Upload folder using huggingface_hub

configuration_bailing_shared_moe_v2.py

@@ -0,0 +1,84 @@
+"""Bailing MoE V2 model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+
+
+class BailingSharedMoeV2Config(PretrainedConfig):
+
+    def __init__(
+        self,
+        vocab_size=157184,
+        hidden_size=2048,
+        intermediate_size=5120,
+        num_hidden_layers=20,
+        num_attention_heads=16,
+        num_key_value_heads=4,
+        hidden_act="silu",
+        use_qkv_bias=False,  # bailing only
+        use_bias=False,  # bailing only
+        rms_norm_eps=1e-06,
+        tie_word_embeddings=False,  # PretrainedConfig key, here change default value.
+        embedding_dropout=0.0,
+        attention_dropout=0.0,
+        output_dropout=0.0,
+        initializer_range=0.02,
+        max_position_embeddings=32768,
+        rope_theta=600000.0,
+        use_cache=True,
+        max_window_layers=20,
+        rope_scaling=None,
+        pad_token_id=156892,
+        eos_token_id=156892,
+        num_experts=256,
+        num_shared_experts=1,
+        num_experts_per_tok=8,
+        n_group=8,
+        topk_group=4,
+        moe_intermediate_size=512,
+        first_k_dense_replace=1,
+        head_dim=128,
+        output_router_logits=False,
+        use_qk_norm=True,
+        num_nextn_predict_layers=0,
+        mtp_loss_scaling_factor=0,
+        moe_router_enable_expert_bias=True,
+        routed_scaling_factor=1.0,
+        **kwargs,
+    ):
+        self.num_hidden_layers = num_hidden_layers
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.use_qkv_bias = use_qkv_bias
+        self.use_bias = use_bias
+        self.rms_norm_eps = rms_norm_eps
+        self.embedding_dropout = embedding_dropout
+        self.attention_dropout = attention_dropout
+        self.output_dropout = output_dropout
+        self.num_nextn_predict_layers = num_nextn_predict_layers
+        self.mtp_loss_scaling_factor = mtp_loss_scaling_factor
+        self.initializer_range = initializer_range
+        self.max_position_embeddings = max_position_embeddings
+        self.rope_theta = rope_theta
+        self.use_cache = use_cache
+        self.max_window_layers = max_window_layers
+        self.head_dim = head_dim or self.hidden_size // self.num_attention_heads
+        self.rope_scaling = rope_scaling
+        self.use_qk_norm = use_qk_norm
+        self.moe_router_enable_expert_bias = moe_router_enable_expert_bias
+        self.routed_scaling_factor = routed_scaling_factor
+
+        # MoE configs
+        self.num_experts = num_experts
+        self.num_shared_experts = num_shared_experts
+        self.num_experts_per_tok = num_experts_per_tok
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.moe_intermediate_size = moe_intermediate_size
+        self.first_k_dense_replace = first_k_dense_replace
+        self.output_router_logits = output_router_logits
+
+        super().__init__(pad_token_id=pad_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs)

convert_hf_to_scm.py

@@ -0,0 +1,137 @@
+import glob
+import re
+import shutil
+import sys
+
+import accelerate
+import torch
+from configuration_bailing_shared_moe_v2 import BailingSharedMoeV2Config
+from modeling_bailing_shared_moe_v2 import BailingSharedMoeV2ForCausalLM
+from configuration_bailing_moe_v2 import BailingMoeV2Config
+from safetensors import safe_open
+
+input_model = sys.argv[1]
+output_model_path = sys.argv[2]
+
+auto_map = {
+    "AutoConfig": "configuration_bailing_shared_moe_v2.BailingSharedMoeV2Config",
+    "AutoModel": "modeling_bailing_shared_moe_v2.BailingSharedMoeV2Model",
+    "AutoModelForCausalLM": "modeling_bailing_shared_moe_v2.BailingSharedMoeV2ForCausalLM"
+}
+
+cfg_standard_moe = BailingMoeV2Config.from_pretrained(input_model)
+cfg_shared_moe = BailingSharedMoeV2Config(
+    auto_map=auto_map,
+    model_type="bailing_shared_moe_v2",
+    vocab_size=cfg_standard_moe.vocab_size,
+    hidden_size=cfg_standard_moe.hidden_size,
+    intermediate_size=cfg_standard_moe.intermediate_size,
+    num_hidden_layers=cfg_standard_moe.num_hidden_layers,
+    num_attention_heads=cfg_standard_moe.num_attention_heads,
+    num_key_value_heads=cfg_standard_moe.num_key_value_heads,
+    hidden_act=cfg_standard_moe.hidden_act,
+    max_position_embeddings=cfg_standard_moe.max_position_embeddings,
+    initializer_range=cfg_standard_moe.initializer_range,
+    rms_norm_eps=cfg_standard_moe.rms_norm_eps,
+    use_cache=cfg_standard_moe.use_cache,
+    tie_word_embeddings=cfg_standard_moe.tie_word_embeddings,
+    rope_theta=cfg_standard_moe.rope_theta,
+    rope_scaling=cfg_standard_moe.rope_scaling,
+    max_window_layers=cfg_standard_moe.max_window_layers,
+    attention_dropout=cfg_standard_moe.attention_dropout,
+    moe_intermediate_size=cfg_standard_moe.moe_intermediate_size,
+    num_experts_per_tok=cfg_standard_moe.num_experts_per_tok,
+    num_experts=cfg_standard_moe.num_experts,
+    num_shared_experts=cfg_standard_moe.num_shared_experts,
+    norm_topk_prob=cfg_standard_moe.norm_topk_prob,
+    output_router_logits=cfg_standard_moe.output_router_logits,
+    shared_expert_intermediate_size=None,
+    head_dim=cfg_standard_moe.head_dim,
+    embedding_dropout=cfg_standard_moe.embedding_dropout,
+    eos_token_id=cfg_standard_moe.eos_token_id,
+    first_k_dense_replace=cfg_standard_moe.first_k_dense_replace,
+    output_dropout=cfg_standard_moe.output_dropout,
+    pad_token_id=cfg_standard_moe.pad_token_id,
+    torch_dtype=cfg_standard_moe.torch_dtype,
+    use_bias=cfg_standard_moe.use_bias,
+    use_qkv_bias=cfg_standard_moe.use_qkv_bias,
+    moe_router_enable_expert_bias=cfg_standard_moe.moe_router_enable_expert_bias,
+    routed_scaling_factor=cfg_standard_moe.routed_scaling_factor,
+    n_group=cfg_standard_moe.n_group,
+    topk_group=cfg_standard_moe.topk_group,
+    use_qk_norm=cfg_standard_moe.use_qk_norm,
+    moe_shared_expert_intermediate_size=cfg_standard_moe.moe_shared_expert_intermediate_size,
+    num_nextn_predict_layers=cfg_standard_moe.num_nextn_predict_layers,
+    score_function=cfg_standard_moe.score_function,
+    router_dtype=cfg_standard_moe.router_dtype,
+    use_rmsnorm=cfg_standard_moe.use_rmsnorm,
+    partial_rotary_factor=cfg_standard_moe.partial_rotary_factor,
+)
+
+num_experts = cfg_standard_moe.num_experts
+
+with accelerate.init_empty_weights():
+    model_shared_moe = BailingSharedMoeV2ForCausalLM(cfg_shared_moe)
+
+model_shared_moe = model_shared_moe.to(torch.bfloat16)
+new_state_dict = {}
+pattern = f"{input_model}/model-*-of-*.safetensors"
+files = sorted(glob.glob(pattern))
+
+if len(files) == 0:
+    raise FileNotFoundError
+tensors = {}
+
+for file_path in files:
+    print(f"processing {file_path}")
+    with safe_open(file_path, framework="pt", device="cpu") as f:
+        for key in f.keys():
+            tensor = f.get_tensor(key)
+            tensors[key] = tensor
+
+for key in tensors:
+    if "experts" not in key or "shared_experts" in key:
+        new_state_dict[key] = tensors[key]
+    elif "experts.0" in key:
+        layer_num = int(re.search(r"\d+", key).group())
+        new_state_dict[
+            f"model.layers.{layer_num}.mlp.moe_mlp.output_experts.weight"
+        ] = torch.stack(
+            [
+                tensors[f"model.layers.{layer_num}.mlp.experts.{i}.down_proj.weight"]
+                for i in range(num_experts)
+            ]
+        )
+        new_state_dict[f"model.layers.{layer_num}.mlp.moe_mlp.experts.weight"] = (
+            torch.stack(
+                [
+                    torch.cat(
+                        [
+                            tensors[
+                                f"model.layers.{layer_num}.mlp.experts.{i}.up_proj.weight"
+                            ],
+                            tensors[
+                                f"model.layers.{layer_num}.mlp.experts.{i}.gate_proj.weight"
+                            ],
+                        ],
+                        dim=0,
+                    )
+                    for i in range(num_experts)
+                ]
+            )
+        )
+model_shared_moe.load_state_dict(new_state_dict, strict=True, assign=True)
+model_shared_moe.save_pretrained(output_model_path)
+cfg_shared_moe.save_pretrained(output_model_path)
+
+
+shutil.copy(
+    "modeling_bailing_shared_moe_v2.py",
+    output_model_path + "/" + "modeling_bailing_shared_moe_v2.py",
+)
+shutil.copy(
+    "configuration_bailing_shared_moe_v2.py",
+    output_model_path + "/" + "configuration_bailing_shared_moe_v2.py",
+)
+for i in ["special_tokens_map.json", "tokenizer_config.json", "tokenizer.json"]:
+    shutil.copy(input_model + "/" + i, output_model_path + "/" + i)

convert_scm_to_hf.py

@@ -0,0 +1,124 @@
+import glob
+import re
+import shutil
+import sys
+
+import accelerate
+import torch
+from safetensors import safe_open
+from configuration_bailing_shared_moe_v2 import BailingSharedMoeV2Config
+from modeling_bailing_moe_v2 import BailingMoeV2ForCausalLM
+from configuration_bailing_moe_v2 import BailingMoeV2Config
+
+input_model = sys.argv[1]
+output_model_path = sys.argv[2]
+
+auto_map = {
+    "AutoConfig": "configuration_bailing_moe_v2.BailingMoeV2Config",
+    "AutoModel": "modeling_bailing_moe_v2.BailingMoeV2Model",
+    "AutoModelForCausalLM": "modeling_bailing_moe_v2.BailingMoeV2ForCausalLM"
+}
+cfg_shared_moe = BailingSharedMoeV2Config.from_pretrained(input_model)
+cfg_standard_moe = BailingMoeV2Config(
+    auto_map=auto_map,
+    vocab_size=cfg_shared_moe.vocab_size,
+    hidden_size=cfg_shared_moe.hidden_size,
+    intermediate_size=cfg_shared_moe.intermediate_size,
+    num_hidden_layers=cfg_shared_moe.num_hidden_layers,
+    num_attention_heads=cfg_shared_moe.num_attention_heads,
+    num_key_value_heads=cfg_shared_moe.num_key_value_heads,
+    hidden_act=cfg_shared_moe.hidden_act,
+    max_position_embeddings=cfg_shared_moe.max_position_embeddings,
+    initializer_range=cfg_shared_moe.initializer_range,
+    rms_norm_eps=cfg_shared_moe.rms_norm_eps,
+    use_cache=cfg_shared_moe.use_cache,
+    tie_word_embeddings=cfg_shared_moe.tie_word_embeddings,
+    rope_theta=cfg_shared_moe.rope_theta,
+    rope_scaling=cfg_shared_moe.rope_scaling,
+    max_window_layers=cfg_shared_moe.max_window_layers,
+    attention_dropout=cfg_shared_moe.attention_dropout,
+    moe_intermediate_size=cfg_shared_moe.moe_intermediate_size,
+    num_experts_per_tok=cfg_shared_moe.num_experts_per_tok,
+    num_experts=cfg_shared_moe.num_experts,
+    num_shared_experts=cfg_shared_moe.num_shared_experts,
+    norm_topk_prob=cfg_shared_moe.norm_topk_prob,
+    output_router_logits=cfg_shared_moe.output_router_logits,
+    shared_expert_intermediate_size=None,
+    head_dim=cfg_shared_moe.head_dim,
+    embedding_dropout=cfg_shared_moe.embedding_dropout,
+    eos_token_id=cfg_shared_moe.eos_token_id,
+    first_k_dense_replace=cfg_shared_moe.first_k_dense_replace,
+    output_dropout=cfg_shared_moe.output_dropout,
+    pad_token_id=cfg_shared_moe.pad_token_id,
+    torch_dtype=cfg_shared_moe.torch_dtype,
+    use_bias=cfg_shared_moe.use_bias,
+    use_qkv_bias=cfg_shared_moe.use_qkv_bias,
+    moe_router_enable_expert_bias=cfg_shared_moe.moe_router_enable_expert_bias,
+    routed_scaling_factor=cfg_shared_moe.routed_scaling_factor,
+    n_group=cfg_shared_moe.n_group,
+    topk_group=cfg_shared_moe.topk_group,
+    use_qk_norm=cfg_shared_moe.use_qk_norm,
+    moe_shared_expert_intermediate_size=cfg_shared_moe.moe_shared_expert_intermediate_size,
+    num_nextn_predict_layers=cfg_shared_moe.num_nextn_predict_layers,
+    score_function=cfg_shared_moe.score_function,
+    router_dtype=cfg_shared_moe.router_dtype,
+    use_rmsnorm=cfg_shared_moe.use_rmsnorm,
+    partial_rotary_factor=cfg_shared_moe.partial_rotary_factor
+)
+num_experts = cfg_standard_moe.num_experts
+
+with accelerate.init_empty_weights():
+    model_standard_moe = BailingMoeV2ForCausalLM(cfg_shared_moe)
+
+model_standard_moe = model_standard_moe.to(torch.bfloat16)
+new_state_dict = {}
+pattern = f"{input_model}/model-*-of-*.safetensors"
+files = sorted(glob.glob(pattern))
+
+if len(files) == 0:
+    raise FileNotFoundError
+tensors = {}
+
+for file_path in files:
+    print(f"processing {file_path}")
+    with safe_open(file_path, framework="pt", device="cpu") as f:
+        for key in f.keys():
+            tensor = f.get_tensor(key)
+            tensors[key] = tensor
+
+for key in tensors:
+    if "moe_mlp" not in key:
+        new_state_dict[key] = tensors[key]
+    elif "moe_mlp.output_experts" in key:
+        layer_num = int(re.search(r"\d+", key).group())
+        for i, tensor in enumerate(torch.unbind(tensors[key])):
+            new_state_dict[
+                f"model.layers.{layer_num}.mlp.experts.{i}.down_proj.weight"
+            ] = tensor.contiguous()
+    elif "moe_mlp.experts" in key:
+        layer_num = int(re.search(r"\d+", key).group())
+        for i, tensor in enumerate(torch.unbind(tensors[key])):
+            (
+                new_state_dict[
+                    f"model.layers.{layer_num}.mlp.experts.{i}.up_proj.weight"
+                ],
+                new_state_dict[
+                    f"model.layers.{layer_num}.mlp.experts.{i}.gate_proj.weight"
+                ],
+            ) = torch.chunk(tensor, 2, dim=0)
+
+model_standard_moe.load_state_dict(new_state_dict, strict=True, assign=True)
+model_standard_moe.save_pretrained(output_model_path)
+cfg_standard_moe.save_pretrained(output_model_path)
+
+shutil.copy(
+    "modeling_bailing_moe_v2.py",
+    output_model_path + "/" + "modeling_bailing_moe_v2.py",
+)
+shutil.copy(
+    "configuration_bailing_moe_v2.py",
+    output_model_path + "/" + "configuration_bailing_moe_v2.py",
+)
+
+for i in ["special_tokens_map.json", "tokenizer_config.json", "tokenizer.json"]:
+    shutil.copy(input_model + "/" + i, output_model_path + "/" + i)

modeling_bailing_shared_moe_v2.py

@@ -0,0 +1,1552 @@
+# coding=utf-8
+# Copyright 2025 Antgroup and The HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch  model."""
+
+import math
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.modeling_attn_mask_utils import (
+    AttentionMaskConverter,
+    _prepare_4d_attention_mask,
+    _prepare_4d_causal_attention_mask,
+    _prepare_4d_causal_attention_mask_for_sdpa,
+)
+from transformers.modeling_outputs import MoeModelOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS, is_torch_greater_or_equal_than_1_13
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.utils.import_utils import is_torch_fx_available
+from .configuration_bailing_shared_moe_v2 import BailingSharedMoeV2Config
+from transformers.generation.utils import GenerationMixin
+from dataclasses import dataclass
+from transformers.utils import ModelOutput
+
+#from cut_cross_entropy.transformers.utils import apply_lce
+from cut_cross_entropy import linear_cross_entropy
+from typing import TypedDict
+
+class CCEPreset(TypedDict):
+    filter_eps: float | str | None
+    accum_e_fp32: bool
+    accum_c_fp32: bool
+    filter_e_grad: bool
+    filter_c_grad: bool
+
+class CCEKwargs(CCEPreset):
+    impl: str
+    reduction: str
+
+@dataclass
+class PatchOptions:
+    impl: str
+    reduction: str
+    filter_eps: float | str | None
+    accum_e_fp32: bool
+    accum_c_fp32: bool
+    filter_e_grad: bool
+    filter_c_grad: bool
+    train_only: bool
+
+    def to_kwargs(self) -> CCEKwargs:
+        return CCEKwargs(
+            impl=self.impl,
+            reduction=self.reduction,
+            filter_eps=self.filter_eps,
+            accum_e_fp32=self.accum_e_fp32,
+            accum_c_fp32=self.accum_c_fp32,
+            filter_e_grad=self.filter_e_grad,
+            filter_c_grad=self.filter_c_grad,
+        )
+
+_PATCH_OPTS = PatchOptions(
+        impl="cce",
+        reduction="mean",
+        filter_eps="auto",
+        accum_e_fp32=False,
+        accum_c_fp32=False,
+        filter_e_grad=True,
+        filter_c_grad=True,
+        train_only=False,
+)
+
+def apply_lce(
+    e: torch.Tensor,
+    c: torch.Tensor,
+    labels: torch.Tensor,
+    opts,
+    bias: torch.Tensor | None = None,
+    **loss_kwargs,
+) -> torch.Tensor:
+    num_items_in_batch = loss_kwargs.get("num_items_in_batch", None)
+    cce_kwargs = opts.to_kwargs()
+    if num_items_in_batch is not None and cce_kwargs["reduction"] == "mean":
+        cce_kwargs["reduction"] = "sum"
+    else:
+        num_items_in_batch = None
+
+    loss = linear_cross_entropy(
+        e,
+        c,
+        labels.to(e.device),
+        bias=bias,
+        shift=True,
+        **cce_kwargs,
+    )
+
+    if num_items_in_batch is not None:
+        loss = loss / num_items_in_batch
+
+    return loss
+
+import scattermoe
+
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+
+# This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
+# It means that the function will not be traced through and simply appear as a node in the graph.
+if is_torch_fx_available():
+    if not is_torch_greater_or_equal_than_1_13:
+        import torch.fx
+
+    _prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask)
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "BailingSharedMoeV2Config"
+
+
+def roll_tensor(tensor, shifts=-1, dims=-1, fill_value=0):
+    """Roll the tensor input along the given dimension(s).
+    Inserted elements are set to be 0.0.
+    """
+    rolled_tensor = torch.roll(tensor, shifts=shifts, dims=dims)
+    rolled_tensor.select(dims, shifts).fill_(fill_value)
+    return rolled_tensor, rolled_tensor.sum()
+
+
+@dataclass
+class MoEV2CausalLMOutputWithPast(ModelOutput):
+    """
+    Base class for causal language model (or autoregressive) outputs as well as Mixture of Expert's router hidden
+    states terms, to train a MoE model.
+
+    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).
+        past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            It is a [`~cache_utils.Cache`] instance. For more details, see our [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache).
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+            `past_key_values` input) to speed up sequential decoding.
+        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.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        z_loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
+            z_loss for the sparse modules.
+        aux_loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
+            aux_loss for the sparse modules.
+        router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_logits=True` is passed or when `config.add_router_probs=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
+
+            Router logits of the encoder model, useful to compute the auxiliary loss and the z_loss for the sparse
+            modules.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    past_key_values: Optional[Cache] = None
+    hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[tuple[torch.FloatTensor, ...]] = None
+    z_loss: Optional[torch.FloatTensor] = None
+    aux_loss: Optional[torch.FloatTensor] = None
+    router_logits: Optional[tuple[torch.FloatTensor]] = None
+    mtp_loss: Optional[torch.FloatTensor] = None
+    mtp_logits: Optional[tuple[torch.FloatTensor, ...]] = None
+
+
+class MoeV2ModelOutputWithPast(MoeModelOutputWithPast):
+
+    def __init__(self, mtp_hidden_states=None, **kwargs):
+        super().__init__(**kwargs)
+        self.mtp_hidden_states = mtp_hidden_states
+
+
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    warnings.warn(
+        "Calling `transformers.models.BailingSharedMoeV2.modeling_BailingSharedMoeV2._prepare_4d_attention_mask` is deprecated and will be removed in v4.37. Use `transformers.modeling_attn_mask_utils._prepare_4d_attention_mask"
+    )
+    return _prepare_4d_attention_mask(mask=mask, dtype=dtype, tgt_len=tgt_len)
+
+
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    warnings.warn(
+        "Calling `transformers.models.BailingSharedMoeV2.modeling_BailingSharedMoeV2._make_causal_mask` is deprecated and will be removed in v4.37. Use `transformers.models.BailingSharedMoeV2.modeling_BailingSharedMoeV2.AttentionMaskConverter._make_causal_mask"
+    )
+    return AttentionMaskConverter._make_causal_mask(
+        input_ids_shape=input_ids_shape, dtype=dtype, device=device, past_key_values_length=past_key_values_length
+    )
+
+
+class BailingSharedMoeV2RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        BailingSharedMoeV2RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+ALL_LAYERNORM_LAYERS.append(BailingSharedMoeV2RMSNorm)
+
+
+class BailingSharedMoeV2RotaryEmbedding(nn.Module):
+    def __init__(self, config: BailingSharedMoeV2Config, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and 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.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    # Keep half or full tensor for later concatenation
+    rotary_dim = cos.shape[-1]
+    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
+    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]
+
+    # Apply rotary embeddings on the first half or full tensor
+    q_embed = (q_rot * cos) + (rotate_half(q_rot) * sin)
+    k_embed = (k_rot * cos) + (rotate_half(k_rot) * sin)
+
+    # Concatenate back to full shape
+    q_embed = torch.cat([q_embed, q_pass], dim=-1)
+    k_embed = torch.cat([k_embed, k_pass], dim=-1)
+    return q_embed, k_embed
+
+
+class BailingSharedMoeV2MLP(nn.Module):
+    def __init__(self, config: BailingSharedMoeV2Config, intermediate_size: int):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class BailingSharedMoeV2Gate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.num_experts = config.num_experts
+
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+
+        # topk selection algorithm
+        self.gating_dim = config.hidden_size
+        self.weight = nn.Parameter(torch.empty((self.num_experts, self.gating_dim)))
+        self.routed_scaling_factor = config.routed_scaling_factor
+
+        self.register_buffer("expert_bias", torch.zeros((self.num_experts)))
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        import torch.nn.init as init
+
+        init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+
+    def group_limited_topk(
+        self,
+        scores: torch.Tensor,
+    ):
+        num_tokens, _ = scores.size()
+        # Organize the experts into groups
+        group_scores = scores.view(num_tokens, self.n_group, -1).topk(2, dim=-1)[0].sum(dim=-1)
+        group_idx = torch.topk(group_scores, k=self.topk_group, dim=-1, sorted=False)[1]
+        group_mask = torch.zeros_like(group_scores)
+        group_mask.scatter_(1, group_idx, 1)
+
+        # Mask the experts based on selection groups
+        score_mask = (
+            group_mask.unsqueeze(-1)
+            .expand(num_tokens, self.n_group, self.num_experts // self.n_group)
+            .reshape(num_tokens, -1)
+        )
+
+        masked_scores = scores.masked_fill(~score_mask.bool(), float('-inf'))
+        probs, top_indices = torch.topk(masked_scores, k=self.top_k, dim=-1)
+
+        return probs, top_indices
+
+    def forward(self, hidden_states):
+        # compute gating score
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        logits = F.linear(hidden_states.type(torch.float32), self.weight.type(torch.float32))
+
+        scores = torch.sigmoid(logits.float()).type_as(logits)
+
+        scores_for_routing = scores + self.expert_bias.to(scores.device)
+        _, topk_idx = self.group_limited_topk(scores_for_routing)
+
+        scores = torch.gather(scores, dim=1, index=topk_idx).type_as(logits)
+
+        topk_weight = scores / (scores.sum(dim=-1, keepdim=True) + 1e-20) if self.top_k > 1 else scores
+        topk_weight = topk_weight * self.routed_scaling_factor
+
+        return topk_idx, topk_weight, logits
+
+
+class BailingSharedMoeV2SparseMoeBlock(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+
+    def __init__(self, config: BailingSharedMoeV2Config):
+        super().__init__()
+        self.config = config
+        self.num_experts_per_tok = config.num_experts_per_tok
+        self.moe_mlp = scattermoe.mlp.GLUMLP(
+            input_size=self.config.hidden_size,
+            hidden_size=self.config.moe_intermediate_size,
+            num_experts=self.config.num_experts,
+            top_k=self.config.num_experts_per_tok,
+            activation=ACT2FN[config.hidden_act],
+        )
+        self.gate = BailingSharedMoeV2Gate(config)
+        if config.num_shared_experts is not None:
+            self.shared_experts = BailingSharedMoeV2MLP(
+                config=config, intermediate_size=config.moe_intermediate_size * config.num_shared_experts
+            )
+
+
+    def forward(self, hidden_states):
+        identity = hidden_states
+        bsz, seq_len, h = hidden_states.shape
+        topk_idx, topk_weight, router_logits = self.gate(hidden_states)
+        hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+        flat_topk_idx = topk_idx.view(-1)
+        y = self.moe_mlp(hidden_states, topk_weight.to(torch.bfloat16), flat_topk_idx)
+        if self.config.num_shared_experts is not None:
+            y = y + self.shared_experts(identity)
+        return y, (router_logits.view(bsz, seq_len, -1), topk_idx.view(bsz, seq_len, -1))
+
+
+# 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 transformers.models.llama.modeling_llama.LlamaAttention with Llama->BailingSharedMoeV2
+class BailingSharedMoeV2Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: BailingSharedMoeV2Config, 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 `layer_idx` is not recommended and will "
+                "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 = config.head_dim or self.hidden_size // self.num_heads
+        partial_rotary_factor = config.partial_rotary_factor if hasattr(config, "partial_rotary_factor") else 1.0
+        self.rope_dim = int(self.head_dim * partial_rotary_factor)
+        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.query_key_value = nn.Linear(
+            self.hidden_size,
+            (self.num_heads + 2 * self.num_key_value_heads) * self.head_dim,
+            bias=config.use_qkv_bias,
+        )
+
+        if self.config.use_qk_norm:
+            self.query_layernorm = BailingSharedMoeV2RMSNorm(self.head_dim, eps=config.rms_norm_eps)
+            self.key_layernorm = BailingSharedMoeV2RMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        self.dense = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.use_bias)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    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,
+        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()
+
+        qkv = self.query_key_value(hidden_states)
+        qkv = qkv.view(bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim)
+
+        query_states, key_states, value_states = qkv.split(
+            [self.num_heads, self.num_key_value_heads, self.num_key_value_heads], dim=-2
+        )
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        if self.config.use_qk_norm:
+            query_states = self.query_layernorm(query_states)
+            key_states = self.key_layernorm(key_states)
+
+        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:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            cache_kwargs = {"sin": sin, "cos": cos}
+            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)
+
+        kv_seq_len = key_states.shape[-2]
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_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.dense(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2 with Llama->BailingSharedMoeV2
+class BailingSharedMoeV2FlashAttention2(BailingSharedMoeV2Attention):
+    """
+    BailingSharedMoeV2 flash attention module. This module inherits from `BailingSharedMoeV2Attention` 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.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        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]]]:
+        # BailingSharedMoeV2FlashAttention2 attention does not support output_attentions
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+
+        # 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
+
+        qkv = self.query_key_value(hidden_states)
+        qkv = qkv.view(bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim)
+
+        query_states, key_states, value_states = qkv.split(
+            [self.num_heads, self.num_key_value_heads, self.num_key_value_heads], dim=-2
+        )
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        if self.config.use_qk_norm:
+            query_states = self.query_layernorm(query_states)
+            key_states = self.key_layernorm(key_states)
+
+        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:
+            cache_kwargs = {"sin": sin, "cos": cos}
+            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
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently cast in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slow down training & inference so it is recommended to not cast the LayerNorms
+        # in fp32. (BailingSharedMoeV2RMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            # Handle the case where the model is quantized
+            if hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            elif torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            else:
+                target_dtype = self.query_key_value.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 = self._flash_attention_forward(
+            query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
+        attn_output = self.dense(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`int`, *optional*):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+            query_length (`int`):
+                The length of the query sequence in terms of tokens. This represents the number of tokens in the
+                `query_states` tensor along the sequence dimension. It is used to determine the effective sequence
+                length for attention computations.
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in BailingSharedMoeV2FlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->BailingSharedMoeV2
+class BailingSharedMoeV2SdpaAttention(BailingSharedMoeV2Attention):
+    """
+    BailingSharedMoeV2 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `BailingSharedMoeV2Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Adapted from BailingSharedMoeV2Attention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        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(
+                "BailingSharedMoeV2Model is using BailingSharedMoeV2SdpaAttention, 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()
+
+        qkv = self.query_key_value(hidden_states)
+        qkv = qkv.view(bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim)
+
+        query_states, key_states, value_states = qkv.split(
+            [self.num_heads, self.num_key_value_heads, self.num_key_value_heads], dim=-2
+        )
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        if self.config.use_qk_norm:
+            query_states = self.query_layernorm(query_states)
+            key_states = self.key_layernorm(key_states)
+
+        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:
+            cache_kwargs = {"sin": sin, "cos": cos}
+            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)
+
+        if attention_mask is not None:
+            kv_seq_len = key_states.shape[-2]
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+
+        # 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()
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            # 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=self.is_causal and attention_mask is None and q_len > 1,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+
+        attn_output = self.dense(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+ATTENTION_CLASSES = {
+    "eager": BailingSharedMoeV2Attention,
+    "flash_attention_2": BailingSharedMoeV2FlashAttention2,
+    "sdpa": BailingSharedMoeV2SdpaAttention,
+}
+
+
+class BailingSharedMoeV2MTPLayer(nn.Module):
+    def __init__(self, config: BailingSharedMoeV2Config, layer_idx: int):
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.input_layernorm = BailingSharedMoeV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.enorm = BailingSharedMoeV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.eh_proj = nn.Linear(config.hidden_size * 2, config.hidden_size, bias=False)
+        self.post_attention_layernorm = BailingSharedMoeV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.attention = ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
+        self.mlp = BailingSharedMoeV2SparseMoeBlock(config)
+
+        self.hnorm = BailingSharedMoeV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.final_layernorm = BailingSharedMoeV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_embeds,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        output_router_logits: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        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]]]:
+        input_embeds = self.enorm(input_embeds)
+        hidden_states = self.hnorm(hidden_states)
+        hidden_states = self.eh_proj(torch.cat([input_embeds, hidden_states], dim=-1))
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.attention(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            position_embeddings=position_embeddings,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        if isinstance(hidden_states, tuple):
+            hidden_states, router_logits = hidden_states
+        else:
+            router_logits = None
+        hidden_states = residual + hidden_states.to(residual.device)
+        hidden_states = self.final_layernorm(hidden_states)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        if output_router_logits:
+            outputs += (router_logits,)
+
+        return outputs
+
+
+class BailingSharedMoeV2DecoderLayer(nn.Module):
+    def __init__(self, config: BailingSharedMoeV2Config, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.attention = ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
+
+        self.mlp = (
+            BailingSharedMoeV2SparseMoeBlock(config)
+            if (config.num_experts is not None and layer_idx >= config.first_k_dense_replace)
+            else BailingSharedMoeV2MLP(config=config, intermediate_size=config.intermediate_size)
+        )
+        self.input_layernorm = BailingSharedMoeV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = BailingSharedMoeV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        output_router_logits: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        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.
+            position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+                config.n_positions - 1]`.
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*):
+                cached past key and value projection states
+            output_attentions (`bool`, *optional*):
+                Whether to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_router_logits (`bool`, *optional*):
+                Whether or not to return the logits of all the routers. They are useful for computing the router loss,
+                and should not be returned during inference.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+        """
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.attention(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            position_embeddings=position_embeddings,
+            use_cache=use_cache,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        if isinstance(hidden_states, tuple):
+            hidden_states, router_logits = hidden_states
+        else:
+            router_logits = None
+        hidden_states = residual + hidden_states.to(residual.device)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        if output_router_logits:
+            outputs += (router_logits,)
+
+        return outputs
+
+
+BAILINGMOEV2_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 ([`BailingSharedMoeV2Config`]):
+            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.
+"""
+
+
+@add_start_docstrings(
+    "The bare BailingSharedMoeV2 Model outputting raw hidden-states without any specific head on top.",
+    BAILINGMOEV2_START_DOCSTRING,
+)
+class BailingSharedMoeV2PreTrainedModel(PreTrainedModel):
+    config_class = BailingSharedMoeV2Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["BailingSharedMoeV2DecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+BAILINGMOEV2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance;
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare BailingSharedMoeV2 Model outputting raw hidden-states without any specific head on top.",
+    BAILINGMOEV2_START_DOCSTRING,
+)
+class BailingSharedMoeV2Model(BailingSharedMoeV2PreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`BailingSharedMoeV2DecoderLayer`]
+
+    Args:
+        config: BailingSharedMoeV2Config
+    """
+
+    def __init__(self, config: BailingSharedMoeV2Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.num_nextn_predict_layers = config.num_nextn_predict_layers
+
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = []
+        for layer_idx in range(config.num_hidden_layers):
+            layer_cls = BailingSharedMoeV2DecoderLayer if layer_idx < config.num_hidden_layers - config.num_nextn_predict_layers else BailingSharedMoeV2MTPLayer
+            self.layers.append(layer_cls(config, layer_idx))
+
+        self.layers = nn.ModuleList(self.layers)
+
+        self._use_sdpa = config._attn_implementation == "sdpa"
+        self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
+        self.norm = BailingSharedMoeV2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = BailingSharedMoeV2RotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.word_embeddings = value
+
+    @add_start_docstrings_to_model_forward(BAILINGMOEV2_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, MoeV2ModelOutputWithPast]:
+        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
+        )
+        output_router_logits = (
+            output_router_logits if output_router_logits is not None else self.config.output_router_logits
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape[:2]
+        elif inputs_embeds is not None:
+            batch_size, seq_length = inputs_embeds.shape[:2]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`transformers."
+                )
+                use_cache = False
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+
+        if position_ids is None:
+            position_ids = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+            position_ids = position_ids.unsqueeze(0)
+
+        if self._use_flash_attention_2:
+            # 2d mask is passed through the layers
+            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+        elif self._use_sdpa and not output_attentions:
+            # output_attentions=True can not be supported when using SDPA, and we fall back on
+            # the manual implementation that requires a 4D causal mask in all cases.
+            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
+                attention_mask,
+                (batch_size, seq_length),
+                inputs_embeds,
+                past_seen_tokens,
+            )
+        else:
+            # 4d mask is passed through the layers
+            attention_mask = _prepare_4d_causal_attention_mask(
+                attention_mask, (batch_size, seq_length), inputs_embeds, past_seen_tokens
+            )
+
+        # embed positions
+        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_logits = () if output_router_logits else None
+        next_decoder_cache = None
+        layers = self.layers[: -self.num_nextn_predict_layers] if self.num_nextn_predict_layers > 0 else self.layers
+        mtp_layers = self.layers[-self.num_nextn_predict_layers :] if self.num_nextn_predict_layers > 0 else None
+
+        for decoder_layer in 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,
+                    attention_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    output_router_logits,
+                    use_cache,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    output_router_logits=output_router_logits,
+                    use_cache=use_cache,
+                    position_embeddings=position_embeddings,
+                )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+            if output_router_logits and layer_outputs[-1] is not None:
+                all_router_logits += (layer_outputs[-1],)
+
+        hidden_states = self.norm(hidden_states)
+        main_hidden_states = hidden_states
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (main_hidden_states,)
+
+        mtp_hidden_states = None
+
+        if mtp_layers:
+            for decoder_layer in mtp_layers:
+                input_ids, _ = roll_tensor(input_ids, shifts=-1, dims=-1)
+                inputs_embeds = self.word_embeddings(input_ids)
+
+                if self.gradient_checkpointing and self.training:
+                    layer_outputs = self._gradient_checkpointing_func(
+                        decoder_layer.__call__,
+                        inputs_embeds,
+                        hidden_states,
+                        attention_mask,
+                        position_ids,
+                        past_key_values,
+                        output_attentions,
+                        output_router_logits,
+                        use_cache,
+                        position_embeddings,
+                    )
+                else:
+                    layer_outputs = decoder_layer(
+                        inputs_embeds,
+                        hidden_states,
+                        attention_mask=attention_mask,
+                        position_ids=position_ids,
+                        past_key_value=past_key_values,
+                        output_attentions=output_attentions,
+                        output_router_logits=output_router_logits,
+                        use_cache=use_cache,
+                        position_embeddings=position_embeddings,
+                    )
+                if mtp_hidden_states is None:
+                    mtp_hidden_states = []
+                hidden_states = layer_outputs[0]
+                mtp_hidden_states.append(hidden_states)
+
+                if output_hidden_states:
+                    all_hidden_states += (hidden_states,)
+
+                if use_cache:
+                    next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+                if output_attentions:
+                    all_self_attns += (layer_outputs[1],)
+
+                if output_router_logits and layer_outputs[-1] is not None:
+                    all_router_logits += (layer_outputs[-1],)
+
+        next_cache = None
+        if use_cache:
+            next_cache = next_decoder_cache
+        if not return_dict:
+            return tuple(
+                v
+                for v in [main_hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits]
+                if v is not None
+            )
+        return MoeV2ModelOutputWithPast(
+            last_hidden_state=main_hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            mtp_hidden_states=mtp_hidden_states,
+            attentions=all_self_attns,
+            router_logits=all_router_logits,
+        )
+
+
+class BailingSharedMoeV2ForCausalLM(BailingSharedMoeV2PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config: BailingSharedMoeV2Config):
+        super().__init__(config)
+        self.model = BailingSharedMoeV2Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.num_nextn_predict_layers = config.num_nextn_predict_layers
+        self.mtp_loss_scaling_factor = config.mtp_loss_scaling_factor
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.model.word_embeddings = 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
+
+    @add_start_docstrings_to_model_forward(BAILINGMOEV2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=MoEV2CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_router_logits: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> Union[Tuple, MoEV2CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer
+
+        >>> model = BailingSharedMoeV2ForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        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
+        # 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,
+            output_router_logits=output_router_logits,
+            return_dict=return_dict,
+            **kwargs,
+        )
+
+        loss = None
+        all_mtp_loss = []
+        aux_loss = None
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        if labels is not None:
+            #loss = self.loss_function(logits, labels, self.config.vocab_size, **kwargs)
+            loss = apply_lce(
+                hidden_states,
+                self.lm_head.weight,
+                labels,
+                _PATCH_OPTS,
+                **kwargs,
+            )
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            if output_router_logits:
+                output = (aux_loss,) + output
+            return (loss,) + output if loss is not None else output
+
+        return MoEV2CausalLMOutputWithPast(
+            loss=loss,
+            mtp_loss=all_mtp_loss,
+            aux_loss=aux_loss,
+            logits=logits,
+            mtp_logits=[],
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_logits=outputs.router_logits,
+        )
+