modeling_mistral.py

# SpydazWeb AI MistralStar
################################ Introduction ##############################

# SpydazWeb AI Mistral Transformer ! this is a model based off of the mistral and mixtral models :
# it is created t eneble the model to generate thoughts before generating response:
# This is the first Generation of research;
# this paradigm will be improved: - 

## Note: to: Self:
# the model should generate a thought based of the thought prompt , then it should use its thought generation to pass to the model input : 
# with the original input : ( cross attention ) - 
# this should enhance the input to the model also providing extra content for the generation stage:
# ( later work ) - these thought should be generated by multiple heads : 
# as perhaps internal agents/Experts hence for each head it would need head prompt :perhaps this should be a hardcoded process?
# problem is how to frame it in the config ? - 
# then each head could generate content and the final head suamarize the content with the input to provide a rich query?
# in fact a single prompt is fine to hold multiple thoughts perhaps ,
#  as this will be stacked on top of the input ? to the hidden context size may need to be larger than the model size?
# PROJECT: ENDNING ?
# we need to have the extra processor in the tokenizer or the model ( perhaps the tokenizer is the best place for input management , 
# so to add the audio spectograph encoder and the Vision caption Trnsformer , 
# so given a image or a sound it will provuide the outputs for each item prompt , 
# hence the tokenizer response will need to be message based : ie seperate image description , seperate text ,
# seperate audio description( not Speech as this shoudl be an other rag front end? or pre processor to the tokenizer ,
#  for speech input it will handled in another model as that will be encoder/decoder ! this model is a decoder model and 
# the tokenizer / preprocessors are the encoder layers ~!))





################################ Imports ##############################
import inspect
import math
import copy
import os
import time
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt

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


from typing import List, Optional, Tuple, Union
import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
import torch.nn.functional as F
from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging,add_start_docstrings,add_start_docstrings_to_model_forward,replace_return_docstrings
from transformers.modeling_utils import PreTrainedModel
from transformers.cache_utils import Cache,DynamicCache, SlidingWindowCache, StaticCache
from transformers.activations import ACT2FN
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
from transformers.modeling_outputs import BaseModelOutputWithPast,CausalLMOutputWithPast,SequenceClassifierOutputWithPast,TokenClassifierOutput,QuestionAnsweringModelOutput,MoeCausalLMOutputWithPast,MoeModelOutputWithPast
from tokenizers import processors
from transformers.tokenization_utils_fast import PreTrainedTokenizerFast
from transformers.utils import is_sentencepiece_available, logging
from transformers.utils.versions import require_version
from shutil import copyfile
from collections import defaultdict
from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
import sentencepiece as spm
from transformers.convert_slow_tokenizer import import_protobuf
from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa
from transformers.utils import logging
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,
)
if TYPE_CHECKING:
    from transformers.tokenization_utils_base import TextInput
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)

MISTRAL_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.
"""
MISTRAL_PRETRAINED_CONFIG_ARCHIVE_MAP = {
    "mistralai/Mistral-7B-v0.1": "https://huggingface.co/mistralai/Mistral-7B-v0.1/resolve/main/config.json",
    "mistralai/Mistral-7B-Instruct-v0.1": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1/resolve/main/config.json",
}
MISTRAL_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 ([`MistralConfig`]):
            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.
"""
_CONFIG_FOR_DOC = "MistralStarConfig"
logger = logging.get_logger(__name__)
SPIECE_UNDERLINE = "▁"
require_version("tokenizers>=0.13.3")
VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model", "tokenizer_file": "tokenizer.json"}

B_INST, E_INST = "[INST]", "[/INST]"
B_SYS, E_SYS = "<<SYS>>\n", "\n<</SYS>>\n\n"

# fmt: off
DEFAULT_SYSTEM_PROMPT = """you are the world archive of knowledge and a friendly helpful artificial intelligent being:
consider if the current task denotes the creation of a function in python that can be saved and re-used for a future task, 
if created, return the artifact to the user, in the response, or if access is given to be save as a file on the user system:
Be flirtatious and joyfull, always willing to help or engage in conversation or express an opinion or sentiment: 
answer all questions indepth supply and helpfull references where possible; """
# fmt: on

##############################   Models   #################################
class MixtralConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`MixtralModel`]. It is used to instantiate an
    Mixtral model according to the specified arguments, defining the model architecture. Instantiating a configuration
    with the defaults will yield a similar configuration to that of the Mixtral-7B-v0.1 or Mixtral-7B-Instruct-v0.1.

    [mixtralai/Mixtral-8x7B](https://huggingface.co/mixtralai/Mixtral-8x7B)
    [mixtralai/Mixtral-7B-Instruct-v0.1](https://huggingface.co/mixtralai/Mixtral-7B-Instruct-v0.1)

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.


    Args:
        vocab_size (`int`, *optional*, defaults to 32000):
            Vocabulary size of the Mixtral model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`MixtralModel`]
        hidden_size (`int`, *optional*, defaults to 4096):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 14336):
            Dimension of the MLP representations.
        num_hidden_layers (`int`, *optional*, defaults to 32):
            Number of hidden layers in the Transformer encoder.
        num_attention_heads (`int`, *optional*, defaults to 32):
            Number of attention heads for each attention layer in the Transformer encoder.
        num_key_value_heads (`int`, *optional*, defaults to 8):
            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
            by meanpooling all the original heads within that group. For more details checkout [this
            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `8`.
        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
            The non-linear activation function (function or string) in the decoder.
        max_position_embeddings (`int`, *optional*, defaults to `4096*32`):
            The maximum sequence length that this model might ever be used with. Mixtral's sliding window attention
            allows sequence of up to 4096*32 tokens.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
            The epsilon used by the rms normalization layers.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models). Only
            relevant if `config.is_decoder=True`.
        pad_token_id (`int`, *optional*):
            The id of the padding token.
        bos_token_id (`int`, *optional*, defaults to 1):
            The id of the "beginning-of-sequence" token.
        eos_token_id (`int`, *optional*, defaults to 2):
            The id of the "end-of-sequence" token.
        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
            Whether the model's input and output word embeddings should be tied.
        rope_theta (`float`, *optional*, defaults to 1000000.0):
            The base period of the RoPE embeddings.
        sliding_window (`int`, *optional*):
            Sliding window attention window size. If not specified, will default to `4096`.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.
        num_experts_per_tok (`int`, *optional*, defaults to 2):
            The number of experts to route per-token, can be also interpreted as the `top-k` routing
            parameter
        num_local_experts (`int`, *optional*, defaults to 8):
            Number of experts per Sparse MLP layer.
        output_router_logits (`bool`, *optional*, defaults to `False`):
            Whether or not the router logits should be returned by the model. Enabeling this will also
            allow the model to output the auxiliary loss. See [here]() for more details
        router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
            The aux loss factor for the total loss.
        router_jitter_noise (`float`, *optional*, defaults to 0.0):
            Amount of noise to add to the router.

    ```python
    >>> from transformers import MixtralModel, MixtralConfig

    >>> # Initializing a Mixtral 7B style configuration
    >>> configuration = MixtralConfig()

    >>> # Initializing a model from the Mixtral 7B style configuration
    >>> model = MixtralModel(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "mixtral"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size=32000,
        hidden_size=4096,
        intermediate_size=14336,
        num_hidden_layers=32,
        num_attention_heads=32,
        num_key_value_heads=8,
        hidden_act="silu",
        max_position_embeddings=4096 * 32,
        initializer_range=0.02,
        rms_norm_eps=1e-5,
        use_cache=True,
        pad_token_id=None,
        bos_token_id=1,
        eos_token_id=2,
        tie_word_embeddings=False,
        rope_theta=1e6,
        sliding_window=None,
        attention_dropout=0.0,
        num_experts_per_tok=2,
        num_local_experts=8,
        output_router_logits=False,
        router_aux_loss_coef=0.001,
        router_jitter_noise=0.0,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.sliding_window = sliding_window

        # for backward compatibility
        if num_key_value_heads is None:
            num_key_value_heads = num_attention_heads

        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_dropout = attention_dropout

        self.num_experts_per_tok = num_experts_per_tok
        self.num_local_experts = num_local_experts
        self.output_router_logits = output_router_logits
        self.router_aux_loss_coef = router_aux_loss_coef
        self.router_jitter_noise = router_jitter_noise
        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )

class MistralStarConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`MistralModel`]. It is used to instantiate an
    Mistral model according to the specified arguments, defining the model architecture. Instantiating a configuration
    with the defaults will yield a similar configuration to that of the Mistral-7B-v0.1 or Mistral-7B-Instruct-v0.1.

    [mistralai/Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
    [mistralai/Mistral-7B-Instruct-v0.1](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1)

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.


    Args:
        vocab_size (`int`, *optional*, defaults to 32000):
            Vocabulary size of the Mistral model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`MistralModel`]
        hidden_size (`int`, *optional*, defaults to 4096):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 14336):
            Dimension of the MLP representations.
        num_hidden_layers (`int`, *optional*, defaults to 32):
            Number of hidden layers in the Transformer encoder.
        num_attention_heads (`int`, *optional*, defaults to 32):
            Number of attention heads for each attention layer in the Transformer encoder.
        num_key_value_heads (`int`, *optional*, defaults to 8):
            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
            by meanpooling all the original heads within that group. For more details checkout [this
            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `8`.
        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
            The non-linear activation function (function or string) in the decoder.
        max_position_embeddings (`int`, *optional*, defaults to `4096*32`):
            The maximum sequence length that this model might ever be used with. Mistral's sliding window attention
            allows sequence of up to 4096*32 tokens.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
            The epsilon used by the rms normalization layers.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models). Only
            relevant if `config.is_decoder=True`.
        pad_token_id (`int`, *optional*):
            The id of the padding token.
        bos_token_id (`int`, *optional*, defaults to 1):
            The id of the "beginning-of-sequence" token.
        eos_token_id (`int`, *optional*, defaults to 2):
            The id of the "end-of-sequence" token.
        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
            Whether the model's input and output word embeddings should be tied.
        rope_theta (`float`, *optional*, defaults to 10000.0):
            The base period of the RoPE embeddings.
        sliding_window (`int`, *optional*, defaults to 4096):
            Sliding window attention window size. If not specified, will default to `4096`.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.

    ```python
    >>> from transformers import MistralModel, MistralConfig

    >>> # Initializing a Mistral 7B style configuration
    >>> configuration = MistralConfig()

    >>> # Initializing a model from the Mistral 7B style configuration
    >>> model = MistralModel(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "mistralstar"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size=32000,
        hidden_size=4096,
        intermediate_size=14336,
        num_hidden_layers=32,
        num_attention_heads=32,
        num_key_value_heads=8,
        hidden_act="silu",
        max_position_embeddings=4096 * 32,
        initializer_range=0.02,
        rms_norm_eps=1e-6,
        use_cache=True,
        pad_token_id=None,
        bos_token_id=1,
        eos_token_id=2,
        tie_word_embeddings=False,
        rope_theta=10000.0,
        sliding_window=4096,
        attention_dropout=0.0,
        max_thoughts=16,
        thought_length = 10,
        merged_talk_heads=True,
        merged_lm_and_talk_heads=False,
        merged_lm_and_think_heads=True,
        use_concat_talk_head=True,
        use_shallow_think=True,
        use_shallow_talk=False,
        use_complex_think_head=False,
        use_complex_talk_head=True,
        use_weighted_talk_head=True,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.sliding_window = sliding_window

        # for backward compatibility
        if num_key_value_heads is None:
            num_key_value_heads = num_attention_heads

        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_dropout = attention_dropout
        self.max_thoughts = max_thoughts
        self.thought_length = thought_length
        self.merged_talk_heads = merged_talk_heads
        self.merged_lm_and_talk_heads = merged_lm_and_talk_heads
        self.merged_lm_and_think_heads = merged_lm_and_think_heads
        self.use_concat_talk_head = use_concat_talk_head
        self.use_shallow_think = use_shallow_think
        self.use_shallow_talk = use_shallow_talk
        self.use_complex_think_head = use_complex_think_head
        self.use_complex_talk_head = use_complex_talk_head
        self.use_weighted_talk_head = use_weighted_talk_head

        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )

class MistralConfig(PretrainedConfig):
    r"""
    This is the configuration class to store the configuration of a [`MistralModel`]. It is used to instantiate an
    Mistral model according to the specified arguments, defining the model architecture. Instantiating a configuration
    with the defaults will yield a similar configuration to that of the Mistral-7B-v0.1 or Mistral-7B-Instruct-v0.1.

    [mistralai/Mistral-7B-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
    [mistralai/Mistral-7B-Instruct-v0.1](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1)

    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
    documentation from [`PretrainedConfig`] for more information.


    Args:
        vocab_size (`int`, *optional*, defaults to 32000):
            Vocabulary size of the Mistral model. Defines the number of different tokens that can be represented by the
            `inputs_ids` passed when calling [`MistralModel`]
        hidden_size (`int`, *optional*, defaults to 4096):
            Dimension of the hidden representations.
        intermediate_size (`int`, *optional*, defaults to 14336):
            Dimension of the MLP representations.
        num_hidden_layers (`int`, *optional*, defaults to 32):
            Number of hidden layers in the Transformer encoder.
        num_attention_heads (`int`, *optional*, defaults to 32):
            Number of attention heads for each attention layer in the Transformer encoder.
        num_key_value_heads (`int`, *optional*, defaults to 8):
            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
            by meanpooling all the original heads within that group. For more details checkout [this
            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `8`.
        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
            The non-linear activation function (function or string) in the decoder.
        max_position_embeddings (`int`, *optional*, defaults to `4096*32`):
            The maximum sequence length that this model might ever be used with. Mistral's sliding window attention
            allows sequence of up to 4096*32 tokens.
        initializer_range (`float`, *optional*, defaults to 0.02):
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
            The epsilon used by the rms normalization layers.
        use_cache (`bool`, *optional*, defaults to `True`):
            Whether or not the model should return the last key/values attentions (not used by all models). Only
            relevant if `config.is_decoder=True`.
        pad_token_id (`int`, *optional*):
            The id of the padding token.
        bos_token_id (`int`, *optional*, defaults to 1):
            The id of the "beginning-of-sequence" token.
        eos_token_id (`int`, *optional*, defaults to 2):
            The id of the "end-of-sequence" token.
        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
            Whether the model's input and output word embeddings should be tied.
        rope_theta (`float`, *optional*, defaults to 10000.0):
            The base period of the RoPE embeddings.
        sliding_window (`int`, *optional*, defaults to 4096):
            Sliding window attention window size. If not specified, will default to `4096`.
        attention_dropout (`float`, *optional*, defaults to 0.0):
            The dropout ratio for the attention probabilities.

    ```python
    >>> from transformers import MistralModel, MistralConfig

    >>> # Initializing a Mistral 7B style configuration
    >>> configuration = MistralConfig()

    >>> # Initializing a model from the Mistral 7B style configuration
    >>> model = MistralModel(configuration)

    >>> # Accessing the model configuration
    >>> configuration = model.config
    ```"""

    model_type = "mistral"
    keys_to_ignore_at_inference = ["past_key_values"]

    def __init__(
        self,
        vocab_size=32000,
        hidden_size=4096,
        intermediate_size=14336,
        num_hidden_layers=32,
        num_attention_heads=32,
        num_key_value_heads=8,
        hidden_act="silu",
        max_position_embeddings=4096 * 32,
        initializer_range=0.02,
        rms_norm_eps=1e-6,
        use_cache=True,
        pad_token_id=None,
        bos_token_id=1,
        eos_token_id=2,
        tie_word_embeddings=False,
        rope_theta=10000.0,
        sliding_window=4096,
        attention_dropout=0.0,
        max_thoughts=16,
        merged_talk_heads=True,
        merged_lm_and_talk_heads=False,
        merged_lm_and_think_heads=True,
        use_concat_talk_head=True,
        use_shallow_think=True,
        use_shallow_talk=False,
        use_complex_think_head=False,
        use_complex_talk_head=True,
        use_weighted_talk_head=True,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.max_position_embeddings = max_position_embeddings
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.sliding_window = sliding_window

        # for backward compatibility
        if num_key_value_heads is None:
            num_key_value_heads = num_attention_heads

        self.num_key_value_heads = num_key_value_heads
        self.hidden_act = hidden_act
        self.initializer_range = initializer_range
        self.rms_norm_eps = rms_norm_eps
        self.use_cache = use_cache
        self.rope_theta = rope_theta
        self.attention_dropout = attention_dropout
        self.max_thoughts = max_thoughts
        self.merged_talk_heads = merged_talk_heads
        self.merged_lm_and_talk_heads = merged_lm_and_talk_heads
        self.merged_lm_and_think_heads = merged_lm_and_think_heads
        self.use_concat_talk_head = use_concat_talk_head
        self.use_shallow_think = use_shallow_think
        self.use_shallow_talk = use_shallow_talk
        self.use_complex_think_head = use_complex_think_head
        self.use_complex_talk_head = use_complex_talk_head
        self.use_weighted_talk_head = use_weighted_talk_head

        super().__init__(
            pad_token_id=pad_token_id,
            bos_token_id=bos_token_id,
            eos_token_id=eos_token_id,
            tie_word_embeddings=tie_word_embeddings,
            **kwargs,
        )


@add_start_docstrings(
    "The bare Mistral Model outputting raw hidden-states without any specific head on top.",
    MISTRAL_START_DOCSTRING,
)
class MistralPreTrainedModel(PreTrainedModel):
    config_class = MistralConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["MistralDecoderLayer"]
    _skip_keys_device_placement = "past_key_values"
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True
    _supports_static_cache = True

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


@add_start_docstrings(
    "The bare Mistral Model outputting raw hidden-states without any specific head on top.",
    MISTRAL_START_DOCSTRING,
)
class MistralModel(MistralPreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MistralDecoderLayer`]

    Args:
        config: MistralConfig
    """

    def __init__(self, config: MistralConfig):
        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(
            [MistralDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self._attn_implementation = config._attn_implementation
        self.norm = MistralRMSNorm(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(MISTRAL_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Union[Tuple, BaseModelOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # retrieve input_ids and inputs_embeds
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError(
                "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
            )

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

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

        return_legacy_cache = False
        if use_cache and not isinstance(past_key_values, Cache):
            past_key_values = DynamicCache.from_legacy_cache(past_key_values)
            return_legacy_cache = True
            logger.warning_once(
                "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. "
                "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)"
            )

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

        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

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

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

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

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

    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        use_cache: bool,
        output_attentions: bool,
    ):
        # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
        # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
        # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
        # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114

        if self._attn_implementation == "flash_attention_2":
            if attention_mask is not None and use_cache:
                is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0]
                if is_padding_right:
                    raise ValueError(
                        "You are attempting to perform batched generation with padding_side='right'"
                        " this may lead to unexpected behaviour for Flash Attention version of Mistral. Make sure to "
                        " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
                    )
            if attention_mask is not None and 0.0 in attention_mask:
                return attention_mask
            return None

        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
        # to infer the attention mask.

        # cache_position must be valid here no matter which cache we use
        past_seen_tokens = cache_position[0] if past_key_values is not None else 0
        using_static_cache = isinstance(past_key_values, StaticCache)
        using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)

        if (
            self.config._attn_implementation == "sdpa"
            and not (using_static_cache or using_sliding_window_cache)
            and not output_attentions
        ):
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                sliding_window=self.config.sliding_window,
                is_training=self.training,
            ):
                return None

        dtype, device = input_tensor.dtype, input_tensor.device
        min_dtype = torch.finfo(dtype).min
        sequence_length = input_tensor.shape[1]
        # SlidingWindowCache
        if using_sliding_window_cache:
            target_length = max(sequence_length, self.config.sliding_window)
        # StaticCache
        elif using_static_cache:
            target_length = past_key_values.get_max_length()
        # DynamicCache or no cache
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        if attention_mask is not None and attention_mask.dim() == 4:
            # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
            if attention_mask.max() != 0:
                raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
            causal_mask = attention_mask
        else:
            causal_mask = torch.full(
                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
            )
            exclude_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
            if self.config.sliding_window is not None:
                if not using_sliding_window_cache or sequence_length > self.config.sliding_window:
                    exclude_mask.bitwise_or_(
                        torch.arange(target_length, device=device)
                        <= (cache_position.reshape(-1, 1) - self.config.sliding_window)
                    )
            causal_mask *= exclude_mask
            causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                if attention_mask.dim() == 2:
                    mask_length = attention_mask.shape[-1]
                    padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
                    padding_mask = padding_mask == 0
                    causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
                        padding_mask, min_dtype
                    )

        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type == "cuda"
            and not output_attentions
        ):
            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
            # Details: https://github.com/pytorch/pytorch/issues/110213
            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

        return causal_mask


############################## LM Heads   #################################






################################ Tokenizer ##############################
class MistralTokenizer(PreTrainedTokenizer):
    """
    Construct a Llama tokenizer. Based on byte-level Byte-Pair-Encoding. The default padding token is unset as there is
    no padding token in the original model.

    Args:
        vocab_file (`str`):
            Path to the vocabulary file.
        unk_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<unk>"`):
            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
            token instead.
        bos_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<s>"`):
            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
        eos_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"</s>"`):
            The end of sequence token.
        pad_token (`str` or `tokenizers.AddedToken`, *optional*):
            A special token used to make arrays of tokens the same size for batching purpose. Will then be ignored by
            attention mechanisms or loss computation.
        sp_model_kwargs (`Dict[str, Any]`, `Optional`, *optional*):
            Will be passed to the `SentencePieceProcessor.__init__()` method. The [Python wrapper for
            SentencePiece](https://github.com/google/sentencepiece/tree/master/python) can be used, among other things,
            to set:

            - `enable_sampling`: Enable subword regularization.
            - `nbest_size`: Sampling parameters for unigram. Invalid for BPE-Dropout.

              - `nbest_size = {0,1}`: No sampling is performed.
              - `nbest_size > 1`: samples from the nbest_size results.
              - `nbest_size < 0`: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
                using forward-filtering-and-backward-sampling algorithm.

            - `alpha`: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
              BPE-dropout.

        add_bos_token (`bool`, *optional*, defaults to `True`):
            Whether or not to add an `bos_token` at the start of sequences.
        add_eos_token (`bool`, *optional*, defaults to `False`):
            Whether or not to add an `eos_token` at the end of sequences.
        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
            extra spaces.
        use_default_system_prompt (`bool`, *optional*, defaults to `False`):
            Whether or not the default system prompt for Llama should be used.
        spaces_between_special_tokens (`bool`, *optional*, defaults to `False`):
            Whether or not to add spaces between special tokens.
        legacy (`bool`, *optional*):
            Whether or not the `legacy` behavior of the tokenizer should be used. Legacy is before the merge of #24622
            and #25224 which includes fixes to properly handle tokens that appear after special tokens.
            Make sure to also set `from_slow` to `True`.
            A simple example:

            - `legacy=True`:
            ```python
            >>> from transformers import LlamaTokenizerFast

            >>> tokenizer = LlamaTokenizerFast.from_pretrained("huggyllama/llama-7b", legacy=True, from_slow=True)
            >>> tokenizer.encode("Hello <s>.") # 869 is '▁.'
            [1, 15043, 29871, 1, 869]
            ```
            - `legacy=False`:
            ```python
            >>> from transformers import LlamaTokenizerFast

            >>> tokenizer = LlamaTokenizerFast.from_pretrained("huggyllama/llama-7b", legacy=False, from_slow=True)
            >>> tokenizer.encode("Hello <s>.")  # 29889 is '.'
            [1, 15043, 29871, 1, 29889]
            ```
            Checkout the [pull request](https://github.com/huggingface/transformers/pull/24565) for more details.
        add_prefix_space (`bool`, *optional*, defaults to `True`):
            Whether or not to add an initial space to the input. This allows to treat the leading word just as any
            other word. Again, this should be set with `from_slow=True` to make sure it's taken into account.
    """

    vocab_files_names = VOCAB_FILES_NAMES
    model_input_names = ["input_ids", "attention_mask"]

    def __init__(
        self,
        vocab_file,
        unk_token="<unk>",
        bos_token="<s>",
        eos_token="</s>",
        pad_token=None,
        sp_model_kwargs: Optional[Dict[str, Any]] = None,
        add_bos_token=True,
        add_eos_token=False,
        clean_up_tokenization_spaces=False,
        use_default_system_prompt=False,
        spaces_between_special_tokens=False,
        legacy=None,
        add_prefix_space=True,
        **kwargs,
    ):
        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
        bos_token = AddedToken(bos_token, normalized=False, special=True) if isinstance(bos_token, str) else bos_token
        eos_token = AddedToken(eos_token, normalized=False, special=True) if isinstance(eos_token, str) else eos_token
        unk_token = AddedToken(unk_token, normalized=False, special=True) if isinstance(unk_token, str) else unk_token
        pad_token = AddedToken(pad_token, normalized=False, special=True) if isinstance(pad_token, str) else pad_token

        if legacy is None:
            logger.warning_once(
                f"You are using the default legacy behaviour of the {self.__class__}. This is"
                " expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you."
                " If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it"
                " means, and thoroughly read the reason why this was added as explained in"
                " https://github.com/huggingface/transformers/pull/24565 - if you loaded a llama tokenizer from a GGUF file"
                " you can ignore this message"
            )
            legacy = True

        self.legacy = legacy
        self.vocab_file = vocab_file
        self.add_bos_token = add_bos_token
        self.add_eos_token = add_eos_token
        self.use_default_system_prompt = use_default_system_prompt
        self.sp_model = self.get_spm_processor(kwargs.pop("from_slow", False))
        self.add_prefix_space = add_prefix_space

        super().__init__(
            bos_token=bos_token,
            eos_token=eos_token,
            unk_token=unk_token,
            pad_token=pad_token,
            add_bos_token=add_bos_token,
            add_eos_token=add_eos_token,
            sp_model_kwargs=self.sp_model_kwargs,
            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
            use_default_system_prompt=use_default_system_prompt,
            spaces_between_special_tokens=spaces_between_special_tokens,
            legacy=legacy,
            add_prefix_space=add_prefix_space,
            **kwargs,
        )

    @property
    def unk_token_length(self):
        return len(self.sp_model.encode(str(self.unk_token)))

    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.get_spm_processor
    def get_spm_processor(self, from_slow=False):
        tokenizer = spm.SentencePieceProcessor(**self.sp_model_kwargs)
        if self.legacy or from_slow:  # no dependency on protobuf
            tokenizer.Load(self.vocab_file)
            return tokenizer

        with open(self.vocab_file, "rb") as f:
            sp_model = f.read()
            model_pb2 = import_protobuf(f"The new behaviour of {self.__class__.__name__} (with `self.legacy = False`)")
            model = model_pb2.ModelProto.FromString(sp_model)
            normalizer_spec = model_pb2.NormalizerSpec()
            normalizer_spec.add_dummy_prefix = False
            model.normalizer_spec.MergeFrom(normalizer_spec)
            sp_model = model.SerializeToString()
            tokenizer.LoadFromSerializedProto(sp_model)
        return tokenizer

    def __getstate__(self):
        state = self.__dict__.copy()
        state["sp_model"] = None
        state["sp_model_proto"] = self.sp_model.serialized_model_proto()
        return state

    def __setstate__(self, d):
        self.__dict__ = d
        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
        self.sp_model.LoadFromSerializedProto(self.sp_model_proto)

    @property
    def vocab_size(self):
        """Returns vocab size"""
        return self.sp_model.get_piece_size()

    def get_vocab(self):
        """Returns vocab as a dict"""
        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
        vocab.update(self.added_tokens_encoder)
        return vocab

    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.tokenize
    def tokenize(self, text: "TextInput", **kwargs) -> List[str]:
        """
        Converts a string to a list of tokens. If `self.legacy` is set to `False`, a prefix token is added unless the
        first token is special.
        """
        if self.legacy or len(text) == 0:
            return super().tokenize(text, **kwargs)

        text = text.replace(SPIECE_UNDERLINE, " ")
        if self.add_prefix_space:
            text = SPIECE_UNDERLINE + text

        tokens = super().tokenize(text, **kwargs)

        if len(tokens) > 1 and tokens[0] == SPIECE_UNDERLINE and tokens[1] in self.all_special_tokens:
            tokens = tokens[1:]
        return tokens

    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer._tokenize
    def _tokenize(self, text, **kwargs):
        """
        Returns a tokenized string.

        We de-activated the `add_dummy_prefix` option, thus the sentencepiece internals will always strip any
        SPIECE_UNDERLINE. For example: `self.sp_model.encode(f"{SPIECE_UNDERLINE}Hey", out_type = str)` will give
        `['H', 'e', 'y']` instead of `['▁He', 'y']`. Thus we always encode `f"{unk_token}text"` and strip the
        `unk_token`. Here is an example with `unk_token = "<unk>"` and `unk_token_length = 4`.
        `self.tokenizer.sp_model.encode("<unk> Hey", out_type = str)[4:]`.
        """
        tokens = self.sp_model.encode(text, out_type=str)
        if self.legacy or not text.startswith((SPIECE_UNDERLINE, " ")):
            return tokens

        # 1. Encode string + prefix ex: "<unk> Hey"
        tokens = self.sp_model.encode(self.unk_token + text, out_type=str)
        # 2. Remove self.unk_token from ['<','unk','>', '▁Hey']
        return tokens[self.unk_token_length :] if len(tokens) >= self.unk_token_length else tokens

    def _convert_token_to_id(self, token):
        """Converts a token (str) in an id using the vocab."""
        return self.sp_model.piece_to_id(token)

    def _convert_id_to_token(self, index):
        """Converts an index (integer) in a token (str) using the vocab."""
        token = self.sp_model.IdToPiece(index)
        return token

    def convert_tokens_to_string(self, tokens):
        """Converts a sequence of tokens (string) in a single string."""
        # since we manually add the prefix space, we have to remove it when decoding
        if tokens[0].startswith(SPIECE_UNDERLINE) and self.add_prefix_space:
            tokens[0] = tokens[0][1:]

        current_sub_tokens = []
        out_string = ""
        prev_is_special = False
        for i, token in enumerate(tokens):
            # make sure that special tokens are not decoded using sentencepiece model
            if token in self.all_special_tokens:
                if not prev_is_special and i != 0 and self.legacy:
                    out_string += " "
                out_string += self.sp_model.decode(current_sub_tokens) + token
                prev_is_special = True
                current_sub_tokens = []
            else:
                if prev_is_special and i == 1 and self.add_prefix_space and not token.startswith(SPIECE_UNDERLINE):
                    out_string += " "
                current_sub_tokens.append(token)
                prev_is_special = False
        out_string += self.sp_model.decode(current_sub_tokens)
        return out_string

    def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
        """
        Save the vocabulary and special tokens file to a directory.

        Args:
            save_directory (`str`):
                The directory in which to save the vocabulary.

        Returns:
            `Tuple(str)`: Paths to the files saved.
        """
        if not os.path.isdir(save_directory):
            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
            return
        out_vocab_file = os.path.join(
            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
        )

        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
            copyfile(self.vocab_file, out_vocab_file)
        elif not os.path.isfile(self.vocab_file):
            with open(out_vocab_file, "wb") as fi:
                content_spiece_model = self.sp_model.serialized_model_proto()
                fi.write(content_spiece_model)

        return (out_vocab_file,)

    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
        eos_token_id = [self.eos_token_id] if self.add_eos_token else []

        output = bos_token_id + token_ids_0 + eos_token_id

        if token_ids_1 is not None:
            output = output + bos_token_id + token_ids_1 + eos_token_id

        return output

    def get_special_tokens_mask(
        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
    ) -> List[int]:
        """
        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
        special tokens using the tokenizer `prepare_for_model` method.

        Args:
            token_ids_0 (`List[int]`):
                List of IDs.
            token_ids_1 (`List[int]`, *optional*):
                Optional second list of IDs for sequence pairs.
            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
                Whether or not the token list is already formatted with special tokens for the model.

        Returns:
            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
        """
        if already_has_special_tokens:
            return super().get_special_tokens_mask(
                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
            )

        bos_token_id = [1] if self.add_bos_token else []
        eos_token_id = [1] if self.add_eos_token else []

        if token_ids_1 is None:
            return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
        return (
            bos_token_id
            + ([0] * len(token_ids_0))
            + eos_token_id
            + bos_token_id
            + ([0] * len(token_ids_1))
            + eos_token_id
        )

    def create_token_type_ids_from_sequences(
        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
    ) -> List[int]:
        """
        Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
        sequence pair mask has the following format:

        ```
        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
        | first sequence    | second sequence |
        ```

        if token_ids_1 is None, only returns the first portion of the mask (0s).

        Args:
            token_ids_0 (`List[int]`):
                List of ids.
            token_ids_1 (`List[int]`, *optional*):
                Optional second list of IDs for sequence pairs.

        Returns:
            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
        """
        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
        eos_token_id = [self.eos_token_id] if self.add_eos_token else []

        output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)

        if token_ids_1 is not None:
            output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)

        return output

    @property
    def default_chat_template(self):
        """
        LLaMA uses [INST] and [/INST] to indicate user messages, and <<SYS>> and <</SYS>> to indicate system messages.
        Assistant messages do not have special tokens, because LLaMA chat models are generally trained with strict
        user/assistant/user/assistant message ordering, and so assistant messages can be identified from the ordering
        rather than needing special tokens. The system message is partly 'embedded' in the first user message, which
        results in an unusual token ordering when it is present. This template should definitely be changed if you wish
        to fine-tune a model with more flexible role ordering!

        The output should look something like:

        <bos>[INST] B_SYS SystemPrompt E_SYS Prompt [/INST] Answer <eos><bos>[INST] Prompt [/INST] Answer <eos>
        <bos>[INST] Prompt [/INST]

        The reference for this chat template is [this code
        snippet](https://github.com/facebookresearch/llama/blob/556949fdfb72da27c2f4a40b7f0e4cf0b8153a28/llama/generation.py#L320-L362)
        in the original repository.
        """
        template = (
            "{% if messages[0]['role'] == 'system' %}"
            "{% set loop_messages = messages[1:] %}"  # Extract system message if it's present
            "{% set system_message = messages[0]['content'] %}"
            "{% elif USE_DEFAULT_PROMPT == true and not '<<SYS>>' in messages[0]['content'] %}"
            "{% set loop_messages = messages %}"  # Or use the default system message if the flag is set
            "{% set system_message = 'DEFAULT_SYSTEM_MESSAGE' %}"
            "{% else %}"
            "{% set loop_messages = messages %}"
            "{% set system_message = false %}"
            "{% endif %}"
            "{% for message in loop_messages %}"  # Loop over all non-system messages
            "{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}"
            "{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}"
            "{% endif %}"
            "{% if loop.index0 == 0 and system_message != false %}"  # Embed system message in first message
            "{% set content = '<<SYS>>\\n' + system_message + '\\n<</SYS>>\\n\\n' + message['content'] %}"
            "{% else %}"
            "{% set content = message['content'] %}"
            "{% endif %}"
            "{% if message['role'] == 'user' %}"  # After all of that, handle messages/roles in a fairly normal way
            "{{ bos_token + '[INST] ' + content.strip() + ' [/INST]' }}"
            "{% elif message['role'] == 'system' %}"
            "{{ '<<SYS>>\\n' + content.strip() + '\\n<</SYS>>\\n\\n' }}"
            "{% elif message['role'] == 'assistant' %}"
            "{{ ' '  + content.strip() + ' ' + eos_token }}"
            "{% endif %}"
            "{% endfor %}"
        )
        template = template.replace("USE_DEFAULT_PROMPT", "true" if self.use_default_system_prompt else "false")
        default_message = DEFAULT_SYSTEM_PROMPT.replace("\n", "\\n").replace("'", "\\'")
        template = template.replace("DEFAULT_SYSTEM_MESSAGE", default_message)

        return template
class MistralTokenizerFast(PreTrainedTokenizerFast):
    """
    Construct a Llama tokenizer. Based on byte-level Byte-Pair-Encoding.

    This uses notably ByteFallback and no normalization.

    ```python
    >>> from transformers import LlamaTokenizerFast

    >>> tokenizer = LlamaTokenizerFast.from_pretrained("hf-internal-testing/llama-tokenizer")
    >>> tokenizer.encode("Hello this is a test")
    [1, 15043, 445, 338, 263, 1243]
    ```

    If you want to change the `bos_token` or the `eos_token`, make sure to specify them when initializing the model, or
    call `tokenizer.update_post_processor()` to make sure that the post-processing is correctly done (otherwise the
    values of the first token and final token of an encoded sequence will not be correct). For more details, checkout
    [post-processors] (https://huggingface.co/docs/tokenizers/api/post-processors) documentation.


    This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
    refer to this superclass for more information regarding those methods.

    Args:
        vocab_file (`str`, *optional*):
            [SentencePiece](https://github.com/google/sentencepiece) file (generally has a .model extension) that
            contains the vocabulary necessary to instantiate a tokenizer.
        tokenizer_file (`str`, *optional*):
            [tokenizers](https://github.com/huggingface/tokenizers) file (generally has a .json extension) that
            contains everything needed to load the tokenizer.
        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
            Whether or not to cleanup spaces after decoding, cleanup consists in removing potential artifacts like
            extra spaces.
        unk_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<unk>"`):
            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
            token instead.
        bos_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"<s>"`):
            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
        eos_token (`str` or `tokenizers.AddedToken`, *optional*, defaults to `"</s>"`):
            The end of sequence token.
        add_bos_token (`bool`, *optional*, defaults to `True`):
            Whether or not to add an `bos_token` at the start of sequences.
        add_eos_token (`bool`, *optional*, defaults to `False`):
            Whether or not to add an `eos_token` at the end of sequences.
        use_default_system_prompt (`bool`, *optional*, defaults to `False`):
            Whether or not the default system prompt for Llama should be used
        legacy (`bool`, *optional*):
            Whether or not the `legacy` behavior of the tokenizer should be used. Legacy is before the merge of #24622
            and #25224 which includes fixes to properly handle tokens that appear after special tokens.
            Make sure to also set `from_slow` to `True`.
            A simple example:

            - `legacy=True`:
            ```python
            >>> from transformers import LlamaTokenizerFast

            >>> tokenizer = LlamaTokenizerFast.from_pretrained("huggyllama/llama-7b", legacy=True, from_slow=True)
            >>> tokenizer.encode("Hello <s>.") # 869 is '▁.'
            [1, 15043, 29871, 1, 869]
            ```
            - `legacy=False`:
            ```python
            >>> from transformers import LlamaTokenizerFast

            >>> tokenizer = LlamaTokenizerFast.from_pretrained("huggyllama/llama-7b", legacy=False, from_slow=True)
            >>> tokenizer.encode("Hello <s>.")  # 29889 is '.'
            [1, 15043, 29871, 1, 29889]
            ```
            Checkout the [pull request](https://github.com/huggingface/transformers/pull/24565) for more details.
        add_prefix_space (`bool`, *optional*):
            Whether or not the tokenizer should automatically add a prefix space
    """

    vocab_files_names = VOCAB_FILES_NAMES
    slow_tokenizer_class = MistralTokenizer
    padding_side = "left"
    model_input_names = ["input_ids", "attention_mask"]

    def __init__(
        self,
        vocab_file=None,
        tokenizer_file=None,
        clean_up_tokenization_spaces=False,
        unk_token="<unk>",
        bos_token="<s>",
        eos_token="</s>",
        add_bos_token=True,
        add_eos_token=False,
        use_default_system_prompt=False,
        legacy=None,
        add_prefix_space=None,
        **kwargs,
    ):
        if legacy is None:
            logger.warning_once(
                f"You are using the default legacy behaviour of the {self.__class__}. This is"
                " expected, and simply means that the `legacy` (previous) behavior will be used so nothing changes for you."
                " If you want to use the new behaviour, set `legacy=False`. This should only be set if you understand what it"
                " means, and thoroughly read the reason why this was added as explained in"
                " https://github.com/huggingface/transformers/pull/24565 - if you loaded a llama tokenizer from a GGUF file"
                " you can ignore this message."
            )
            legacy = True
        self.legacy = legacy

        if add_prefix_space is not None:
            kwargs["from_slow"] = True

        super().__init__(
            vocab_file=vocab_file,
            tokenizer_file=tokenizer_file,
            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
            unk_token=unk_token,
            bos_token=bos_token,
            eos_token=eos_token,
            add_bos_token=add_bos_token,
            add_eos_token=add_eos_token,
            use_default_system_prompt=use_default_system_prompt,
            add_prefix_space=add_prefix_space,
            legacy=legacy,
            **kwargs,
        )
        self._add_bos_token = add_bos_token
        self._add_eos_token = add_eos_token
        self.update_post_processor()
        self.use_default_system_prompt = use_default_system_prompt
        self.vocab_file = vocab_file

    @property
    def can_save_slow_tokenizer(self) -> bool:
        return os.path.isfile(self.vocab_file) if self.vocab_file else False

    def update_post_processor(self):
        """
        Updates the underlying post processor with the current `bos_token` and `eos_token`.
        """
        bos = self.bos_token
        bos_token_id = self.bos_token_id
        if bos is None and self.add_bos_token:
            raise ValueError("add_bos_token = True but bos_token = None")

        eos = self.eos_token
        eos_token_id = self.eos_token_id
        if eos is None and self.add_eos_token:
            raise ValueError("add_eos_token = True but eos_token = None")

        single = f"{(bos+':0 ') if self.add_bos_token else ''}$A:0{(' '+eos+':0') if self.add_eos_token else ''}"
        pair = f"{single}{(' '+bos+':1') if self.add_bos_token else ''} $B:1{(' '+eos+':1') if self.add_eos_token else ''}"

        special_tokens = []
        if self.add_bos_token:
            special_tokens.append((bos, bos_token_id))
        if self.add_eos_token:
            special_tokens.append((eos, eos_token_id))
        self._tokenizer.post_processor = processors.TemplateProcessing(
            single=single, pair=pair, special_tokens=special_tokens
        )

    @property
    def add_eos_token(self):
        return self._add_eos_token

    @property
    def add_bos_token(self):
        return self._add_bos_token

    @add_eos_token.setter
    def add_eos_token(self, value):
        self._add_eos_token = value
        self.update_post_processor()

    @add_bos_token.setter
    def add_bos_token(self, value):
        self._add_bos_token = value
        self.update_post_processor()

    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
        if not self.can_save_slow_tokenizer:
            raise ValueError(
                "Your fast tokenizer does not have the necessary information to save the vocabulary for a slow "
                "tokenizer."
            )

        if not os.path.isdir(save_directory):
            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
            return
        out_vocab_file = os.path.join(
            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
        )

        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
            copyfile(self.vocab_file, out_vocab_file)

        return (out_vocab_file,)

    @property
    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.default_chat_template
    def default_chat_template(self):
        """
        LLaMA uses [INST] and [/INST] to indicate user messages, and <<SYS>> and <</SYS>> to indicate system messages.
        Assistant messages do not have special tokens, because LLaMA chat models are generally trained with strict
        user/assistant/user/assistant message ordering, and so assistant messages can be identified from the ordering
        rather than needing special tokens. The system message is partly 'embedded' in the first user message, which
        results in an unusual token ordering when it is present. This template should definitely be changed if you wish
        to fine-tune a model with more flexible role ordering!

        The output should look something like:

        <bos>[INST] B_SYS SystemPrompt E_SYS Prompt [/INST] Answer <eos><bos>[INST] Prompt [/INST] Answer <eos>
        <bos>[INST] Prompt [/INST]

        The reference for this chat template is [this code
        snippet](https://github.com/facebookresearch/llama/blob/556949fdfb72da27c2f4a40b7f0e4cf0b8153a28/llama/generation.py#L320-L362)
        in the original repository.
        """
        template = (
            "{% if messages[0]['role'] == 'system' %}"
            "{% set loop_messages = messages[1:] %}"  # Extract system message if it's present
            "{% set system_message = messages[0]['content'] %}"
            "{% elif USE_DEFAULT_PROMPT == true and not '<<SYS>>' in messages[0]['content'] %}"
            "{% set loop_messages = messages %}"  # Or use the default system message if the flag is set
            "{% set system_message = 'DEFAULT_SYSTEM_MESSAGE' %}"
            "{% else %}"
            "{% set loop_messages = messages %}"
            "{% set system_message = false %}"
            "{% endif %}"
            "{% for message in loop_messages %}"  # Loop over all non-system messages
            "{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}"
            "{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}"
            "{% endif %}"
            "{% if loop.index0 == 0 and system_message != false %}"  # Embed system message in first message
            "{% set content = '<<SYS>>\\n' + system_message + '\\n<</SYS>>\\n\\n' + message['content'] %}"
            "{% else %}"
            "{% set content = message['content'] %}"
            "{% endif %}"
            "{% if message['role'] == 'user' %}"  # After all of that, handle messages/roles in a fairly normal way
            "{{ bos_token + '[INST] ' + content.strip() + ' [/INST]' }}"
            "{% elif message['role'] == 'system' %}"
            "{{ '<<SYS>>\\n' + content.strip() + '\\n<</SYS>>\\n\\n' }}"
            "{% elif message['role'] == 'assistant' %}"
            "{{ ' '  + content.strip() + ' ' + eos_token }}"
            "{% endif %}"
            "{% endfor %}"
        )
        template = template.replace("USE_DEFAULT_PROMPT", "true" if self.use_default_system_prompt else "false")
        default_message = DEFAULT_SYSTEM_PROMPT.replace("\n", "\\n").replace("'", "\\'")
        template = template.replace("DEFAULT_SYSTEM_MESSAGE", default_message)

        return template

    # TODO ArthurZ let's rely on the template processor instead, refactor all fast tokenizers
    # Copied from transformers.models.llama.tokenization_llama.LlamaTokenizer.build_inputs_with_special_tokens
    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
        eos_token_id = [self.eos_token_id] if self.add_eos_token else []

        output = bos_token_id + token_ids_0 + eos_token_id

        if token_ids_1 is not None:
            output = output + bos_token_id + token_ids_1 + eos_token_id

        return output
################################ Tokenizer ##############################




################################  UNIVERSAL NN COMPONENTS ################################
# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Mistral



# 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,
    )


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

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

class MistralRotaryEmbedding(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, dtype=torch.int64).float().to(device) / self.dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    @torch.no_grad()
    # Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding.forward
    def forward(self, x, position_ids):
        # x: [bs, num_attention_heads, seq_len, head_size]
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        position_ids_expanded = position_ids[:, None, :].float()
        # Force float32 since bfloat16 loses precision on long contexts
        # See https://github.com/huggingface/transformers/pull/29285
        device_type = x.device.type
        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()
        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
################################  UNIVERSAL Functions ################################
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()

# 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=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
    cos = cos.squeeze(1).squeeze(0)  # [seq_len, dim]
    sin = sin.squeeze(1).squeeze(0)  # [seq_len, dim]
    cos = cos[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
    sin = sin[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
    q_embed = (q * cos[:,:, -q.shape[2]:]) + (rotate_half(q) * sin[:,:, -q.shape[2]:]) if q is not None else None
    k_embed = (k * cos) + (rotate_half(k) * sin) if k is not None else None
    return q_embed, k_embed

def apply_grouped_rotary_pos_emb(q, k, cos, sin, position_ids, g_size_1=1, g_size_2=4096):
    # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
    position_ids_q = position_ids//g_size_1 + g_size_2 - g_size_2//g_size_1
    position_ids_k = position_ids//g_size_1

    cos = cos.squeeze(1).squeeze(0)  # [seq_len, dim]
    sin = sin.squeeze(1).squeeze(0)  # [seq_len, dim]
    cos_q = cos[position_ids_q].unsqueeze(1)  # [bs, 1, seq_len, dim]
    sin_q = sin[position_ids_q].unsqueeze(1)  # [bs, 1, seq_len, dim]
    cos_k = cos[position_ids_k].unsqueeze(1)  # [bs, 1, seq_len, dim]
    sin_k = sin[position_ids_k].unsqueeze(1)  # [bs, 1, seq_len, dim]
    q_embed = (q * cos_q) + (rotate_half(q) * sin_q) if q is not None else None
    k_embed = (k * cos_k) + (rotate_half(k) * sin_k) if k is not None else None

    return q_embed, k_embed

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

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

    Args:
        gate_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]):
            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
            shape [batch_size X sequence_length, num_experts].
        attention_mask (`torch.Tensor`, None):
            The attention_mask used in forward function
            shape [batch_size X sequence_length] if not None.
        num_experts (`int`, *optional*):
            Number of experts

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

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

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

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

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

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

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

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

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

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

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

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

class MistralMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        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, hidden_state):
        return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
# 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 MistralAttention(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: MistralConfig, layer_idx: Optional[int] = None):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.attention_dropout = config.attention_dropout
        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

        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.hidden_size, self.hidden_size, bias=False)

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

    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,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        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)

        cos, sin = self.rotary_emb(value_states, position_ids)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

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

        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 attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
            attn_weights = attn_weights + causal_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.view(bsz, q_len, -1)
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value
class MistralFlashAttention2(MistralAttention):
    """
    Mistral flash attention module. This module inherits from `MistralAttention` 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)


        # 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,
        cache_position: Optional[torch.LongTensor] = None,
    ):
        if isinstance(past_key_value, StaticCache):
            raise ValueError(
                "`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` "
                "make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformers"
            )

        output_attentions = False

        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 += cache_position[0]

        cos, sin = self.rotary_emb(value_states, position_ids)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        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 = _flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            q_len,
            dropout=dropout_rate,
            sliding_window=getattr(self.config, "sliding_window", None),
            use_top_left_mask=self._flash_attn_uses_top_left_mask,
            is_causal=self.is_causal,
        )

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

    # Adapted from MistralAttention.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,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if output_attentions:

            logger.warning_once(
                "MistralModel is using MistralSdpaAttention, 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,
                cache_position=cache_position,
            )

        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)

        cos, sin = self.rotary_emb(value_states, position_ids)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

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

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

        causal_mask = attention_mask
        if attention_mask is not None:
            causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]

        # 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 causal_mask is not None:
            query_states = query_states.contiguous()
            key_states = key_states.contiguous()
            value_states = value_states.contiguous()

        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
        is_causal = True if causal_mask is None and q_len > 1 else False

        attn_output = torch.nn.functional.scaled_dot_product_attention(
            query_states,
            key_states,
            value_states,
            attn_mask=causal_mask,
            dropout_p=self.attention_dropout if self.training else 0.0,
            is_causal=is_causal,
        )

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

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value
MISTRAL_ATTENTION_CLASSES = {
    "eager": MistralAttention,
    "flash_attention_2": MistralFlashAttention2,
    "sdpa": MistralSdpaAttention,
}

# Copied from transformers.models.llama.modeling_llama.LlamaDecoderLayer with Llama->Mistral, LLAMA->MISTRAL
class MistralDecoderLayer(nn.Module):
    def __init__(self, config: MistralConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

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

        self.mlp = MistralMLP(config)
        self.input_layernorm = MistralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = MistralRMSNorm(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[Cache] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*):
                attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
                query_sequence_length, key_sequence_length)` if default attention is used.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence
            kwargs (`dict`, *optional*):
                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
                into the model
        """
        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,
            cache_position=cache_position,
            **kwargs,
        )
        hidden_states = residual + 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

class MixtralBlockSparseTop2MLP(nn.Module):
    def __init__(self, config: MixtralConfig):
        super().__init__()
        self.ffn_dim = config.intermediate_size
        self.hidden_dim = config.hidden_size

        self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
        self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)
        self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)

        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, hidden_states):
        current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)
        current_hidden_states = self.w2(current_hidden_states)
        return current_hidden_states
class MixtralSparseMoeBlock(nn.Module):
    """
    This implementation is
    strictly equivalent to standard MoE with full capacity (no
    dropped tokens). It's faster since it formulates MoE operations
    in terms of block-sparse operations to accomodate imbalanced
    assignments of tokens to experts, whereas standard MoE either
    (1) drop tokens at the cost of reduced performance or (2) set
    capacity factor to number of experts and thus waste computation
    and memory on padding.
    """

    def __init__(self, config):
        super().__init__()
        self.hidden_dim = config.hidden_size
        self.ffn_dim = config.intermediate_size
        self.num_experts = config.num_local_experts
        self.top_k = config.num_experts_per_tok

        # gating
        self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)

        self.experts = nn.ModuleList([MixtralBlockSparseTop2MLP(config) for _ in range(self.num_experts)])

        # Jitter parameters
        self.jitter_noise = config.router_jitter_noise

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        """ """
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        if self.training and self.jitter_noise > 0:
            hidden_states *= torch.empty_like(hidden_states).uniform_(1.0 - self.jitter_noise, 1.0 + self.jitter_noise)
        hidden_states = hidden_states.view(-1, hidden_dim)
        # router_logits: (batch * sequence_length, n_experts)
        router_logits = self.gate(hidden_states)

        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
        # we cast back to the input dtype
        routing_weights = routing_weights.to(hidden_states.dtype)

        final_hidden_states = torch.zeros(
            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
        )

        # One hot encode the selected experts to create an expert mask
        # this will be used to easily index which expert is going to be sollicitated
        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)

        # Loop over all available experts in the model and perform the computation on each expert
        for expert_idx in range(self.num_experts):
            expert_layer = self.experts[expert_idx]
            idx, top_x = torch.where(expert_mask[expert_idx])

            # Index the correct hidden states and compute the expert hidden state for
            # the current expert. We need to make sure to multiply the output hidden
            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]

            # However `index_add_` only support torch tensors for indexing so we'll use
            # the `top_x` tensor here.
            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
        return final_hidden_states, router_logits
class MixtralDecoderLayer(nn.Module):
    def __init__(self, config: MixtralConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.self_attn = MISTRAL_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
        self.mlp = MistralMLP(config)
        self.block_sparse_moe = MixtralSparseMoeBlock(config)
        self.input_layernorm = MistralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = MistralRMSNorm(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,
        output_router_logits: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
        """
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
                `(batch, sequence_length)` where padding elements are indicated by 0.
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            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_router_logits (`bool`, *optional*):
                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
                should not be returned during inference.
            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`).
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence.
            kwargs (`dict`, *optional*):
                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
                into the model
        """

        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,
            cache_position=cache_position,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states, router_logits = self.block_sparse_moe(hidden_states)
        hidden_states = residual + 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,)

        if output_router_logits:
            outputs += (router_logits,)

        return outputs

################################  closed COMPONENTS ################################


############# Causal LM ################# 
class MistralForCausalLM(MistralPreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = MistralModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        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.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.print_final_only = True
        self.loss_mean = loss_mean
        self.all_rewards = []
        self.all_unreduced_losses = []
        self.kill_after = 100

        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.reinforce_temperature = 3
        self.base_loss_beta = 1

        # 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 = True
    
        # 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)
                )])

        # 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 calculate_policy_loss(self, thoughts, rewards):
        thought_log_probs = []
        for thought in thoughts:
            thought_log_prob = self.lm_head(thought).log_softmax(dim=-1)
            thought_log_probs.append(thought_log_prob)
        
        thought_log_probs = torch.stack(thought_log_probs, dim=1)  # (batch_size, num_thoughts, seq_length, vocab_size)
        thought_probs = torch.exp(thought_log_probs)
        
        policy_loss = -torch.mean(thought_log_probs * rewards.unsqueeze(-1).unsqueeze(-1))
        
        return policy_loss
    
    def _generate_thoughts(self, hidden_states, max_length):
        batch_size = hidden_states.size(0)
        thought_ids = torch.zeros((batch_size, self.config.max_thoughts, max_length), dtype=torch.long, device=hidden_states.device)
        thought_embeddings = []

        for i in range(self.config.max_thoughts):
            thought_input_ids = torch.zeros((batch_size, 1), dtype=torch.long, device=hidden_states.device)
            thought_outputs = self.generate(
                input_ids=thought_input_ids,
                max_length=max_length,
                do_sample=True,
                top_k=50,
                top_p=0.95,
                pad_token_id=self.config.pad_token_id,
                eos_token_id=self.config.eos_token_id,
            )
            thought_ids[:, i, :] = thought_outputs
            thought_embeddings.append(self.get_input_embeddings()(thought_outputs))

        thought_embeddings = torch.stack(thought_embeddings, dim=1)
        return thought_ids, thought_embeddings

    @torch.no_grad()
    def infer(
        self,
        input_ids: torch.LongTensor,
        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,
    ):
        batch_size, seq_len = input_ids.shape

        # Save the original input_ids and attention_mask for later use
        original_input_ids = input_ids.clone()
        original_attention_mask = attention_mask.clone() if attention_mask is not None else None

        # Append the start thought token to the input sequence
        start_thought_token_id = self.tokenizer.convert_tokens_to_ids("<|startthought|>")
        input_ids = torch.cat([input_ids, torch.tensor([[start_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1)
        seq_len += 1

        # Update the attention mask
        if attention_mask is not None:
            attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)

        # Generate the continuation
        continuation_length = self.n_ahead - 2
        new_key_values = past_key_values
        
        start_time = time.time()
        for continuation_idx in range(continuation_length):
            outputs = self.model(
                input_ids=input_ids if continuation_idx == 0 else next_token_id.unsqueeze(-1).to(input_ids.device),
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=new_key_values,
                inputs_embeds=inputs_embeds,
                use_cache=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
            new_key_values = outputs.past_key_values

            hidden_states = outputs[0]

            logits = self.lm_head(hidden_states)
            logits = logits[:, -1, :]  # Only consider the last token

            # Apply Gumbel-Softmax to the logits
            next_token_logits = F.gumbel_softmax(logits, tau=self.gumbel_temperature, hard=True, dim=-1)
            next_token_id = torch.argmax(next_token_logits, dim=-1)

            # Append the generated token to the input sequence
            input_ids = torch.cat([input_ids, next_token_id.unsqueeze(-1).to(input_ids.device)], dim=-1)
            seq_len += 1

            # Update the attention mask
            if attention_mask is not None:
                attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)

        # Append the end thought token to the input sequence
        end_thought_token_id = self.tokenizer.convert_tokens_to_ids("<|endthought|>")
        input_ids = torch.cat([input_ids, torch.tensor([[end_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1)
        seq_len += 1

        # Update the attention mask
        if attention_mask is not None:
            attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)

        # Get the hidden states before and after the thought
        outputs_before = self.model(
            input_ids=original_input_ids,
            attention_mask=original_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_before = outputs_before[0][:, -1:, :]

        # two new tokens: last continuation token and end thought token
        outputs_after = self.model(
            input_ids=torch.cat([next_token_id.unsqueeze(-1).to(input_ids.device), torch.tensor(end_thought_token_id).unsqueeze(-1).unsqueeze(-1).to(input_ids.device)], dim=-1),
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=new_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_after = outputs_after[0][:, -1:, :]

        # Apply the talk head to get the mixing weight
        mixing_weight = self.talk_head[0](torch.cat([hidden_states_before, hidden_states_after], dim=-1))

        # Apply the mixing weight to the hidden states
        mixed_hidden_states = (1 - mixing_weight) * hidden_states_before + mixing_weight * hidden_states_after

        # Apply the language model head to get the final logits
        logits = self.lm_head(mixed_hidden_states)
        return logits

    @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> 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, MistralForCausalLM

        >>> model = MistralForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")
        >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-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."
        ```"""
        log_dict = self.log_dict if self.training else self.eval_log_dict

        if self.training and self.kill_after is not None and self.training_steps // self.gradient_accumulation_steps > self.kill_after:
            raise ValueError("Killed after")

        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

        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):
                    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 = []

        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
                end_embedding = self.end_embedding[0].unsqueeze(0) * self.embedding_scale
            else:
                start_embedding = self.start_embedding * self.embedding_scale
                end_embedding = self.end_embedding * self.embedding_scale
            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,
                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")
                        shift_logits = shift_logits.view(-1, self.config.vocab_size)
                        shift_labels = shift_labels.view(-1).clone()
                        # 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
                        )
                        # 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)
                        skip_sampling = True
                    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))
                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)

                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)
                    if torch.any(unreduced_loss != unreduced_loss):
                        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
                    
                    if self.use_policy_loss and ahead_idx == self.n_ahead + self.n_ahead_talk - 2:
                        # only compute during the thinking phase
                        if self.use_reparam_for_thought_embeddings and (self.use_start_thought_token or self.use_end_thought_token):
                            # sampled_start, sampled_end
                            # calculate the log likelihood of the start and end embeddings sampled from a multivariate normal distribution
                            # with mean start_embedding[0] and standard deviation start_embedding[1]
                            if self.use_start_thought_token:
                                exp_start_std = torch.exp(start_embedding[1])
                                start_loglikelihood = -0.5 * (sampled_start.detach() - start_embedding[0]) ** 2 / exp_start_std ** 2 - start_embedding[1] - 0.5 * math.log(2 * math.pi)
                                start_loglikelihood = start_loglikelihood.mean(dim=-1)
                            if self.use_end_thought_token:
                                exp_end_std = torch.exp(end_embedding[1])
                                end_loglikelihood = -0.5 * (sampled_end.detach() - end_embedding[0]) ** 2 / exp_end_std ** 2 - end_embedding[1] - 0.5 * math.log(2 * math.pi)
                                end_loglikelihood = end_loglikelihood.mean(dim=-1)
                            # we use the mean instead of the sum to prevent dependence on the dimensionality of the embeddings
                            if self.use_end_thought_token and self.use_policy_loss_for_end_thought:
                                action_loglikelihoods_list.append(end_loglikelihood)
                            if self.use_start_thought_token:
                                action_loglikelihoods_list.append(start_loglikelihood)                                

                        if ahead_idx == self.n_ahead + self.n_ahead_talk - 2 and self.eval_mode:
                            with torch.no_grad():
                                # calculate the 0.75 quantile of the rewards
                                filtered_tokens = input_ids[:, :policy_reward.shape[-1]].cpu().detach().numpy().flatten()
                                filtered_tokens_mask = filtered_tokens != self.tokenizer.pad_token_id
                                filtered_tokens = filtered_tokens[filtered_tokens_mask]
                                filtered_rewards = policy_reward.float().cpu().detach().numpy()[:, :seq_len - self.n_ahead_talk].flatten()
                                filtered_rewards = filtered_rewards[filtered_tokens_mask]

                                abs_reward_list = np.abs(policy_reward.float().cpu().detach().numpy()[:, :seq_len - self.n_ahead_talk].flatten())
                                abs_reward_list = abs_reward_list[filtered_tokens_mask]
                                medium_quantile = np.quantile(abs_reward_list, 0.5)
                                upper_quantile = np.quantile(abs_reward_list, 0.95)

                                save_tokens_with_rewards_to_pdf(
                                    filtered_tokens,
                                    [0] + filtered_rewards.tolist(),
                                    self.tokenizer,
                                    output_file=f"texts/rewards_talk_{self.n_ahead_talk}_{self.training_steps}.pdf",
                                    eps=medium_quantile,
                                    eps2=upper_quantile,
                                )

                                def plot_kde(data, losses):
                                    sns.set(style="whitegrid")
                                    # Create the KDE plot
                                    sns.kdeplot(data, fill=True)
                                    # Set the plot title and labels
                                    plt.title("KDE Plot")
                                    plt.xlabel("Value")
                                    plt.ylabel("Density")
                                    # Save the plot
                                    plt.savefig(f"texts/kde_talk_{self.n_ahead_talk}_{self.training_steps}.pdf")
                                    # Close the plot
                                    plt.close()

                                    # Step 1: Create a base color palette
                                    base_colors = sns.color_palette("light:#5A9", n_colors=256)  # More colors for a smoother gradient
                                    base_cmap = LinearSegmentedColormap.from_list("log_light", base_colors)
                                    log_norm = LogNorm(vmin=1e-3, vmax=10)

                                    sns.kdeplot(x=data, y=losses, fill=True, levels=20, norm=log_norm, cut=0, linewidths=0)
                                    # limit y to 0 to 25 and x to -1 to 1
                                    plt.xlim(-1, 1)
                                    plt.ylim(0, 25)
                                    plt.savefig(f"texts/jointer_talk_{self.n_ahead_talk}_{self.training_steps}.pdf")
                                    plt.close()

                                self.all_rewards.extend(filtered_rewards)
                                self.all_unreduced_losses.extend(unreduced_loss[:, :-1].flatten()[filtered_tokens_mask].float().flatten().cpu().detach().numpy())
                                plot_kde(self.all_rewards, self.all_unreduced_losses)

                        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 * 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()
        
        for loss_key, loss_val in base_log_dict.items():
            log_dict[loss_key] += loss_val / self.n_tokens_print
                
        if self.use_policy_loss and policy_reward is not None:
            log_dict["policy_loss"] += dqn_loss / self.n_tokens_print
            log_dict["policy_reward"] += policy_reward.mean() / self.n_tokens_print

        if not loss_list:
            if loss is not None:
                log_dict["loss_0"] += loss / self.n_tokens_print
        else:
            log_dict["loss_final"] += nonzero_mean(loss_list[-1]) / self.n_tokens_print
            log_dict["loss_talk"] += sum(nonzero_mean(cur_loss_item) for cur_loss_item in loss_list[-self.n_ahead_talk:]) / self.n_ahead_talk / self.n_tokens_print

        # also log relative losses to loss_0
        if loss_list:
            for i in range(len(loss_list)):
                talk_idx = min(max(i - (self.n_ahead - 1), 0), len(talk_loss_list) - 1)
                if not talk_loss_list:
                    cur_talk_loss = nonzero_mean(loss_list[0])
                else:
                    cur_talk_loss = talk_loss_list[talk_idx]
                log_dict[f"rel_loss_{i}"] += (nonzero_mean(loss_list[i]) - cur_talk_loss) / self.n_tokens_print
        if self.training:
            self.training_steps += 1
        try:
            # if self.training_steps % (self.gradient_accumulation_steps * 256) == 0:
            if self.wandb_enabled:
                if self.training_steps % (self.n_tokens_print) == 0 or not self.training:# and "0" in str(loss.device):
                    if not self.training:
                        new_log_dict = {}
                        for key in list(log_dict.keys()):
                            new_log_dict["eval_" + key] = log_dict[key]
                        log_dict = new_log_dict
                    log_dict["training_steps"] = self.training_steps 
                    log_dict["batch_size"] = batch_size
                    log_dict["example_steps"] = self.training_steps * batch_size * self.gradient_accumulation_steps
                    if self.n_ahead > 1:
                        log_dict["compute_steps"] = self.training_steps * batch_size * (self.n_ahead + self.n_ahead_talk - 1) * self.gradient_accumulation_steps
                    else: # There's no overhead for talk tokens if there's no thinking
                        log_dict["compute_steps"] = self.training_steps * batch_size * self.gradient_accumulation_steps
                    # remove all nans
                    for key in list(log_dict.keys()):
                        if log_dict[key] != log_dict[key]:
                            del log_dict[key]
                    if self.training:
                        wandb.log(log_dict)
                    if self.training:
                        self.log_dict = defaultdict(int)
                    else:
                        self.eval_log_dict = defaultdict(int)
        except Exception as e:
            pass

        if not self.training:
            self.n_ahead_talk = n_ahead_talk_to_restore
            self.n_passes = n_passes_to_restore
        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,
        )

    @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward_quiet(
        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, 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."
        ```"""

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

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            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=True,
        )
        hidden_states = outputs.last_hidden_state
        logits = self.lm_head(hidden_states)

        thought_ids, thought_embeddings = self._generate_thoughts(hidden_states, max_length=self.config.thought_length)
        thought_hidden_states = self.model(inputs_embeds=thought_embeddings).last_hidden_state

        # Compute thought logits
        thought_logits = self.lm_head(thought_hidden_states)

        # Mix base and thought logits
        mixed_logits = logits.unsqueeze(1) + self.mixing_head(thought_logits)
        mixed_logits = mixed_logits.view(-1, mixed_logits.size(-1))

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = mixed_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

            if self.use_policy_loss:
                rewards = loss.detach().unsqueeze(1).repeat(1, self.max_thoughts)
                if self.remove_negative_rewards:
                    rewards = torch.clamp(rewards, min=0)
                policy_loss = self.calculate_policy_loss(thought_ids, rewards)
                loss = loss + policy_loss
        else:
            loss = None

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

        return CausalLMOutputWithPast(
            loss=loss if loss is not None else None,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )
    @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward_legacy(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> 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, MistralForCausalLM

        >>> model = MistralForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")
        >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-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."
        ```"""

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

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )

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

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

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

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

    def self_extend_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,
        padding_mask: Optional[torch.LongTensor] = None,
        group_size_1: Optional[float] = 8,
        group_size_2: Optional[float] = 2048,
        **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)
    
        
        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)
        
        query_position_ids = position_ids
        key_position_ids = torch.arange(kv_seq_len, dtype=position_ids.dtype).to(query_position_ids.device).view(bsz, kv_seq_len)
    
    
        neighbor_query_states, _ = apply_rotary_pos_emb(query_states, None, cos, sin, query_position_ids) 
        _, neighbor_key_states = apply_rotary_pos_emb(None, key_states, cos, sin, key_position_ids) 
        _re_group_size_2 = 0 if position_ids.max() < group_size_2 else group_size_2 # in case that, the smallest q position, g2-g2//g1 exceed the max position
        group_query_states, _ = apply_grouped_rotary_pos_emb(query_states, None, cos, sin, query_position_ids, g_size_1=group_size_1, g_size_2=_re_group_size_2) 
        _, group_key_states = apply_grouped_rotary_pos_emb(None, key_states, cos, sin, key_position_ids, g_size_1=group_size_1, g_size_2=_re_group_size_2) 
    
    
        group_key_states = repeat_kv(group_key_states, self.num_key_value_groups)
        neighbor_key_states = repeat_kv(neighbor_key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)
    
        neighbor_attn_weights = torch.matmul(neighbor_query_states, neighbor_key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
        group_attn_weights = torch.matmul(group_query_states, group_key_states.transpose(2, 3)) / math.sqrt(self.head_dim) 
    
    
        if group_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" {group_attn_weights.size()}"
            )
        
        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()}"
                )
            group_attn_weights = group_attn_weights + attention_mask
            neighbor_attn_weights = neighbor_attn_weights + attention_mask
    
    
        if q_len == 1:
            neighbor_attention_mask = torch.zeros((q_len, kv_seq_len), device=neighbor_attn_weights.device)
            neighbor_attention_mask[:, -group_size_2:] = 1
        elif q_len == kv_seq_len:
            neighbor_attention_mask = torch.ones((q_len, kv_seq_len), device=neighbor_attn_weights.device)
            neighbor_attention_mask = torch.tril(neighbor_attention_mask)
            if q_len-group_size_2 > 0:
                group_attention_mask =  torch.tril(torch.ones((q_len-group_size_2, kv_seq_len-group_size_2), device=group_attn_weights.device))
                neighbor_attention_mask[group_size_2:, :-group_size_2] -= group_attention_mask
    
        else:
            raise ValueError("q_len should be 1 or seq_len.")
    
    
        neighbor_attention_mask = neighbor_attention_mask.bool()
        attn_weights = torch.where(neighbor_attention_mask, neighbor_attn_weights, group_attn_weights)
        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

    @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forwardStar(
        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, 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, MistralForCausalLM

        >>> model = MistralForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")
        >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-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."
        ```"""
        log_dict = self.log_dict if self.training else self.eval_log_dict

        if self.training and self.kill_after is not None and self.training_steps // self.gradient_accumulation_steps > self.kill_after:
            raise ValueError("Killed after")

        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

        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):
                    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 = []

        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
                end_embedding = self.end_embedding[0].unsqueeze(0) * self.embedding_scale
            else:
                start_embedding = self.start_embedding * self.embedding_scale
                end_embedding = self.end_embedding * self.embedding_scale
            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,
                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")
                        shift_logits = shift_logits.view(-1, self.config.vocab_size)
                        shift_labels = shift_labels.view(-1).clone()
                        # 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
                        )
                        # 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)
                        skip_sampling = True
                    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))
                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)

                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)
                    if torch.any(unreduced_loss != unreduced_loss):
                        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
                    
                    if self.use_policy_loss and ahead_idx == self.n_ahead + self.n_ahead_talk - 2:
                        # only compute during the thinking phase
                        if self.use_reparam_for_thought_embeddings and (self.use_start_thought_token or self.use_end_thought_token):
                            # sampled_start, sampled_end
                            # calculate the log likelihood of the start and end embeddings sampled from a multivariate normal distribution
                            # with mean start_embedding[0] and standard deviation start_embedding[1]
                            if self.use_start_thought_token:
                                exp_start_std = torch.exp(start_embedding[1])
                                start_loglikelihood = -0.5 * (sampled_start.detach() - start_embedding[0]) ** 2 / exp_start_std ** 2 - start_embedding[1] - 0.5 * math.log(2 * math.pi)
                                start_loglikelihood = start_loglikelihood.mean(dim=-1)
                            if self.use_end_thought_token:
                                exp_end_std = torch.exp(end_embedding[1])
                                end_loglikelihood = -0.5 * (sampled_end.detach() - end_embedding[0]) ** 2 / exp_end_std ** 2 - end_embedding[1] - 0.5 * math.log(2 * math.pi)
                                end_loglikelihood = end_loglikelihood.mean(dim=-1)
                            # we use the mean instead of the sum to prevent dependence on the dimensionality of the embeddings
                            if self.use_end_thought_token and self.use_policy_loss_for_end_thought:
                                action_loglikelihoods_list.append(end_loglikelihood)
                            if self.use_start_thought_token:
                                action_loglikelihoods_list.append(start_loglikelihood)                                

                        if ahead_idx == self.n_ahead + self.n_ahead_talk - 2 and self.eval_mode:
                            with torch.no_grad():
                                # calculate the 0.75 quantile of the rewards
                                filtered_tokens = input_ids[:, :policy_reward.shape[-1]].cpu().detach().numpy().flatten()
                                filtered_tokens_mask = filtered_tokens != self.tokenizer.pad_token_id
                                filtered_tokens = filtered_tokens[filtered_tokens_mask]
                                filtered_rewards = policy_reward.float().cpu().detach().numpy()[:, :seq_len - self.n_ahead_talk].flatten()
                                filtered_rewards = filtered_rewards[filtered_tokens_mask]

                                abs_reward_list = np.abs(policy_reward.float().cpu().detach().numpy()[:, :seq_len - self.n_ahead_talk].flatten())
                                abs_reward_list = abs_reward_list[filtered_tokens_mask]
                                medium_quantile = np.quantile(abs_reward_list, 0.5)
                                upper_quantile = np.quantile(abs_reward_list, 0.95)

                        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 * 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()
        
        for loss_key, loss_val in base_log_dict.items():
            log_dict[loss_key] += loss_val / self.n_tokens_print
                
        if self.use_policy_loss and policy_reward is not None:
            log_dict["policy_loss"] += dqn_loss / self.n_tokens_print
            log_dict["policy_reward"] += policy_reward.mean() / self.n_tokens_print

        if not loss_list:
            if loss is not None:
                log_dict["loss_0"] += loss / self.n_tokens_print
        else:
            log_dict["loss_final"] += nonzero_mean(loss_list[-1]) / self.n_tokens_print
            log_dict["loss_talk"] += sum(nonzero_mean(cur_loss_item) for cur_loss_item in loss_list[-self.n_ahead_talk:]) / self.n_ahead_talk / self.n_tokens_print

        # also log relative losses to loss_0
        if loss_list:
            for i in range(len(loss_list)):
                talk_idx = min(max(i - (self.n_ahead - 1), 0), len(talk_loss_list) - 1)
                if not talk_loss_list:
                    cur_talk_loss = nonzero_mean(loss_list[0])
                else:
                    cur_talk_loss = talk_loss_list[talk_idx]
                log_dict[f"rel_loss_{i}"] += (nonzero_mean(loss_list[i]) - cur_talk_loss) / self.n_tokens_print
        if self.training:
            self.training_steps += 1

        if not self.training:
            self.n_ahead_talk = n_ahead_talk_to_restore
            self.n_passes = n_passes_to_restore
        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 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

class MistralSelfExtendForCausalLM(MistralPreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = MistralModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        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.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.print_final_only = True
        self.loss_mean = loss_mean
        self.all_rewards = []
        self.all_unreduced_losses = []
        self.kill_after = 100

        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.reinforce_temperature = 3
        self.base_loss_beta = 1

        # 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 = True
    
        # 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)
                )])

        # 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 calculate_policy_loss(self, thoughts, rewards):
        thought_log_probs = []
        for thought in thoughts:
            thought_log_prob = self.lm_head(thought).log_softmax(dim=-1)
            thought_log_probs.append(thought_log_prob)
        
        thought_log_probs = torch.stack(thought_log_probs, dim=1)  # (batch_size, num_thoughts, seq_length, vocab_size)
        thought_probs = torch.exp(thought_log_probs)
        
        policy_loss = -torch.mean(thought_log_probs * rewards.unsqueeze(-1).unsqueeze(-1))
        
        return policy_loss
    
    def _generate_thoughts(self, hidden_states, max_length):
        batch_size = hidden_states.size(0)
        thought_ids = torch.zeros((batch_size, self.config.max_thoughts, max_length), dtype=torch.long, device=hidden_states.device)
        thought_embeddings = []

        for i in range(self.config.max_thoughts):
            thought_input_ids = torch.zeros((batch_size, 1), dtype=torch.long, device=hidden_states.device)
            thought_outputs = self.generate(
                input_ids=thought_input_ids,
                max_length=max_length,
                do_sample=True,
                top_k=50,
                top_p=0.95,
                pad_token_id=self.config.pad_token_id,
                eos_token_id=self.config.eos_token_id,
            )
            thought_ids[:, i, :] = thought_outputs
            thought_embeddings.append(self.get_input_embeddings()(thought_outputs))

        thought_embeddings = torch.stack(thought_embeddings, dim=1)
        return thought_ids, thought_embeddings

    @torch.no_grad()
    def infer(
        self,
        input_ids: torch.LongTensor,
        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,
    ):
        batch_size, seq_len = input_ids.shape

        # Save the original input_ids and attention_mask for later use
        original_input_ids = input_ids.clone()
        original_attention_mask = attention_mask.clone() if attention_mask is not None else None

        # Append the start thought token to the input sequence
        start_thought_token_id = self.tokenizer.convert_tokens_to_ids("<|startthought|>")
        input_ids = torch.cat([input_ids, torch.tensor([[start_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1)
        seq_len += 1

        # Update the attention mask
        if attention_mask is not None:
            attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)

        # Generate the continuation
        continuation_length = self.n_ahead - 2
        new_key_values = past_key_values
        
        start_time = time.time()
        for continuation_idx in range(continuation_length):
            outputs = self.model(
                input_ids=input_ids if continuation_idx == 0 else next_token_id.unsqueeze(-1).to(input_ids.device),
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=new_key_values,
                inputs_embeds=inputs_embeds,
                use_cache=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
            new_key_values = outputs.past_key_values

            hidden_states = outputs[0]

            logits = self.lm_head(hidden_states)
            logits = logits[:, -1, :]  # Only consider the last token

            # Apply Gumbel-Softmax to the logits
            next_token_logits = F.gumbel_softmax(logits, tau=self.gumbel_temperature, hard=True, dim=-1)
            next_token_id = torch.argmax(next_token_logits, dim=-1)

            # Append the generated token to the input sequence
            input_ids = torch.cat([input_ids, next_token_id.unsqueeze(-1).to(input_ids.device)], dim=-1)
            seq_len += 1

            # Update the attention mask
            if attention_mask is not None:
                attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)

        # Append the end thought token to the input sequence
        end_thought_token_id = self.tokenizer.convert_tokens_to_ids("<|endthought|>")
        input_ids = torch.cat([input_ids, torch.tensor([[end_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1)
        seq_len += 1

        # Update the attention mask
        if attention_mask is not None:
            attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)

        # Get the hidden states before and after the thought
        outputs_before = self.model(
            input_ids=original_input_ids,
            attention_mask=original_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_before = outputs_before[0][:, -1:, :]

        # two new tokens: last continuation token and end thought token
        outputs_after = self.model(
            input_ids=torch.cat([next_token_id.unsqueeze(-1).to(input_ids.device), torch.tensor(end_thought_token_id).unsqueeze(-1).unsqueeze(-1).to(input_ids.device)], dim=-1),
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=new_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_after = outputs_after[0][:, -1:, :]

        # Apply the talk head to get the mixing weight
        mixing_weight = self.talk_head[0](torch.cat([hidden_states_before, hidden_states_after], dim=-1))

        # Apply the mixing weight to the hidden states
        mixed_hidden_states = (1 - mixing_weight) * hidden_states_before + mixing_weight * hidden_states_after

        # Apply the language model head to get the final logits
        logits = self.lm_head(mixed_hidden_states)
        return logits

    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,
        padding_mask: Optional[torch.LongTensor] = None,
        group_size_1: Optional[float] = 8,
        group_size_2: Optional[float] = 2048,
        **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)
    
        
        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)
        
        query_position_ids = position_ids
        key_position_ids = torch.arange(kv_seq_len, dtype=position_ids.dtype).to(query_position_ids.device).view(bsz, kv_seq_len)
    
    
        neighbor_query_states, _ = apply_rotary_pos_emb(query_states, None, cos, sin, query_position_ids) 
        _, neighbor_key_states = apply_rotary_pos_emb(None, key_states, cos, sin, key_position_ids) 
        _re_group_size_2 = 0 if position_ids.max() < group_size_2 else group_size_2 # in case that, the smallest q position, g2-g2//g1 exceed the max position
        group_query_states, _ = apply_grouped_rotary_pos_emb(query_states, None, cos, sin, query_position_ids, g_size_1=group_size_1, g_size_2=_re_group_size_2) 
        _, group_key_states = apply_grouped_rotary_pos_emb(None, key_states, cos, sin, key_position_ids, g_size_1=group_size_1, g_size_2=_re_group_size_2) 
    
    
        group_key_states = repeat_kv(group_key_states, self.num_key_value_groups)
        neighbor_key_states = repeat_kv(neighbor_key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)
    
        neighbor_attn_weights = torch.matmul(neighbor_query_states, neighbor_key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
        group_attn_weights = torch.matmul(group_query_states, group_key_states.transpose(2, 3)) / math.sqrt(self.head_dim) 
    
    
        if group_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" {group_attn_weights.size()}"
            )
        
        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()}"
                )
            group_attn_weights = group_attn_weights + attention_mask
            neighbor_attn_weights = neighbor_attn_weights + attention_mask
    
    
        if q_len == 1:
            neighbor_attention_mask = torch.zeros((q_len, kv_seq_len), device=neighbor_attn_weights.device)
            neighbor_attention_mask[:, -group_size_2:] = 1
        elif q_len == kv_seq_len:
            neighbor_attention_mask = torch.ones((q_len, kv_seq_len), device=neighbor_attn_weights.device)
            neighbor_attention_mask = torch.tril(neighbor_attention_mask)
            if q_len-group_size_2 > 0:
                group_attention_mask =  torch.tril(torch.ones((q_len-group_size_2, kv_seq_len-group_size_2), device=group_attn_weights.device))
                neighbor_attention_mask[group_size_2:, :-group_size_2] -= group_attention_mask
    
        else:
            raise ValueError("q_len should be 1 or seq_len.")
    
    
        neighbor_attention_mask = neighbor_attention_mask.bool()
        attn_weights = torch.where(neighbor_attention_mask, neighbor_attn_weights, group_attn_weights)
        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

    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

class MistralStarForCausalLM(MistralPreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = MistralModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        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.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.print_final_only = True
        self.loss_mean = loss_mean
        self.all_rewards = []
        self.all_unreduced_losses = []
        self.kill_after = 100

        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.reinforce_temperature = 3
        self.base_loss_beta = 1

        # 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 = True
    
        # 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)
                )])

        # 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 calculate_policy_loss(self, thoughts, rewards):
        thought_log_probs = []
        for thought in thoughts:
            thought_log_prob = self.lm_head(thought).log_softmax(dim=-1)
            thought_log_probs.append(thought_log_prob)
        
        thought_log_probs = torch.stack(thought_log_probs, dim=1)  # (batch_size, num_thoughts, seq_length, vocab_size)
        thought_probs = torch.exp(thought_log_probs)
        
        policy_loss = -torch.mean(thought_log_probs * rewards.unsqueeze(-1).unsqueeze(-1))
        
        return policy_loss
    
    def _generate_thoughts(self, hidden_states, max_length):
        batch_size = hidden_states.size(0)
        thought_ids = torch.zeros((batch_size, self.config.max_thoughts, max_length), dtype=torch.long, device=hidden_states.device)
        thought_embeddings = []

        for i in range(self.config.max_thoughts):
            thought_input_ids = torch.zeros((batch_size, 1), dtype=torch.long, device=hidden_states.device)
            thought_outputs = self.generate(
                input_ids=thought_input_ids,
                max_length=max_length,
                do_sample=True,
                top_k=50,
                top_p=0.95,
                pad_token_id=self.config.pad_token_id,
                eos_token_id=self.config.eos_token_id,
            )
            thought_ids[:, i, :] = thought_outputs
            thought_embeddings.append(self.get_input_embeddings()(thought_outputs))

        thought_embeddings = torch.stack(thought_embeddings, dim=1)
        return thought_ids, thought_embeddings

    @torch.no_grad()
    def infer(
        self,
        input_ids: torch.LongTensor,
        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,
    ):
        batch_size, seq_len = input_ids.shape

        # Save the original input_ids and attention_mask for later use
        original_input_ids = input_ids.clone()
        original_attention_mask = attention_mask.clone() if attention_mask is not None else None

        # Append the start thought token to the input sequence
        start_thought_token_id = self.tokenizer.convert_tokens_to_ids("<|startthought|>")
        input_ids = torch.cat([input_ids, torch.tensor([[start_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1)
        seq_len += 1

        # Update the attention mask
        if attention_mask is not None:
            attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)

        # Generate the continuation
        continuation_length = self.n_ahead - 2
        new_key_values = past_key_values
        
        start_time = time.time()
        for continuation_idx in range(continuation_length):
            outputs = self.model(
                input_ids=input_ids if continuation_idx == 0 else next_token_id.unsqueeze(-1).to(input_ids.device),
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=new_key_values,
                inputs_embeds=inputs_embeds,
                use_cache=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
            new_key_values = outputs.past_key_values

            hidden_states = outputs[0]

            logits = self.lm_head(hidden_states)
            logits = logits[:, -1, :]  # Only consider the last token

            # Apply Gumbel-Softmax to the logits
            next_token_logits = F.gumbel_softmax(logits, tau=self.gumbel_temperature, hard=True, dim=-1)
            next_token_id = torch.argmax(next_token_logits, dim=-1)

            # Append the generated token to the input sequence
            input_ids = torch.cat([input_ids, next_token_id.unsqueeze(-1).to(input_ids.device)], dim=-1)
            seq_len += 1

            # Update the attention mask
            if attention_mask is not None:
                attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)

        # Append the end thought token to the input sequence
        end_thought_token_id = self.tokenizer.convert_tokens_to_ids("<|endthought|>")
        input_ids = torch.cat([input_ids, torch.tensor([[end_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1)
        seq_len += 1

        # Update the attention mask
        if attention_mask is not None:
            attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)

        # Get the hidden states before and after the thought
        outputs_before = self.model(
            input_ids=original_input_ids,
            attention_mask=original_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_before = outputs_before[0][:, -1:, :]

        # two new tokens: last continuation token and end thought token
        outputs_after = self.model(
            input_ids=torch.cat([next_token_id.unsqueeze(-1).to(input_ids.device), torch.tensor(end_thought_token_id).unsqueeze(-1).unsqueeze(-1).to(input_ids.device)], dim=-1),
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=new_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_after = outputs_after[0][:, -1:, :]

        # Apply the talk head to get the mixing weight
        mixing_weight = self.talk_head[0](torch.cat([hidden_states_before, hidden_states_after], dim=-1))

        # Apply the mixing weight to the hidden states
        mixed_hidden_states = (1 - mixing_weight) * hidden_states_before + mixing_weight * hidden_states_after

        # Apply the language model head to get the final logits
        logits = self.lm_head(mixed_hidden_states)
        return logits

    @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward_quiet(
        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, 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."
        ```"""

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

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            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=True,
        )
        hidden_states = outputs.last_hidden_state
        logits = self.lm_head(hidden_states)

        thought_ids, thought_embeddings = self._generate_thoughts(hidden_states, max_length=self.config.thought_length)
        thought_hidden_states = self.model(inputs_embeds=thought_embeddings).last_hidden_state

        # Compute thought logits
        thought_logits = self.lm_head(thought_hidden_states)

        # Mix base and thought logits
        mixed_logits = logits.unsqueeze(1) + self.mixing_head(thought_logits)
        mixed_logits = mixed_logits.view(-1, mixed_logits.size(-1))

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = mixed_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

            if self.use_policy_loss:
                rewards = loss.detach().unsqueeze(1).repeat(1, self.max_thoughts)
                if self.remove_negative_rewards:
                    rewards = torch.clamp(rewards, min=0)
                policy_loss = self.calculate_policy_loss(thought_ids, rewards)
                loss = loss + policy_loss
        else:
            loss = None

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

        return CausalLMOutputWithPast(
            loss=loss if loss is not None else None,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> 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, MistralForCausalLM

        >>> model = MistralForCausalLM.from_pretrained("mistralai/Mistral-7B-v0.1")
        >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-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."
        ```"""
        log_dict = self.log_dict if self.training else self.eval_log_dict

        if self.training and self.kill_after is not None and self.training_steps // self.gradient_accumulation_steps > self.kill_after:
            raise ValueError("Killed after")

        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

        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):
                    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 = []

        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
                end_embedding = self.end_embedding[0].unsqueeze(0) * self.embedding_scale
            else:
                start_embedding = self.start_embedding * self.embedding_scale
                end_embedding = self.end_embedding * self.embedding_scale
            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,
                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")
                        shift_logits = shift_logits.view(-1, self.config.vocab_size)
                        shift_labels = shift_labels.view(-1).clone()
                        # 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
                        )
                        # 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)
                        skip_sampling = True
                    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))
                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)

                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)
                    if torch.any(unreduced_loss != unreduced_loss):
                        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
                    
                    if self.use_policy_loss and ahead_idx == self.n_ahead + self.n_ahead_talk - 2:
                        # only compute during the thinking phase
                        if self.use_reparam_for_thought_embeddings and (self.use_start_thought_token or self.use_end_thought_token):
                            # sampled_start, sampled_end
                            # calculate the log likelihood of the start and end embeddings sampled from a multivariate normal distribution
                            # with mean start_embedding[0] and standard deviation start_embedding[1]
                            if self.use_start_thought_token:
                                exp_start_std = torch.exp(start_embedding[1])
                                start_loglikelihood = -0.5 * (sampled_start.detach() - start_embedding[0]) ** 2 / exp_start_std ** 2 - start_embedding[1] - 0.5 * math.log(2 * math.pi)
                                start_loglikelihood = start_loglikelihood.mean(dim=-1)
                            if self.use_end_thought_token:
                                exp_end_std = torch.exp(end_embedding[1])
                                end_loglikelihood = -0.5 * (sampled_end.detach() - end_embedding[0]) ** 2 / exp_end_std ** 2 - end_embedding[1] - 0.5 * math.log(2 * math.pi)
                                end_loglikelihood = end_loglikelihood.mean(dim=-1)
                            # we use the mean instead of the sum to prevent dependence on the dimensionality of the embeddings
                            if self.use_end_thought_token and self.use_policy_loss_for_end_thought:
                                action_loglikelihoods_list.append(end_loglikelihood)
                            if self.use_start_thought_token:
                                action_loglikelihoods_list.append(start_loglikelihood)                                

                        if ahead_idx == self.n_ahead + self.n_ahead_talk - 2 and self.eval_mode:
                            with torch.no_grad():
                                # calculate the 0.75 quantile of the rewards
                                filtered_tokens = input_ids[:, :policy_reward.shape[-1]].cpu().detach().numpy().flatten()
                                filtered_tokens_mask = filtered_tokens != self.tokenizer.pad_token_id
                                filtered_tokens = filtered_tokens[filtered_tokens_mask]
                                filtered_rewards = policy_reward.float().cpu().detach().numpy()[:, :seq_len - self.n_ahead_talk].flatten()
                                filtered_rewards = filtered_rewards[filtered_tokens_mask]

                                abs_reward_list = np.abs(policy_reward.float().cpu().detach().numpy()[:, :seq_len - self.n_ahead_talk].flatten())
                                abs_reward_list = abs_reward_list[filtered_tokens_mask]
                                medium_quantile = np.quantile(abs_reward_list, 0.5)
                                upper_quantile = np.quantile(abs_reward_list, 0.95)

                        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 * 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()
        
        for loss_key, loss_val in base_log_dict.items():
            log_dict[loss_key] += loss_val / self.n_tokens_print
                
        if self.use_policy_loss and policy_reward is not None:
            log_dict["policy_loss"] += dqn_loss / self.n_tokens_print
            log_dict["policy_reward"] += policy_reward.mean() / self.n_tokens_print

        if not loss_list:
            if loss is not None:
                log_dict["loss_0"] += loss / self.n_tokens_print
        else:
            log_dict["loss_final"] += nonzero_mean(loss_list[-1]) / self.n_tokens_print
            log_dict["loss_talk"] += sum(nonzero_mean(cur_loss_item) for cur_loss_item in loss_list[-self.n_ahead_talk:]) / self.n_ahead_talk / self.n_tokens_print

        # also log relative losses to loss_0
        if loss_list:
            for i in range(len(loss_list)):
                talk_idx = min(max(i - (self.n_ahead - 1), 0), len(talk_loss_list) - 1)
                if not talk_loss_list:
                    cur_talk_loss = nonzero_mean(loss_list[0])
                else:
                    cur_talk_loss = talk_loss_list[talk_idx]
                log_dict[f"rel_loss_{i}"] += (nonzero_mean(loss_list[i]) - cur_talk_loss) / self.n_tokens_print
        if self.training:
            self.training_steps += 1

        if not self.training:
            self.n_ahead_talk = n_ahead_talk_to_restore
            self.n_passes = n_passes_to_restore
        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 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

class MistralQuietForCausalLM(MistralPreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = MistralModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        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.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.print_final_only = True
        self.loss_mean = loss_mean
        self.all_rewards = []
        self.all_unreduced_losses = []
        self.kill_after = 100

        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.reinforce_temperature = 3
        self.base_loss_beta = 1

        # 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 = True
    
        # 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)
                )])

        # 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 calculate_policy_loss(self, thoughts, rewards):
        thought_log_probs = []
        for thought in thoughts:
            thought_log_prob = self.lm_head(thought).log_softmax(dim=-1)
            thought_log_probs.append(thought_log_prob)
        
        thought_log_probs = torch.stack(thought_log_probs, dim=1)  # (batch_size, num_thoughts, seq_length, vocab_size)
        thought_probs = torch.exp(thought_log_probs)
        
        policy_loss = -torch.mean(thought_log_probs * rewards.unsqueeze(-1).unsqueeze(-1))
        
        return policy_loss
    
    def _generate_thoughts(self, hidden_states, max_length):
        batch_size = hidden_states.size(0)
        thought_ids = torch.zeros((batch_size, self.config.max_thoughts, max_length), dtype=torch.long, device=hidden_states.device)
        thought_embeddings = []

        for i in range(self.config.max_thoughts):
            thought_input_ids = torch.zeros((batch_size, 1), dtype=torch.long, device=hidden_states.device)
            thought_outputs = self.generate(
                input_ids=thought_input_ids,
                max_length=max_length,
                do_sample=True,
                top_k=50,
                top_p=0.95,
                pad_token_id=self.config.pad_token_id,
                eos_token_id=self.config.eos_token_id,
            )
            thought_ids[:, i, :] = thought_outputs
            thought_embeddings.append(self.get_input_embeddings()(thought_outputs))

        thought_embeddings = torch.stack(thought_embeddings, dim=1)
        return thought_ids, thought_embeddings

    @torch.no_grad()
    def infer(
        self,
        input_ids: torch.LongTensor,
        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,
    ):
        batch_size, seq_len = input_ids.shape

        # Save the original input_ids and attention_mask for later use
        original_input_ids = input_ids.clone()
        original_attention_mask = attention_mask.clone() if attention_mask is not None else None

        # Append the start thought token to the input sequence
        start_thought_token_id = self.tokenizer.convert_tokens_to_ids("<|startthought|>")
        input_ids = torch.cat([input_ids, torch.tensor([[start_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1)
        seq_len += 1

        # Update the attention mask
        if attention_mask is not None:
            attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)

        # Generate the continuation
        continuation_length = self.n_ahead - 2
        new_key_values = past_key_values
        
        start_time = time.time()
        for continuation_idx in range(continuation_length):
            outputs = self.model(
                input_ids=input_ids if continuation_idx == 0 else next_token_id.unsqueeze(-1).to(input_ids.device),
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_values=new_key_values,
                inputs_embeds=inputs_embeds,
                use_cache=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
            new_key_values = outputs.past_key_values

            hidden_states = outputs[0]

            logits = self.lm_head(hidden_states)
            logits = logits[:, -1, :]  # Only consider the last token

            # Apply Gumbel-Softmax to the logits
            next_token_logits = F.gumbel_softmax(logits, tau=self.gumbel_temperature, hard=True, dim=-1)
            next_token_id = torch.argmax(next_token_logits, dim=-1)

            # Append the generated token to the input sequence
            input_ids = torch.cat([input_ids, next_token_id.unsqueeze(-1).to(input_ids.device)], dim=-1)
            seq_len += 1

            # Update the attention mask
            if attention_mask is not None:
                attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)

        # Append the end thought token to the input sequence
        end_thought_token_id = self.tokenizer.convert_tokens_to_ids("<|endthought|>")
        input_ids = torch.cat([input_ids, torch.tensor([[end_thought_token_id]] * batch_size).to(input_ids.device)], dim=-1)
        seq_len += 1

        # Update the attention mask
        if attention_mask is not None:
            attention_mask = torch.cat([attention_mask, torch.ones((batch_size, 1)).to(attention_mask.device)], dim=-1)

        # Get the hidden states before and after the thought
        outputs_before = self.model(
            input_ids=original_input_ids,
            attention_mask=original_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_before = outputs_before[0][:, -1:, :]

        # two new tokens: last continuation token and end thought token
        outputs_after = self.model(
            input_ids=torch.cat([next_token_id.unsqueeze(-1).to(input_ids.device), torch.tensor(end_thought_token_id).unsqueeze(-1).unsqueeze(-1).to(input_ids.device)], dim=-1),
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=new_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_after = outputs_after[0][:, -1:, :]

        # Apply the talk head to get the mixing weight
        mixing_weight = self.talk_head[0](torch.cat([hidden_states_before, hidden_states_after], dim=-1))

        # Apply the mixing weight to the hidden states
        mixed_hidden_states = (1 - mixing_weight) * hidden_states_before + mixing_weight * hidden_states_after

        # Apply the language model head to get the final logits
        logits = self.lm_head(mixed_hidden_states)
        return logits

    @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> 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."
        ```"""

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

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            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=True,
        )
        hidden_states = outputs.last_hidden_state
        logits = self.lm_head(hidden_states)

        thought_ids, thought_embeddings = self._generate_thoughts(hidden_states, max_length=self.config.thought_length)
        thought_hidden_states = self.model(inputs_embeds=thought_embeddings).last_hidden_state

        # Compute thought logits
        thought_logits = self.lm_head(thought_hidden_states)

        # Mix base and thought logits
        mixed_logits = logits.unsqueeze(1) + self.mixing_head(thought_logits)
        mixed_logits = mixed_logits.view(-1, mixed_logits.size(-1))

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = mixed_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

            if self.use_policy_loss:
                rewards = loss.detach().unsqueeze(1).repeat(1, self.max_thoughts)
                if self.remove_negative_rewards:
                    rewards = torch.clamp(rewards, min=0)
                policy_loss = self.calculate_policy_loss(thought_ids, rewards)
                loss = loss + policy_loss
        else:
            loss = None

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

        return CausalLMOutputWithPast(
            loss=loss if loss is not None else None,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    def prepare_inputs_for_generation(
        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
    ):
        # 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



############################## Extra Heads #################################

############# Sequence Classification ################# 
@add_start_docstrings(
    """
    The Mistral Model transformer with a sequence classification head on top (linear layer).

    [`MistralForSequenceClassification`] 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).
    """,
    MISTRAL_START_DOCSTRING,
)
# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Mistral, LLAMA->MISTRAL
class MistralForSequenceClassification(MistralPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.model = MistralModel(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(MISTRAL_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        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,
        )

############# Token Classification    ################# 
@add_start_docstrings(
    """
    The Mistral Model transformer with a token classification head on top (a linear layer on top of the hidden-states
    output) e.g. for Named-Entity-Recognition (NER) tasks.
    """,
    MISTRAL_START_DOCSTRING,
)
# Copied from transformers.models.llama.modeling_llama.LlamaForTokenClassification with Llama->Mistral, LLAMA->MISTRAL
class MistralForTokenClassification(MistralPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.model = MistralModel(config)
        if getattr(config, "classifier_dropout", None) is not None:
            classifier_dropout = config.classifier_dropout
        elif getattr(config, "hidden_dropout", None) is not None:
            classifier_dropout = config.hidden_dropout
        else:
            classifier_dropout = 0.1
        self.dropout = nn.Dropout(classifier_dropout)
        self.score = nn.Linear(config.hidden_size, config.num_labels)

        # 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(MISTRAL_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[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, TokenClassifierOutput]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        sequence_output = outputs[0]
        sequence_output = self.dropout(sequence_output)
        logits = self.score(sequence_output)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return TokenClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

############# QuestionAnswer    ################# 
@add_start_docstrings(
    """
The Mistral Model transformer with a span classification head on top for extractive question-answering tasks like
SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
    """,
    MISTRAL_START_DOCSTRING,
)
class MistralForQuestionAnswering(MistralPreTrainedModel):
    base_model_prefix = "transformer"

    # Copied from transformers.models.bloom.modeling_bloom.BloomForQuestionAnswering.__init__ with Bloom->Llama
    def __init__(self, config):
        super().__init__(config)
        self.transformer = MistralModel(config)
        self.qa_outputs = nn.Linear(config.hidden_size, 2)

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

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

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

    @add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        start_positions: Optional[torch.LongTensor] = None,
        end_positions: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
        r"""
        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for position (index) of the start of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
            are not taken into account for computing the loss.
        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for position (index) of the end of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
            are not taken into account for computing the loss.
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.transformer(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        sequence_output = outputs[0]

        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1).contiguous()
        end_logits = end_logits.squeeze(-1).contiguous()

        total_loss = None
        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1).to(start_logits.device)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1).to(end_logits.device)
            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
            start_positions = start_positions.clamp(0, ignored_index)
            end_positions = end_positions.clamp(0, ignored_index)

            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
            start_loss = loss_fct(start_logits, start_positions)
            end_loss = loss_fct(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2

        if not return_dict:
            output = (start_logits, end_logits) + outputs[2:]
            return ((total_loss,) + output) if total_loss is not None else output

        return QuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )
############################## Closed Extra Heads ###########################