modeling_switchllama.py

# copyright idek

from transformers.models.llama.modeling_llama import *
from torch import nn
import torch
from .configuration_switchllama import SwitchLlamaConfig


def router_z_loss_func(router_logits: torch.Tensor) -> float:
    r"""
    Compute the router z-loss implemented in PyTorch.

    The router z-loss was introduced in [Designing Effective Sparse Expert Models](https://arxiv.org/abs/2202.08906).
    It encourages router logits to remain small in an effort to improve stability.

    Args:
        router_logits (`float`):
            Input logits of shape [batch_size, sequence_length, num_experts]

    Returns:
        Scalar router z-loss.
    """
    num_groups, tokens_per_group, _ = router_logits.shape
    log_z = torch.logsumexp(router_logits, dim=-1)
    z_loss = log_z**2
    return torch.sum(z_loss) / (num_groups * tokens_per_group)


def load_balancing_loss_func(router_probs: torch.Tensor, expert_indices: torch.Tensor) -> 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:
        router_probs (`torch.Tensor`):
            Probability assigned to each expert per token. Shape: [batch_size, seqeunce_length, num_experts].
        expert_indices (`torch.Tensor`):
            Indices tensor of shape [batch_size, seqeunce_length] identifying the selected expert for a given token.

    Returns:
        The auxiliary loss.
    """
    num_experts = router_probs.shape[-1]

    # cast the expert indices to int64, otherwise one-hot encoding will fail
    if expert_indices.dtype != torch.int64:
        expert_indices = expert_indices.to(torch.int64)

    if len(expert_indices.shape) == 2:
        expert_indices = expert_indices.unsqueeze(2)

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

    # For a given token, determine if it was routed to a given expert.
    expert_mask = torch.max(expert_mask, axis=-2).values

    # cast to float32 otherwise mean will fail
    expert_mask = expert_mask.to(torch.float32)
    tokens_per_group_and_expert = torch.mean(expert_mask, axis=-2)

    router_prob_per_group_and_expert = torch.mean(router_probs, axis=-2)
    return torch.mean(tokens_per_group_and_expert * router_prob_per_group_and_expert) * (num_experts**2)


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

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

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

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

    inverted_mask = 1.0 - expanded_mask

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



class SwitchLlamaTop1Router(nn.Module):
    """
    Router using tokens choose top-1 experts assignment.

    This router uses the same mechanism as in Switch Transformer (https://arxiv.org/abs/2101.03961) and V-MoE
    (https://arxiv.org/abs/2106.05974): tokens choose their top experts. Items are sorted by router_probs and then
    routed to their choice of expert until the expert's expert_capacity is reached. **There is no guarantee that each
    token is processed by an expert**, or that each expert receives at least one token.

    """

    def __init__(self, config: SwitchLlamaConfig):
        super().__init__()
        self.num_experts = config.num_experts
        self.expert_capacity = config.expert_capacity
        self.classifier = nn.Linear(config.hidden_size, self.num_experts, bias=config.router_bias)
        self.jitter_noise = config.router_jitter_noise
        self.ignore_padding_tokens = config.router_ignore_padding_tokens
        
    def _compute_router_probabilities(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        r"""
        Computes router probabilities from input hidden states.

        Args:
            hidden_states (`torch.Tensor`):
                (batch_size, sequence_length, hidden_dim) from which router probabilities are computed.
        Returns:
            router_probabilities (`torch.Tensor`):
                Tensor of shape (batch_size, sequence_length, num_experts) corresponding to the probabilities for each
                token and expert. Used for routing tokens to experts.
            router_logits (`torch.Tensor`):
                Logits tensor of shape (batch_size, sequence_length, num_experts) corresponding to raw router logits.
                This is used later for computing router z-loss.
        """
        if self.jitter_noise > 0:
            # Get the lower and upper bound of the uniform distribution
            # Adapted from: https://stackoverflow.com/questions/44328530/how-to-get-a-uniform-distribution-in-a-range-r1-r2-in-pytorch
            distrib_lower_bound = 1.0 - self.jitter_noise
            distrib_upper_bound = 1.0 + self.jitter_noise

            uniform_distrib = torch.rand(hidden_states.shape, device=hidden_states.device, dtype=hidden_states.dtype)
            uniform_distrib = uniform_distrib * (distrib_lower_bound - distrib_upper_bound)

            uniform_distrib = uniform_distrib + distrib_upper_bound
            # Multiply the token inputs by the uniform distribution - adding some noise
            hidden_states *= uniform_distrib

        # Shape: [num_groups, tokens_per_group, num_experts]
        router_logits = self.classifier(hidden_states)

        # Apply Softmax
        router_probabilities = nn.functional.softmax(router_logits, dim=-1)
        return router_probabilities, router_logits
        
    def forward(self, hidden_states: torch.Tensor) -> Tuple:
        r"""
        Generic forward function for every Router class. Each Router expects to have the same input hidden states
        (`hidden_states`) corresponding to the hidden states for each token, the `expert_capacity` corresponding to the
        number of tokens the Router will send to each expert, some Routers can send up to few tokens to each expert.

        Each Router works as the following: it expects the hidden states for each token, gets the `router_probs` and
        `router_logits` from the `router_weights`. This will assign for each token, the raw probability to be assigned
        to an expert. Then each Router class will have to define its own `_compute_routing_instructions`.

        Args:
            hidden_states (`torch.Tensor`) :
                [num_groups, tokens_per_group, hidden_dim] inputs to send to experts.
        Returns:
            Tuple[`torch.Tensor`, `torch.Tensor`, `torch.Tensor`] Tuple containing the expert index, the router probs
            and the router logits. The router probabilities and logits are required to compute the loss.
        """
        router_probs, router_logits = self._compute_router_probabilities(hidden_states)

        expert_index = torch.argmax(router_probs, dim=-1)
        expert_index = torch.nn.functional.one_hot(expert_index, num_classes=self.num_experts)

        # Mask tokens outside expert capacity. Sum over each sequence
        token_priority = torch.cumsum(expert_index, dim=-2)
        # mask if the token routed to to the expert will overflow
        expert_capacity_mask = token_priority <= self.expert_capacity
        expert_index = expert_index * expert_capacity_mask

        router_probs = torch.max(router_probs, dim=-1).values.unsqueeze(-1)
        return expert_index, router_probs, router_logits

class SwitchLlamaSparseMLP(nn.Module):
    r"""
    Implementation of the Switch Transformers Sparse MLP module.
    """

    def __init__(self, config: SwitchLlamaConfig, expert_class: nn.Module = LlamaMLP):
        super().__init__()
        # Step 1: Get the correct router according to its class
        self.router = SwitchLlamaTop1Router(config)

        # Step 2: Get the experts
        self.experts = nn.ModuleDict()
        for idx in range(config.num_experts):
            self.experts[f"expert_{idx}"] = expert_class(config)

    def forward(self, hidden_states):
        r"""
        Hold on, this will be slightly tricky to understand In the correct order, a MoE layer does the following:

        1- Gets the `router_mask` from the router. The shape of the mask is `(batch_size, sequence_length, num_expert)`
        and corresponds to the argmax of the `router_probs`. The probabilities are needed in the computation of the
        hidden states : they are broadcasted to the hidden states values (can be interpreted as a scaling factor).

        2- Dispatch the tokens to its associated experts. We do a classic for loop over the experts and assign for each
        expert the corresponding hidden states.

        """
        # Step 1: Get the router_mask from the router as wel as the probabilities
        router_mask, router_probs, router_logits = self.router(hidden_states)
        expert_index = torch.argmax(router_mask, dim=-1)

        # The routers introduced might not always map all the tokens, to a router, which means that some hidden states
        # can be unchanged from one layer to another. That is why the hidden states are cloned before updating only the seleced ones.

        next_states = hidden_states.clone()
        for idx, expert in enumerate(self.experts.values()):
            token_indices = router_mask[:, :, idx].bool()
            next_states[token_indices] = expert(hidden_states[token_indices])

        hidden_states = router_probs * next_states
        return hidden_states, (router_logits, expert_index)

class SwitchLlamaLayerFF(nn.Module):
    r"""
    Switch Transformers Feed Forward layer module. This is a wrapper around the Mixture of Experts module.

    Parameters:
        config : ([`SwitchTransformersConfig`]): 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.
        is_sparse (`bool`):
            Whether the MLP layer is a `Sparse` layer (contains a Mixture of Experts) or not
    """

    def __init__(self, config: SwitchLlamaConfig, is_sparse=True):
        super().__init__()
        self.is_sparse = is_sparse

        # Check if it is a sparse layer, if not then it is a dense layer
        if not self.is_sparse:
            self.mlp = LlamaMLP(config)
        else:
            self.mlp = SwitchLlamaSparseMLP(config)

        # self.layer_norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.dropout = nn.Dropout(config.dropout_rate)

    def forward(self, hidden_states, output_router_logits=False):
        # forwarded_states = self.layer_norm(hidden_states)
        forwarded_states = self.mlp(hidden_states)

        if isinstance(forwarded_states, tuple):
            forwarded_states, router_tuple = forwarded_states
        else:
            router_tuple = None

        output = hidden_states + self.dropout(forwarded_states)

        if output_router_logits and router_tuple is not None:
            output = (output, router_tuple)

        return output

class SwitchLlamaDecoderLayer(nn.Module):
    def __init__(self, config: SwitchLlamaConfig):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = LlamaAttention(config=config)
        self.mlp = SwitchLlamaLayerFF(config)
        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        output_router_logits = True
    ) -> 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, 1, tgt_len, src_len)` where padding elements are indicated by very large negative 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.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
        """

        residual = hidden_states

        hidden_states = self.input_layernorm(hidden_states)

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

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states, output_router_logits=output_router_logits)
        if type(hidden_states)==tuple:
            hidden_states, router_tuple = hidden_states
        else:
            router_tuple = (torch.tensor([0], device=hidden_states.device),)
        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_tuple,)
        return outputs + (router_tuple,)

class SwitchLlamaPreTrainedModel(PreTrainedModel):
    config_class = SwitchLlamaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["SwitchLlamaDecoderLayer"]
    _skip_keys_device_placement = "past_key_values"

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

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


class SwitchLlamaModel(SwitchLlamaPreTrainedModel):
    """
    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]

    Args:
        config: SwitchLlamaConfig
    """

    def __init__(self, config: SwitchLlamaConfig):
        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([SwitchLlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
        self.norm = LlamaRMSNorm(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

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

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

        return combined_attention_mask

    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        output_router_logits = False
    ) -> 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
        all_router_probs = () if output_router_logits else None
        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
        elif input_ids is not None:
            batch_size, seq_length = input_ids.shape
        elif inputs_embeds is not None:
            batch_size, seq_length, _ = inputs_embeds.shape
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

        seq_length_with_past = seq_length
        past_key_values_length = 0

        if past_key_values is not None:
            past_key_values_length = past_key_values[0][0].shape[2]
            seq_length_with_past = seq_length_with_past + past_key_values_length

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

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

        hidden_states = inputs_embeds

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

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

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

            past_key_value = past_key_values[idx] if past_key_values is not None else None

            if self.gradient_checkpointing and self.training:

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

                    return custom_forward

                layer_outputs = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(decoder_layer),
                    hidden_states,
                    attention_mask,
                    position_ids,
                )
            else:
                layer_outputs = decoder_layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    past_key_value=past_key_value,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    output_router_logits=output_router_logits
                )

            hidden_states = layer_outputs[0]
            router_probs = layer_outputs[-1]
            
            if use_cache:
                next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)

            if output_attentions:
                all_self_attns += (layer_outputs[1],)
                
            if output_router_logits:
                all_router_probs = all_router_probs + (router_probs,)
        hidden_states = self.norm(hidden_states)

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

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

        from transformers.models.switch_transformers.modeling_switch_transformers import MoEModelOutputWithPastAndCrossAttentions
        return MoEModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            router_probs=all_router_probs,
        )

class SwitchLlamaForCausalLM(SwitchLlamaPreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = SwitchLlamaModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        self.router_z_loss_coef = config.router_z_loss_coef
        self.router_aux_loss_coef = config.router_aux_loss_coef
        # Initialize weights and apply final processing
        self.post_init()
    def _unpack_router_logits(self, router_outputs):
        total_router_logits = []
        total_expert_indexes = []
        for router_output in router_outputs:
            if len(router_output[0].shape) > 1:
                router_logits, expert_indexes = router_output
                total_router_logits.append(router_logits)
                total_expert_indexes.append(expert_indexes)
        return torch.cat(total_router_logits, dim=1), torch.cat(total_expert_indexes, dim=1)


    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 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,
        output_router_logits = False,
    ) -> 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, LlamaForCausalLM

        >>> model = LlamaForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)

        >>> 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,
            output_router_logits=output_router_logits
        )

        hidden_states = outputs[0]
        if self.config.pretraining_tp > 1:
            lm_head_slices = self.lm_head.weight.split(self.vocab_size // self.config.pretraining_tp, dim=0)
            logits = [F.linear(hidden_states, lm_head_slices[i]) for i in range(self.config.pretraining_tp)]
            logits = torch.cat(logits, dim=-1)
        else:
            logits = self.lm_head(hidden_states)
        logits = logits.float()

        loss = None
        decoder_z_loss = None
        decoder_aux_loss = None

        if output_router_logits:
            decoder_router_logits, decoder_expert_indexes = self._unpack_router_logits(outputs[-1])
            decoder_z_loss = router_z_loss_func(decoder_router_logits)
            decoder_router_probs = nn.Softmax(dim=-1)(decoder_router_logits)
            decoder_aux_loss = load_balancing_loss_func(decoder_router_probs, decoder_expert_indexes)
            
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Enable model parallelism
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

            ##########################
            if output_router_logits:
                z_loss = self.router_z_loss_coef * decoder_z_loss
                aux_loss = self.router_aux_loss_coef * decoder_aux_loss
                loss = loss + z_loss + aux_loss
            #########################
        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

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

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

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

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

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

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