modeling_ddllama.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.
import inspect
import os
import math
from einops import repeat
from typing import List, Optional, Tuple, Union, Dict, Callable

import torch
import torch.distributed as dist
from torch.distributions.uniform import Uniform
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.generation import GenerationMixin
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs, _flash_attention_forward
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutputWithPast,
    TokenClassifierOutput,
)
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
from transformers.modeling_utils import PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS
from transformers.utils import (
    LossKwargs,
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_flash_attn_greater_or_equal_2_10,
    logging,
    replace_return_docstrings,
)
import wandb
from .configuration_ddllama import DDLlamaConfig


logger = logging.get_logger(__name__)

# implementation from: https://github.com/microsoft/SparseMixer
"""
@inproceedings{liu2023bridging,
  title={Sparse Backpropagation for MoE Training},
  author = {Liu, Liyuan and Gao, Jianfeng and Chen, Weizhu},
  booktitle = {arXiv:2310.00811 [cs]},
  year={2023}
}
@inproceedings{liu2023bridging,
  title={Bridging Discrete and Backpropagation: Straight-Through and Beyond},
  author = {Liu, Liyuan and Dong, Chengyu and Liu, Xiaodong and Yu, Bin and Gao, Jianfeng},
  booktitle = {arXiv:2304.08612 [cs]},
  year={2023}
}
"""
uniform_map: Dict[torch.device, Callable] = {}
def multiplicative_jitter(input, epsilon, training):

    if epsilon == 0 or not training:
        return input

    uniform = uniform_map.get(input.device)

    if uniform is None:
        uniform = Uniform(low=torch.tensor(1.0 - epsilon, device=input.device, dtype=input.dtype),
                          high=torch.tensor(1.0 + epsilon, device=input.device, dtype=input.dtype)
                ).rsample
        uniform_map[input.device] = uniform

    return input * uniform(input.shape)

class v2core(torch.autograd.Function):
    @staticmethod
    def forward(
        ctx, 
        scores: torch.Tensor, 
        multiplier: torch.Tensor, 
        selected_experts: torch.Tensor,
        masked_gates: torch.Tensor,
        mask_for_one: torch.Tensor,
    ):
        ctx.save_for_backward(multiplier, selected_experts, masked_gates)
        return multiplier * mask_for_one
        
    @staticmethod
    def backward(
        ctx, 
        grad_at_output: torch.Tensor, 
    ):
        multiplier, selected_experts, masked_gates = ctx.saved_tensors
        
        grad_at_output = grad_at_output * multiplier
        
        grad_at_scores_expaned = masked_gates * grad_at_output.mul(-1)
        grad_at_scores_expaned.scatter_add_(
            dim=-1,
            index=selected_experts,
            src=grad_at_output,
        )
        
        return (
            grad_at_scores_expaned, 
            None, 
            None, 
            None, 
            None, 
        )

def sparsemixerv2_routing(scores, top_k, jitter_eps, training):
    assert top_k in [1, 2], "only top-1/2 gating has been tested!"
    
    original_gates = torch.softmax(scores, dim=-1)
    ################ first expert ################
    
    with torch.no_grad():
        # compute mask for sparsity
        mask_logits_threshold, max_ind = scores.max(dim=-1, keepdim=True)
        factor = scores.abs().clamp(min=mask_logits_threshold)
        mask_logits_threshold = (
            (mask_logits_threshold - scores) / factor
        ) > (2 * jitter_eps)

    # apply mask 
    masked_gates = scores.masked_fill(mask_logits_threshold, float('-inf'))
    if training:
        selected_experts = (
            masked_gates - torch.empty_like(masked_gates, memory_format=torch.legacy_contiguous_format).exponential_().log()
        ).max(dim=-1)[1].unsqueeze(-1) # gumbel sampling, more robust than than the multinomial method
    else:
        selected_experts = max_ind
        
    # compute scores for gradients
    masked_gates = torch.softmax(masked_gates, dim=-1)
    
    # compute midpoint mask 
    max_scores, max_ind = masked_gates.max(dim=-1, keepdim=True)
    mask_for_one = torch.logical_or(
        selected_experts == max_ind,
        torch.rand_like(max_scores) > 0.75 # Heun's third-order method: f(x) - f(0) = .25 f'(x) + .75 f'(x/3.)
    ) 
    # 1 -> 1.0 & 0 -> 1./3: lambda x: (x + 0.5) / 1.5
    mask_for_one = torch.add(0.3333, mask_for_one, alpha=0.6667).type_as(masked_gates)

    multiplier_o = masked_gates.gather(dim=-1, index=selected_experts)
    multiplier = v2core.apply(
        scores, 
        multiplier_o, 
        selected_experts, 
        masked_gates, 
        mask_for_one,
    )
    
    ################ second expert ################
    if top_k > 1:
        # masked out first expert 
        masked_scores = torch.scatter(
            scores,
            -1,
            selected_experts,
            float('-inf'),
        )
        with torch.no_grad():
            # compute mask for sparsity
            mask_logits_threshold, max_ind = masked_scores.max(dim=-1, keepdim=True)
            factor = scores.abs().clamp(min=mask_logits_threshold)
            mask_logits_threshold = (
                (mask_logits_threshold - scores) / factor
            ) > (2 * jitter_eps)

        # apply mask 
        masked_gates_top2 = masked_scores.masked_fill(mask_logits_threshold, float('-inf'))
        if training:
            selected_experts_top2 = (
                masked_gates_top2 - torch.empty_like(masked_gates_top2, memory_format=torch.legacy_contiguous_format).exponential_().log()
            ).max(dim=-1)[1].unsqueeze(-1) # gumbel sampling, more robust than than the multinomial method
        else:
            selected_experts_top2 = max_ind
        # compute scores for gradients
        masked_gates_top2 = torch.softmax(masked_gates_top2, dim=-1)
        
        # compute midpoint mask 
        max_scores, max_ind = masked_gates_top2.max(dim=-1, keepdim=True)
        mask_for_one_top2 = torch.logical_or(
            selected_experts_top2 == max_ind,
            torch.rand_like(max_scores).uniform_() > 0.75 # Heun's third-order method: f(x) - f(0) = .25 f'(x) + .75 f'(x/3.)
        ) 
        # 1 -> 1.0 & 0 -> 1./3: lambda x: (x + 0.5) / 1.5
        mask_for_one_top2 = torch.add(0.3333, mask_for_one_top2, alpha=0.6667).type_as(masked_gates_top2)

        multiplier_top2_o = masked_gates_top2.gather(dim=-1, index=selected_experts_top2)
        multiplier_top2 = v2core.apply(
            scores, 
            multiplier_top2_o, 
            selected_experts_top2, 
            masked_gates_top2, 
            mask_for_one_top2,
        )
        
        multiplier = torch.concat((multiplier, multiplier_top2), dim=-1)
        selected_experts = torch.concat((selected_experts, selected_experts_top2), dim=-1)
            
    return (
        multiplier, 
        original_gates, 
        selected_experts,
    )

class SparseMixer(nn.Module):
    def __init__(self, num_experts, embed_dim, jitter_eps=0.1):
        super(SparseMixer, self).__init__()
        self.num_experts = num_experts
        self.jitter_eps = jitter_eps

    def forward(self, logits):
        multiplier, original_gates, sample = sparsemixerv2_routing(logits, 1, self.jitter_eps, self.training)
        return sample, multiplier

class DDModelOutputWithPast(BaseModelOutputWithPast):
    def __init__(self, last_hidden_state, past_key_values=None, hidden_states=None, attentions=None, router_weights=None, router_masks=None):
        super().__init__(last_hidden_state=last_hidden_state, past_key_values=past_key_values, hidden_states=hidden_states, attentions=attentions)
        self.router_weights = router_weights
        self.router_masks = router_masks

class DDLlamaRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        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)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"


ALL_LAYERNORM_LAYERS.append(DDLlamaRMSNorm)


class DDLlamaRotaryEmbedding(nn.Module):
    def __init__(
        self,
        dim=None,
        max_position_embeddings=2048,
        base=10000,
        device=None,
        scaling_factor=1.0,
        rope_type="default",
        config: Optional[DDLlamaConfig] = None,
    ):
        super().__init__()
        # TODO (joao): remove the `if` below, only used for BC
        self.rope_kwargs = {}
        if config is None:
            logger.warning_once(
                "`MoDLlamaRotaryEmbedding` can now be fully parameterized by passing the model config through the "
                "`config` argument. All other arguments will be removed in v4.46"
            )
            self.rope_kwargs = {
                "rope_type": rope_type,
                "factor": scaling_factor,
                "dim": dim,
                "base": base,
                "max_position_embeddings": max_position_embeddings,
            }
            self.rope_type = rope_type
            self.max_seq_len_cached = max_position_embeddings
            self.original_max_seq_len = max_position_embeddings
        else:
            # BC: "rope_type" was originally "type"
            if config.rope_scaling is not None:
                self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
            else:
                self.rope_type = "default"
            self.max_seq_len_cached = config.max_position_embeddings
            self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, **self.rope_kwargs)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    def _dynamic_frequency_update(self, position_ids, device):
        """
        dynamic RoPE layers should recompute `inv_freq` in the following situations:
        1 - growing beyond the cached sequence length (allow scaling)
        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
        """
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.max_seq_len_cached:  # growth
            inv_freq, self.attention_scaling = self.rope_init_fn(
                self.config, device, seq_len=seq_len, **self.rope_kwargs
            )
            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
            self.max_seq_len_cached = seq_len

        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
            self.max_seq_len_cached = self.original_max_seq_len

    @torch.no_grad()
    def forward(self, x, position_ids):
        if "dynamic" in self.rope_type:
            self._dynamic_frequency_update(position_ids, device=x.device)

        # Core RoPE block
        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 (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()

        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
        cos = cos * self.attention_scaling
        sin = sin * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


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)


def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


class DDLlamaMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
        self.act_fn = ACT2FN[config.hidden_act]

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


class DDLlamaRouter(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.reduction_factor = config.router_reduction_factor
        self.router_enc = nn.Linear(config.hidden_size, config.hidden_size // self.reduction_factor, bias=config.mlp_bias)
        self.router_norm = DDLlamaRMSNorm(config.hidden_size // self.reduction_factor, eps=config.rms_norm_eps)
        self.router_act = nn.Tanh()
        self.router_dec = nn.Linear(config.hidden_size // self.reduction_factor, config.hidden_size, bias=config.mlp_bias)
        self.router_head = nn.Linear(config.hidden_size, 1, bias=config.mlp_bias)

    def forward(self, hidden_states):
        router_logits = self.router_head(self.router_dec(self.router_act(self.router_norm(self.router_enc(hidden_states)))))
        return router_logits


class DDLlamaProj(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.reduction_factor = config.proj_reduction_factor
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size // self.reduction_factor, bias=config.mlp_bias)
        self.down_proj = nn.Linear(self.hidden_size, self.intermediate_size // self.reduction_factor, bias=config.mlp_bias)
        self.up_proj = nn.Linear(self.intermediate_size // self.reduction_factor, self.hidden_size, bias=config.mlp_bias)
        self.act_fn = ACT2FN[config.hidden_act]

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

def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    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 DDLlamaAttention(nn.Module):
    def __init__(self, config: DDLlamaConfig, 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 = getattr(config, "head_dim", self.hidden_size // self.num_heads)
        self.num_key_value_heads = config.num_key_value_heads
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.is_causal = True

        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias)

    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,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **kwargs,
    ) -> 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)

        # use -1 to infer num_heads and num_key_value_heads as they may vary if tensor parallel is used
        query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)

        cos, sin = position_embeddings
        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.reshape(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 DDLlamaFlashAttention2(DDLlamaAttention):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        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.LongTensor] = 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,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        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)

        # Flash attention requires the input to have the shape
        # batch_size x seq_length x head_dim x hidden_dim
        # therefore we just need to keep the original shape
        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 = position_embeddings
        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)

        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
        # to be able to avoid many of these transpose/reshape/view.
        query_states = query_states.transpose(1, 2)
        key_states = key_states.transpose(1, 2)
        value_states = value_states.transpose(1, 2)

        dropout_rate = self.attention_dropout if self.training else 0.0

        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)

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

        attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


class DDLlamaSdpaAttention(DDLlamaAttention):
    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,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **kwargs,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if output_attentions:
            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
            logger.warning_once(
                "DDLlamaModel is using DDLlamaSdpaAttention, 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,
                position_embeddings=position_embeddings,
            )

        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)

        # use -1 to infer num_heads and num_key_value_heads as they may vary if tensor parallel is used
        query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)

        cos, sin = position_embeddings
        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


DDLLAMA_ATTENTION_CLASSES = {
    "eager": DDLlamaAttention,
    "flash_attention_2": DDLlamaFlashAttention2,
    "sdpa": DDLlamaSdpaAttention,
}


class DDLlamaDecoderLayer(nn.Module):
    def __init__(self, config: DDLlamaConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = DDLLAMA_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
        self.mlp = DDLlamaMLP(config)
        self.input_layernorm = DDLlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = DDLlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        self.routing = layer_idx in config.router_layers
        if self.routing:
            self.router = DDLlamaRouter(config)
            self.mixer = SparseMixer(2, config.hidden_size, 0.1)
            self.router_proj = DDLlamaProj(config)

    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,
        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **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
            position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                with `head_dim` being the embedding dimension of each attention head.
            kwargs (`dict`, *optional*):
                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
                into the model
        """
        bsz, seq_len, hidden_size = hidden_states.shape
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)

        # Our implementation is based on masking the hidden_states skip this layer
        # This approach is faster than fetching the indices of the hidden_states for conditional computation
        # Further improvement could be done by implmenting a hardware-friendly skipping
        if self.routing:
            router_logits = self.router(hidden_states.view(-1, hidden_size))
            router_probs = torch.sigmoid(router_logits).squeeze(dim=-1)
            router_logits = torch.stack([1 - router_probs, router_probs], dim=-1)
            router_mask, router_weights = self.mixer(router_logits)
            router_mask = router_mask.view(bsz, seq_len, 1)
            router_weights = router_weights.view(bsz, seq_len, 1)
        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,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        if self.routing:
            # masking the hidden_states skip this layer
            hidden_states = residual + hidden_states * router_mask
        else:
            hidden_states = residual + hidden_states
        
        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        if self.routing:
            # masking the hidden_states skip this layer
            preserved_hidden_states = self.mlp(hidden_states) * router_mask * router_weights
            skipped_hidden_states = self.router_proj(hidden_states) * (1-router_mask) * (1-router_weights)
            hidden_states = residual + preserved_hidden_states + skipped_hidden_states
        else:
            hidden_states = self.mlp(hidden_states)
            hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if self.routing:
            outputs += (router_probs.reshape(bsz, seq_len),)
            outputs += (router_mask.reshape(bsz, seq_len),)

        if output_attentions:
            outputs += (self_attn_weights,)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


class DDLlamaPreTrainedModel(PreTrainedModel):
    config_class = DDLlamaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["DDLlamaDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True
    _supports_quantized_cache = 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_()

class DDLlamaModel(DDLlamaPreTrainedModel):
    config_class = DDLlamaConfig
    def __init__(self, config: DDLlamaConfig):
        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(
            [DDLlamaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = DDLlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = DDLlamaRotaryEmbedding(config=config)

        self.gradient_checkpointing = False
        if getattr(config, "pretraining_tp", 1) != 1:
            logger.warn("`pretraining_tp` is deprecated, please use `model.tensor_parallel` instead.")

        # 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[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,
        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
    ) -> 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 must specify exactly one of input_ids or inputs_embeds")

        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)

        # kept for BC (non `Cache` `past_key_values` inputs)
        return_legacy_cache = False
        if use_cache and not isinstance(past_key_values, Cache):
            return_legacy_cache = True
            if past_key_values is None:
                past_key_values = DynamicCache()
            else:
                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
                logger.warning_once(
                    "We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and "
                    "will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class "
                    "(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)"
                )

        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, output_attentions
        )
        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

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

        for decoder_layer in self.layers[: self.config.num_hidden_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,
                    position_embeddings,
                )
            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,
                    position_embeddings=position_embeddings,
                    **flash_attn_kwargs,
                )

            hidden_states = layer_outputs[0]

            output_idx = 1
            if decoder_layer.routing:
                all_router_weights += (layer_outputs[output_idx],)
                output_idx += 1
                all_router_masks += (layer_outputs[output_idx],)
                output_idx += 1

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

            if use_cache:
                next_decoder_cache = layer_outputs[output_idx]

        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 DDModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            router_weights=all_router_weights,
            router_masks=all_router_masks,
        )

    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool,
    ):
        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

        # 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.
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        using_static_cache = isinstance(past_key_values, StaticCache)

        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                is_training=self.training,
            ):
                return None

        dtype, device = input_tensor.dtype, input_tensor.device
        sequence_length = input_tensor.shape[1]
        if using_static_cache:
            target_length = past_key_values.get_max_cache_shape()
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            device=device,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
        )

        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
            min_dtype = torch.finfo(dtype).min
            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)

        return causal_mask

    @staticmethod
    def _prepare_4d_causal_attention_mask_with_cache_position(
        attention_mask: torch.Tensor,
        sequence_length: int,
        target_length: int,
        dtype: torch.dtype,
        device: torch.device,
        cache_position: torch.Tensor,
        batch_size: int,
        **kwargs,
    ):
        """
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
                `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache,
                to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            device (`torch.device`):
                The device to plcae the 4D attention mask on.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        """
        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.
            causal_mask = attention_mask
        else:
            min_dtype = torch.finfo(dtype).min
            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(batch_size, 1, -1, -1)
            if attention_mask is not None:
                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
                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
                )

        return causal_mask


class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...


class DDLlamaForCausalLM(DDLlamaPreTrainedModel, GenerationMixin):
    config_class = DDLlamaConfig
    _tied_weights_keys = ["lm_head.weight"]
    _tp_plan = {"lm_head": "colwise_rep"}

    def __init__(self, config):
        super().__init__(config)
        self.model = DDLlamaModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        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[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,
        num_logits_to_keep: int = 0,
        **kwargs: Unpack[KwargsForCausalLM],
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            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,
            **kwargs,
        )

        hidden_states = outputs[0]
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])

        loss = None
        if labels is not None:
            logits = logits.float()
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].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)
            shift_loss = loss_fct(shift_logits, shift_labels)
            loss = torch.mean(shift_loss)

        router_weights = outputs.router_weights
        router_masks = outputs.router_masks
        if len(router_weights) > 0:
            router_weights = [weight.to(hidden_states.device) for weight in router_weights]
            router_weights = torch.stack(router_weights, dim=-1).float()

            router_masks = [mask.to(hidden_states.device) for mask in router_masks]
            router_masks = torch.stack(router_masks, dim=-1).float()

            # print the number of layers used
            n_layers = torch.mean(torch.sum(router_masks, dim=2))
            print(16+int(n_layers.item()), end='')

            if self.training and labels is not None:
                router_weights *= router_masks
                shift_router_weights = router_weights[:, :-1, :].contiguous()
                shift_router_weights = shift_router_weights.view(-1, shift_router_weights.shape[-1])
                router_penalty_loss = torch.sum(shift_router_weights, dim=1)
                router_penalty_loss = 1e-3 * torch.mean((router_penalty_loss) ** 2)
                loss += router_penalty_loss
                # normalize the loss by your gradient accumulation steps
                loss /= 4

        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,
        )