modeling_swimbaII_v1.py

# coding=utf-8
# Copyright 2024 HuggingFace Inc. team.
# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
#
# 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 SwimbaI model."""

import math
from dataclasses import dataclass
from typing import Any, Dict, Optional, Tuple, Union

import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss

from transformers.activations import ACT2FN
from transformers.cache_utils import DynamicCache  # we need __iter__ and __len__ of pkv
from transformers.generation import GenerationMixin
from transformers.modeling_attn_mask_utils import (
    AttentionMaskConverter,
)
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import (
    ModelOutput,
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
)
from transformers.utils.import_utils import (
    is_causal_conv1d_available,
    is_flash_attn_2_available,
    is_flash_attn_greater_or_equal_2_10,
    is_mamba_2_ssm_available,
)
from configuration_swimba import SwimbaIConfig


logger = logging.get_logger(__name__)


# Copied from transformers.models.mixtral.modeling_mixtral.load_balancing_loss_func
def load_balancing_loss_func(
    gate_logits: Union[torch.Tensor, tuple[torch.Tensor], None],
    num_experts: Optional[int] = None,
    top_k=2,
    attention_mask: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, int]:
    r"""
    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.

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

    Args:
        gate_logits:
            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
            shape [batch_size X sequence_length, num_experts].
        num_experts:
            Number of experts
        top_k:
            The number of experts to route per-token, can be also interpreted as the `top-k` routing
            parameter.
        attention_mask (`torch.Tensor`, *optional*):
            The attention_mask used in forward function
            shape [batch_size X sequence_length] if not None.

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

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

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

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

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

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

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

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

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

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

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

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


# Copied from transformers.models.mamba.modeling_mamba2.modeling_mamba2.py with MAMBA2->NEMOTRONH,Mamba2->SwimbaI
# For Mamba2 components Mamba2->SwimbaIMamba2
if is_mamba_2_ssm_available():
    from mamba_ssm.ops.triton.selective_state_update import selective_state_update
    from mamba_ssm.ops.triton.ssd_combined import mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined
else:
    mamba_chunk_scan_combined, mamba_split_conv1d_scan_combined, selective_state_update = None, None, None

try:
    #from mamba_ssm.ops.triton.layernorm_gated import RMSNorm as RMSNormGated
    from mamba_ssm.ops.triton.layernorm_gated import rmsnorm_fn
except ImportError:
    raise ImportError("mamba-ssm is required by the Mamba model but cannot be imported")

if is_causal_conv1d_available():
    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
else:
    causal_conv1d_update, causal_conv1d_fn = None, None

if is_flash_attn_2_available():
    from transformers.modeling_flash_attention_utils import _flash_attention_forward

is_fast_path_available = all(
    (
        selective_state_update,
        mamba_chunk_scan_combined,
        mamba_split_conv1d_scan_combined,
        causal_conv1d_fn,
        causal_conv1d_update,
    )
)


_CHECKPOINT_FOR_DOC = "nvidia/SwimbaI-56B-Base-8K"
_CONFIG_FOR_DOC = "SwimbaIConfig"


# Helper methods for segment sum computation


def pad_tensor_by_size(input_tensor: torch.Tensor, pad_size: int):
    """
    Padding x tensor with `pad_size` on the seq_len dim (dim=1)

    Assumes that we only have tensors of either size 4 or 3
    """
    pad_shape = (0, 0, 0, 0, 0, pad_size, 0, 0) if len(input_tensor.shape) == 4 else (0, 0, 0, pad_size, 0, 0)

    return torch.nn.functional.pad(input_tensor, pad_shape, mode="constant", value=0)


def reshape_into_chunks(input_tensor, pad_size, chunk_size):
    """
    Padding input_tensor with `pad_size` on the seq_len dim (dim=1) and
    simultaneously splitting it into chunk sequences.

    Assumes that we only have tensors of either size 4 or 3
    """
    # [bsz, seq_len, ...] -> [bsz, seq_len multiple of chunk_size, ...]
    input_tensor = pad_tensor_by_size(input_tensor, pad_size)

    if len(input_tensor.shape) == 3:
        # [bsz, seq_len multiple of chunk_size, num_heads] -> [bsz, -1, chunk_size, num_heads]
        return input_tensor.reshape(input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2])
    else:
        # [bsz, seq_len multiple of chunk_size, num_heads, head_dim or state_size] -> [bsz, -1, chunk_size, num_heads, head_dim or state_size]
        return input_tensor.reshape(
            input_tensor.shape[0], -1, chunk_size, input_tensor.shape[2], input_tensor.shape[3]
        )


def segment_sum(input_tensor):
    """
    More stable segment sum calculation. Uses cumulative sums and masking instead of direct subtractions.
    """
    chunk_size = input_tensor.size(-1)
    # 1. expand input tensor to have an additional dimension and repeat along that dimension
    # [..., chunk_size] -> [..., chunk_size, chunk_size]
    input_tensor = input_tensor[..., None].expand(*input_tensor.size(), chunk_size)
    # 2. create a lower triangular mask with the diagonal set to 0 to 0 out elements above diag
    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=-1)
    input_tensor = input_tensor.masked_fill(~mask, 0)
    # 3. compute actual cumsum
    tensor_segsum = torch.cumsum(input_tensor, dim=-2)

    # 4. apply mask to keep only the lower triangular part of the cumulative sum result (incl diagonal this time)
    mask = torch.tril(torch.ones(chunk_size, chunk_size, device=input_tensor.device, dtype=torch.bool), diagonal=0)
    tensor_segsum = tensor_segsum.masked_fill(~mask, -torch.inf)
    return tensor_segsum


def apply_mask_to_padding_states(hidden_states, attention_mask):
    """
    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
    """
    if attention_mask is not None and attention_mask.shape[1] > 1 and attention_mask.shape[0] > 1:
        dtype = hidden_states.dtype
        hidden_states = (hidden_states * attention_mask[:, :, None]).to(dtype)

    return hidden_states

# Copied from NemotronH
# Copied from https://github.com/huggingface/transformers/blob/main/src/transformers/models/jamba/modeling_jamba.py
class HybridMambaAttentionDynamicCache(DynamicCache):
    """
    A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache
    (which has a constant shape regardless of seq_len).

    This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states`
    and `ssm_states` for mamba cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor
    For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`,
    while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors).
    For mamba layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors),
    while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`,
    and `ssm_states` represents the ssm state and has a shape of `(batch_size, d_inner, d_state)`.
    """

    def __init__(self, config, batch_size, dtype=torch.float16, device=None):
        super().__init__()
        self.dtype = dtype
        self.hybrid_override_pattern = config.hybrid_override_pattern
        self.has_previous_state = False  # only used by mamba
        intermediate_size = config.expand * config.hidden_size
        ssm_state_size = config.ssm_state_size
        conv_kernel_size = config.conv_kernel
        self.conv_states = []
        self.ssm_states = []
        self.transformer_layers = []
        for i in range(config.num_hidden_layers):
            if self.hybrid_override_pattern[i] == "M":
                # Mamba layer
                self.conv_states += [
                    torch.zeros(batch_size, intermediate_size, conv_kernel_size, device=device, dtype=dtype)
                ]
                self.ssm_states += [
                    torch.zeros(batch_size, intermediate_size, ssm_state_size, device=device, dtype=dtype)
                ]
            else:
                # Attention or MLP layer
                self.conv_states += [torch.tensor([[]] * batch_size, device=device)]
                self.ssm_states += [torch.tensor([[]] * batch_size, device=device)]
                self.transformer_layers.append(i)

        self.key_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]
        self.value_cache = [torch.tensor([[]] * batch_size, device=device) for _ in range(config.num_hidden_layers)]

    def update(
        self,
        key_states: torch.Tensor,
        value_states: torch.Tensor,
        layer_idx: int,
        cache_kwargs: Optional[Dict[str, Any]] = None,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        # Update the cache
        if self.key_cache[layer_idx].shape[-1] == 0:
            self.key_cache[layer_idx] = key_states
            self.value_cache[layer_idx] = value_states
        else:
            self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=2)
            self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=2)

        return self.key_cache[layer_idx], self.value_cache[layer_idx]

    def reorder_cache(self, beam_idx: torch.LongTensor):
        """Reorders the cache for beam search, given the selected beam indices."""
        for layer_idx in range(len(self.key_cache)):
            device = self.key_cache[layer_idx].device
            self.key_cache[layer_idx] = self.key_cache[layer_idx].index_select(0, beam_idx.to(device))
            device = self.value_cache[layer_idx].device
            self.value_cache[layer_idx] = self.value_cache[layer_idx].index_select(0, beam_idx.to(device))

            device = self.conv_states[layer_idx].device
            self.conv_states[layer_idx] = self.conv_states[layer_idx].index_select(0, beam_idx.to(device))
            device = self.ssm_states[layer_idx].device
            self.ssm_states[layer_idx] = self.ssm_states[layer_idx].index_select(0, beam_idx.to(device))

    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
        # take any layer that contains cache and not empty tensor
        layer_idx = self.transformer_layers[0] if layer_idx not in self.transformer_layers else layer_idx
        if len(self.key_cache) <= layer_idx:
            return 0
        return self.key_cache[layer_idx].shape[-2]

    def to_legacy_cache(self) -> Tuple[Tuple[torch.Tensor], Tuple[torch.Tensor]]:
        raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")

    @classmethod
    def from_legacy_cache(cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None) -> "DynamicCache":
        raise NotImplementedError("HybridMambaAttentionDynamicCache does not have a legacy cache equivalent.")

    # Copied from modeling_mamba2.py
    def update_conv_state(
        self, layer_idx: int, new_conv_state: torch.Tensor, cache_init: bool = False
    ) -> torch.Tensor:
        if cache_init:
            self.conv_states[layer_idx] = new_conv_state.to(self.conv_states.device)
        else:
            self.conv_states[layer_idx] = self.conv_states[layer_idx].roll(shifts=-1, dims=-1)
            self.conv_states[layer_idx][:, :, -1] = new_conv_state[:, 0, :].to(self.conv_states.device)
        return self.conv_states[layer_idx]

    def update_ssm_state(self, layer_idx: int, new_ssm_state: torch.Tensor):
        self.ssm_states[layer_idx] = new_ssm_state.to(self.ssm_states.device)
        return self.ssm_states[layer_idx]

    def reset(self):
        self.conv_states.zero_()
        self.ssm_states.zero_()

# Copied from NemotronH
class MambaRMSNormGated(torch.nn.Module):
    def __init__(self, hidden_size, group_size, eps=1e-5):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps
        self.group_size = group_size

    # jan28b version
    def forward(self, hidden_states, gate=None):
        return rmsnorm_fn(x=hidden_states,
                          weight=self.weight,
                          bias=None, # No bias
                          z=gate,
                          eps=self.variance_epsilon,
                          group_size=self.group_size,
                          norm_before_gate=False
        )

class SwimbaISparseMoEMamba2Mixer(nn.Module):
    """
    A modified Mamba2 mixer that incorporates MoE for the input projection.
    """

    def __init__(self, config: SwimbaIConfig, layer_idx: int):
        super().__init__()
        self.num_heads = config.mamba_num_heads
        self.hidden_size = config.hidden_size
        self.ssm_state_size = config.ssm_state_size
        self.conv_kernel_size = config.conv_kernel
        self.intermediate_size = config.mamba_num_heads * config.mamba_head_dim
        self.layer_idx = layer_idx
        self.use_conv_bias = config.use_conv_bias
        self.activation = config.mamba_hidden_act
        self.act = ACT2FN[config.mamba_hidden_act]
        self.layer_norm_epsilon = config.layer_norm_epsilon
        self.n_groups = config.n_groups
        self.head_dim = config.mamba_head_dim
        self.chunk_size = config.chunk_size
        self.time_step_limit = config.time_step_limit
        self.time_step_min = config.time_step_min
        self.time_step_max = config.time_step_max
        self.conv_dim = self.intermediate_size + 2 * self.n_groups * self.ssm_state_size
        self.use_bias = config.use_bias

        # MoE specific attributes
        self.num_experts = config.num_experts
        self.top_k = config.num_experts_per_tok
        self.norm_topk_prob = getattr(config, "norm_topk_prob", False)
        self.gate = nn.Linear(self.hidden_size, self.num_experts, bias=False)
        self.projection_size = self.intermediate_size + self.conv_dim + self.num_heads
        self.experts = nn.ModuleList(
            [nn.Linear(self.hidden_size, self.projection_size, bias=self.use_bias) for _ in range(self.num_experts)]
        )

        self.conv1d = nn.Conv1d(
            in_channels=self.conv_dim,
            out_channels=self.conv_dim,
            bias=config.use_conv_bias,
            kernel_size=config.conv_kernel,
            groups=self.conv_dim,
            padding=config.conv_kernel - 1,
        )
        self.dt_bias = nn.Parameter(torch.ones(self.num_heads))
        A = torch.arange(1, self.num_heads + 1)
        self.A_log = nn.Parameter(torch.log(A))
        self.A_log._no_weight_decay = True
        self.norm = MambaRMSNormGated(self.intermediate_size, eps=self.layer_norm_epsilon, group_size=self.intermediate_size // self.n_groups)
        self.D = nn.Parameter(torch.ones(self.num_heads))
        self.D._no_weight_decay = True
        self.out_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.use_bias)

    # Copied from OLMoE
    def _moe_forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
        # router_logits: (batch * sequence_length, n_experts)
        router_logits = self.gate(hidden_states)

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

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

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

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

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

            # However `index_add_` only support torch tensors for indexing so we'll use
            # the `top_x` tensor here.
            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, self.projection_size)
        return final_hidden_states, router_logits

    def cuda_kernels_forward(
        self,
        hidden_states: torch.Tensor,
        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
    ):
        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
        projected_states, router_logits = self._moe_forward(hidden_states)

        batch_size, seq_len, _ = hidden_states.shape
        groups_time_state_size = self.n_groups * self.ssm_state_size
        d_mlp = (
            projected_states.shape[-1]
            - 2 * self.intermediate_size
            - 2 * self.n_groups * self.ssm_state_size
            - self.num_heads
        ) // 2

        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
            _, _, gate, hidden_states_B_C, dt = projected_states.squeeze(1).split(
                [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
            )
            hidden_states_B_C = causal_conv1d_update(
                hidden_states_B_C,
                cache_params.conv_states[self.layer_idx],
                self.conv1d.weight.squeeze(1),
                self.conv1d.bias,
                self.activation,
            )
            hidden_states, B, C = torch.split(
                hidden_states_B_C,
                [self.intermediate_size, groups_time_state_size, groups_time_state_size],
                dim=-1,
            )
            A = -torch.exp(self.A_log.float())
            A = A[:, None, ...][:, :, None].expand(-1, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
            dt = dt[:, :, None].expand(-1, -1, self.head_dim)
            dt_bias = self.dt_bias[:, None, ...].expand(-1, self.head_dim)
            D = self.D[:, None, ...].expand(-1, self.head_dim)
            B = B.view(batch_size, self.n_groups, B.shape[1] // self.n_groups)
            C = C.view(batch_size, self.n_groups, C.shape[1] // self.n_groups)
            hidden_states_reshaped = hidden_states.view(batch_size, self.num_heads, self.head_dim)
            hidden_states = selective_state_update(
                cache_params.ssm_states[self.layer_idx],
                hidden_states_reshaped,
                dt,
                A,
                B,
                C,
                D,
                z=None,
                dt_bias=dt_bias,
                dt_softplus=True,
            )
            hidden_states = hidden_states.view(batch_size, self.num_heads * self.head_dim)
            hidden_states = self.norm(hidden_states, gate)
            out = self.out_proj(hidden_states)[:, None, ...]
        else:
            A = -torch.exp(self.A_log.float())
            dt_limit_kwargs = {} if self.time_step_limit == (0.0, float("inf")) else {"dt_limit": self.time_step_limit}
            if self.training and cache_params is None:
                out = mamba_split_conv1d_scan_combined(
                    projected_states,
                    self.conv1d.weight.squeeze(1),
                    self.conv1d.bias,
                    self.dt_bias,
                    A,
                    D=self.D,
                    chunk_size=self.chunk_size,
                    seq_idx=None,
                    activation=self.activation,
                    rmsnorm_weight=self.norm.weight,
                    rmsnorm_eps=self.norm.variance_epsilon,
                    outproj_weight=self.out_proj.weight,
                    outproj_bias=self.out_proj.bias,
                    headdim=self.head_dim,
                    ngroups=self.n_groups,
                    norm_before_gate=False,
                    return_final_states=False,
                    **dt_limit_kwargs,
                )
            else:
                _, _, gate, hidden_states_B_C, dt = projected_states.split(
                    [d_mlp, d_mlp, self.intermediate_size, self.conv_dim, self.num_heads], dim=-1
                )
                if cache_params is not None:
                    hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
                    conv_states = nn.functional.pad(
                        hidden_states_B_C_transposed,
                        (cache_params.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0),
                    )
                    cache_params.update_conv_state(
                        layer_idx=self.layer_idx, new_conv_state=conv_states, cache_init=True
                    )
                if self.activation not in ["silu", "swish"]:
                    hidden_states_B_C = self.act(
                        self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2)
                    )
                else:
                    hidden_states_B_C = causal_conv1d_fn(
                        x=hidden_states_B_C.transpose(1, 2),
                        weight=self.conv1d.weight.squeeze(1),
                        bias=self.conv1d.bias,
                        activation=self.activation,
                    ).transpose(1, 2)
                hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
                hidden_states, B, C = torch.split(
                    hidden_states_B_C,
                    [self.intermediate_size, groups_time_state_size, groups_time_state_size],
                    dim=-1,
                )
                scan_output, ssm_state = mamba_chunk_scan_combined(
                    hidden_states.view(batch_size, seq_len, -1, self.head_dim),
                    dt,
                    A,
                    B.view(batch_size, seq_len, self.n_groups, -1),
                    C.view(batch_size, seq_len, self.n_groups, -1),
                    chunk_size=self.chunk_size,
                    D=self.D,
                    z=None,
                    seq_idx=None,
                    return_final_states=True,
                    dt_bias=self.dt_bias,
                    dt_softplus=True,
                    **dt_limit_kwargs,
                )
                if ssm_state is not None and cache_params is not None:
                    cache_params.update_ssm_state(layer_idx=self.layer_idx, new_ssm_state=ssm_state)
                scan_output = scan_output.view(batch_size, seq_len, -1)
                scan_output = self.norm(scan_output, gate)
                out = self.out_proj(scan_output)
        return out, router_logits

    def torch_forward(self, hidden_states, cache_params: Optional[HybridMambaAttentionDynamicCache]=None, cache_position:Optional[torch.LongTensor]=None, attention_mask: Optional[torch.Tensor]=None):
        batch_size, seq_len, _ = hidden_states.shape
        dtype = hidden_states.dtype
        hidden_states = apply_mask_to_padding_states(hidden_states, attention_mask)
        projected_states, router_logits = self._moe_forward(hidden_states)
        d_mlp = (projected_states.shape[-1] - 2 * self.intermediate_size - 2 * self.n_groups * self.ssm_state_size-self.num_heads) // 2
        _, _, gate, hidden_states_B_C, dt = projected_states.split(
                [d_mlp, d_mlp, self.intermediate_size,  self.conv_dim, self.num_heads], dim=-1
        )

        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
            cache_params.update_conv_state(layer_idx=self.layer_idx, new_conv_state=hidden_states_B_C, cache_init=False)
            conv_states = cache_params.conv_states[self.layer_idx].to(device=self.conv1d.weight.device)
            hidden_states_B_C = torch.sum(
                conv_states * self.conv1d.weight.squeeze(1), dim=-1
            )
            if self.use_conv_bias:
                hidden_states_B_C = hidden_states_B_C + self.conv1d.bias
            hidden_states_B_C = self.act(hidden_states_B_C)
        else:
            if cache_params is not None:
                hidden_states_B_C_transposed = hidden_states_B_C.transpose(1, 2)
                conv_states = nn.functional.pad(
                    hidden_states_B_C_transposed, (cache_params.conv_kernel_size - hidden_states_B_C_transposed.shape[-1], 0)
                )
                cache_params.update_conv_state(layer_idx=self.layer_idx, new_conv_state=conv_states, cache_init=True)
            hidden_states_B_C = self.act(self.conv1d(hidden_states_B_C.transpose(1, 2))[..., :seq_len].transpose(1, 2))

        hidden_states_B_C = apply_mask_to_padding_states(hidden_states_B_C, attention_mask)
        hidden_states, B, C = torch.split(
            hidden_states_B_C,
            [self.intermediate_size, self.n_groups * self.ssm_state_size, self.n_groups * self.ssm_state_size],
            dim=-1
        )
        A = -torch.exp(self.A_log.float())
        if cache_params is not None and cache_position is not None and cache_position[0] > 0:
            cache_device = cache_params.ssm_states.device
            dt = dt[:, 0, :][:, None, ...]
            dt = dt.transpose(1, 2).expand(batch_size, dt.shape[-1], self.head_dim)
            dt_bias = self.dt_bias[..., None].expand(self.dt_bias.shape[0], self.head_dim)
            dt = torch.nn.functional.softplus(dt + dt_bias.to(dt.dtype))
            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
            A = A[..., None, None].expand(self.num_heads, self.head_dim, self.ssm_state_size).to(dtype=torch.float32)
            dA = (torch.exp(dt[..., None] * A)).to(device=cache_device)
            B = B.reshape(batch_size, self.n_groups, -1)[..., None, :]
            B = B.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, B.shape[-1]).contiguous()
            B = B.reshape(batch_size, -1, B.shape[-1])
            dB = dt[..., None] * B[..., None, :]
            hidden_states = hidden_states.reshape(batch_size, -1, self.head_dim)
            dBx = (dB * hidden_states[..., None]).to(device=cache_device)
            cache_params.update_ssm_state(
                layer_idx=self.layer_idx,
                new_ssm_state=cache_params.ssm_states[self.layer_idx] * dA + dBx
            )
            C = C.reshape(batch_size, self.n_groups, -1)[..., None, :]
            C = C.expand(batch_size, self.n_groups, self.num_heads // self.n_groups, C.shape[-1]).contiguous()
            C = C.reshape(batch_size, -1, C.shape[-1])
            ssm_states = cache_params.ssm_states[self.layer_idx].to(device=C.device, dtype=C.dtype)
            ssm_states_reshaped = ssm_states.view(batch_size * self.num_heads, self.head_dim, self.ssm_state_size)
            C_reshaped = C.view(batch_size * self.num_heads, self.ssm_state_size, 1)
            y = torch.bmm(ssm_states_reshaped, C_reshaped)
            y = y.view(batch_size, self.num_heads, self.head_dim)
            D = self.D[..., None].expand(self.D.shape[0], self.head_dim)
            y = (y + hidden_states * D).to(y.dtype)
            y = y.reshape(batch_size, -1)[:, None, ...]
        else:
            dt = nn.functional.softplus(dt + self.dt_bias)
            dt = torch.clamp(dt, self.time_step_limit[0], self.time_step_limit[1])
            hidden_states = hidden_states.reshape(batch_size, seq_len, -1, self.head_dim).float()
            B = B.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
            C = C.reshape(batch_size, seq_len, -1, self.ssm_state_size).float()
            B = B.repeat(1, 1, self.num_heads // self.n_groups, 1)
            C = C.repeat(1, 1, self.num_heads // self.n_groups, 1)
            pad_size = (self.chunk_size - seq_len % self.chunk_size) % self.chunk_size
            D_residual = self.D[..., None] * pad_tensor_by_size(hidden_states, pad_size)
            hidden_states = hidden_states * dt[..., None]
            A = A.to(hidden_states.dtype) * dt
            hidden_states, A, B, C = [reshape_into_chunks(t, pad_size, self.chunk_size) for t in (hidden_states, A, B, C)]
            A = A.permute(0, 3, 1, 2)
            A_cumsum = torch.cumsum(A, dim=-1)
            L = torch.exp(segment_sum(A))
            G_intermediate = C[:, :, :, None, :, :] * B[:, :, None, :, :, :]
            G = G_intermediate.sum(dim=-1)
            M_intermediate = G[..., None] * L.permute(0, 2, 3, 4, 1)[..., None]
            M = M_intermediate.sum(dim=-1)
            Y_diag = (M[..., None] * hidden_states[:, :, None]).sum(dim=3)
            decay_states = torch.exp((A_cumsum[:, :, :, -1:] - A_cumsum))
            B_decay = B * decay_states.permute(0, -2, -1, 1)[..., None]
            states = (B_decay[..., None, :] * hidden_states[..., None]).sum(dim=2)
            if cache_params is not None and cache_position is not None and cache_position[0] > 0:
                previous_states = cache_params.ssm_states[self.layer_idx][:, None, ...].to(device=states.device)
            else:
                previous_states = torch.zeros_like(states[:, :1])
            states = torch.cat([previous_states, states], dim=1)
            decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))
            decay_chunk = decay_chunk.transpose(1, 3)
            new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
            states, ssm_state = new_states[:, :-1], new_states[:, -1]
            state_decay_out = torch.exp(A_cumsum)
            C_times_states = (C[..., None, :] * states[:, :, None, ...])
            state_decay_out_permuted = state_decay_out.permute(0, 2, 3, 1)
            Y_off = (C_times_states.sum(-1) * state_decay_out_permuted[..., None])
            y = Y_diag + Y_off
            y = y.reshape(batch_size, -1, self.num_heads, self.head_dim)
            y = y + D_residual
            if pad_size > 0:
                y = y[:, :seq_len, :, :]
            y = y.reshape(batch_size, seq_len, -1)
            if ssm_state is not None and cache_params is not None:
                cache_params.update_ssm_state(layer_idx=self.layer_idx, new_ssm_state=ssm_state)
        scan_output = self.norm(y, gate)
        contextualized_states = self.out_proj(scan_output.to(dtype))
        return contextualized_states, router_logits

    def forward(
        self,
        hidden_states,
        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
    ):
        if is_fast_path_available and "cuda" in self.gate.weight.device.type:
            return self.cuda_kernels_forward(hidden_states, cache_params, cache_position, attention_mask)
        return self.torch_forward(hidden_states, cache_params, cache_position, attention_mask)


# Copied from NemotronH
class SwimbaIRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        SwimbaIRMSNorm is equivalent to T5LayerNorm and LlamaRMSNorm
        """
        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)
        # Weights are in float32
        return (self.weight.to(torch.float32) * hidden_states).to(input_dtype)


# Copied from transformers.models.nemotron.modeling_nemotron Nemotron->SwimbaI
# Copied from NemotronH
class SwimbaIMLP(nn.Module):
    def __init__(self, config, 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.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        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.mlp_hidden_act]

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

# Copied from NemotronH
class SwimbaIBlock(nn.Module):
    def __init__(self, config, layer_idx):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.residual_in_fp32 = config.residual_in_fp32
        self.norm = SwimbaIRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

        # M: Mamba2, *: Attention, -: MLP
        self.block_type = config.layers_block_type[layer_idx]
        if self.block_type == "mamba":
            self.mixer = SwimbaISparseMoEMamba2Mixer(config, layer_idx=layer_idx)
        elif self.block_type == "attention":
            self.mixer = SWIMBAI_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
        elif self.block_type == "mlp":
            self.mixer = SwimbaIMLP(config)
        else:
            raise ValueError(f"Invalid layer pattern {config.hybrid_override_pattern[layer_idx]}")

    def forward(
        self,
        hidden_states,
        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        output_router_logits: bool = False,
        use_cache: bool = False,
    ):
        with torch.cuda.stream(torch.cuda.default_stream(hidden_states.device)):
            # * Use torch.cuda.stream() to avoid NaN issues when using multiple GPUs
            residual = hidden_states
            hidden_states = self.norm(hidden_states.to(dtype=self.norm.weight.dtype))
            if self.residual_in_fp32:
                residual = residual.to(torch.float32)

            router_logits = None
            if self.block_type == "mamba":
                hidden_states, router_logits = self.mixer(
                    hidden_states, cache_params=cache_params, cache_position=cache_position
                )
            elif self.block_type == "attention":
                hidden_states, _, _ = self.mixer(
                    hidden_states, cache_position=cache_position
                )
            elif self.block_type == "mlp":
                hidden_states = self.mixer(
                    hidden_states
                )
            else:
                raise ValueError(f"Invalid block_type: {self.block_type}")

            hidden_states = residual + hidden_states

            outputs = (hidden_states,)
            if output_router_logits:
                outputs += (router_logits,)
            return outputs


# Copied from transformers.models.llama.modeling_llama.repeat_kv
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


# Copied from NemotronH
class SwimbaIAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: SwimbaIConfig, 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
        if config.attention_head_dim is not None:
            self.head_dim = config.attention_head_dim
        else:
            self.head_dim = config.hidden_size // config.num_attention_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.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.head_dim * self.num_heads, self.hidden_size, bias=config.attention_bias)

    def forward(
        self,
        hidden_states: torch.Tensor,
        # position_embeddings: Tuple[torch.Tensor, torch.Tensor], #TODO
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

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

        if past_key_value is not None:
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)

        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:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]

        if query_states.device.type == "cuda" and attention_mask is not None:
            query_states = query_states.contiguous()
            key_states = key_states.contiguous()
            value_states = value_states.contiguous()

        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, self.hidden_size)
        attn_output = attn_output.view(bsz, q_len, self.num_heads * self.head_dim)

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value


# Adapted from transformers.models.mistral.modeling_mistral.MistralFlashAttention2 with Mistral->Jamba
#class JambaFlashAttention2(JambaAttention):
# Copied from NemotronH
class SwimbaIFlashAttention2(SwimbaIAttention):
    """
    Jamba flash attention module. This module inherits from `JambaAttention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    """
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

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

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs,
    ):
        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)
        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)

        if past_key_value is not None:
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)

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

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

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

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

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

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

        #attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim).contiguous()
        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value


# Adapted from transformers.models.mistral.modeling_mistral.MistralSdpaAttention with Mistral->Jamba
#class JambaSdpaAttention(JambaAttention):
# Copied from NemotronH
class SwimbaISdpaAttention(SwimbaIAttention):
    """
    Jamba attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
    `JambaAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
    SDPA API.
    """

    # Adapted from SwimbaIAttention.forward
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[HybridMambaAttentionDynamicCache] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
        cache_position: Optional[torch.LongTensor] = None,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        if output_attentions:
            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
            logger.warning_once(
                "SwimbaIModel is using SwimbaISdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
            )
            return super().forward(
                hidden_states=hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
            )

        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

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

        if past_key_value is not None:
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)

        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 attention_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.
        # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
        is_causal = True if self.is_causal and 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, self.hidden_size)

        attn_output = self.o_proj(attn_output)

        return attn_output, None, past_key_value


SWIMBAI_ATTENTION_CLASSES = {
    "eager": SwimbaIAttention,
    "flash_attention_2": SwimbaIFlashAttention2,
    "sdpa": SwimbaISdpaAttention,
}

# Copied from transformers.models.mamba.modeling_mamba2.Mamba2PreTrainedModel
# Copied from NemotronH
class SwimbaIPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = SwimbaIConfig
    base_model_prefix = "backbone"
    _no_split_modules = ["SwimbaIBlock"]
    supports_gradient_checkpointing = True
    _is_stateful = True

    def _init_weights(self, module):
        """Initialize the weights."""
        if isinstance(module, SwimbaISparseMoEMamba2Mixer):
            module.A_log._no_weight_decay = True
            module.D._no_weight_decay = True

            dt = torch.exp(
                torch.rand(self.config.mamba_num_heads)
                * (math.log(self.config.time_step_max) - math.log(self.config.time_step_min))
                + math.log(self.config.time_step_min)
            ).clamp(min=self.config.time_step_floor)

            # # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
            inv_dt = dt + torch.log(-torch.expm1(-dt))
            with torch.no_grad():
                module.dt_bias.copy_(inv_dt)
            module.dt_bias._no_reinit = True

        if isinstance(module, nn.Linear):
            if module.bias is not None:
                if not getattr(module.bias, "_no_reinit", False):
                    nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            nn.init.normal_(module.weight, std=self.config.initializer_range)

        # TODO: Check
        if self.config.rescale_prenorm_residual:
            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
            #
            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
            for name, p in module.named_parameters():
                if name in ["out_proj.weight"]:
                    # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
                    # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
                    # We need to reinit p since this code could be called multiple times
                    # Having just p *= scale would repeatedly scale it down
                    nn.init.kaiming_uniform_(p, a=math.sqrt(5))
                    with torch.no_grad():
                        p /= math.sqrt(self.config.num_hidden_layers)


@dataclass
# Copied from transformers.models.mamba.modeling_mamba2.Mamba2Output with MAMBA2->SwimbaI,Mamba2->SwimbaI
# Copied from NemotronH
class SwimbaIOutput(ModelOutput):
    """
    Class for the SwimbaI model outputs.

    Args:
        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
            Sequence of hidden-states at the output of the last layer of the model.
        cache_params (`HybridMambaAttentionDynamicCache`):
            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
            avoid providing the old `input_ids`.

            Includes both the State space model state matrices after the selective scan, and the Convolutional states
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
    """

    last_hidden_state: Optional[torch.FloatTensor] = None
    cache_params: Optional[HybridMambaAttentionDynamicCache] = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None
    router_logits: Optional[Tuple[torch.FloatTensor]] = None


@dataclass
# Copied from transformers.models.mamba2.modeling_mamba2.MambaCausalLMOutput with Mamba2->SwimbaI
# Copied from NemotronH
class SwimbaICausalLMOutput(ModelOutput):
    """
    Base class for causal language model (or autoregressive) outputs.

    Args:
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
            Language modeling loss (for next-token prediction).
        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
        cache_params (`HybridMambaAttentionDynamicCache`):
            The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to
            avoid providing the old `input_ids`.

            Includes both the State space model state matrices after the selective scan, and the Convolutional states
        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.

            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
    """

    loss: Optional[torch.FloatTensor] = None
    logits: Optional[torch.FloatTensor] = None
    cache_params: Optional[HybridMambaAttentionDynamicCache] = None
    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[torch.FloatTensor]] = None
    aux_loss: Optional[torch.FloatTensor] = None
    router_logits: Optional[Tuple[torch.FloatTensor]] = None


SWIMBAI_START_DOCSTRING = r"""

    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)

    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.

    Parameters:
        config ([`SwimbaIConfig`]): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the
            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""

SWIMBAI_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
            Indices of input sequence tokens in the vocabulary.

            If `cache_params.seqlen_offset>0`, only `input_ids` that do not have their past calculated should be passed as
            `input_ids`.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
            model's internal embedding lookup matrix.
        position_ids (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings.
        cache_params (`HybridMambaAttentionDynamicCache`, *optional*):
            If passed along, the model uses the previous state in all the blocks (which will give the output for the
            `input_ids` provided as if the model add `state_input_ids + input_ids` as context).
        use_cache (`bool`, *optional*):
            If set to `True`, the `cache_params` is returned and can be used to quickly generate the next logits.
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers.
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
            more detail.
        return_dict (`bool`, *optional*):
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
        cache_position (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            The position of the current input in the cache. This is used to ensure that the cache is correctly updated.
            If `cache_params` is passed, `cache_position` should also be passed.
        attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

            [What are attention masks?](../glossary#attention-mask)
"""


@add_start_docstrings(
    "The bare SwimbaI Model transformer outputting raw hidden-states without any specific head on top.",
    SWIMBAI_START_DOCSTRING,
)
# Copied from NemotronH
class SwimbaIModel(SwimbaIPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
        self.layers = nn.ModuleList([SwimbaIBlock(config, layer_idx=idx) for idx in range(config.num_hidden_layers)])

        self.gradient_checkpointing = False
        self.norm_f = SwimbaIRMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
        # Initialize weights and apply final processing
        self._register_load_state_dict_pre_hook(self.load_hook)
        self.post_init()

    def load_hook(self, state_dict, prefix, *args):
        for k in state_dict:
            if "embedding." in k:
                state_dict[k.replace("embedding.", "embeddings.")] = state_dict.pop(k)
                break

    def get_input_embeddings(self):
        return self.embeddings

    def set_input_embeddings(self, new_embeddings):
        self.embeddings = new_embeddings

    @add_start_docstrings_to_model_forward(SWIMBAI_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=SwimbaIOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_router_logits: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Union[Tuple, SwimbaIOutput]:
        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 if not self.training else False)
        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )

        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):  # ^ is python for xor
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

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

        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

        # From zamba_modeling.py
        if use_cache and cache_params is None:
            logger.warning_once(
                "SwimbaI requires an initialized `SwimbaIHybridDynamicCache` to return a cache. None was "
                "provided, so no cache will be returned."
            )

        hidden_states = inputs_embeds

        if cache_position is None:
            cache_position = torch.arange(hidden_states.shape[1], device=hidden_states.device)
        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position)
        mamba_mask = self._update_mamba_mask(attention_mask, cache_position)

        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

        for layer_idx, mixer_block in enumerate(self.layers):
            # Depending on the layer type we opt for 2D base attention mask (Mamba) or 4D causal mask (Attention)
            if mixer_block.block_type == "mamba":
                layer_mask = mamba_mask
            elif mixer_block.block_type == "attention":
                layer_mask = causal_mask
            elif mixer_block.block_type == "mlp":
                layer_mask = None
            else:
                raise ValueError(f"Invalid block_type: {self.block_type}")

            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    mixer_block.__call__, hidden_states, cache_params, cache_position, layer_mask, output_router_logits, use_cache
                )
            else:
                layer_outputs = mixer_block(
                    hidden_states,
                    cache_params=cache_params,
                    cache_position=cache_position,
                    attention_mask=layer_mask,
                    output_router_logits=output_router_logits,
                    use_cache=use_cache,
                )
            
            hidden_states = layer_outputs[0]
            if output_router_logits and layer_outputs[1] is not None:
                all_router_logits += (layer_outputs[1],)

        hidden_states = self.norm_f(hidden_states)

        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, cache_params, all_hidden_states, all_self_attns, all_router_logits] if v is not None)

        return SwimbaIOutput(
            last_hidden_state=hidden_states,
            cache_params=cache_params if use_cache else None,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            router_logits=all_router_logits,
        )

    # Copied from transformers.models.jamba.modeling_jamba.JambaModel._update_causal_mask
    def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
        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]
        target_length = cache_position[-1] + 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)

        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

    def _update_mamba_mask(self, attention_mask, cache_position):
        """
        No need for zeroing states when
            1. Cached forward
            2. Attending to all inputs
        """
        mamba_mask = attention_mask
        if cache_position[0] > 0 or (attention_mask is not None and torch.all(attention_mask == 1)):
            mamba_mask = None
        return mamba_mask


@add_start_docstrings(
    """
    The SWIMBAI Model transformer with a language modeling head on top (linear layer with weights not tied to the input
    embeddings).
    """,
    SWIMBAI_START_DOCSTRING,
)
# Copied from NemotronH
class SwimbaIForCausalLM(SwimbaIPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]

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

        # MoE config
        self.router_aux_loss_coef = config.router_aux_loss_coef
        self.num_experts = config.num_experts
        self.num_experts_per_tok = config.num_experts_per_tok

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

    def get_input_embeddings(self):
        return self.backbone.get_input_embeddings()

    def set_input_embeddings(self, new_embeddings):
        return self.backbone.set_input_embeddings(new_embeddings)

    def get_output_embeddings(self):
        return self.lm_head

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

    def get_decoder(self):
        return self.model

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

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        cache_position=None,
        position_ids=None,
        use_cache=True,
        **kwargs,
    ):
        # Copy from https://github.com/huggingface/transformers/blob/main/src/transformers/models/jamba/modeling_jamba.py
        # Overwitten -- uses `cache_params` as opposed to `past_key_values`
        empty_past_kv = past_key_values is None

        # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
        # Exception 1: when passing input_embeds, input_ids may be missing entries
        # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
        # Exception 3: with synced GPUs cache_position may go out of bounds, but we only want dummy token in that case.
        #              (we can't check exception 3 while compiling)
        if not empty_past_kv:
            if (
                inputs_embeds is not None  # Exception 1
                or cache_position[-1] >= input_ids.shape[1]  # Exception 3
            ):
                input_ids = input_ids[:, -cache_position.shape[0] :]
            elif input_ids.shape[1] != cache_position.shape[0]:  # Default case (the "else", a no op, is Exception 2)
                input_ids = input_ids[:, cache_position]
        else:
            past_key_values = HybridMambaAttentionDynamicCache(
                self.config, input_ids.shape[0], self.dtype, device=self.device
            )

        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 not empty_past_kv:
                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 empty_past_kv:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids.contiguous()}  # `contiguous()` needed for compilation use cases

        model_inputs.update(
            {
                "position_ids": position_ids,
                "past_key_values": past_key_values,
                "use_cache": use_cache,
                "attention_mask": attention_mask,
                "logits_to_keep": self.config.num_logits_to_keep,
                "cache_position": cache_position,
            }
        )
        return model_inputs

    @add_start_docstrings_to_model_forward(SWIMBAI_INPUTS_DOCSTRING)
    @add_code_sample_docstrings(
        checkpoint=_CHECKPOINT_FOR_DOC,
        output_type=SwimbaICausalLMOutput,
        config_class=_CONFIG_FOR_DOC,
    )
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        cache_params: Optional[HybridMambaAttentionDynamicCache] = None,
        labels: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_router_logits: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        **kwargs,  # for now we need this for generation
    ) -> Union[Tuple, SwimbaICausalLMOutput]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_router_logits = (
            output_router_logits if output_router_logits is not None else self.config.output_router_logits
        )
        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

        swimba_i_outputs = self.backbone(
            input_ids,
            cache_params=cache_params,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            output_router_logits=output_router_logits,
            return_dict=return_dict,
            use_cache=use_cache,
            cache_position=cache_position,
            attention_mask=attention_mask,
        )
        hidden_states = swimba_i_outputs[0]

        # TODO: Check zamba_modeling.py: https://github.com/huggingface/transformers/blob/d7188ba600e36d3fd191b12e19f1b3bb81a8404f/src/transformers/models/zamba/modeling_zamba.py#L1284C1-L1286C2
        #logits = self.lm_head(hidden_states.to(self.lm_head.weight.dtype)).float()
        logits = self.lm_head(hidden_states.to(self.lm_head.weight.dtype)).float()

        loss = None
        if labels is not None:
            # move labels to correct device to enable model parallelism
            labels = labels.to(logits.device)
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

        aux_loss = None
        if output_router_logits:
            # Copied from Olmoe
            aux_loss = load_balancing_loss_func(
                swimba_i_outputs.router_logits if return_dict else swimba_i_outputs[-1],
                self.num_experts,
                self.num_experts_per_tok,
                attention_mask,
            )
            if labels is not None:
                loss += self.router_aux_loss_coef * aux_loss.to(loss.device)

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

        return SwimbaICausalLMOutput(
            loss=loss,
            logits=logits,
            cache_params=swimba_i_outputs.cache_params,
            hidden_states=swimba_i_outputs.hidden_states,
            attentions=swimba_i_outputs.attentions,
            aux_loss=aux_loss,
            router_logits=swimba_i_outputs.router_logits,
        )