modeling_llama.py

# coding=utf-8
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch LLaMA model."""
import json
import math
import warnings
from typing import List, Optional, Tuple, Union

import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache, StaticCache
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
    MoeModelOutputWithPast, MoeCausalLMOutputWithPast,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutputWithPast,
)
from transformers.modeling_utils import PreTrainedModel
from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS
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,
)
# we just reuse everything we don't modified
from transformers.models.llama.modeling_llama import (
    LlamaRMSNorm, LlamaRotaryEmbedding,
    LlamaLinearScalingRotaryEmbedding,
    LlamaDynamicNTKScalingRotaryEmbedding,
    LlamaAttention,
    LlamaMLP,
    LlamaFlashAttention2,
    LlamaSdpaAttention,
    LlamaDecoderLayer
)
from .configuration_llama import LlamaMoDConfig


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


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "LlamaMoDConfig"


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.int32), (1, 0))
    return (
        indices,
        cu_seqlens,
        max_seqlen_in_batch,
    )


ALL_LAYERNORM_LAYERS.append(LlamaRMSNorm)

LLAMA_ATTENTION_CLASSES = {
    "eager": LlamaAttention,
    "flash_attention_2": LlamaFlashAttention2,
    "sdpa": LlamaSdpaAttention,
}

def dual_router_aux_loss(
    gate_logits: torch.Tensor, num_experts: torch.Tensor = 40, capacity_load=5, attention_mask: Optional[torch.Tensor] = None
) -> float:
    r"""
    Computes auxiliary load balancing loss as in Layer wise mode - implemented in Pytorch.

    Modified from Switch Transformer (https://arxiv.org/abs/2101.03961), I mean mixtral model.
    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`):
            Number of layers
        top_k (`int`):
            Number of experts (capacility load * num_experts)

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

    bce_loss = nn.BCEWithLogitsLoss()
    if isinstance(gate_logits, tuple):
        compute_device = gate_logits[0].device
        concatenated_gate_logits = torch.cat([layer_gate.unsqueeze(-1).to(compute_device) for layer_gate in gate_logits], dim=-1)
    seq_len = concatenated_gate_logits.shape[1]
    top_k = int(capacity_load*seq_len)
    bs = concatenated_gate_logits.shape[0]
    # concatenated_gate_logits : bs x seq_len x [ route logits, mlp router pred ] x layers
    routing_weights = torch.nn.functional.softmax(concatenated_gate_logits[:, :, 0, :], dim=1)
    # routing_weights = routing_weights
    _, selected_tokens = torch.topk(routing_weights, top_k, dim=1)
    pred_logits = concatenated_gate_logits[:, :, 1, :]
    router_targets = torch.zeros_like(pred_logits).view(-1)
    router_targets[selected_tokens.view(-1)] = 1.0
    loss = bce_loss(pred_logits, router_targets.view(bs, seq_len, -1))
    return loss

class LlamaMoDDuaRouter(nn.Module):
    # implement the prediction inside this instead to make sure weights are transferable
    # Implement method 1

    def __init__(self, config: LlamaMoDConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.layer_idx = layer_idx
        self.capacity_factor = config.capacity_load
        self.dynamic_skip = config.setup_layer_mod[layer_idx]
        if self.dynamic_skip:
            self.mod_router = nn.Linear(self.hidden_size, 1, bias=False)
            # used in inference instead
            self.mlp_router = nn.Sequential(
                nn.Linear(self.hidden_size, self.hidden_size//2),
                nn.SiLU(),
                nn.Linear(self.hidden_size//2, 1, bias=False)
            )

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

        self.mlp = LlamaMLP(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,
        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
        """
        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.`"
            )

        initial_residual = hidden_states
        if self.dynamic_skip:
            residual = hidden_states
            seq_len = hidden_states.shape[1]
            route = torch.softmax(self.mod_router(hidden_states), dim=1)
            mlp_router_logits = self.mlp_router(hidden_states)
            if not self.training:
                # use mlp for during inference
                _old_router = route
                route = torch.sigmoid(mlp_router_logits)
            hidden_states = self.input_layernorm(hidden_states)

            if not self.training and hidden_states.shape[1] == 1 and hidden_states.shape[0] == 1:
                # TODO: fix this broke when batch is > 1
                if route[-1] > 0.5:
                    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,
                        **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
                    hidden_states *= route
                else:
                    hidden_states = initial_residual
                    self_attn_weights = None
                    present_key_value = None
            else:
                # There should be a sigmoid version where we sample the route based on > 0.5
                # but this doesn't really work mainly in batch inference
                acc_route_choice = torch.cumsum(route > 0.5, dim=-1)
                min_top_k = max(int(self.capacity_factor*seq_len), 2)
                top_k = max(torch.max(acc_route_choice), min_top_k)
                weights, selected_tokens = torch.topk(route, top_k, dim=1, sorted=False)
                # reorder back to original position?
                selected_tokens, index = torch.sort(selected_tokens, dim=1)
                weights = torch.gather(weights, dim=1, index=index)
                indices_expanded = selected_tokens.expand(-1, -1, self.hidden_size)
                sub_hidden_states = torch.gather(hidden_states, 1, indices_expanded)
                sub_position_ids = position_ids[: , :top_k]
                if len(attention_mask.shape) == 4:
                    sub_attention_mask = attention_mask[:, :, :top_k,:top_k]
                else:
                    sub_attention_mask = attention_mask[:, :top_k]

                residual = sub_hidden_states

                sub_hidden_states, attn_weights, present_key_value = self.self_attn(
                    hidden_states=sub_hidden_states,
                    attention_mask=sub_attention_mask,
                    position_ids=sub_position_ids,
                    past_key_value=past_key_value,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    **kwargs,
                )
                if not self.training:
                    print(int(top_k), seq_len, self.layer_idx)
                    print(selected_tokens.flatten())
                    print(_old_router.flatten())
                    print(route.flatten())

                sub_hidden_states = residual + sub_hidden_states
                sub_residual = sub_hidden_states
                # MLP
                sub_hidden_states = self.post_attention_layernorm(sub_hidden_states)
                sub_hidden_states = self.mlp(sub_hidden_states)
                sub_hidden_states = sub_residual + sub_hidden_states

                hidden_states = torch.scatter(
                    initial_residual,
                    dim=1,
                    index=indices_expanded,
                    src=sub_hidden_states * weights,
                )
        else:
            residual = initial_residual
            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,
                **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

        # this section must be modified during inference, otherwise no speedup
        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        if self.dynamic_skip:
            outputs += (torch.concat([route, mlp_router_logits], dim=-1), )

        return outputs

def load_balancing_loss_func(
    gate_logits: torch.Tensor, num_experts: torch.Tensor = 40, capacity_load=0.125, attention_mask: Optional[torch.Tensor] = None
) -> float:
    r"""
    Computes auxiliary load balancing loss as in Layer wise mode - implemented in Pytorch.

    The original paper of mixture of depth didn't specify beyond one word : use aux loss

    I would assume its from this:

    Modified from Switch Transformer (https://arxiv.org/abs/2101.03961), I mean mixtral model.
    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`):
            Number of layers
        top_k (`int`):
            Number of experts (capacility load * num_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=-1)
    batch_size, sequence_length, num_hidden_layers = concatenated_gate_logits.shape
    top_k = int(capacity_load*sequence_length)

    # bs x seq_length x layers
    routing_weights = torch.nn.functional.sigmoid(concatenated_gate_logits)
    # bs x layers x seq_length
    routing_weights = routing_weights.permute(0, 2, 1)
    _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
    expert_mask = torch.nn.functional.one_hot(selected_experts, sequence_length)
    expert_mask = expert_mask.reshape(-1, top_k, sequence_length)
    # bs x num_layers x top_k x sequence length
    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:
        # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
        expert_attention_mask = (
            attention_mask[:, :, None, None]
            .expand((batch_size, sequence_length, top_k, num_experts))
            .permute(0, 3, 2, 1)
            .reshape(-1, top_k, sequence_length)
            .to(compute_device)
        )

        # Compute the percentage of tokens routed to each position id
        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]
            .expand((batch_size, sequence_length, num_experts))
            .reshape(-1, sequence_length)
            .to(compute_device)
        )
        # Compute the average probability of routing to these experts
        router_prob_per_expert = torch.sum(routing_weights.reshape(-1, sequence_length) * 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



class LlamaMoDBalanceAux(nn.Module):
    # implement the prediction inside this instead to make sure weights are transferable
    # Implement method 1

    def __init__(self, config: LlamaMoDConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.layer_idx = layer_idx
        self.capacity_factor = config.capacity_load
        self.dynamic_skip = config.setup_layer_mod[layer_idx]
        if self.dynamic_skip:
            self.mod_router = nn.Linear(self.hidden_size, 1, bias=True)

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

        self.mlp = LlamaMLP(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,
        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
        """
        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.`"
            )

        initial_residual = hidden_states
        if self.dynamic_skip:
            residual = hidden_states
            seq_len = hidden_states.shape[1]
            route = torch.sigmoid(self.mod_router(hidden_states))

            if not self.training and (hidden_states.shape[1] == 1 and hidden_states.shape[0] == 1):
                # TODO: fix this broke when batch is > 1
                # single inference mode
                if route[-1] > 0.5:
                    hidden_states = self.input_layernorm(hidden_states)
                    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,
                        **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
                else:
                    hidden_states = initial_residual
                    self_attn_weights = None
                    present_key_value = None

            else:
                acc_route_choice = torch.cumsum(route > 0.5, dim=1)
                min_top_k = max(int(self.capacity_factor*seq_len), 2)
                top_k = max(torch.max(acc_route_choice), min_top_k)
                # with open('experiments_bias_0.1_aux_0.001.jsonl', 'a') as fout:
                #     fout.write(json.dumps({'idx': self.layer_idx, 'top_k': int(top_k), 'seq_len': seq_len})+'\n')
                weights, selected_tokens = torch.topk(route, top_k, dim=1, sorted=False)
                # reorder back to original position?
                selected_tokens, index = torch.sort(selected_tokens, dim=1)
                if not self.training:
                    print(int(top_k), seq_len, self.layer_idx)
                    print(selected_tokens.flatten())

                weights = torch.gather(weights, dim=1, index=index)
                indices_expanded = selected_tokens.expand(-1, -1, self.hidden_size)
                sub_hidden_states = torch.gather(hidden_states, 1, indices_expanded)
                sub_position_ids = position_ids[: , :top_k]
                if len(attention_mask.shape) == 4:
                    sub_attention_mask = attention_mask[:, :, :top_k,:top_k]
                else:
                    sub_attention_mask = attention_mask[:, :top_k]


                residual = sub_hidden_states
                sub_hidden_states = self.input_layernorm(sub_hidden_states)
                sub_hidden_states, attn_weights, present_key_value = self.self_attn(
                    hidden_states=sub_hidden_states,
                    attention_mask=sub_attention_mask,
                    position_ids=sub_position_ids,
                    past_key_value=past_key_value,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    **kwargs,
                )

                sub_hidden_states = residual + sub_hidden_states
                sub_residual = sub_hidden_states
                # MLP
                sub_hidden_states = self.post_attention_layernorm(sub_hidden_states)
                sub_hidden_states = self.mlp(sub_hidden_states)
                sub_hidden_states = sub_residual + sub_hidden_states
                hidden_states = sub_hidden_states
                hidden_states = torch.scatter(
                    initial_residual,
                    dim=1,
                    index=indices_expanded,
                    src=sub_hidden_states,
                )
        else:
            residual = initial_residual
            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,
                **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

        # this section must be modified during inference, otherwise no speedup
        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        if self.dynamic_skip:
            outputs += (route, )

        return outputs


class LlamaPreTrainedModel(PreTrainedModel):
    config_class = LlamaMoDConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["LlamaDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True

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

    def _setup_cache(self, cache_cls, max_batch_size, max_cache_len: Optional[int] = None):
        if self.config._attn_implementation == "flash_attention_2" and cache_cls == 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"
            )

        for layer in self.model.layers:
            device = layer.input_layernorm.weight.device
            if hasattr(self.config, "_pre_quantization_dtype"):
                dtype = self.config._pre_quantization_dtype
            else:
                dtype = layer.self_attn.o_proj.weight.dtype
            layer.self_attn.past_key_value = cache_cls(
                self.config, max_batch_size, max_cache_len, device=device, dtype=dtype
            )

    def _reset_cache(self):
        for layer in self.model.layers:
            layer.self_attn.past_key_value = None


LLAMA_DECODER_LAYER = {
    'none': LlamaDecoderLayer,
    'mod_1aux': LlamaMoDBalanceAux,
    'mod_dual': LlamaMoDDuaRouter
}

AUX_LOSS = {
    'mod_1aux': load_balancing_loss_func,
    'mod_dual': dual_router_aux_loss
}

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

    Args:
        config: LlamaMoDConfig
    """

    def __init__(self, config: LlamaMoDConfig):
        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.has_router = config.mod_method != 'none'
        self.layers = nn.ModuleList(
            [LLAMA_DECODER_LAYER[config.mod_method](config, layer_idx) for layer_idx 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

    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,
        cache_position: Optional[torch.LongTensor] = None,
        output_router_logits: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        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)

        past_seen_tokens = 0
        if use_cache:  # kept for BC (cache positions)
            if not isinstance(past_key_values, StaticCache):
                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
                past_seen_tokens = past_key_values.get_seq_length()

        if cache_position is None:
            if isinstance(past_key_values, StaticCache):
                raise ValueError("cache_position is a required argument when using StaticCache.")
            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_seen_tokens + inputs_embeds.shape[1]
        )

        # embed positions
        hidden_states = inputs_embeds

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        all_router_logits = () if output_router_logits 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],)
            if self.has_router and decoder_layer.dynamic_skip and output_router_logits:
                all_router_logits += (layer_outputs[-1],)

        hidden_states = self.norm(hidden_states)

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

        next_cache = None
        if use_cache:
            next_cache = (
                next_decoder_cache.to_legacy_cache() if isinstance(next_decoder_cache, Cache) else next_decoder_cache
            )
        if not return_dict:
            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
        return MoeModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            router_logits=all_router_logits
        )

    # 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
    def _update_causal_mask(self, attention_mask, input_tensor, cache_position, current_length):
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and 0.0 in attention_mask:
                return attention_mask
            return None

        dtype, device = input_tensor.dtype, input_tensor.device
        min_dtype = torch.finfo(dtype).min
        sequence_length = input_tensor.shape[1]
        if hasattr(getattr(self.layers[0], "self_attn", {}), "past_key_value"):  # static cache
            target_length = self.config.max_position_embeddings
        else:  # dynamic cache
            target_length = (
                attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else current_length + 1
            )

        causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
        if sequence_length != 1:
            causal_mask = torch.triu(causal_mask, diagonal=1)
        causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
        causal_mask = causal_mask[None, None, :, :].expand(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].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
                causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
            elif attention_mask.dim() == 4:
                # backwards compatibility: we allow passing a 4D attention mask shorter than the input length with
                # cache. In that case, the 4D attention mask attends to the newest tokens only.
                if attention_mask.shape[-2] < cache_position[0] + sequence_length:
                    offset = cache_position[0]
                else:
                    offset = 0
                mask_shape = attention_mask.shape
                mask_slice = (attention_mask.eq(0.0)).to(dtype=dtype) * min_dtype
                causal_mask[
                    : mask_shape[0], : mask_shape[1], offset : mask_shape[2] + offset, : mask_shape[3]
                ] = mask_slice

        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type == "cuda"
        ):
            # 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


class LlamaMoDForCausalLM(LlamaPreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = LlamaMoDModel(config)
        self.vocab_size = config.vocab_size
        self.route_method = config.mod_method
        self.router_aux_loss_coef = config.router_aux_loss_coef
        if config.mod_method != 'none':
            self.num_experts = sum(config.setup_layer_mod)
            self.capacity_load = config.capacity_load
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

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

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

    def get_output_embeddings(self):
        return self.lm_head

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

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

    def get_decoder(self):
        return self.model

    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,
        cache_position: Optional[torch.LongTensor] = None,
        output_router_logits: 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, LlamaForCausalLM

        >>> model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")

        >>> 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,
            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
        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:
            aux_loss = AUX_LOSS[self.route_method](
                outputs.router_logits if return_dict else outputs[-1],
                self.num_experts,
                self.capacity_load,
                attention_mask,
            )
            if labels is not None:
                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)


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

        return MoeCausalLMOutputWithPast(
            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, cache_position=None, **kwargs
    ):
        # With static cache, the `past_key_values` is None
        # TODO joao: standardize interface for the different Cache classes and remove of this if
        has_static_cache = False
        if past_key_values is None:
            past_key_values = getattr(getattr(self.model.layers[0], "self_attn", {}), "past_key_value", None)
            has_static_cache = past_key_values is not None

        past_length = 0
        if past_key_values is not None:
            if isinstance(past_key_values, Cache):
                past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length()
                max_cache_length = (
                    torch.tensor(past_key_values.get_max_length(), device=input_ids.device)
                    if past_key_values.get_max_length() is not None
                    else None
                )
                cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length)
            # TODO joao: remove this `else` after `generate` prioritizes `Cache` objects
            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 input_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:
            # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise
            # recompiles graphs as the stride of the inputs is a guard. Ref: https://github.com/huggingface/transformers/pull/29114
            # TODO: use `next_tokens` directly instead.
            model_inputs = {"input_ids": input_ids.contiguous()}

        input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1]
        if cache_position is None:
            cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device)
        else:
            cache_position = cache_position[-input_length:]

        if has_static_cache:
            past_key_values = None

        model_inputs.update(
            {
                "position_ids": position_ids,
                "cache_position": cache_position,
                "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 LlamaForSequenceClassification(LlamaPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.model = LlamaModel(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

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

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

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

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

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

        loss = None
        if labels is not None:
            labels = labels.to(logits.device)
            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

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

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


class LlamaForQuestionAnswering(LlamaPreTrainedModel):
    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 = LlamaModel(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

    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[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,
        )