads,
                                                                                                       1, 1)
        else:
            scale = torch.ones_like(decay)

        past_key_value["prev_key_value"] = kv
        past_key_value["scale"] = scale

        output = torch.sum(qr * kv, dim=3)
        return output, past_key_value

    def parallel_forward(self, qr, kr, v, mask):
        bsz, tgt_len, embed_dim = v.size()

        vr = v.view(bsz, tgt_len, self.num_heads, self.expansion_dim // self.num_heads).transpose(1, 2)

        qk_mat = qr @ kr.transpose(-1, -2)  # bsz * m * tgt_len * tgt_len
        qk_mat = qk_mat * mask
        # invariant after normalization
        qk_mat = qk_mat / qk_mat.detach().abs().sum(dim=-1, keepdim=True).clamp(min=1, max=5e4)

        output = torch.matmul(qk_mat, vr)
        output = output.transpose(1, 2)
        return output


class DenseGauRetNetDecoderLayer(nn.Module):
    def __init__(self, config: DenseGauRetNetConfig):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = MultiScaleGauRetention(config=config)
        self.input_layernorm = DenseGauRetNetRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        if config.deepnorm:
            self.alpha = math.pow(2.0 * config.num_hidden_layers, 0.25)
        else:
            self.alpha = 1.0

    def forward(
            self,
            hidden_states: torch.Tensor,
            rel_pos=None,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[Tuple[torch.Tensor]] = None,
            output_attentions: Optional[bool] = False,
            k_features: Optional[List] = [],
            v_features: Optional[List] = [],
            dense=False,
            dense_layers=0,
            forward_impl: str = 'parallel',
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:

        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states, past_key_value, k_curr, v_curr = self.self_attn(

            hidden_states=hidden_states,
            rel_pos=rel_pos,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            k_features=k_features,
            v_features=v_features,
            dense=dense,
            dense_layers=dense_layers,
            forward_impl=forward_impl,

        )
        hidden_states = residual + hidden_states

        return hidden_states, past_key_value, k_curr, v_curr


DenseGauRetNet_START_DOCSTRING = r"""
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a Pytorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular Pytorch Module and refer to the Pytorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`DenseGauRetNetConfig`]):
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""


@add_start_docstrings(
    "The bare DenseGauRetNet Model outputting raw hidden-states without any specific head on top.",
    DenseGauRetNet_START_DOCSTRING,
)
class DenseGauRetNetPreTrainedModel(PreTrainedModel):
    config_class = DenseGauRetNetConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["DenseGauRetNetDecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]

    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_()

    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, DenseGauRetNetModel):
            module.gradient_checkpointing = value


DenseGauRetNet_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 `decoder_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 (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
            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)`) and 2 additional tensors of shape
            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.

            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.

            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
            `decoder_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.
"""


class RetNetRelPos(nn.Module):
    def __init__(self, decoder_embed_dim, decoder_retention_heads, query_key_dim):
        super().__init__()
        angle = 1.0 / (10000 ** torch.linspace(0, 1, (query_key_dim // decoder_retention_heads) // 2))
        angle = angle.unsqueeze(-1).repeat(1, 2).flatten()
        decay = torch.log(1 - 2 ** (-5 - torch.arange(decoder_retention_heads, dtype=torch.float)))
        self.register_buffer("angle", angle)
        self.register_buffer("decay", decay)

    def forward(self, slen, activate_recurrent=False):
        if activate_recurrent:
            sin = torch.sin(self.angle * (slen - 1))
            cos = torch.cos(self.angle * (slen - 1))
            retention_rel_pos = ((sin, cos), self.decay.exp())
        else:
            index = torch.arange(slen).to(self.decay)
            sin = torch.sin(index[:, None] * self.angle[None, :])
            cos = torch.cos(index[:, None] * self.angle[None, :])
            mask = torch.tril(torch.ones(slen, slen).to(self.decay))
            mask = torch.masked_fill(index[:, None] - index[None, :], ~mask.bool(), float("inf"))
            mask = torch.exp(mask * self.decay[:, None, None])
            mask = torch.nan_to_num(mask)
            mask = mask / mask.sum(dim=-1, keepdim=True).sqrt()
            retention_rel_pos = ((sin, cos), mask)

        return retention_rel_pos


@add_start_docstrings(
    "The bare DenseGauRetNet Model outputting raw hidden-states without any specific head on top.",
    DenseGauRetNet_START_DOCSTRING,
)
class DenseGauRetNetModel(DenseGauRetNetPreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`DenseGauRetNetDecoderLayer`]

    Args:
        config: DenseGauRetNetConfig
    """

    def __init__(self, config: DenseGauRetNetConfig):
        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([DenseGauRetNetDecoderLayer(config) for _ in range(config.num_hidden_layers)])
        self.norm = DenseGauRetNetRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()
        self.retnet_rel_pos = RetNetRelPos(config.hidden_size, config.num_attention_heads,
                                           config.query_key_dim)

        if config.deepnorm:
            init_scale = math.pow(8.0 * config.num_hidden_layers, 0.25)
            for name, p in self.named_parameters():

                if (
                        "fc1" in name
                        or "fc2" in name
                        or "gate_proj" in name
                        or "down_proj" in name
                        or "up_proj" in name
                        or "out_proj" in name
                        or "v_proj" in name
                        or "to_hidden" in name
                        or "to_output" in name

                ):
                    p.data.div_(init_scale)

    def get_input_embeddings(self):
        return self.embed_tokens

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

    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
        # create causal mask
        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
        combined_attention_mask = None
        if input_shape[-1] > 1:
            combined_attention_mask = _make_causal_mask(
                input_shape,
                inputs_embeds.dtype,
                device=inputs_embeds.device,
                past_key_values_length=past_key_values_length,
            )

        if attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
                inputs_embeds.device
            )
            combined_attention_mask = (
                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
            )

        return combined_attention_mask

    def is_first_step(self, incremental_state):
        if incremental_state is None:
            return False
        return incremental_state.get("is_first_step", False)

    @add_start_docstrings_to_model_forward(DenseGauRetNet_INPUTS_DOCSTRING)
    def forward(
            self,
            forward_impl: Optional[str] = 'parallel',
            input_ids: torch.LongTensor = None,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_values: Optional[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,
            sequence_offset=0,
    ) -> 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

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
        elif input_ids is not None:
            batch_size, seq_length = input_ids.shape
        elif inputs_embeds is not None:
            batch_size, seq_length, _ = inputs_embeds.shape
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

        seq_length_with_past = seq_length
        past_key_values_length = 0

        if position_ids is None:
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(
                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
            )
            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
        else:
            position_ids = position_ids.view(-1, seq_length).long()

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)
        # embed positions
        if attention_mask is None:
            attention_mask = torch.ones(
                (batch_size, seq_length_with_past), dtype=torch.bool, device=inputs_embeds.device
            )
        attention_mask = self._prepare_decoder_attention_mask(
            attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
        )

        hidden_states = inputs_embeds

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

        # decoder layer
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        next_decoder_cache = [] if use_cache else None

        #
        k_features = []
        v_features = []
        dense_layers = 0
        for idx, decoder_layer in enumerate(self.layers):
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            past_key_value = past_key_values[idx] if past_key_values is not None and len(
                past_key_values) != 0 else {}

            slen = input_ids.size(1)
            if forward_impl == 'recurrent':
                slen = sequence_offset
            rel_pos = self.retnet_rel_pos(slen, forward_impl == 'recurrent',)

            if self.gradient_checkpointing and self.training:

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs, output_attentions, None)

                    return custom_forward

                hidden_states = layer_outputslayer_outputs = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(decoder_layer),
                    hidden_states,
                    attention_mask,
                    position_ids,
                    None,
                )
            else:
                dense = False
                if idx >= 1:
                    dense = True

                layer_outputs, past_key_value, k_curr, v_curr = decoder_layer(
                    hidden_states,
                    rel_pos,
                    forward_impl=forward_impl,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_value,
                    output_attentions=output_attentions,
                    k_features=k_features,
                    v_features=v_features,
                    dense=dense,
                    dense_layers=dense_layers,
                )
                dense_layers += 1
                k_features.append(k_curr)
                v_features.append(v_curr)
                if len(k_features) > self.config.dense_block_layers:
                    k_features.pop(0)
                if len(v_features) > self.config.dense_block_layers:
                    v_features.pop(0)

            hidden_states = layer_outputs  # used to be 3 ele,tmp  1


            if use_cache:
                next_decoder_cache.append(past_key_value)

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

        hidden_states = self.norm(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 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,
        )


class DenseGauRetNetForCausalLM(DenseGauRetNetPreTrainedModel):
    _auto_class = "AutoModelForCausalLM"
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = DenseGauRetNetModel(config)

        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.embed_tokens

    def set_input_embeddings(self, value):
        self.model.embed_tokens = 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(DenseGauRetNet_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[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,
            forward_impl: str = 'parallel',
            sequence_offset=0,

    ) -> Union[Tuple, CausalLMOutputWithPast]:
        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 outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            forward_impl=forward_impl,
            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,
            sequence_offset=sequence_offset,
        )

        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Enable model parallelism
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

        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,
        )

    def prepare_inputs_for_generation(
            self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
    ):
        if past_key_values:
            input_ids = input_ids[:, -1:]

        position_ids = kwargs.get("position_ids", None)
        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past_key_values:
                position_ids = position_ids[:, -1].unsqueeze(-1)

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
            }
        )
        return model_inputs

    @staticmethod
    def _reorder_cache(past_key_values, beam_idx):
        reordered_past = ()
        for layer_past in past_key_values:
            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
        return reordered_past

    # infer mode

    def sample_token(self, logit, do_sample=False, top_k=1, top_p=1.0, temperature=1.0):
        if not do_sample:
            return torch.argmax(logit, dim=-1, keepdim=True)
        filtered = top_k_top_p_filtering(logit / temperature, top_k=top_k, top_p=top_p)
        return torch.multinomial(torch.softmax(filtered, dim=-1), num_samples=1)

    @torch.inference_mode()
    def generate(
            self,
            input_ids: Optional[torch.Tensor] = None,
            parallel_compute_prompt=False,
            generation_config: Optional[GenerationConfig] = None,
            **kwargs,
    ):
        # breakpoint()
        past_key_values = {}
        for p_i in range(input_ids.shape[1] - 1):
            outputs = self(input_ids[:, p_i:p_i + 1],
                           forward_impl='recurrent',
                           past_key_values=past_key_values,
                           sequence_offset=p_i,
                           return_dict=True,
                           use_cache=True)
            past_key_values = outputs.past_key_values
        generated = input_ids[:, -1].unsqueeze(-1)  # [B, 1]
        for i in range(generation_config.max_new_tokens):
            outputs = self(generated[:,-1:],
                           forward_impl='recurrent',
                           past_key_values=past_key_values,
                           use_cache=True,
                           return_dict=True,
                           sequence_offset=input_ids.shape[-1]+generated.shape[-1]-2  #1
                           )
            logit = outputs.logits[:, -1, :] # [batch_size, vocab_size]
            past_key_values = outputs.past_key_values
            token = self.sample_token(logit,
                                      do_sample=generation_config.do_sample,
                                      temperature=generation_config.temperature)

            generated = torch.cat([generated, token], dim=-1)
        return generated[:,1:]


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

    [`DenseGauRetNetForSequenceClassification`] 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).
    """,
    DenseGauRetNet_START_DOCSTRING,
)
class DenseGauRetNetForSequenceClassification(DenseGauRetNetPreTrainedModel):
    _keys_to_ignore_on_load_missing = [r"lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.model = DenseGauRetNetModel(config)
        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)

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

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

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

    @add_start_docstrings_to_model_forward(DenseGauRetNet_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[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

        transformer_outputs = self.model(
            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 = transformer_outputs[0]
        logits = self.score(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:
                sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).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:
            labels = labels.to(logits.device)
            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.num_labels == 1:
                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(pooled_logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(pooled_logits, labels)
        if not return_dict:
            output = (pooled_logits,) + transformer_outputs[1:]
            return ((loss,) + output) if loss is not None else output

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