diff --git "a/modeling_motif.py" "b/modeling_motif.py"
new file mode 100644--- /dev/null
+++ "b/modeling_motif.py"
@@ -0,0 +1,2271 @@
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_flash_attention_utils import _flash_attention_forward
+from transformers.modeling_outputs import (
+    CausalLMOutputWithPast,
+    ModelOutput,
+)
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
+from transformers.modeling_utils import PreTrainedModel
+from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_greater_or_equal_2_10,
+    is_flash_attn_2_available,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_motif import MotifConfig
+from dataclasses import dataclass
+
+import torch.nn.functional as F
+import time
+
+logger = logging.get_logger(__name__)
+
+if is_flash_attn_2_available():
+    from transformers.modeling_flash_attention_utils import _flash_attention_forward
+
+try:
+    moreh_ops = torch.ops.moreh
+    MorehRMSNorm = moreh_ops.T5LayerNorm
+    ScaledDotProductAttention = moreh_ops.scaled_dot_product_attention
+    MorehFlashAttention = moreh_ops.flash_attention
+    logger.warning_once("Using moreh ops")
+except AttributeError:
+    MorehRMSNorm = None
+    ScaledDotProductAttention = None
+    MorehFlashAttention = None
+    logger.warning_once("Failed to import moreh ops")
+
+
+
+# DEBUG = False
+# logger.info(f"DEBUG: {DEBUG} : will log timing")
+# def log_timing(obj):
+#     """Decorator to log timing of function or class execution"""
+#     if isinstance(obj, type):
+#         # If decorating a class
+#         class TimedClass(obj):
+#             def __getattribute__(self, name):
+#                 attr = super().__getattribute__(name)
+#                 if callable(attr) and not name.startswith('__'):
+#                     def timed_method(*args, **kwargs):
+#                         if not DEBUG:
+#                             return attr(*args, **kwargs)
+#                         if name != "forward":
+#                             return attr(*args, **kwargs)
+
+#                         start_time = time.time()
+#                         logger.info(f"Entering {obj.__name__}.{name}")
+#                         result = attr(*args, **kwargs)
+#                         end_time = time.time()
+#                         logger.info(f"Exiting {obj.__name__}.{name}, took {end_time - start_time:.4f} seconds")
+#                         return result
+#                     return timed_method
+#                 return attr
+#         return TimedClass
+#     else:
+#         # If decorating a function
+#         def wrapper(*args, **kwargs):
+#             if not DEBUG:
+#                 return obj(*args, **kwargs)
+
+#             start_time = time.time()
+#             logger.info(f"Entering {obj.__name__}")
+#             result = obj(*args, **kwargs)
+#             end_time = time.time()
+#             logger.info(f"Exiting {obj.__name__}, took {end_time - start_time:.4f} seconds")
+#             return result
+#         return wrapper
+
+
+
+#_CHECKPOINT_FOR_DOC = "moreh/Motif-102B"
+_CONFIG_FOR_DOC = "MotifConfig"
+
+#from .moreh_moe import MorehMoeMLP, MorehMoeFusedMLP
+
+import torch
+from transformers.activations import ACT2CLS as _ACT2CLS
+from transformers.activations import ClassInstantier
+moreh_ops = torch.ops.moreh
+
+from typing import Callable, Dict, List, Tuple
+
+import torch
+
+
+# @log_timing
+def multi_head_forward_backward(shared_activation: torch.Tensor,
+                                head_fns: List[Callable[[torch.Tensor], Dict[str, torch.Tensor]]],
+                                return_keys=("loss", ),
+                                return_only_first_head=True) -> Tuple[torch.Tensor, ...]:
+    """
+    The forward-backward pattern introduced in the paper https://arxiv.org/abs/2404.19737
+    to reduce memory overhead due to activations from multiple heads.
+
+    Args:
+    - shared_activation: the shared activation across all heads
+    - head_fns: the head-wise forward computations that start from `shared_activation`.
+        it should output a dictionary of tensors with keys matching `return_keys`
+    - return_keys: the keys to return in order
+    - return_only_first_head: whether to return only the values from the first head
+
+    Returns:
+    - a tuple of return tensors
+
+    Side effect:
+    - (only when `torch.is_grad_enabled()`)
+        the gradients accumulated as if `sum(head_fn(shared_activation)["loss"] for head_fn in head_fns).backward()` had been called
+    """
+    if not return_only_first_head:
+        raise NotImplementedError
+
+    return_key_set = set(return_keys)
+    if "loss" not in return_key_set:
+        raise Exception("'loss' is a required return key.")
+
+    detached_shared_activation = shared_activation.detach()
+    detached_shared_activation.requires_grad = True
+    return_values = {key: None for key in return_keys}
+    for head_idx, head_fn in enumerate(head_fns):
+        if head_idx > 0 and not torch.is_grad_enabled():
+            continue
+
+        # forward pass for the head
+        headwise_outputs = head_fn(detached_shared_activation)
+        if set(headwise_outputs.keys()) != return_key_set:
+            raise Exception(f"Headwise output keys {headwise_outputs.keys()} do not match return keys {return_keys}.")
+
+        # backward pass for the head
+        # effect 1: the parameters of the head
+        # effect 2: gradient accumulated in `detached_shared_activation.grad`
+        if torch.is_grad_enabled():
+            headwise_loss = headwise_outputs["loss"]
+            headwise_loss.backward(
+            )  # NOTE: You do not need to retain graph since no graph is shared across backward passes
+
+        if head_idx == 0:
+            for key in return_keys:
+                return_values[key] = headwise_outputs[key]
+
+    assert all(value is not None for value in return_values.values())
+
+    # backward pass for the shared part
+    if torch.is_grad_enabled():
+        shared_activation.backward(detached_shared_activation.grad)
+
+    return tuple(return_values[key] for key in return_keys)
+
+
+class PolyNorm(torch.nn.Module):
+    """
+    A trainable activation function introduced in https://arxiv.org/html/2411.03884v1.
+    The code is copied from https://github.com/BryceZhuo/PolyCom?tab=readme-ov-file/README.md,
+    with the change `* torch.rsqrt` => `/ torch.sqrt` for potential MAF incompatibility.
+    """
+
+    def __init__(self, eps=1e-6):
+        super(PolyNorm, self).__init__()
+        self.weight = torch.nn.Parameter(torch.ones(3) / 3)
+        self.bias = torch.nn.Parameter(torch.zeros(1))
+        self.eps = eps
+
+    def _norm(self, x):
+        return x / torch.sqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x):
+        return self.weight[0] * self._norm(x ** 3) + self.weight[1] * self._norm(
+            x ** 2) + self.weight[2] * self._norm(x) + self.bias
+
+
+class PolyNorm_Test(torch.nn.Module):
+    """
+    A trainable activation function introduced in https://arxiv.org/html/2411.03884v1.
+    The code is copied from https://github.com/BryceZhuo/PolyCom?tab=readme-ov-file/README.md,
+    with the change `* torch.rsqrt` => `/ torch.sqrt` for potential MAF incompatibility.
+    """
+
+    def __init__(self, eps=1e-6):
+        super(PolyNorm_Test, self).__init__()
+        self.weight = torch.nn.Parameter(torch.ones(3) / 3)
+        self.bias = torch.nn.Parameter(torch.zeros(1))
+        self.eps = eps
+
+    def forward(self, x):
+
+        #return torch.nn.SiLU(x)
+        return moreh_ops.poly_norm(x, self.weight, self.bias)
+
+
+CUSTOM_ACT2CLS = {"poly_norm": PolyNorm_Test, "poly_norm_test": PolyNorm_Test}
+ACT2CLS = {**_ACT2CLS, **CUSTOM_ACT2CLS}
+ACT2FN = ClassInstantier(ACT2CLS)
+
+
+
+class MotifRMSNorm(nn.Module):
+
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        MotifRMSNorm 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)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+ALL_LAYERNORM_LAYERS.append(MotifRMSNorm if MorehRMSNorm is None else MorehRMSNorm)
+
+
+class MotifRotaryEmbeddingWithCache(nn.Module):
+    """
+    Rotary positional embedding module with caching for efficiency.
+
+    Args:
+        dim (int): Dimensionality of the embedding.
+        max_position_embeddings (int): Maximum sequence length for caching. Default is 2048.
+        base (int): Base for computing inverse frequency. Default is 10000.
+        device (torch.device, optional): Device for tensor storage.
+
+    Methods:
+        forward(x, seq_len=None):
+            Computes cosine and sine embeddings for input sequence length.
+            Automatically updates cache if `seq_len` exceeds cached length.
+
+    Attributes:
+        inv_freq (torch.Tensor): Inverse frequency tensor for position encoding.
+        cos_cached (torch.Tensor): Cached cosine embeddings.
+        sin_cached (torch.Tensor): Cached sine embeddings.
+    """
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+        # Build here to make `torch.jit.trace` work.
+        self._set_cos_sin_cache(seq_len=max_position_embeddings,
+                                device=self.inv_freq.device,
+                                dtype=torch.get_default_dtype())
+
+    def _set_cos_sin_cache(self, seq_len, device, dtype):
+        self.max_seq_len_cached = seq_len
+        t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype)
+
+        freqs = torch.outer(t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
+        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if seq_len > self.max_seq_len_cached:
+            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
+
+        return (
+            self.cos_cached[None, :seq_len].to(dtype=x.dtype),
+            self.sin_cached[None, :seq_len].to(dtype=x.dtype),
+        )
+
+
+# @log_timing
+class MotifRotaryEmbedding(nn.Module):
+
+    def __init__(
+        self,
+        dim=None,
+        max_position_embeddings=2048,
+        base=10000,
+        device=None,
+        scaling_factor=1.0,
+        rope_type="default",
+        config: Optional[MotifConfig] = None,
+    ):
+        super().__init__()
+        # TODO (joao): remove the `if` below, only used for BC
+        self.rope_kwargs = {}
+        if config is None:
+            logger.warning_once(
+                "`MotifRotaryEmbedding` can now be fully parameterized by passing the model config through the "
+                "`config` argument. All other arguments will be removed in v4.46")
+            self.rope_kwargs = {
+                "rope_type": rope_type,
+                "factor": scaling_factor,
+                "dim": dim,
+                "base": base,
+                "max_position_embeddings": max_position_embeddings,
+            }
+            self.rope_type = rope_type
+            self.max_seq_len_cached = max_position_embeddings
+            self.original_max_seq_len = max_position_embeddings
+        else:
+            # BC: "rope_type" was originally "type"
+            if config.rope_scaling is not None:
+                self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+            else:
+                self.rope_type = "default"
+            self.max_seq_len_cached = config.max_position_embeddings
+            self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, **self.rope_kwargs)
+
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(self.config,
+                                                                 device,
+                                                                 seq_len=seq_len,
+                                                                 **self.rope_kwargs)
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+def rotate_half(x):
+    """
+    Rotates half of the dimensions of the input tensor using torch.roll and in-place negation.
+
+    Args:
+    x (torch.Tensor): The input tensor.
+
+    Returns:
+    torch.Tensor: A tensor where the latter half of the dimensions are negated
+                  and moved before the first half.
+    """
+    half_size = x.shape[-1] // 2
+    rotated_tensor = torch.roll(x, shifts=-half_size, dims=-1)
+    rotated_tensor[..., :half_size] *= -1
+
+    return rotated_tensor
+
+
+# @log_timing
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1, fused_rope=True):
+    """
+    Applies rotary position embeddings to the input tensors.
+
+    Args:
+        q (torch.Tensor): Query tensor of shape (B, NH, S, D_KV).
+        k (torch.Tensor): Key tensor of shape (B, NH, S, D_KV).
+        cos (torch.Tensor): Cosine values for rotary embedding.
+        sin (torch.Tensor): Sine values for rotary embedding.
+        unsqueeze_dim (int, optional): Dimension along which `cos` and `sin` are unsqueezed.
+            Defaults to 1.
+        fused_rope (bool, optional): If True, applies fused rotary embeddings using
+            `moreh_ops.apply_rotary_emb`. If False, computes rotary embeddings manually.
+            Defaults to False.
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: Returns transformed query and key tensors after applying rotary embeddings.
+    """
+    '''
+    # (B, NH, S, D_KV) -> (B, S, NH, D_KV)
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    '''
+    #cos = cos[position_ids]
+    #sin = sin[position_ids]
+
+    #cos = cos[position_ids].unsqueeze(unsqueeze_dim)  # [bs, 1, seq_len, dim]
+    #sin = sin[position_ids].unsqueeze(unsqueeze_dim)  # [bs, 1, seq_len, dim]
+
+    q = q.transpose(1, 2)
+    k = k.transpose(1, 2)
+
+    # Expand 'batch' dim
+    cos = cos.expand(q.shape[0], *cos.shape[1:])
+    sin = sin.expand(q.shape[0], *sin.shape[1:])
+
+    q_embed = moreh_ops.apply_rotary_emb(q, cos, sin, opcode=1)
+    k_embed = moreh_ops.apply_rotary_emb(k, cos, sin, opcode=1)
+
+    # (B, S, NH, D_KV) -> (B, NH, S, D_KV)
+    q_embed = q_embed.transpose(1, 2)
+    k_embed = k_embed.transpose(1, 2)
+
+    return q_embed, k_embed
+
+
+# @log_timing
+class MotifMLP(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.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]
+
+        if config.wesar_weights:
+            self.gate_up_proj_alpha = nn.Parameter(torch.tensor(1) *config.gate_up_proj_alpha)
+            self.down_proj_alpha = nn.Parameter(torch.tensor(1) * config.down_proj_alpha)
+        else:
+            self.gate_up_proj_alpha=1
+            self.down_proj_alpha=1
+        if config.muP:
+            self.down_proj.__do_scale_tager__ = True
+            self.gate_proj.__do_scale_tager_mu_dim_model__  = True
+            self.up_proj.__do_scale_tager_mu_dim_model__ = True
+            self.down_proj.__do_scale_tager_mu_ffn__ = True
+
+
+    def forward(self, hidden_state):
+        hidden_state = hidden_state*self.gate_up_proj_alpha
+        #hidden_state = self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))*
+        return self.down_proj_alpha*self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
+
+
+class MorehMoeFusedMLP(nn.Module):
+    def __init__(self,
+                 ffn_dim,
+                 hidden_dim,
+                 hidden_act_moe,
+                 num_experts,
+                 num_groups=1,
+                 device=None,
+                 continual_training=False):
+        super().__init__()
+        self.ffn_dim = ffn_dim
+        self.hidden_dim = hidden_dim
+        self.hidden_act_moe = hidden_act_moe
+
+        self.num_experts = num_experts
+        self.num_groups = num_groups
+
+        assert self.num_experts % self.num_groups == 0
+        self.num_experts_per_group = self.num_experts // self.num_groups
+
+        ## bsz, seq, group size, 2*ffn_size
+
+        moreh_ops = torch.ops.moreh
+        self.w13 = nn.ModuleList([
+            moreh_ops.MoeFanInLinear(self.hidden_dim,
+                                     self.ffn_dim * 2,
+                                     bias=False,
+                                     num_experts=self.num_experts_per_group,
+                                     device=device)
+            for _ in range(self.num_groups)
+        ])
+
+        self.w2 = nn.ModuleList([
+            moreh_ops.MoeFanOutLinear(self.ffn_dim,
+                                      self.hidden_dim,
+                                      bias=False,
+                                      num_experts=self.num_experts_per_group,
+                                      device=device)
+            for _ in range(self.num_groups)
+        ])
+
+        ## use silu?
+        self.act_fn = ACT2FN[self.hidden_act_moe]
+
+        if continual_training:
+            logger.info('two optipons 1. zero init all weights, 2. add scaling param to moe output.')
+            self._zero_init()
+
+    def _zero_init(self):
+        for module in self.w2:
+            for n,param in module.named_parameters():
+                logger.info(f'{n} {param.shape}')
+                param.data.zero_()
+
+
+    def forward(self, hidden_states, selected_experts, routing_weights):
+        w13_final_output = None
+        for group_idx in range(self.num_groups):
+            w13_output_in_group = self._get_w13_output(hidden_states,
+                                                       selected_experts,
+                                                       group_idx)
+            if w13_final_output is None:
+                w13_final_output = w13_output_in_group
+            else:
+                w13_final_output += w13_output_in_group
+
+        current_hidden_states = self.act_fn(
+            w13_final_output[:, :, :, :self.ffn_dim]
+        ) * w13_final_output[:, :, :, self.ffn_dim:]
+
+        final_hidden_states = None
+        for group_idx in range(self.num_groups):
+            w2_output_in_group = self._get_w2_output(current_hidden_states,
+                                                     selected_experts,
+                                                     routing_weights, group_idx)
+            if final_hidden_states is None:
+                final_hidden_states = w2_output_in_group
+            else:
+                final_hidden_states += w2_output_in_group
+        return final_hidden_states
+
+    def _get_w13_output(self, hidden_states, selected_experts, group_idx):
+        selected_experts_in_group = selected_experts - (
+            group_idx * self.num_experts_per_group)
+
+        w13_output = self.w13[group_idx](hidden_states,
+                                         selected_experts_in_group)
+        return w13_output
+
+    def _get_w2_output(self, hidden_states, selected_experts, routing_weights,
+                       group_idx):
+        selected_experts_in_group = selected_experts - (
+            group_idx * self.num_experts_per_group)
+        output = self.w2[group_idx](hidden_states, selected_experts_in_group,
+                                    routing_weights)
+        return output
+
+
+class MoEGate(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.n_routed_experts = config.n_routed_experts
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.scoring_func = config.scoring_func
+        self.seq_aux = config.seq_aux
+        self.topk_method = config.topk_method
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+
+        # topk selection algorithm
+        self.norm_topk_prob = config.norm_topk_prob
+        self.gating_dim = config.hidden_size
+        self.weight = nn.Parameter(
+            torch.empty((self.n_routed_experts, self.gating_dim)))
+        if self.topk_method == "noaux_tc":
+            self.e_score_correction_bias = nn.Parameter(
+                torch.empty((self.n_routed_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 forward(self, hidden_states):
+        bsz, seq_len, h = hidden_states.shape
+        ### compute gating score
+        hidden_states = hidden_states.view(-1, h)
+        logits = F.linear(hidden_states.type(torch.float32),
+                          self.weight.type(torch.float32), None)
+        if self.scoring_func == "sigmoid":
+            scores = logits.sigmoid()
+        else:
+            raise NotImplementedError(
+                f"insupportable scoring function for MoE gating: {self.scoring_func}"
+            )
+
+        ### select top-k experts
+        if self.topk_method == "greedy":
+            topk_weight, topk_idx = torch.topk(scores,
+                                               k=self.top_k,
+                                               dim=-1,
+                                               sorted=False)
+        elif self.topk_method == "group_limited_greedy":
+            group_scores = (scores.view(bsz * seq_len, self.n_group,
+                                        -1).max(dim=-1).values)  # [n, n_group]
+            group_idx = torch.topk(group_scores,
+                                   k=self.topk_group,
+                                   dim=-1,
+                                   sorted=False)[1]  # [n, top_k_group]
+            group_mask = torch.zeros_like(group_scores)  # [n, n_group]
+            group_mask.scatter_(1, group_idx, 1)  # [n, n_group]
+            score_mask = (group_mask.unsqueeze(-1).expand(
+                bsz * seq_len, self.n_group,
+                self.n_routed_experts // self.n_group).reshape(
+                    bsz * seq_len, -1))  # [n, e]
+            tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0)  # [n, e]
+            topk_weight, topk_idx = torch.topk(tmp_scores,
+                                               k=self.top_k,
+                                               dim=-1,
+                                               sorted=False)
+        elif self.topk_method == "noaux_tc":
+            ###  will be used. ###
+            scores_for_choice = scores.view(
+                bsz * seq_len, -1) + self.e_score_correction_bias.unsqueeze(0)
+            group_scores = (scores_for_choice.view(
+                bsz * seq_len, self.n_group,
+                -1).topk(2, dim=-1)[0].sum(dim=-1))  # [n, n_group]
+            group_idx = torch.topk(group_scores,
+                                   k=self.topk_group,
+                                   dim=-1,
+                                   sorted=False)[1]  # [n, top_k_group]
+            group_mask = torch.zeros_like(group_scores)  # [n, n_group]
+            group_mask.scatter_(1, group_idx, 1)  # [n, n_group]
+            score_mask = (group_mask.unsqueeze(-1).expand(
+                bsz * seq_len, self.n_group,
+                self.n_routed_experts // self.n_group).reshape(
+                    bsz * seq_len, -1))  # [n, e]
+            tmp_scores = scores_for_choice.masked_fill(~score_mask.bool(),
+                                                       0.0)  # [n, e]
+            _, topk_idx = torch.topk(tmp_scores,
+                                     k=self.top_k,
+                                     dim=-1,
+                                     sorted=False)
+            topk_weight = scores.gather(1, topk_idx)
+        else:
+            raise NotImplementedError(
+                f"insupportable TopK function for MoE gating: {self.topk_method}"
+            )
+
+        ### norm gate to sum 1
+        if self.top_k > 1 and self.norm_topk_prob:
+            denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weight = topk_weight / denominator
+        topk_weight = topk_weight * self.routed_scaling_factor  # must multiply the scaling factor
+
+        return topk_idx, topk_weight
+
+
+class MotifMoE(nn.Module):
+    """
+    A mixed expert module containing shared experts.
+    """
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.num_experts_per_tok = config.num_experts_per_tok
+        self.use_moreh_moe = config.use_moreh_moe
+        self.use_fused_mlp = config.use_fused_mlp
+
+        if hasattr(config, "ep_size") and config.ep_size > 1:
+            assert config.ep_size == dist.get_world_size()
+            assert not config.use_moreh_moe
+            self.ep_size = config.ep_size
+            self.experts_per_rank = config.n_routed_experts // config.ep_size
+            self.ep_rank = dist.get_rank()
+            self.experts = nn.ModuleList([
+                (DeepseekV3MLP(config,
+                               intermediate_size=config.moe_intermediate_size)
+                 if i >= self.ep_rank * self.experts_per_rank and i <
+                 (self.ep_rank + 1) * self.experts_per_rank else None)
+                for i in range(config.n_routed_experts)
+            ])
+        else:
+            self.ep_size = 1
+            self.experts_per_rank = config.n_routed_experts
+            self.ep_rank = 0
+            if self.use_moreh_moe:
+                if not self.use_fused_mlp:
+                    self.experts = MorehMoeMLP(
+                        ffn_dim=config.moe_intermediate_size,
+                        hidden_dim=config.hidden_size,
+                        hidden_act_moe=config.hidden_act_moe,
+                        num_experts=config.n_routed_experts,
+                        device=None)
+                else:
+                    ## group expert.
+                    self.experts = MorehMoeFusedMLP(
+                        ffn_dim=config.moe_intermediate_size,
+                        hidden_dim=config.hidden_size,
+                        hidden_act_moe=config.hidden_act_moe,
+                        num_experts=config.n_routed_experts,
+                        num_groups=config.n_group,
+                        device=None,
+                        continual_training=config.continual_training,
+                        )
+            else:
+                self.experts = nn.ModuleList([
+                    DeepseekV3MLP(
+                        config, intermediate_size=config.moe_intermediate_size)
+                    for i in range(config.n_routed_experts)
+                ])
+
+        self.gate = MoEGate(config)
+
+    def forward(self, hidden_states):
+        identity = hidden_states
+        orig_shape = hidden_states.shape
+        topk_idx, topk_weight = self.gate(hidden_states)
+        if self.use_moreh_moe:
+            y = self.experts(hidden_states, topk_idx.view(*orig_shape[:-1], -1),
+                             topk_weight.view(*orig_shape[:-1], -1))
+            y = y.type(hidden_states.dtype)
+        else:
+            hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+            flat_topk_idx = topk_idx.view(-1)
+            if self.training:
+                hidden_states = hidden_states.repeat_interleave(
+                    self.num_experts_per_tok, dim=0)
+                y = torch.empty_like(hidden_states)
+                for i, expert in enumerate(self.experts):
+                    y[flat_topk_idx == i] = expert(
+                        hidden_states[flat_topk_idx == i])
+                y = (y.view(*topk_weight.shape, -1) *
+                     topk_weight.unsqueeze(-1)).sum(dim=1)
+                y = y.type(hidden_states.dtype)
+                y = y.view(*orig_shape)
+                # y = AddAuxiliaryLoss.apply(y, aux_loss)
+            else:
+                y = self.moe_infer(hidden_states, topk_idx,
+                                   topk_weight).view(*orig_shape)
+        return y, identity
+
+    @torch.no_grad()
+    def moe_infer(self, x, topk_ids, topk_weight):
+        cnts = topk_ids.new_zeros((topk_ids.shape[0], len(self.experts)))
+        cnts.scatter_(1, topk_ids, 1)
+        tokens_per_expert = cnts.sum(dim=0)
+        idxs = topk_ids.view(-1).argsort()
+        sorted_tokens = x[idxs // topk_ids.shape[1]]
+        sorted_tokens_shape = sorted_tokens.shape
+        if self.ep_size > 1:
+            tokens_per_ep_rank = tokens_per_expert.view(self.ep_size,
+                                                        -1).sum(dim=1)
+            tokens_per_expert_group = tokens_per_expert.new_empty(
+                tokens_per_expert.shape[0])
+            dist.all_to_all_single(tokens_per_expert_group, tokens_per_expert)
+            output_splits = (tokens_per_expert_group.view(
+                self.ep_size, -1).sum(1).cpu().numpy().tolist())
+            gathered_tokens = sorted_tokens.new_empty(
+                tokens_per_expert_group.sum(dim=0).cpu().item(),
+                sorted_tokens.shape[1])
+            input_split_sizes = tokens_per_ep_rank.cpu().numpy().tolist()
+            dist.all_to_all(
+                list(gathered_tokens.split(output_splits)),
+                list(sorted_tokens.split(input_split_sizes)),
+            )
+            tokens_per_expert_post_gather = tokens_per_expert_group.view(
+                self.ep_size, self.experts_per_rank).sum(dim=0)
+            gatherd_idxs = np.zeros(shape=(gathered_tokens.shape[0],),
+                                    dtype=np.int32)
+            s = 0
+            for i, k in enumerate(tokens_per_expert_group.cpu().numpy()):
+                gatherd_idxs[s:s + k] = i % self.experts_per_rank
+                s += k
+            gatherd_idxs = gatherd_idxs.argsort()
+            sorted_tokens = gathered_tokens[gatherd_idxs]
+            tokens_per_expert = tokens_per_expert_post_gather
+        tokens_per_expert = tokens_per_expert.cpu().numpy()
+
+        outputs = []
+        start_idx = 0
+        for i, num_tokens in enumerate(tokens_per_expert):
+            end_idx = start_idx + num_tokens
+            if num_tokens == 0:
+                continue
+            expert = self.experts[i + self.ep_rank * self.experts_per_rank]
+            tokens_for_this_expert = sorted_tokens[start_idx:end_idx]
+            expert_out = expert(tokens_for_this_expert)
+            outputs.append(expert_out)
+            start_idx = end_idx
+
+        outs = torch.cat(outputs,
+                         dim=0) if len(outputs) else sorted_tokens.new_empty(0)
+        if self.ep_size > 1:
+            new_x = torch.empty_like(outs)
+            new_x[gatherd_idxs] = outs
+            gathered_tokens = new_x.new_empty(*sorted_tokens_shape)
+            dist.all_to_all(
+                list(gathered_tokens.split(input_split_sizes)),
+                list(new_x.split(output_splits)),
+            )
+            outs = gathered_tokens
+
+        new_x = torch.empty_like(outs)
+        new_x[idxs] = outs
+        final_out = (new_x.view(
+            *topk_ids.shape, -1).type(topk_weight.dtype).mul_(
+                topk_weight.unsqueeze(dim=-1)).sum(dim=1).type(new_x.dtype))
+        return final_out
+
+
+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)
+    """
+
+    return torch.repeat_interleave(hidden_states, dim=1, repeats=n_rep)
+
+
+# @log_timing
+class MotifAttention(nn.Module):
+    """
+    Differential Attention (DiffAttention) module.
+
+    Implements the Differential Attention from
+    "DIFFERENTIAL TRANSFORMER" (https://arxiv.org/pdf/2410.05258).
+
+    Overview
+        Standard transformers often over-allocate attention to irrelevant context.
+        DiffAttention addresses this by computing attention as the difference between
+        two separate softmax attention maps, effectively canceling noise and promoting
+        sparse, structured attention patterns.
+
+    Reference Implementation
+        https://github.com/microsoft/unilm/tree/master/Diff-Transformer
+
+    Args
+        The differential attention mechanism computes attention as the difference of two softmax attention scores, weighted by a learnable scalar λ.
+        λ is re-parameterized as λ = exp(λ_q1 · λ_k1) − exp(λ_q2 · λ_k2) + λ_init.
+        - lambda_q1, lambda_q2 (nn.Parameter): Learnable vectors used to compute the first and second components of λ for query transformations.
+        - lambda_k1, lambda_k2 (nn.Parameter): Learnable vectors used to compute the first and second components of λ for key transformations.
+        - lambda_init (float): A constant used for initializing λ, typically set as λ_init = 0.8 − 0.6 × exp(−0.3 × (layer_index − 1)).
+
+    """
+
+    def __init__(self, config: MotifConfig, 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.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.is_causal = True
+        self.attention_dropout = config.attention_dropout
+        try:
+            self.batch_num = config.batch_num
+            logger.info(f'self.batcn_num : {self.batch_num}')
+        except:
+            self.batch_num = None
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                             f" and `num_heads`: {self.num_heads}).")
+
+        self.num_heads //= 2
+        self.num_key_value_heads //= 2
+        self.n_rep = self.num_heads // self.num_key_value_heads
+
+        ##mix attn
+
+        self.mix_attn = config.mix_attn
+
+        if self.mix_attn:
+
+            self.cq, self.ck = 6, 11
+            self.ch = 2
+
+            self.key_query_conv = nn.Conv2d(
+                in_channels=self.num_heads*2,
+                out_channels=self.num_heads*2,
+                kernel_size=(self.cq, self.ck),
+                padding="same",
+                groups=self.num_heads*2
+            )
+
+            self.head_conv = nn.Conv1d(
+                in_channels=self.num_heads,
+                out_channels=self.num_heads,
+                kernel_size=1,
+                padding=0,
+                groups=self.num_heads // self.ch
+            )
+
+            self.group_norm = nn.GroupNorm(num_groups=self.num_heads, num_channels=self.num_heads)
+
+
+
+        # re-init projections
+        self.q_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.hidden_size // self.n_rep, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.hidden_size // self.n_rep, bias=False)
+        self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+
+        # init lambdas
+        for name in ["lambda_q1", "lambda_k1", "lambda_q2", "lambda_k2"]:
+            setattr(self, name, nn.Parameter(torch.zeros(self.head_dim, dtype=torch.float32)))
+            getattr(self, name).data.normal_(mean=0.0, std=0.1)
+
+        # Uses same norm as motif norm, without elementwise_affine option
+        self.subln = MotifRMSNorm(2 * self.head_dim, eps=1e-5)
+        self.lambda_init = 0.8 - 0.6 * math.exp(-0.3 * (layer_idx - 1))
+
+        self.rotary_emb = MotifRotaryEmbedding(self.head_dim,
+                                                max_position_embeddings=self.max_position_embeddings,
+                                                base=self.rope_theta)
+
+        for param in ["q_proj_alpha", "k_proj_alpha", "v_proj_alpha", "o_proj_alpha"]:
+            setattr(
+                self, param,
+                nn.Parameter(torch.tensor(getattr(config, param, 1.0), dtype=torch.float))
+                if config.wesar_weights else 1.0)
+
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_value: Optional[Cache] = None,
+            output_attentions: bool = False,
+            use_cache: bool = False,
+            cache_position: Optional[torch.LongTensor] = None,
+            position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states) * self.q_proj_alpha
+        key_states = self.k_proj(hidden_states) * self.k_proj_alpha
+        value_states = self.v_proj(hidden_states) * self.v_proj_alpha
+
+        ## bsz, seq, n_heads, head_dim
+
+        query_states = query_states.view(bsz, q_len, 2 * self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, 2 * self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, 2 * self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if position_embeddings is None:
+            logger.warning_once(
+                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
+                "removed and `position_embeddings` will be mandatory.")
+            cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        else:
+            cos, sin = (self.rotary_emb(value_states, q_len + past_key_value.get_usable_length(q_len, self.layer_idx))
+                        if use_cache else position_embeddings)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states,
+                                                        key_states,
+                                                        cos,
+                                                        sin,
+                                                        fused_rope=self.config.fused_rope)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # 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)
+
+        ## bsz, #haead, q_len, head_dim -> bsz, head, q_len, q_len
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        kv_seq_len = key_states.shape[-2]
+        offset = kv_seq_len - q_len
+
+        attention_mask = torch.triu(
+            torch.full((q_len, kv_seq_len), float("-inf"), dtype=attn_weights.dtype, device=attn_weights.device),
+            1 + offset)
+        ##attn weights conv2d, softmax and add attention_mask
+        if self.mix_attn:
+            ## condition mask==0, value : 0
+            attn_weights = attn_weights.masked_fill( attention_mask == 0, 0)
+            attn_weights = self.key_query_conv(attn_weights)
+            attn_weights = attn_weights[:, :, :kv_seq_len, :kv_seq_len]
+
+        ###add attn
+        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)
+
+        # differential transformer lambdas
+        lambda_1 = torch.exp(torch.sum(self.lambda_q1 * self.lambda_k1, dim=-1).float()).type_as(attn_weights)
+        lambda_2 = torch.exp(torch.sum(self.lambda_q2 * self.lambda_k2, dim=-1).float()).type_as(attn_weights)
+        lambda_full = lambda_1 - lambda_2 + self.lambda_init
+        attn_weights = attn_weights.view(bsz, self.num_heads, 2, q_len, -1)
+        attn_weights = attn_weights[:, :, 0] - lambda_full * attn_weights[:, :, 1]
+        ##head_conv
+        if self.mix_attn:
+            attn_weights = attn_weights.view(bsz, self.num_heads, -1).contiguous()
+            attn_weights = self.head_conv(attn_weights)
+            attn_weights = attn_weights.view(bsz, self.num_heads, q_len, -1).contiguous()
+
+        ##shape : bsz, #heads, seq, head_dim
+        attn_output = torch.matmul(attn_weights, value_states)
+
+
+        attn_output = self.subln(attn_output)
+        attn_output = attn_output * (1 - self.lambda_init)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim * 2):
+            raise ValueError(f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                             f" {attn_output.size()}")
+        if self.mix_attn:
+            attn_output = self.group_norm(attn_output)
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output) * self.o_proj_alpha
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+# @log_timing
+class MotifFlashAttention2(MotifAttention):
+    """
+    Motif flash attention module, following Motif attention module. This module inherits from `MotifAttention`
+    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. Additionally, for sliding window attention, we apply SWA only to the bottom
+    config.max_window_layers layers.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        logger.info(f'flash attention True')
+
+        # 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 _reshape_heads(self, tensor, batch_size, seq_len):
+        """2-way head split tensor reshape"""
+        return tensor.reshape(batch_size, seq_len, self.num_heads, 2, self.head_dim)
+
+    def _restore_shape(self, tensor, batch_size, seq_len):
+        """restore tensor"""
+        return tensor.reshape(batch_size, seq_len, self.num_heads, self.head_dim)
+
+    def _compute_attention(self, query_states, key_states, value_states, attention_mask, q_len, position_ids,
+                           dropout_rate, sliding_window, is_moreh_attention, batch_num):
+        """Flash Attention 2 implements"""
+        if is_moreh_attention:
+            scale_factor = 1.0 / math.sqrt(self.head_dim)
+            # Copied from _flash_attention_forward
+            if not self._flash_attn_uses_top_left_mask:
+                causal = self.is_causal
+            else:
+                causal = self.is_causal and q_len != 1
+
+            bsz = query_states.shape[0]
+
+            if batch_num:
+                query_states = query_states.reshape(bsz*q_len,self.num_heads,self.head_dim)
+                key_states = key_states.reshape(bsz*q_len,self.num_heads,self.head_dim)
+                value_states = value_states.reshape(bsz*q_len,self.num_heads,self.head_dim)
+
+                attn_out = moreh_ops.flash_attention_varlen_dp(query_states,
+                                                        key_states,
+                                                        value_states,
+                                                        attention_mask,
+                                                        attention_mask,
+                                                        max_seqlen_q=q_len,
+                                                        max_seqlen_kv=q_len,
+                                                        dropout_p=dropout_rate,
+                                                        softmax_scale=scale_factor,
+                                                        is_causal=causal,
+                                                        batch_num=batch_num)
+                attn_out = attn_out.reshape(bsz, q_len, self.num_heads, -1)
+            else:
+                return MorehFlashAttention(query_states,
+                                        key_states,
+                                        value_states,
+                                        padding_mask=attention_mask,
+                                        dropout_p=dropout_rate,
+                                        softmax_scale=scale_factor,
+                                        causal=causal)
+            return attn_out
+        else:
+            return _flash_attention_forward(query_states,
+                                            key_states,
+                                            value_states,
+                                            attention_mask,
+                                            q_len,
+                                            position_ids=position_ids,
+                                            dropout=dropout_rate,
+                                            sliding_window=sliding_window,
+                                            is_causal=self.is_causal,
+                                            use_top_left_mask=self._flash_attn_uses_top_left_mask)
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_value: Optional[Cache] = None,
+            output_attentions: bool = False,
+            use_cache: bool = False,
+            cache_position: Optional[torch.LongTensor] = None,
+            position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+    ):
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states) * self.q_proj_alpha
+        key_states = self.k_proj(hidden_states) * self.k_proj_alpha
+        value_states = self.v_proj(hidden_states) * self.v_proj_alpha
+
+        query_states = query_states.view(bsz, q_len, 2 * self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, 2 * self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, 2 * self.head_dim).transpose(1, 2)
+        kv_seq_len = key_states.shape[-2]
+        if position_embeddings is None:
+            logger.warning_once(
+                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
+                "removed and `position_embeddings` will be mandatory.")
+            cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        else:
+            cos, sin = (self.rotary_emb(value_states, q_len + past_key_value.get_usable_length(q_len, self.layer_idx))
+                        if use_cache else position_embeddings)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states,
+                                                        key_states,
+                                                        cos,
+                                                        sin,
+                                                        fused_rope=True)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # 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 and MorehFlashAttention is None:
+            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)
+
+        q_len = query_states.shape[-2]
+        kv_seq_len = key_states.shape[-2]
+
+        # Reashape to the expected shape for Flash Attention
+        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_sliding_window and getattr(self.config, "sliding_window", None) is not None
+                and self.layer_idx >= self.config.max_window_layers and MorehFlashAttention is None):
+            sliding_window = self.config.sliding_window
+        else:
+            sliding_window = None
+
+        q = self._reshape_heads(query_states, bsz, q_len)
+        k = self._reshape_heads(key_states, bsz, kv_seq_len)
+        v = self._reshape_heads(value_states, bsz, kv_seq_len)
+
+        q1, q2 = q[..., 0, :], q[..., 1, :]
+        k1, k2 = k[..., 0, :], k[..., 1, :]
+        v1, v2 = v[..., 0, :], v[..., 1, :]
+
+        q1, q2, k1, k2, v1, v2 = map(lambda x: self._restore_shape(x, bsz, q_len if x is q1 or x is q2 else kv_seq_len),
+                                     (q1, q2, k1, k2, v1, v2))
+
+        q1, q2 = q1.contiguous(), q2.contiguous()
+        k1, k2 = k1.contiguous(), k2.contiguous()
+        v1, v2 = v1.contiguous(), v2.contiguous()
+
+        is_moreh_attention = MorehFlashAttention is not None
+
+        attn11, attn12 = self._compute_attention(q1, k1, v1, attention_mask, q_len, position_ids, dropout_rate, sliding_window, is_moreh_attention, self.batch_num), \
+                            self._compute_attention(q1, k1, v2, attention_mask, q_len, position_ids, dropout_rate, sliding_window, is_moreh_attention, self.batch_num)
+        attn21, attn22 = self._compute_attention(q2, k2, v1, attention_mask, q_len, position_ids, dropout_rate, sliding_window, is_moreh_attention, self.batch_num), \
+                            self._compute_attention(q2, k2, v2, attention_mask, q_len, position_ids, dropout_rate, sliding_window, is_moreh_attention, self.batch_num)
+
+        attn1, attn2 = torch.cat([attn11, attn12], dim=-1), torch.cat([attn21, attn22], dim=-1)
+
+        lambda_q1 = self.lambda_q1.unsqueeze(0).expand([bsz, self.lambda_q1.shape[0]])  # bsz, num_head
+        lambda_q2 = self.lambda_q2.unsqueeze(0).expand([bsz, self.lambda_q2.shape[0]])  # bsz, num_head
+
+        lambda_1 = torch.exp(torch.sum(lambda_q1 * self.lambda_k1, dim=-1).float()).type_as(attn1)  # bsz
+        lambda_2 = torch.exp(torch.sum(lambda_q2 * self.lambda_k2, dim=-1).float()).type_as(attn2)  # bsz
+
+        lambda_full = lambda_1 - lambda_2 + self.lambda_init
+
+        attn_output = attn1 - lambda_full.view([bsz, 1, 1, 1]) * attn2
+
+        attn_output = self.subln(attn_output)
+        attn_output = attn_output * (1 - self.lambda_init)
+
+        if attn_output.size() != (bsz, q_len, self.num_heads, self.head_dim * 2):
+            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.reshape(bsz, q_len, self.hidden_size)
+        attn_output = self.o_proj(attn_output) * self.o_proj_alpha
+
+        return attn_output, None, past_key_value
+
+
+# @log_timing
+class MotifSdpaAttention(MotifAttention):
+    """
+    Motif attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `MotifAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Adapted from MotifAttention.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,
+            cache_position: Optional[torch.LongTensor] = None,
+            position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+    ) -> 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(
+                "MotifModel is using MotifSdpaAttention, 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)
+        kv_seq_len = key_states.shape[-2]
+        if position_embeddings is None:
+            logger.warning_once(
+                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
+                "removed and `position_embeddings` will be mandatory.")
+            cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        else:
+            cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states,
+                                                        key_states,
+                                                        cos,
+                                                        sin,
+                                                        fused_rope=self.config.fused_rope)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        query_states = query_states.transpose(1, 2).reshape(bsz, q_len, self.hidden_size)
+        key_states = key_states.transpose(1, 2).reshape(bsz, q_len, self.hidden_size // self.num_key_value_groups)
+        value_states = value_states.transpose(1, 2).reshape(bsz, q_len, self.hidden_size // self.num_key_value_groups)
+
+        batch, query_length, key_length = query_states.size(0), query_states.size(-2), key_states.size(-2)
+        masked_bias = attention_mask.expand(batch, self.num_heads, query_length, key_length)
+
+        # Compute Scale Factor
+        scale_factor = 1.0
+        scale_factor /= float(self.head_dim) ** 0.5
+
+        attn_output = ScaledDotProductAttention(query_states,
+                                                key_states,
+                                                value_states,
+                                                masked_bias,
+                                                dropout_rate=0.0,
+                                                training=self.training,
+                                                attn_weight_scale_factor=scale_factor,
+                                                num_kv_groups=self.num_key_value_groups,
+                                                recompute_mode=False)
+        attn_output = attn_output.to(hidden_states.dtype)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, None, past_key_value
+
+
+MOTIF_ATTENTION_CLASSES = {
+    "eager": MotifAttention,
+    "flash_attention_2": MotifFlashAttention2,
+    "sdpa": MotifSdpaAttention,
+}
+
+
+# @log_timing
+class MotifDecoderLayer(nn.Module):
+
+    def __init__(self, config: MotifConfig, moe_layer: bool, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        if config.use_moreh_attention:
+            config._attn_implementation = "flash_attention_2"
+        if config.sliding_window and config._attn_implementation != "flash_attention_2":
+            logger.warning_once(
+                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
+                "unexpected results may be encountered.")
+        if not config.mix_attn:
+            self.self_attn = MOTIF_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
+        else:
+            self.self_attn = MOTIF_ATTENTION_CLASSES["eager"](config, layer_idx)
+        self.mlp = MotifMLP(config)
+        ### moe
+        self.moe = None
+        if moe_layer:
+            self.moe = MotifMoE(config)
+
+        RMSNorm = MorehRMSNorm if MorehRMSNorm is not None else MotifRMSNorm
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        if config.wesar_weights and config.use_norm_alpha:
+            self.input_layernorm_alpha = nn.Parameter(torch.tensor(1).float())
+        else:
+            self.input_layernorm_alpha = 1
+
+        if config.wesar_weights and config.use_norm_alpha :
+            self.post_attention_layernorm_alpha = nn.Parameter(torch.tensor(1).float())
+        else:
+            self.post_attention_layernorm_alpha = 1
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+        **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, sequence_length)` where padding elements are indicated by 0.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
+                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
+                with `head_dim` being the embedding dimension of each attention head.
+            kwargs (`dict`, *optional*):
+                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
+                into the model
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states) * self.input_layernorm_alpha
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states) * self.post_attention_layernorm_alpha
+
+        if self.moe is not None:
+            hidden_states, identity = self.moe(hidden_states)
+            ## add output of shared expert and output of small moe experts.
+            ## hidden state must be zero tensor (for first forward)
+            hidden_states += self.mlp(identity)
+        else:
+            hidden_states = self.mlp(hidden_states)
+
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states, )
+
+        if output_attentions:
+            outputs += (self_attn_weights, )
+
+        if use_cache:
+            outputs += (present_key_value, )
+
+        return outputs
+
+
+MOTIF_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 ([`MotifConfig`]):
+            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 Motif Model outputting raw hidden-states without any specific head on top.",
+    MOTIF_START_DOCSTRING,
+)
+class MotifPreTrainedModel(PreTrainedModel):
+    config_class = MotifConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["MotifDecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+
+    def _init_weights(self, module):
+        module_std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            if getattr(module, "__do_scale_tager__", False):
+                module_std = module_std / self.config.init_scale_o
+
+            if getattr(module, "__do_scale_tager_mu_o__", False):
+                if self.config.dim_model_base_init is not None:
+                    module_std = module_std / math.sqrt(2*(self.config.hidden_size / self.config.dim_model_base_init)*self.config.num_hidden_layers)
+                else:
+                    module_std = module_std
+            elif getattr(module, "__do_scale_tager_mu_ffn__", False):
+                if self.config.dim_model_base_init is not None:
+                    module_std = module_std = module_std / math.sqrt(2*(self.config.hidden_size / self.config.dim_model_base_init)*self.config.num_hidden_layers)
+                else:
+                    module_std = module_std
+            elif getattr(module, "__do_scale_tager_mu_dim_model__", False):
+                if self.config.dim_model_base_init is not None:
+                    module_std = module_std / math.sqrt(self.config.hidden_size / self.config.dim_model_base_init)
+                else:
+                    module_std = module_std
+            elif getattr(module, "__do_scale_tager_mu_dim_base_model__", False):
+                module_std = module_std / math.sqrt(self.config.dim_model_base_lmh) ### lmhead.. 1
+            else:
+                module_std = module_std
+
+            torch.nn.init.trunc_normal_(module.weight.data, mean=0.0, std=module_std, a=-3*module_std, b=3*module_std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+
+        elif isinstance(module, nn.Embedding):
+            torch.nn.init.trunc_normal_(module.weight.data, mean=0.0, std=module_std, a=-3*module_std, b=3*module_std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+@dataclass
+class MotifModelOutputWithPast(ModelOutput):
+    """
+    This augments `BaseModelOutputWithPast` in `transformers.modeling_outputs` with new optional keys: `causal_mask`, `position_embeddings`.
+    The optional keys are currently used in the following ways:
+    - pass information to the token-wise last attention layers in multi-token training
+    """
+    last_hidden_state: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
+    causal_mask: Optional[torch.Tensor] = None
+    position_embeddings: Optional[torch.FloatTensor] = None
+
+
+MOTIF_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 `decoder_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, see our
+            [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
+            - 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.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+# @log_timing
+@add_start_docstrings(
+    "The bare Motif Model outputting raw hidden-states without any specific head on top.",
+    MOTIF_START_DOCSTRING,
+)
+class MotifModel(MotifPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MotifDecoderLayer`]
+
+    Args:
+        config: MotifConfig
+    """
+
+    def __init__(self, config: MotifConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.multi_token_heads = config.multi_token_heads
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        # NOTE: For multi-token models, the last decoder layers (one for each token index)
+        # are implemented as a part of `MotifModelForCausalLM` to enable a custom forward-backward procedure.
+
+        num_hidden_layers = config.num_hidden_layers if self.multi_token_heads is None else config.num_hidden_layers - 1
+        if config.moe:
+            moe_layer = [True for i in range(num_hidden_layers)]
+        else:
+            moe_layer = [False for i in range(num_hidden_layers)]
+        logger.info(f'current_moe layer { moe_layer }')
+        self.layers = nn.ModuleList([MotifDecoderLayer(config = config, moe_layer= moe_layer[layer_idx],
+                                    layer_idx=layer_idx) for layer_idx in range(num_hidden_layers)])
+        self._attn_implementation = config._attn_implementation
+        RMSNorm = MorehRMSNorm if MorehRMSNorm is not None else MotifRMSNorm
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.rotary_emb = MotifRotaryEmbeddingWithCache(self.head_dim,
+                                                max_position_embeddings=self.max_position_embeddings,
+                                                base=self.rope_theta)
+
+        self.gradient_checkpointing = False
+        self.post_init()
+
+        self.use_pipeline = config.use_pipeline
+        if self.use_pipeline:
+            logger.info('use reinforced pp..')
+            if config.num_stages==2:
+                ### moe version
+                if config.decontam_attn:
+                    self.split_layers = [15]
+                else:
+                    if num_hidden_layers == 32:
+                        self.split_layers = [15] # 14: 15,17 # 13: 14:18
+                    else:
+                        self.split_layers = [6]
+            elif config.num_stages==3:
+                self.split_layers = [9,20] ## 11, 11, 10
+            elif config.num_stages==4:
+                self.split_layers = [7,15,23] #7,9,9,7
+            elif config.num_stages==16:
+                self.split_layers = [1,3,5,7,9,11,13,15,17,19,21,23,25,27,29]
+            logger.info(f' check the split layers (moe): {self.split_layers}')
+
+        self.scale_emb = 1
+
+        # Reparameterization <|_1_|>
+        if config.wesar_weights :
+            logger.info(f'config.wesar_weights {config.wesar_weights}')
+            self.norm_alpha = nn.Parameter(torch.tensor(1).float())
+            self.scale_emb = 10
+        else:
+            self.norm_alpha = 1
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(MOTIF_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        outputs_include_causal_mask: bool = False,
+        outputs_include_position_embeddings: bool = False,
+    ) -> Union[Tuple, MotifModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (output_hidden_states
+                                if output_hidden_states is not None else self.config.output_hidden_states)
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...")
+                use_cache = False
+
+        # kept for BC (non `Cache` `past_key_values` inputs)
+        return_legacy_cache = False
+        if use_cache and not isinstance(past_key_values, Cache):
+            return_legacy_cache = True
+            if past_key_values is None:
+                past_key_values = DynamicCache()
+            else:
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+                logger.warning_once(
+                    "We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and "
+                    "will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class "
+                    "(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids) * self.scale_emb
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(past_seen_tokens,
+                                          past_seen_tokens + inputs_embeds.shape[1],
+                                          device=inputs_embeds.device)
+        position_ids = None
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_key_values,
+                                               output_attentions)
+
+        hidden_states = inputs_embeds
+        bsz, q_len, _ = hidden_states.size()
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, seq_len=q_len)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+
+        for idx, decoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                all_hidden_states += (hidden_states, )
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                )
+
+            hidden_states = layer_outputs[0]
+
+
+            if self.use_pipeline and idx in self.split_layers:
+                hidden_states = torch.moreh.pipeline_assign(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], )
+
+        # <|_2_|>
+        hidden_states = self.norm(hidden_states)* self.norm_alpha
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states, )
+
+        next_cache = next_decoder_cache if use_cache else None
+        if return_legacy_cache:
+            next_cache = next_cache.to_legacy_cache()
+
+        causal_mask_output = causal_mask if outputs_include_causal_mask else None
+        position_embeddings_output = position_embeddings if outputs_include_position_embeddings else None
+        if not return_dict:
+            return tuple(v for v in [
+                hidden_states, next_cache, all_hidden_states, all_self_attns, causal_mask_output,
+                position_embeddings_output
+            ] if v is not None)
+        return MotifModelOutputWithPast(last_hidden_state=hidden_states,
+                                        past_key_values=next_cache,
+                                        hidden_states=all_hidden_states,
+                                        attentions=all_self_attns,
+                                        causal_mask=causal_mask_output,
+                                        position_embeddings=position_embeddings_output)
+
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if MorehFlashAttention is not None:
+                return attention_mask
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+        using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if (self.config._attn_implementation == "sdpa" and not (using_static_cache or using_sliding_window_cache)
+                and not output_attentions):
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                    attention_mask,
+                    inputs_embeds=input_tensor,
+                    past_key_values_length=past_seen_tokens,
+                    sliding_window=self.config.sliding_window,
+                    is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        # SlidingWindowCache or StaticCache
+        if using_sliding_window_cache or using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        # DynamicCache or no cache
+        else:
+            target_length = (attention_mask.shape[-1]
+                             if isinstance(attention_mask, torch.Tensor) else past_seen_tokens + sequence_length + 1)
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+            config=self.config,
+            past_key_values=past_key_values,
+        )
+
+        if (self.config._attn_implementation == "sdpa" and attention_mask is not None
+                and attention_mask.device.type == "cuda" and not output_attentions):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        config: MotifConfig,
+        past_key_values: Cache,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            device (`torch.device`):
+                The device to plcae the 4D attention mask on.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+            config (`MotifConfig`):
+                The model's configuration class
+            past_key_values (`Cache`):
+                The cache class that is being used currently to generate
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device)
+            diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+            if config.sliding_window is not None:
+                # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
+                # the check is needed to verify is current checkpoint was trained with sliding window or not
+                if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length:
+                    sliding_attend_mask = torch.arange(
+                        target_length, device=device) <= (cache_position.reshape(-1, 1) - config.sliding_window)
+                    diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
+            causal_mask *= diagonal_attend_mask
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                if attention_mask.shape[-1] > target_length:
+                    attention_mask = attention_mask[:, :target_length]
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype)
+        return causal_mask
+
+
+# @log_timing
+class MotifForCausalLM(MotifPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config: MotifConfig):
+        super().__init__(config)
+        self.model = MotifModel(config)
+        self.vocab_size = config.vocab_size
+        self.multi_token_heads = config.multi_token_heads
+
+        if self.multi_token_heads is None:
+            self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        else:
+            self.tokenwise_last_layers = nn.ModuleList(
+                [MotifDecoderLayer(config, config.num_hidden_layers - 1) for _ in range(self.multi_token_heads)])
+            self.tokenwise_lm_heads = nn.ModuleList(
+                [nn.Linear(config.hidden_size, config.vocab_size, bias=False) for _ in range(self.multi_token_heads)])
+        self.should_skip_separate_backward_pass = self.multi_token_heads is not None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+        # <|_3_|>
+        if config.muP:
+            self.lm_head.__do_scale_tager_mu_dim_base_model__=True
+
+        # <|_4_|>
+        self.lm_head_alpha = 1
+        if config.wesar_weights:
+            self.lm_head_alpha = nn.Parameter(torch.tensor(1).float())
+
+        if getattr(config, "tie_word_embeddings", True):
+            logger.info('tie embeddings')
+            self.tie_weights()
+        else:
+            # <|_5_|>
+            self.lm_head.__do_scale_tager_mu_dim_base_model__ = False
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    def multi_token_forward_backward(self,
+                                     hidden_states: torch.FloatTensor,
+                                     outputs: MotifModelOutputWithPast,
+                                     labels: torch.LongTensor,
+                                     position_ids: Optional[torch.LongTensor],
+                                     output_attentions: Optional[bool],
+                                     use_cache: Optional[bool],
+                                     cache_position: Optional[torch.LongTensor],
+                                     return_dict: Optional[bool],
+                                     num_logits_to_keep: int = 0) -> CausalLMOutputWithPast:
+        """
+        This implements the main forward-backward procedure for multi-token model training proposed in
+        the paper https://arxiv.org/abs/2404.19737.
+        Essentially,
+        - The multi-token model tries to predict n (instead of 1) tokens at a time.
+        - Applying this only during training and using first-token prediction during inference is still helpful.
+        - The change in architecture: when using n-token prediction, each token index (between 1 and n) has its own
+            (1) last attention layer and (2) lm head.
+        - The change in loss: sum of cross-entropy losses corresponding to each token index.
+        - Custom forward-backward procedure for memory efficiency: refer to the implementation of `multi_head_forward_backward`.
+        """
+        if not return_dict:
+            raise NotImplementedError("return_dict must be True for multi-token training")
+
+        past_key_values = outputs.past_key_values
+        causal_mask = outputs.causal_mask
+        position_embeddings = outputs.position_embeddings
+
+        if labels is not None:
+            labels = labels.to(hidden_states.device)
+
+        def _tokenwise_forward(hidden_states: torch.Tensor, token_idx):
+            ## Model forward
+            layer = self.tokenwise_last_layers[token_idx]
+            lm_head = self.tokenwise_lm_heads[token_idx]
+
+            layer_outputs = layer(
+                hidden_states,
+                attention_mask=causal_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,  # TODO: update past_key_values?
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+            )
+            last_hidden_states = layer_outputs[0]
+            if num_logits_to_keep > 0:
+                assert labels is None
+                last_hidden_states = last_hidden_states[:, -num_logits_to_keep:, :]
+            tokenwise_logits = lm_head(last_hidden_states)
+
+            if labels is None:
+                return {
+                    "loss": None,
+                    "logits": tokenwise_logits,
+                }
+
+            ## Compute loss
+            shift_n = token_idx + 1
+            shift_logits = tokenwise_logits[..., :-shift_n, :].contiguous()
+            shift_labels = labels[..., shift_n:].contiguous()
+
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+
+            tokenwise_loss = loss_fct(shift_logits, shift_labels)
+
+            return {
+                "loss": tokenwise_loss,
+                "logits": tokenwise_logits,
+            }
+
+        head_fns = [
+            lambda hidden_states, token_idx=token_idx: _tokenwise_forward(hidden_states, token_idx)
+            for token_idx in range(self.multi_token_heads)
+        ]
+        loss, logits = multi_head_forward_backward(hidden_states,
+                                                   head_fns,
+                                                   return_keys=("loss", "logits"),
+                                                   return_only_first_head=True)
+
+        if not return_dict:
+            output = (logits, ) + outputs[1:]
+            return (loss, ) + output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    @add_start_docstrings_to_model_forward(MOTIF_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, 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,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        num_logits_to_keep: int = 0,
+        **loss_kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+            num_logits_to_keep (`int`, *optional*):
+                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, MotifForCausalLM
+
+        >>> model = MotifForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (output_hidden_states
+                                if output_hidden_states is not None else self.config.output_hidden_states)
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs_include_causal_mask = self.multi_token_heads is not None
+        outputs_include_position_embeddings = self.multi_token_heads is not None
+        outputs: MotifModelOutputWithPast = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            outputs_include_causal_mask=outputs_include_causal_mask,
+            outputs_include_position_embeddings=outputs_include_position_embeddings,
+        )
+
+        hidden_states = outputs[0]
+
+        if self.multi_token_heads is not None:
+            return self.multi_token_forward_backward(hidden_states,
+                                                     outputs,
+                                                     labels,
+                                                     position_ids,
+                                                     output_attentions,
+                                                     use_cache,
+                                                     cache_position,
+                                                     return_dict,
+                                                     num_logits_to_keep=num_logits_to_keep)
+
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        hidden_states = hidden_states * self.lm_head_alpha
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            logits = logits
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits, ) + outputs[1:]
+            return (loss, ) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
\ No newline at end of file