modeling_quiet.py

# coding=utf-8
# Copyright 2023 Quiet AI 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 Quiet model."""
import inspect
import math
import copy
import os
import time
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
import wandb
from termcolor import colored
from tqdm import tqdm
import random
import numpy as np
from matplotlib.colors import LinearSegmentedColormap, LogNorm
import warnings
from collections import defaultdict
from typing import List, Optional, Tuple, Union

import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from transformers.generation.utils import GenerationMixin
from transformers.generation.stopping_criteria import StoppingCriteriaList, validate_stopping_criteria
from transformers import TextStreamer

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
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,
    is_flash_attn_2_available,
    is_flash_attn_greater_or_equal_2_10,
    logging,
    replace_return_docstrings,
)
from .configuration_quiet import QuietConfig


if is_flash_attn_2_available():
    from flash_attn import flash_attn_func, flash_attn_varlen_func
    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa

    _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "QuietConfig"

from reportlab.pdfgen import canvas
from reportlab.lib.pagesizes import letter
from reportlab.lib.colors import HexColor


def _prepare_4d_causal_attention_mask_for_sdpa(attention_mask, input_shape, inputs_embeds, past_key_values_length):
    # Compute the attention mask correctly
    bsz, tgt_len = input_shape

    # Create a 4D attention mask from a 2D tensor mask.
    # The shape of the output attention mask is (batch_size, 1, tgt_len, src_len)
    # The values are either 0 or 1, where 0 means padding and 1 means non-padding.
    combined_attention_mask = None
    if attention_mask is not None:
        # What if attention_mask is not None and has a shape of (batch_size, 1, tgt_len, src_len)
        # In this case, we can just use it directly.
        if attention_mask.dim() == 4:
            combined_attention_mask = attention_mask
        # What if attention_mask is not None and has a shape of (batch_size, 1, tgt_len)
        # In this case, we need to expand it to (batch_size, 1, tgt_len, src_len)
        elif attention_mask.dim() == 3:
            expanded_attn_mask = attention_mask[:, None, :, :]
            combined_attention_mask = expanded_attn_mask
        # What if attention_mask is not None and has a shape of (batch_size, tgt_len)
        # In this case, we need to expand it to (batch_size, 1, tgt_len, src_len)
        elif attention_mask.dim() == 2:
            # Provided a padding mask of dimensions [batch_size, seq_length]
            # - if the model is a decoder, apply a causal mask in addition to the padding mask
            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
            if past_key_values_length > 0:
                attention_mask = attention_mask.to(dtype=torch.long)
                attention_mask = attention_mask[:, past_key_values_length:]
            expanded_attn_mask = attention_mask[:, None, None, :]
            combined_attention_mask = expanded_attn_mask
        else:
            raise ValueError(
                "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
                    input_shape, attention_mask.shape
                )
            )

    # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
    # masked positions, this operation will create a tensor which is 0.0 for
    # positions we want to attend and -10000.0 for masked positions.
    # Since we are adding it to the raw scores before the softmax, this is
    # effectively the same as removing these entirely.
    if combined_attention_mask is not None:
        # Ensure the attention mask values are within a reasonable range
        combined_attention_mask = combined_attention_mask.clamp(min=0, max=1)

        # Convert the attention mask to bfloat16
        combined_attention_mask = combined_attention_mask.to(torch.bfloat16)

        # Normalize the attention mask values to be between 0 and 1
        combined_attention_mask = (1.0 - combined_attention_mask) * -10000.0
    else:
        combined_attention_mask = torch.zeros(
            (bsz, 1, tgt_len, tgt_len), dtype=torch.bfloat16, device=inputs_embeds.device
        )

    return combined_attention_mask


# Copied from transformers.models.llama.modeling_llama._get_unpad_data
def _get_unpad_data(attention_mask):
    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
    max_seqlen_in_batch = seqlens_in_batch.max().item()
    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
    return (
        indices,
        cu_seqlens,
        max_seqlen_in_batch,
    )


# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Quiet
class QuietRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        QuietRMSNorm 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 hidden_states.to(input_dtype) * self.weight.to(hidden_states.device)


# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Quiet
class QuietRotaryEmbedding(nn.Module):
    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
        super().__init__()

        self.dim = dim
        self.max_position_embeddings = max_position_embeddings
        self.base = base
        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

        # Build here to make `torch.jit.trace` work.
        self._set_cos_sin_cache(
            seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
        )

    def _set_cos_sin_cache(self, seq_len, device, dtype):
        self.max_seq_len_cached = seq_len
        t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)

        freqs = torch.outer(t, self.inv_freq)
        # Different from paper, but it uses a different permutation in order to obtain the same calculation
        emb = torch.cat((freqs, freqs), dim=-1)
        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)

    def forward(self, x, seq_len=None):
        # x: [bs, num_attention_heads, seq_len, head_size]
        if seq_len > self.max_seq_len_cached:
            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)

        return (
            self.cos_cached[:seq_len].to(dtype=x.dtype),
            self.sin_cached[:seq_len].to(dtype=x.dtype),
        )


# Copied from transformers.models.llama.modeling_llama.rotate_half
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)


# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, position_ids, 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`):
            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
            used to pass offsetted position ids when working with a KV-cache.
        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[position_ids].unsqueeze(unsqueeze_dim)
    sin = sin[position_ids].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 QuietMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))


# Copied from transformers.models.llama.modeling_llama.repeat_kv
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)


class QuietAttention(nn.Module):
    """
    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
    and "Generating Long Sequences with Sparse Transformers".
    """

    def __init__(self, config: QuietConfig, 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 `layer_idx` is not recommended and will "
                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.is_causal = True
        self.attention_dropout = config.attention_dropout
        self._attn_implementation = config._attn_implementation

        if (self.head_dim * self.num_heads) != self.hidden_size:
            raise ValueError(
                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                f" and `num_heads`: {self.num_heads})."
            )
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)

        self.rotary_emb = QuietRotaryEmbedding(
            self.head_dim,
            max_position_embeddings=self.max_position_embeddings,
            base=self.rope_theta,
        )

    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()

    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: bool = False,
        use_cache: bool = False,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )
        bsz, q_len, _ = hidden_states.size()

        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_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError(
                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                    "with a layer index."
                )
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        # repeat k/v heads if n_kv_heads < n_heads
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)

        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
                f" {attn_weights.size()}"
            )
        if self._attn_implementation == "flash_attention_2":
            # Prepare attention mask for flash-attn
            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
        elif self._attn_implementation == "sdpa":
            # Prepare attention mask for SDPA
            if attention_mask is None or attention_mask.dim() == 2:
                attention_mask = _prepare_4d_causal_attention_mask(
                    attention_mask,
                    (batch_size, seq_length),
                    inputs_embeds,
                    past_key_values_length,
                    sliding_window=self.config.sliding_window,
                )
        else:
            # Prepare attention mask for other implementations
            if attention_mask is None or attention_mask.dim() == 2:
                attention_mask = _prepare_4d_causal_attention_mask(
                    attention_mask,
                    (batch_size, seq_length),
                    inputs_embeds,
                    past_key_values_length,
                    sliding_window=self.config.sliding_window,
                )
                
        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                )

            attn_weights = attn_weights + attention_mask

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

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

        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


class QuietFlashAttention2(QuietAttention):
    """
    Quiet flash attention module. This module inherits from `QuietAttention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    """

    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()

    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: bool = False,
        use_cache: bool = False,
        **kwargs,
    ):
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )

            # overwrite attention_mask with padding_mask
            attention_mask = kwargs.pop("padding_mask")
        bsz, q_len, _ = hidden_states.size()

        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_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            if self.layer_idx is None:
                raise ValueError(
                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
                    "with a layer index."
                )
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)

        # Because the input can be padded, the absolute sequence length depends on the max position id.
        rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
        cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len)

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

        use_sliding_windows = (
            _flash_supports_window_size
            and getattr(self.config, "sliding_window", None) is not None
            and kv_seq_len > self.config.sliding_window
        )

        if not _flash_supports_window_size:
            logger.warning_once(
                "The current flash attention version does not support sliding window attention, for a more memory efficient implementation"
                " make sure to upgrade flash-attn library."
            )

        if past_key_value is not None:
            # Activate slicing cache only if the config has a value `sliding_windows` attribute
            cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
            if (
                getattr(self.config, "sliding_window", None) is not None
                and kv_seq_len > self.config.sliding_window
                and cache_has_contents
            ):
                slicing_tokens = 1 - self.config.sliding_window

                past_key = past_key_value[self.layer_idx][0]
                past_value = past_key_value[self.layer_idx][1]

                past_key = past_key[:, :, slicing_tokens:, :].contiguous()
                past_value = past_value[:, :, slicing_tokens:, :].contiguous()

                if past_key.shape[-2] != self.config.sliding_window - 1:
                    raise ValueError(
                        f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
                        f" {past_key.shape}"
                    )

                if attention_mask is not None:
                    attention_mask = attention_mask[:, slicing_tokens:]
                    attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)

            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        # repeat k/v heads if n_kv_heads < n_heads
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)
        dropout_rate = 0.0 if not self.training else self.attention_dropout

        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
        # therefore the input hidden states gets silently casted in float32. Hence, we need
        # cast them back in float16 just to be sure everything works as expected.
        input_dtype = query_states.dtype
        if input_dtype == torch.float32:
            if torch.is_autocast_enabled():
                target_dtype = torch.get_autocast_gpu_dtype()
            # Handle the case where the model is quantized
            elif hasattr(self.config, "_pre_quantization_dtype"):
                target_dtype = self.config._pre_quantization_dtype
            else:
                target_dtype = self.q_proj.weight.dtype

            logger.warning_once(
                f"The input hidden states seems to be silently casted in float32, this might be related to"
                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
                f" {target_dtype}."
            )

            query_states = query_states.to(target_dtype)
            key_states = key_states.to(target_dtype)
            value_states = value_states.to(target_dtype)

        # Reashape to the expected shape for Flash Attention
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

        attn_output = self._flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=dropout_rate,
            use_sliding_windows=use_sliding_windows,
        )

        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value

    def _flash_attention_forward(
        self,
        query_states,
        key_states,
        value_states,
        attention_mask,
        query_length,
        dropout=0.0,
        softmax_scale=None,
        use_sliding_windows=False,
    ):
        """
        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
        first unpad the input, then computes the attention scores and pad the final attention scores.
        Args:
            query_states (`torch.Tensor`):
                Input query states to be passed to Flash Attention API
            key_states (`torch.Tensor`):
                Input key states to be passed to Flash Attention API
            value_states (`torch.Tensor`):
                Input value states to be passed to Flash Attention API
            attention_mask (`torch.Tensor`):
                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
                position of padding tokens and 1 for the position of non-padding tokens.
            dropout (`int`, *optional*):
                Attention dropout
            softmax_scale (`float`, *optional*):
                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
            use_sliding_windows (`bool`, *optional*):
                Whether to activate sliding window attention.
        """
        if not self._flash_attn_uses_top_left_mask:
            causal = self.is_causal
        else:
            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
            causal = self.is_causal and query_length != 1
            
        # Ensure attention_mask has the correct shape and values
        if attention_mask is not None:
            if attention_mask.dim() == 4:
                # Convert 4D attention mask to 2D
                attention_mask = attention_mask.squeeze(1).squeeze(1)
            elif attention_mask.dim() != 2:
                raise ValueError(
                    f"Invalid attention mask dimension: {attention_mask.dim()}. Expected 2D or 4D mask."
                )
            
            # Ensure attention_mask has values of 0 and 1
            attention_mask = attention_mask.to(torch.bool).to(torch.int32)

        # Contains at least one padding token in the sequence
        if attention_mask is not None:
            batch_size = query_states.shape[0]
            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
                query_states, key_states, value_states, attention_mask, query_length
            )

            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens

            if not use_sliding_windows:
                attn_output_unpad = flash_attn_varlen_func(
                    query_states,
                    key_states,
                    value_states,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_k=cu_seqlens_k,
                    max_seqlen_q=max_seqlen_in_batch_q,
                    max_seqlen_k=max_seqlen_in_batch_k,
                    dropout_p=dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                )
            else:
                attn_output_unpad = flash_attn_varlen_func(
                    query_states,
                    key_states,
                    value_states,
                    cu_seqlens_q=cu_seqlens_q,
                    cu_seqlens_k=cu_seqlens_k,
                    max_seqlen_q=max_seqlen_in_batch_q,
                    max_seqlen_k=max_seqlen_in_batch_k,
                    dropout_p=dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                    window_size=(self.config.sliding_window, self.config.sliding_window),
                )

            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
        else:
            if not use_sliding_windows:
                attn_output = flash_attn_func(
                    query_states,
                    key_states,
                    value_states,
                    dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                )
            else:
                attn_output = flash_attn_func(
                    query_states,
                    key_states,
                    value_states,
                    dropout,
                    softmax_scale=softmax_scale,
                    causal=causal,
                    window_size=(self.config.sliding_window, self.config.sliding_window),
                )

        return attn_output

    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
        batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape

        # On the first iteration we need to properly re-create the padding mask
        # by slicing it on the proper place
        if kv_seq_len != attention_mask.shape[-1]:
            attention_mask_num_tokens = attention_mask.shape[-1]
            attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]

        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)

        key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
        value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)

        if query_length == kv_seq_len:
            query_layer = index_first_axis(
                query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
            )
            cu_seqlens_q = cu_seqlens_k
            max_seqlen_in_batch_q = max_seqlen_in_batch_k
            indices_q = indices_k
        elif query_length == 1:
            max_seqlen_in_batch_q = 1
            cu_seqlens_q = torch.arange(
                batch_size + 1, dtype=torch.int32, device=query_layer.device
            )  # There is a memcpy here, that is very bad.
            indices_q = cu_seqlens_q[:-1]
            query_layer = query_layer.squeeze(1)
        else:
            # The -q_len: slice assumes left padding.
            attention_mask = attention_mask[:, -query_length:]
            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)

        return (
            query_layer,
            key_layer,
            value_layer,
            indices_q,
            (cu_seqlens_q, cu_seqlens_k),
            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
        )


# Copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Quiet
class QuietSdpaAttention(QuietAttention):
    """
    Quiet attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
    `QuietAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
    SDPA API.
    """

    # Adapted from QuietAttention.forward
    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: bool = False,
        use_cache: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if output_attentions:
            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
            logger.warning_once(
                "QuietModel is using QuietSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
            )
            return super().forward(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
            )

        bsz, q_len, _ = hidden_states.size()

        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_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)

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

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                )

        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
        # Reference: https://github.com/pytorch/pytorch/issues/112577.
        if query_states.device.type == "cuda" and attention_mask is not None:
            query_states = query_states.contiguous()
            key_states = key_states.contiguous()
            value_states = value_states.contiguous()

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=attention_mask.to(query_states.device) if attention_mask is not None else None,
            dropout_p=self.attention_dropout if self.training else 0.0,
            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
            is_causal=self.is_causal and attention_mask is None and q_len > 1,
        )

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

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value


QUIET_ATTENTION_CLASSES = {
    "eager": QuietAttention,
    "flash_attention_2": QuietFlashAttention2,
    "sdpa": QuietSdpaAttention,
}


class QuietDecoderLayer(nn.Module):
    def __init__(self, config: QuietConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.self_attn = QUIET_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)

        self.mlp = QuietMLP(config)
        self.input_layernorm = QuietRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = QuietRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        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,
        use_cache: Optional[bool] = False,
        **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        if "padding_mask" in kwargs:
            warnings.warn(
                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
            )
        """
        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, sequence_length)` where padding elements are indicated by 0.
            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
        """

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

        # Self Attention
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
        )
        hidden_states = residual.to(hidden_states.device) + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


QUIET_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 ([`QuietConfig`]):
            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 Quiet Model outputting raw hidden-states without any specific head on top.",
    QUIET_START_DOCSTRING,
)
class QuietPreTrainedModel(PreTrainedModel):
    config_class = QuietConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["QuietDecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = 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_()


QUIET_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 (`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;
            - 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.
"""


@add_start_docstrings(
    "The bare Quiet Model outputting raw hidden-states without any specific head on top.",
    QUIET_START_DOCSTRING,
)
class QuietModel(QuietPreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`QuietDecoderLayer`]
    Args:
        config: QuietConfig
    """

    def __init__(self, config: QuietConfig):
        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(
            [QuietDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self._attn_implementation = config._attn_implementation
        self.norm = QuietRMSNorm(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.embed_tokens

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

    @add_start_docstrings_to_model_forward(QUIET_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,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = 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

        # 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")

        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

        past_key_values_length = 0

        if use_cache:
            use_legacy_cache = not isinstance(past_key_values, Cache)
            if use_legacy_cache:
                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
            past_key_values_length = past_key_values.get_usable_length(seq_length)

        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)

        if self._attn_implementation == "flash_attention_2":
            # 2d mask is passed through the layers
            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
        elif self._attn_implementation == "sdpa" and not output_attentions and attention_mask.dim() == 2 and False:
            # output_attentions=True can not be supported when using SDPA, and we fall back on
            # the manual implementation that requires a 4D causal mask in all cases.
            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
                attention_mask,
                (batch_size, seq_length),
                inputs_embeds,
                past_key_values_length,
            )
        elif attention_mask is None or attention_mask.dim() == 2:
            # 4d mask is passed through the layers
            attention_mask = _prepare_4d_causal_attention_mask(
                attention_mask,
                (batch_size, seq_length),
                inputs_embeds,
                past_key_values_length,
                sliding_window=self.config.sliding_window,
            )

        hidden_states = inputs_embeds

        # 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,
                    attention_mask,
                    position_ids,
                    past_key_values,
                    output_attentions,
                    use_cache,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_values,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                )

            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.norm(hidden_states)

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

        next_cache = None
        if use_cache:
            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_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 nonzero_mean(x, axis=None):
    if axis is not None:
        return x.sum(axis) / (x != 0).sum(axis)
    return x.sum() / (x != 0).sum()

def loss_mean(x):
    return x.sum() / (x != 0).sum()

class QuietForCausalLM(QuietPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = QuietModel(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_experts
        # self.num_experts_per_tok = config.num_experts_per_tok
        self.max_thoughts = config.max_thoughts
        self.merged_lm_and_talk_heads = config.merged_lm_and_talk_heads
        self.use_concat_talk_head = config.use_concat_talk_head
        self.use_shallow_talk = config.use_shallow_talk
        self.use_complex_talk_head = config.use_complex_talk_head
        self.use_weighted_talk_head = config.use_weighted_talk_head
        # the weighted head will output a single value, so it can't be passed to the lm head
        assert not (self.use_weighted_talk_head and self.use_shallow_talk)

        self.n_ahead = 1
        self.n_ahead_talk = 1
        self.n_passes = 1
        self.n_tokens_print = 1
        self.gradient_accumulation_steps = 1
        self.training_steps = 0
        self.tokenizer = None
        self.start_token_id = None
        self.end_token_id = None
        self.rm_initialized = False
        self.residual_talk_head = True
        self.thought_init_std_scale = 1e-2

        self.final_only_mode = False
        self.first_and_last_mode = True
        self.first_only = False
        self.original_loss_weight = 0.5

        self.cumulative_residual = False
        self.clever_residual = False
        self.skip_residual = False
        self.no_residual = True

        self.optimize_lm_head_only_at_start = False
        self.optimize_model_only_at_start = False

        if self.optimize_model_only_at_start:
            raise NotImplementedError
        self.train_only_thinking_embedding = False
        self.weighted_embeddings = False
        self.use_start_thought_token = True
        self.use_end_thought_token = True
        self.initialize_thought_embedding_to_normal = False
        self.initial_start_token = "---"
        self.initial_end_token = "---"
        self.output_logits_at_the_end = True

        self.wandb_enabled = False
        self.gumbel_temperature = 0.001

        self.use_policy_loss = True
        self.include_policy_loss = True
        self.trice_mode = True
        self.remove_negative_rewards = True
        self.use_policy_loss_for_end_thought = True
        
        self.base_original_mode = False
        self.original_mode = False

        self.thought_prefix = "(Let's think step by step"
        self.tokenized_thought_prefix = None
        self.log_dict = defaultdict(int)
        self.eval_log_dict = defaultdict(int)
        self.loss_mean = loss_mean

        self.start_embedding = nn.Parameter(torch.zeros(2, self.model.config.hidden_size))
        self.end_embedding = nn.Parameter(torch.zeros(2, self.model.config.hidden_size))

        self.policy_loss_beta = 1e6
        self.embedding_scale = 1e2
        self.temperature = nn.Parameter(torch.ones(1))
        self.max_temperature = config.max_temperature
        self.complexity_factor = config.complexity_factor
        self.reinforce_temperature = 3
        self.base_loss_beta = 1
        self.thinking_usefulness_head = nn.Linear(self.model.config.hidden_size, 1)
        self.thinking_threshold = 0.5
        self.thinking_usefulness_loss_weight = 1e-2

        # Not used in the paper:
        self.use_thought_prefix = False
        self.use_reparam_for_thought_embeddings = False
        self.use_upper_triangular = False
        self.subtract_mean_reward = False
        self.comparison_mode = False
        self.gumbel_detach = False
    
        # For visualization
        self.eval_mode = False

        num_talk = 1
        talk_input_dim = config.hidden_size if not self.use_concat_talk_head else config.hidden_size * 2
        if self.use_weighted_talk_head:
            talk_output_dim = 1
        else:
            talk_output_dim = config.hidden_size if self.use_shallow_talk else config.vocab_size

        if not self.merged_lm_and_talk_heads:
            if self.use_complex_talk_head:
                self.talk_head = nn.ModuleList([nn.Sequential(
                    nn.Linear(talk_input_dim, config.hidden_size),
                    nn.ReLU(),
                    nn.Linear(config.hidden_size, config.hidden_size),
                    nn.ReLU(),
                    nn.Linear(config.hidden_size, talk_output_dim, bias=False)
                )])
            else:
                self.talk_head = nn.ModuleList([nn.Sequential(
                    nn.Linear(talk_input_dim, talk_output_dim, bias=False)
                )])
                
        self.apply(self._init_weights)

        # Add dropout regularization
        self.dropout = nn.Dropout(config.hidden_dropout_prob)

        # 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
    
    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            nn.init.xavier_uniform_(module.weight)
            if module.bias is not None:
                nn.init.constant_(module.bias, 0)
        elif isinstance(module, nn.Embedding):
            nn.init.xavier_uniform_(module.weight)
            
    @torch.no_grad()
    def generate(self, input_ids, attention_mask=None, streamer=None, **kwargs):
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids)
        
        max_length = kwargs.get("max_length", 20)
        temp = kwargs.get("temperature", 1.0)

        with torch.no_grad():
            finished_generating = torch.zeros(len(input_ids), dtype=torch.bool, device=input_ids.device)
            for cur_token_idx in range(max_length):
                # Sample the next token
                new_ids = self(
                    input_ids[~finished_generating],
                    attention_mask=attention_mask[~finished_generating]
                )['logits']
                # Mask out the start and end thought tokens so we don't accidentally sample them
                new_ids[:, :, self.tokenizer.vocab_size:] = -float("inf")
                for list_idx, answer_idx in enumerate((~finished_generating).nonzero(as_tuple=True)[0]):
                    # Find the index of the last token that is not padding
                    base_answer_ids = input_ids[answer_idx]
                    new_answer_ids = new_ids[list_idx]
                    last_token_idx = (base_answer_ids != self.tokenizer.pad_token_id).nonzero(as_tuple=True)[0].max()

                    new_ids_sampled = torch.multinomial(
                            torch.nn.functional.softmax(new_answer_ids[last_token_idx] / temp, dim=-1), 1)
                    # Assign the new id to the last token
                    if last_token_idx + 1 >= len(base_answer_ids):
                        # Add padding everywhere
                        new_padding = torch.full((len(input_ids), 1), self.tokenizer.pad_token_id, dtype=torch.long,
                                                 device=input_ids.device)
                        input_ids = torch.cat([input_ids, new_padding], dim=-1)
                        attention_mask = torch.cat([attention_mask, torch.zeros_like(new_padding)], dim=-1)
                    attention_mask[answer_idx, last_token_idx + 1] = 1
                    input_ids[answer_idx, last_token_idx + 1] = new_ids_sampled
                    if new_ids_sampled == self.tokenizer.eos_token_id or new_ids_sampled == self.tokenizer.bos_token_id or new_ids_sampled == self.tokenizer.pad_token_id:
                        finished_generating[answer_idx] = 1
                if finished_generating.all():
                    break
        return input_ids, attention_mask

    @add_start_docstrings_to_model_forward(QUIET_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_router_logits: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> 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]`.
        Returns:
        Example:
        ```python
        >>> from transformers import AutoTokenizer, QuietForCausalLM
        >>> model = QuietForCausalLM.from_pretrained("quietai/Quiet-7B-v0.1")
        >>> tokenizer = AutoTokenizer.from_pretrained("quietai/Quiet-7B-v0.1")
        >>> 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."
        ```"""

        if not self.training:
            n_ahead_talk_to_restore = self.n_ahead_talk
            n_passes_to_restore = self.n_passes
            self.n_ahead_talk = 1
            self.n_passes = 1
        
        # aux_loss = None
        # output_router_logits = output_router_logits if output_router_logits is not None else self.config.output_router_logits
        # if output_router_logits:
        #     router_logits = outputs.router_logits if return_dict else outputs[-1]
        #     if router_logits is not None:
        #         aux_loss = load_balancing_loss_func(
        #             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)
                    
        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

        assert self.cumulative_residual or self.clever_residual or self.skip_residual or self.no_residual
        assert not (self.skip_residual and self.use_policy_loss)

        if self.tokenized_thought_prefix is None and self.use_thought_prefix:
            self.tokenized_thought_prefix = self.tokenizer(self.thought_prefix, return_tensors="pt", add_special_tokens=False)["input_ids"]

        def apply_head(head, states, detach=False):
            if detach:
                head_weight = head.weight.detach()
            else:
                head_weight = head.weight
            head_weight = head_weight.to(states.device)
            return (head_weight @ states.transpose(-1, -2)).transpose(-1, -2).contiguous()
    
        def idx_if_sequential(head, idx=0):
            if isinstance(head, nn.Sequential) or isinstance(head, nn.ModuleList):
                return idx_if_sequential(head[idx], idx=idx)
            return head

        def none_repeat_interleave(x, n):
            if x is None:
                return x
            return x.repeat_interleave(n, dim=0)

        if self.n_passes > 1:
            input_ids = none_repeat_interleave(input_ids, self.n_passes)
            attention_mask = none_repeat_interleave(attention_mask, self.n_passes)
            position_ids = none_repeat_interleave(position_ids, self.n_passes)
            inputs_embeds = none_repeat_interleave(inputs_embeds, self.n_passes)
            labels = none_repeat_interleave(labels, self.n_passes)
            if past_key_values is not None:
                past_key_values = [none_repeat_interleave(p, self.n_passes) for p in past_key_values]
        cur_token_indices = torch.arange(input_ids.shape[1], device=input_ids.device)

        self.tokenizer_has_start_thought_token = True
        self.tokenizer_has_end_thought_token = True
        if self.start_token_id is None:
            self.start_token_id = self.tokenizer.convert_tokens_to_ids("<|startthought|>")
            if self.start_token_id == 0:
                self.start_token_id = self.tokenizer.bos_token_id
                self.tokenizer_has_start_thought_token = False
            elif self.use_start_thought_token:
                # base_start_id = self.tokenizer.convert_tokens_to_ids(self.initial_start_token)
                base_start_id = self.tokenizer.encode(self.initial_start_token, add_special_tokens=False)[0]
                if self.initialize_thought_embedding_to_normal:
                    self.start_embedding.data = torch.zeros_like(self.start_embedding.data)
                else:
                    self.start_embedding.data[0] = self.model.embed_tokens.weight.data[base_start_id].clone().detach() / self.embedding_scale
                self.start_embedding.data[1] = torch.log(self.model.embed_tokens.weight.data.std(dim=0) * self.thought_init_std_scale / self.embedding_scale)
        if self.end_token_id is None:
            self.end_token_id = self.tokenizer.convert_tokens_to_ids("<|endthought|>")
            if self.end_token_id == 0:
                self.end_token_id = self.tokenizer.eos_token_id
                self.tokenizer_has_end_thought_token = False
            elif self.use_end_thought_token:
                # base_end_id = self.tokenizer.convert_tokens_to_ids(self.initial_end_token)
                base_end_id = self.tokenizer.encode(self.initial_end_token, add_special_tokens=False)[0]
                if self.initialize_thought_embedding_to_normal:
                    self.end_embedding.data = torch.zeros_like(self.end_embedding.data)
                else:
                    self.end_embedding.data[0] = self.model.embed_tokens.weight.data[base_end_id].clone().detach() / self.embedding_scale
                self.end_embedding.data[1] = torch.log(self.model.embed_tokens.weight.data.std(dim=0) * self.thought_init_std_scale / self.embedding_scale)

        if not self.rm_initialized and (self.n_ahead > 1 or not self.base_original_mode):
            self.rm_initialized = True                        
            if not self.use_shallow_talk:
                head = self.talk_head[0]
                cur_head = head[-1] if isinstance(head, nn.Sequential) else head
                talk_input_dim = cur_head.weight.data.shape[1]
                talk_output_dim = 1 if self.use_weighted_talk_head else self.lm_head.weight.data.shape[0]
                cur_head.weight.data = torch.zeros(talk_output_dim, talk_input_dim, device=cur_head.weight.device, dtype=cur_head.weight.dtype)
            else:
                # convert to identity transform
                def lambda_transform(cur_head):
                    # pdb.set_trace()
                    if cur_head.weight.data.shape[0] != cur_head.weight.data.shape[1]:
                        return torch.cat([
                        torch.eye(
                            cur_head.weight.data.shape[0],
                            device=cur_head.weight.device,
                            dtype=cur_head.weight.dtype
                        ),
                        torch.zeros(
                            cur_head.weight.data.shape[0],
                            cur_head.weight.data.shape[1] - cur_head.weight.data.shape[0],
                            device=cur_head.weight.device,
                            dtype=cur_head.weight.dtype
                        )], dim=1)
                    return torch.eye(
                        cur_head.weight.data.shape[0],
                        device=cur_head.weight.device,
                        dtype=cur_head.weight.dtype
                    )
                if isinstance(self.talk_head[0], nn.Sequential):
                    for cur_head in self.talk_head[0]:
                        # if it has weights
                        if hasattr(cur_head, "weight"):
                            cur_head.weight.data = lambda_transform(cur_head)
                else:
                    self.talk_head[-1].weight.data = lambda_transform(self.talk_head[0])

        loss = None
        prev_rm_tokens = None
        cur_rm_tokens = None
        prev_rm_logits = None
        prev_sample_probs = None
        did_skip_sampling = None
        skip_sampling = None
        sample_probs = None
        hidden_states = None
        logits = None
        talk_kl_penalty = None
        rm_logits = None
        residual_logits = None
        probabilities_2d = None
        prev_probabilities_2d = None
        policy_reward = None
        logits_to_output = None
        batch_size, seq_len = input_ids.shape
        base_input_ids = input_ids.clone()
        loss_list = []
        dqn_loss_list = []
        sampled_token_history = []
        sample_probs_history = []
        action_loglikelihoods_list = []

        complexity_scores = self.compute_complexity_scores(input_ids, attention_mask)
        temperature = self.temperature * complexity_scores.unsqueeze(-1)

        if self.use_end_thought_token or self.use_start_thought_token:
            if not self.use_reparam_for_thought_embeddings:
                start_embedding = self.start_embedding[0].unsqueeze(0) * self.embedding_scale * temperature
                end_embedding = self.end_embedding[0].unsqueeze(0) * self.embedding_scale * temperature
            else:
                start_embedding = self.start_embedding * self.embedding_scale * temperature
                end_embedding = self.end_embedding * self.embedding_scale * temperature
            base_embeddings = self.model.embed_tokens.weight
            if self.train_only_thinking_embedding:
                base_embeddings = base_embeddings.detach()
                
        # # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        fwd_iters = 1 if self.original_mode else self.n_ahead + self.n_ahead_talk - 1
        for ahead_idx in range(fwd_iters):
            past_key_values_length = 0
            if past_key_values is not None:
                use_legacy_cache = not isinstance(past_key_values, Cache)
                if use_legacy_cache:
                    past_key_values = DynamicCache.from_legacy_cache(past_key_values)
                past_key_values_length = past_key_values.get_usable_length(seq_len)

            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_len + past_key_values_length, dtype=torch.long, device=device
                )
                position_ids = position_ids.unsqueeze(0).view(-1, seq_len)
            else:
                position_ids = position_ids.view(-1, seq_len).long()

            if inputs_embeds is None:
                contains_start = self.use_start_thought_token and (input_ids == self.start_token_id).any()
                contains_end = self.use_end_thought_token and (input_ids == self.end_token_id).any()
                contains_thought = contains_start or contains_end
                if contains_thought:
                    thought_id = self.start_token_id if contains_start else self.end_token_id
                    cur_thought_embedding = start_embedding if contains_start else end_embedding
                    if self.use_reparam_for_thought_embeddings:
                        inputs_embeds = torch.randn(batch_size, seq_len, self.model.config.hidden_size, device=input_ids.device, dtype=cur_thought_embedding.dtype)
                        inputs_embeds = inputs_embeds.detach() * torch.exp(cur_thought_embedding[1]) + cur_thought_embedding[0]
                        if contains_start:
                            sampled_start = inputs_embeds.clone().detach()
                        if contains_end:
                            sampled_end = inputs_embeds.clone().detach()
                    else:
                        inputs_embeds = cur_thought_embedding.unsqueeze(0).repeat(batch_size, seq_len, 1)
                else:
                    with torch.set_grad_enabled(not self.train_only_thinking_embedding):
                        inputs_embeds = self.model.embed_tokens(input_ids)
            
            if self.n_ahead != 1 or self.n_ahead_talk != 1 or self.comparison_mode:
                if attention_mask is None:
                    base_attention_mask = torch.triu(torch.ones(seq_len, seq_len), diagonal=0).to(input_ids.device)
                    base_attention_mask = base_attention_mask.view(1, 1, seq_len, seq_len)
                    base_attention_mask = base_attention_mask.repeat(input_ids.shape[0], 1, 1, 1)
                    attention_mask = base_attention_mask
                    breakpoint()
                elif attention_mask.dim() == 2:
                    if seq_len + past_key_values_length != attention_mask.shape[-1]:
                        breakpoint()
                        attention_mask = torch.cat(
                            [torch.ones((attention_mask.shape[0], past_key_values_length), dtype=attention_mask.dtype, device=attention_mask.device), attention_mask],
                            dim=-1
                        )
                    # # if the attention mask 
                    attention_mask = _prepare_4d_causal_attention_mask(
                        attention_mask,
                        (batch_size, seq_len),
                        inputs_embeds,
                        past_key_values_length,
                        sliding_window=self.config.sliding_window,
                    )

            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,
                # output_router_logits=output_router_logits,
                return_dict=return_dict,
            )

            prev_hidden_states = hidden_states
            hidden_states = outputs[0]
            prev_rm_logits = rm_logits  # for policy gradient
            prev_rm_tokens = cur_rm_tokens  # for policy gradient

            if ahead_idx == 0:
                hidden_states_lm = hidden_states
                logits = self.lm_head(hidden_states_lm)
                base_hidden_states = hidden_states.clone()
                initial_loss_logits = logits.clone()
                if self.optimize_lm_head_only_at_start or self.optimize_model_only_at_start:
                    logits = logits.detach()
                    base_hidden_states = base_hidden_states.detach()
                if self.optimize_model_only_at_start:
                    hidden_states = hidden_states.detach()
                base_logits = logits.clone()
            else:
                talk_hidden_states = hidden_states
                if self.merged_lm_and_talk_heads:
                    assert self.no_residual
                    residual_logits = self.lm_head(hidden_states)
                    talk_hidden_states = hidden_states
                else:
                    if ahead_idx > self.n_ahead - 1:
                        cur_base_hidden = torch.cat([
                            base_hidden_states[..., ahead_idx - self.n_ahead + 1:, :],
                            base_hidden_states[..., :ahead_idx - self.n_ahead + 1, :]
                        ], dim=-2)
                    else:
                        cur_base_hidden = base_hidden_states

                    if self.use_concat_talk_head:
                        # concatenate the hidden states with the original hidden states
                        head_input_hidden_states = torch.cat([cur_base_hidden, talk_hidden_states], dim=-1)
                    else:
                        head_input_hidden_states = talk_hidden_states

                    residual_logits = self.talk_head[0](head_input_hidden_states)
                    if self.use_shallow_talk:
                        residual_logits = apply_head(self.lm_head, residual_logits, detach=self.optimize_lm_head_only_at_start)                        
                    residual_logits = residual_logits.to(logits.device)
                    if self.use_weighted_talk_head:
                        # combine the cur_base_hidden with the talk_hidden_states according to the weighted head
                        residual_logits = cur_base_hidden * (1 - residual_logits) + talk_hidden_states * residual_logits
                        residual_logits = apply_head(self.lm_head, residual_logits, detach=self.optimize_lm_head_only_at_start)

                assert sum([self.cumulative_residual, self.clever_residual, self.skip_residual, self.no_residual]) == 1
                if self.clever_residual:
                    if ahead_idx >= self.n_ahead - 1:
                        # get the logits shifted according to the current talk ahead
                        cur_base_logits = torch.cat([
                            base_logits[..., ahead_idx - self.n_ahead + 1:, :],
                            base_logits[..., :ahead_idx - self.n_ahead + 1, :]
                        ], dim=-2)
                        if self.optimize_lm_head_only_at_start:
                            cur_base_logits = cur_base_logits.detach()
                        logits = cur_base_logits + residual_logits
                    else:
                        logits += residual_logits / self.n_ahead
                elif self.cumulative_residual:
                    if self.residual_talk_head:
                        if ahead_idx < self.n_ahead:
                            logits += residual_logits
                        else:
                            # get the logits shifted according to the current talk ahead
                            cur_base_logits = torch.cat([
                                base_logits[..., ahead_idx - self.n_ahead + 1:, :],
                                base_logits[..., :ahead_idx - self.n_ahead + 1, :]
                            ], dim=-2)
                            if self.optimize_lm_head_only_at_start:
                                cur_base_logits = cur_base_logits.detach()
                            logits = cur_base_logits + residual_logits
                    else:
                        if ahead_idx < self.n_ahead:
                            logits += residual_logits
                        else:
                            logits = residual_logits
                elif self.skip_residual:
                    if ahead_idx >= self.n_ahead:
                        # get the logits shifted according to the current talk ahead
                        cur_base_logits = torch.cat([
                            base_logits[..., ahead_idx - self.n_ahead + 1:, :],
                            base_logits[..., :ahead_idx - self.n_ahead + 1, :]
                        ], dim=-2)
                        if self.optimize_lm_head_only_at_start:
                            cur_base_logits = cur_base_logits.detach()
                        logits = cur_base_logits
                elif self.no_residual:
                    logits = residual_logits
                else:
                    logits = base_logits + residual_logits

            attempted = False
            talk_loss_list = []
            if self.original_mode or (self.n_ahead == 1) or (self.comparison_mode and ahead_idx == 0):# or (self.optimize_lm_head_only_at_start and ahead_idx == 0):
                loss = None
                attempted = True

                if labels is not None:
                    for shift_amount in range(self.n_ahead_talk):
                        # Shift so that tokens < n predict n
                        #  ab[cde]f
                        # abc[def]
                        if ahead_idx == 0 and self.optimize_lm_head_only_at_start:
                            loss_logits = initial_loss_logits
                        else:
                            loss_logits = logits
                        shift_logits = loss_logits[..., shift_amount:-1, :].contiguous()
                        shift_labels = labels[..., 1 + shift_amount:].contiguous()
                        # Flatten the tokens
                        loss_fct = CrossEntropyLoss(reduction="none")
                        print("Shift logits before:", shift_logits)
                        shift_logits = shift_logits.view(-1, self.config.vocab_size)
                        shift_labels = shift_labels.view(-1).clone()
                        print("shift logits after:", shift_logits)
                        # Enable model parallelism
                        shift_labels[shift_labels == self.tokenizer.pad_token_id] = -100
                        shift_labels = shift_labels.to(shift_logits.device)
                        loss = loss_fct(shift_logits, shift_labels)
                        if not self.comparison_mode and not (self.optimize_lm_head_only_at_start and (self.n_ahead + self.n_ahead_talk > 2)) or self.original_mode:
                            loss_list.append(loss)
                        talk_loss_list.append(nonzero_mean(loss).detach())
            
            if not attempted or self.comparison_mode:
                rm_hidden_states = hidden_states
                # print("Magnitude of RM hidden states before RM head", rm_hidden_states.norm())
                rm_logits = apply_head(self.lm_head, rm_hidden_states, detach=self.optimize_lm_head_only_at_start)
                    
                # don't allow it to predict the thinking token
                if self.tokenizer_has_start_thought_token:                    
                    rm_logits[..., self.start_token_id] = -1e10
                if self.tokenizer_has_end_thought_token:
                    rm_logits[..., self.end_token_id] = -1e10
                probabilities = rm_logits
                if probabilities_2d is not None:
                    prev_probabilities_2d = probabilities_2d.clone()
                probabilities_2d = probabilities.view(-1, probabilities.size(-1))

                did_skip_sampling = skip_sampling
                skip_sampling = False
                if ahead_idx == 0 and self.use_start_thought_token:
                    override_token = self.start_token_id
                elif self.use_thought_prefix and ahead_idx < self.tokenized_thought_prefix.shape[-1]:
                    override_token = self.tokenized_thought_prefix[..., ahead_idx]
                elif ahead_idx == self.n_ahead - 2 and self.use_end_thought_token:
                    override_token = self.end_token_id
                else:
                    override_token = None
                if override_token is not None and self.n_ahead > 1:
                    # always start with the start token
                    probabilities_2d = torch.zeros_like(probabilities_2d)
                    probabilities_2d[:, override_token] = 1.0
                    skip_sampling = True
                elif ahead_idx >= self.n_ahead - 1:
                    if labels is not None:  # we're in the talk phase
                        cur_talk_n = ahead_idx - (self.n_ahead - 1) + 1
                        # print("Setting rm to labels", cur_talk_n, "during", ahead_idx)
                        shift_labels = labels[..., cur_talk_n:].contiguous().to(probabilities_2d.device)
                        padding = torch.full_like(
                            labels[..., :cur_talk_n],
                            self.tokenizer.pad_token_id,
                            dtype=torch.long,
                            device=shift_labels.device
                        )
                        new_rm_tokens = torch.cat(
                            [shift_labels, padding],
                            dim=-1
                        )
                        
                        # print((new_rm_tokens > self.vocab_size - 1).any().item())
                        new_rm_tokens = torch.clamp(new_rm_tokens, 0, self.vocab_size - 1)

                        # Now safely convert rm tokens to one-hot
                        probabilities_2d = F.one_hot(new_rm_tokens, num_classes=self.vocab_size).reshape(-1, self.vocab_size).to(probabilities_2d.dtype)
                    else:
                        continue
                temperature = self.gumbel_temperature if self.training else 0.001
                prev_sample_probs = sample_probs
                sample_probs = probabilities_2d
                if ahead_idx < self.n_ahead - 1 and not skip_sampling:
                    probabilities_2d = F.gumbel_softmax(sample_probs, tau=temperature, hard=True, dim=-1)
                    if self.gumbel_detach:
                        probabilities_2d = probabilities_2d.detach()
                sampled_token_history.append(probabilities_2d.argmax(dim=-1).detach().cpu())
                # convert rm logits directly to embeddings
                contains_start = self.use_start_thought_token and (probabilities_2d[..., self.start_token_id].sum() > 0)
                contains_end = self.use_end_thought_token and (probabilities_2d[..., self.end_token_id].sum() > 0)
                contains_thought = contains_start or contains_end


                if not contains_thought:
                    with torch.set_grad_enabled(not self.train_only_thinking_embedding):
                        inputs_embeds = probabilities_2d @ (self.model.embed_tokens.weight.to(probabilities.device).to(probabilities.dtype) * temperature)
                else:
                    thought_id = self.start_token_id if contains_start else self.end_token_id
                    cur_thought_embedding = start_embedding if contains_start else end_embedding
                    if self.use_reparam_for_thought_embeddings:
                        inputs_embeds = torch.randn(batch_size, seq_len, self.model.config.hidden_size, device=input_ids.device, dtype=cur_thought_embedding.dtype)
                        inputs_embeds = inputs_embeds * torch.exp(cur_thought_embedding[1]) + cur_thought_embedding[0]
                        if contains_start:
                            sampled_start = inputs_embeds.clone().detach()
                        else:
                            sampled_end = inputs_embeds.clone().detach()
                    else:
                        inputs_embeds = cur_thought_embedding.unsqueeze(0).repeat(batch_size, seq_len, 1)
                inputs_embeds = inputs_embeds.view(probabilities.size(0), probabilities.size(1), -1).to(self.model.embed_tokens.weight.dtype)
                inputs_embeds = inputs_embeds.view(probabilities.size(0), probabilities.size(1), -1).to(self.model.embed_tokens.weight.dtype)
                
                # Predict the usefulness of thinking at each token position
                thinking_usefulness = self.thinking_usefulness_head(hidden_states).squeeze(-1)

                # Apply a threshold to decide where to generate thoughts
                generate_thought_mask = thinking_usefulness > self.thinking_threshold

                # Compute the regularization loss for thinking usefulness prediction
                thinking_usefulness_loss = torch.mean(thinking_usefulness * (1 - generate_thought_mask.float()))

                # Add the regularization loss to the total loss
                if loss is not None:
                    loss = loss + self.thinking_usefulness_loss_weight * thinking_usefulness_loss
                else:
                    loss = self.thinking_usefulness_loss_weight * thinking_usefulness_loss


                if len(attention_mask.shape) == 2:
                    breakpoint()
                else:
                    original_attention = attention_mask[..., :attention_mask.shape[-2]]
                    if self.use_upper_triangular:
                        new_attention = original_attention
                    else:
                        original_attention = original_attention == attention_mask.max()
                        # because eye isn't implemented for BF16, we need to handle the case
                        if not attention_mask.dtype == torch.bfloat16:
                            new_attention = torch.eye(
                                seq_len, dtype=attention_mask.dtype, device=attention_mask.device
                            )
                        else:
                            new_attention = torch.eye(
                                seq_len, dtype=torch.float32, device=attention_mask.device
                            ).to(attention_mask.dtype)

                        new_attention = new_attention.view(1, 1, seq_len, seq_len).repeat(input_ids.shape[0], 1, 1, 1)
                        new_attention = new_attention * original_attention
                        new_attention[new_attention == 0] = attention_mask.min()
                        new_attention[new_attention == 1] = attention_mask.max()
                    attention_mask = torch.cat([attention_mask, new_attention], dim=-1)
                past_key_values = outputs.past_key_values
                position_ids = position_ids + 1

                if labels is not None and (self.n_ahead > 1 or not self.base_original_mode):
                    # Shift so that tokens < n predict n
                    # logits: abcdef -> bcdef? -> cdef??
                    # labels: abcdef -> ?bcdef -> ??cdef
                    if ahead_idx == 0 and self.optimize_lm_head_only_at_start:
                        loss_logits = initial_loss_logits
                    else:
                        loss_logits = logits
                    shift_idx = 1 + max(0, ahead_idx - (self.n_ahead - 1))
                    shift_logits = loss_logits[..., :-shift_idx, :].contiguous()
                    shift_labels = labels[..., shift_idx:].contiguous()
                    # Flatten the tokens
                    loss_fct = CrossEntropyLoss(reduction="none")
                    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)
                    # if shift_labels.min() == self.tokenizer.pad_token_id:
                    shift_labels = torch.where(shift_labels == self.tokenizer.pad_token_id, -100, shift_labels)
                    unreduced_loss = loss_fct(shift_logits, shift_labels)
                    # print("Loss:", unreduced_loss.item())  # Print the loss before checking for NaN values
                    if torch.any(unreduced_loss != unreduced_loss):
                        # pdb.set_trace()
                        raise ValueError("NaN loss")
                    unreduced_loss = unreduced_loss.reshape(logits.shape[0], -1)
                    loss_list.append(unreduced_loss)


                    if self.use_policy_loss and ahead_idx > 0 and (ahead_idx > 1 or not self.use_start_thought_token):
                        # we treat the change in loss as the reward
                        previous_loss = loss_list[-2]
                        # for example, suppose n_ahead = 3 and n_ahead_talk = 2
                        # note that we end at self.n_ahead + self.n_ahead_talk - 2
                        # in this case, 5 - 2 = 3, so we end at ahead_idx = 3
                        # we also predict the next token at ahead_idx = 2
                        # when we get to ahead_idx = 2, we predict ahead
                        # so we shift by 1
                        # note that this is ahead_idx = n_ahead - 1
                        # when we get to ahead_idx = 3, we predict ahead
                        # so we shift by 2
                        # note that this is ahead_idx = n_ahead
                        if ahead_idx < self.n_ahead - 1:
                            shift_amount = 0
                            original_dqn_reward = (previous_loss - unreduced_loss).detach()
                            if self.first_and_last_mode:
                                original_dqn_reward = original_dqn_reward * 0.0
                        else:
                            # logits vs cur_policy_shift_logits
                            # let's look at rm_logits and prev_rm_logits
                            shift_amount = max(0, ahead_idx - (self.n_ahead - 1))
                            # let's say shift_amount = 2
                            # abcdefg -> bcdefg? -> cdefg??
                            # logits = [a b]c d e f[g]
                            # labels = [a b c]d e f g
                            cur_policy_shift_logits = initial_loss_logits[..., shift_amount:-1, :].contiguous().detach()
                            cur_policy_shift_labels = labels[..., 1 + shift_amount:].contiguous()
                            # Flatten the tokens
                            cur_policy_loss_fct = CrossEntropyLoss(reduction="none")
                            cur_policy_shift_logits = cur_policy_shift_logits.view(-1, self.config.vocab_size)
                            cur_policy_shift_labels = cur_policy_shift_labels.view(-1).clone()
                            # Enable model parallelism
                            cur_policy_shift_labels[cur_policy_shift_labels == self.tokenizer.pad_token_id] = -100
                            cur_policy_shift_labels = cur_policy_shift_labels.to(cur_policy_shift_labels.device)
                            cur_policy_reward_base_loss = loss_fct(
                                cur_policy_shift_logits, cur_policy_shift_labels.to(cur_policy_shift_logits.device)
                            ).reshape(logits.shape[0], -1)
                            original_dqn_reward = cur_policy_reward_base_loss.detach() - unreduced_loss
                                
                        if not did_skip_sampling:
                            nonzero_indices = prev_probabilities_2d.nonzero()
                            action_loglikelihoods = F.log_softmax(prev_sample_probs / self.reinforce_temperature, dim=-1)[nonzero_indices[:, 0], nonzero_indices[:, 1]]
                            action_loglikelihoods_2d = action_loglikelihoods.reshape(batch_size, -1)[:, :-1 - shift_amount]
                            action_loglikelihoods_list.append(action_loglikelihoods_2d)
                        if policy_reward is None:
                            policy_reward = original_dqn_reward[:, :-(self.n_ahead_talk - shift_amount)]
                        else:
                            if self.n_ahead_talk > shift_amount:
                                added_reward = original_dqn_reward[:, :-(self.n_ahead_talk - shift_amount)]
                            else:
                                added_reward = original_dqn_reward
                            policy_reward += added_reward

                        for action_loglikelihoods_2d in action_loglikelihoods_list:
                            train_policy_reward = policy_reward

                            # discard rewards below the mean
                            if self.trice_mode and self.n_passes > 1:
                                batched_policy_reward = train_policy_reward.reshape(-1, self.n_passes, train_policy_reward.shape[-1])
                                # average over the passes
                                train_policy_reward = batched_policy_reward - batched_policy_reward.mean(dim=1, keepdim=True)
                                train_policy_reward = train_policy_reward.reshape(-1, train_policy_reward.shape[-1])
                                
                            if self.subtract_mean_reward:
                                train_policy_reward = train_policy_reward - train_policy_reward.mean()
                            if self.remove_negative_rewards:
                                fixed_policy_reward = train_policy_reward.detach().clamp(min=0)
                            else:
                                fixed_policy_reward = train_policy_reward.detach()
                            actor_loss = -fixed_policy_reward * action_loglikelihoods_2d[:, :policy_reward.shape[-1]].to(policy_reward.device)
                            if action_loglikelihoods_2d.mean() < -1e4 and not self.use_policy_loss_just_for_thoughts:
                                # This will only happen when we force the next token to be the end of thought token
                                break
                            dqn_loss_list.append(actor_loss.mean())

        if loss_list:
            if self.first_and_last_mode:
                loss = sum(
                    self.loss_mean(loss_list[-(i + 1)]) for i in range(self.n_ahead_talk)
                ) * (1 - self.original_loss_weight) / self.n_ahead_talk
                loss = loss + self.loss_mean(loss_list[0]) * self.original_loss_weight
                # Let's NaN out the others
                # e.g. if n_ahead_talk = 2 and the list is 5 long, we want to NaN out 1, 2 but keep 0, 3, 4
                for i in range(1, len(loss_list) - self.n_ahead_talk):
                    loss_list[i] = loss_list[i] * math.nan
            elif self.first_only:
                loss = self.loss_mean(loss_list[0])
            elif self.final_only_mode:
                loss = sum(
                    self.loss_mean(loss_list[-i]) for i in range(1, self.n_ahead_talk + 1)
                ) / self.n_ahead_talk   
            else:
                loss = None
                for i in range(len(loss_list)):
                    cur_loss = self.loss_mean(loss_list[i])
                    if loss is not None:
                        loss = loss + cur_loss.to(loss.device)
                    else:
                        loss = cur_loss
                loss = loss / len(loss_list)
                loss = loss + thinking_usefulness_loss
            
            loss = loss * self.base_loss_beta

        if dqn_loss_list:
            dqn_loss = sum(dqn_loss_list) / len(dqn_loss_list)
            if self.include_policy_loss:
                if loss is not None:
                    loss += dqn_loss * self.policy_loss_beta
                else:
                    loss = dqn_loss * self.policy_loss_beta

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output
    
        base_log_dict = {
            f"loss_{i}": nonzero_mean(loss_list[i]) for i in range(len(loss_list))
        }

        if loss is not None:
            base_log_dict["loss_train"] = loss.item()

        if not self.training:
            self.n_ahead_talk = n_ahead_talk_to_restore
            self.n_passes = n_passes_to_restore
        
        del start_embedding
        del end_embedding
        torch.cuda.empty_cache()
        

        return CausalLMOutputWithPast(
            loss=loss if loss is not None else None,
            logits=(rm_logits if self.n_ahead > 1 else logits) if not self.output_logits_at_the_end else logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )
        
    
        
    def compute_complexity_scores(self, input_ids, attention_mask):
        # Compute complexity scores based on input sequence characteristics
        # Example: Normalize sequence lengths and consider the presence of rare tokens
        seq_lengths = torch.sum(attention_mask, dim=-1)
        max_length = torch.max(seq_lengths)
        length_scores = seq_lengths / max_length

        # Compute the proportion of rare tokens in each sequence
        rare_token_ids = self.get_rare_token_ids()
        rare_token_mask = torch.isin(input_ids, rare_token_ids)
        rare_token_counts = torch.sum(rare_token_mask, dim=-1)
        rare_token_scores = rare_token_counts / seq_lengths

        # Combine length scores and rare token scores
        complexity_scores = self.complexity_factor * length_scores + (1 - self.complexity_factor) * rare_token_scores
        return complexity_scores

    def get_rare_token_ids(self):
        # Get the IDs of rare tokens based on a predefined frequency threshold
        frequency_threshold = 1e-4
        token_counts = torch.bincount(self.model.embed_tokens.weight.argmax(dim=-1))
        total_tokens = torch.sum(token_counts)
        rare_token_mask = token_counts / total_tokens < frequency_threshold
        rare_token_ids = torch.nonzero(rare_token_mask).squeeze(-1)
        return rare_token_ids


    def prepare_inputs_for_generation(
        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
    ):
        # Omit tokens covered by past_key_values
        if past_key_values is not None:
            if isinstance(past_key_values, Cache):
                cache_length = past_key_values.get_seq_length()
                past_length = past_key_values.seen_tokens
                max_cache_length = past_key_values.get_max_length()
            else:
                cache_length = past_length = past_key_values[0][0].shape[2]
                max_cache_length = None

            # Keep only the unprocessed tokens:
            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
            # some of the inputs are exclusively passed as part of the cache (e.g. when passing inputs_embeds as
            # input)
            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
            # input_ids based on the past_length.
            elif past_length < input_ids.shape[1]:
                input_ids = input_ids[:, past_length:]
            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.

            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
            if (
                max_cache_length is not None
                and attention_mask is not None
                and cache_length + input_ids.shape[1] > max_cache_length
            ):
                attention_mask = attention_mask[:, -max_cache_length:]

        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[:, -input_ids.shape[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.to(past_state.device)) for past_state in layer_past),
            )
        return reordered_past
    
    


@add_start_docstrings(
    """
    The Quiet Model transformer with a sequence classification head on top (linear layer).
    [`QuietForSequenceClassification`] 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).
    """,
    QUIET_START_DOCSTRING,
)
# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Quiet, LLAMA->QUIET
class QuietForSequenceClassification(QuietPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.model = QuietModel(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(QUIET_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:
                # 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:
            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,
        )