modeling_modernalbert.py

import copy
import math
from dataclasses import dataclass
from typing import Optional, Tuple, Union

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

import transformers as ts
from transformers.modeling_outputs import *
from transformers.modeling_rope_utils import dynamic_rope_update
from transformers.activations import ACT2FN

# Optional FlashAttention import
try:
    from flash_attn.bert_padding import unpad_input, pad_input
    from flash_attn.flash_attn_interface import flash_attn_varlen_qkvpacked_func
    from flash_attn.layers.rotary import RotaryEmbedding
    from flash_attn.ops.triton.rotary import apply_rotary
    FLASH_ATTN_AVAILABLE = True
    print("✅ FlashAttention is available.")
except ImportError:
    FLASH_ATTN_AVAILABLE = False
    print("❌ FlashAttention is not available. Using PyTorch SDPA fallback.")

from .configuration_modernalbert import ModernALBERTConfig

# --- Shared FFN (Unchanged) ---
class SharedLoraFFN(nn.Module):
    """
    A shared Feed-Forward Network modified by LoRA weights.
    The forward pass accepts pre-merged LoRA weights.
    """
    def __init__(self, config):
        super().__init__()
        dim = config.hidden_size
        intermediate_dim = config.expert_intermediate_size
        
        self.linear1 = nn.Linear(dim, intermediate_dim)
        self.act = nn.GELU()
        self.linear2 = nn.Linear(intermediate_dim, dim)
        self.lora_scaling = config.lora_alpha / config.lora_rank

    def forward(self, x, lora_A1, lora_B1, lora_A2, lora_B2):
        # Apply the merged LoRA weights
        # Formula: (x @ A.T @ B.T) * scale
        expanded = self.linear1(x) + (x @ lora_A1.T @ lora_B1.T) * self.lora_scaling
        activated = self.act(expanded)
        contracted = self.linear2(activated) + (activated @ lora_A2.T @ lora_B2.T) * self.lora_scaling
        return contracted

# --- 1. The Router (Fixed for Flash Attention / Unpadded Inputs) ---
class SwitchRouterTopK(nn.Module):
    """
    Calculates the EMA weights for expert merging.
    Optimized for unpadded (Flash Attention) inputs where shape is (total_nnz, dim).
    """
    def __init__(self, config):
        super().__init__()
        self.config = config
        
        self.num_experts = config.num_experts
        # Use a slightly lower decay during training if starting from scratch (e.g., 0.9)
        self.ema_decay = getattr(config, "router_ema_decay", 0.99) 
        
        self.layer = nn.Linear(config.hidden_size, config.num_experts, bias=False)
        self.k = config.top_k
        self.jitter_noise = config.router_jitter_noise
        
        # Buffer for inference (frozen stats)
        self.register_buffer("ema_weights", torch.ones(config.num_experts) / config.num_experts)
        
    def forward(self, hidden_states):
        # hidden_states shape: (total_nnz, hidden_size)
        if self.config.routing_strategy == "ema":
            # 1. Compute Router Probabilities
            logits = self.layer(hidden_states) # Shape: (total_nnz, num_experts)
            probs = F.softmax(logits, dim=-1)
            
            if self.training:
                # 2. Compute batch-level routing vector r_b
                # Since inputs are unpadded (Batch * Seq flattened to dim 0), 
                # we simply average across all tokens to get the global batch stats.
                r_b = probs.mean(dim=0) # Shape: (num_experts,)
    
                # 3. Calculate the weight to USE for this step (Allow Gradients!)
                # We mix history (detached) with current (with grad) to stabilize training.
                weights_for_forward = self.ema_decay * self.ema_weights.detach() + (1 - self.ema_decay) * r_b
                
                # 4. Update the buffer in the background (No Gradients needed for storage)
                new_ema_value = weights_for_forward.detach()
                self.ema_weights.copy_(new_ema_value)
                
                # Normalize to ensure sum is 1
                self.ema_weights.div_(self.ema_weights.sum() + 1e-9)
    
                return weights_for_forward
    
            # During inference, return the frozen stable weights
            return self.ema_weights
        else:
            num_tokens = hidden_states.shape[0]
    
            # if self.training and self.jitter_noise > 0:
            #     noise = torch.randn_like(hidden_states) * self.jitter_noise
            #     hidden_states = hidden_states + noise
    
            logits = self.layer(hidden_states)
            probs = F.softmax(logits, dim=-1, dtype=torch.float32)
            topk_probs, topk_indices = torch.topk(probs, k=self.k, dim=-1)
            topk_probs_normalized = topk_probs / torch.sum(topk_probs, dim=-1, keepdim=True)
            
            # Load Balancing for K = 1
            
            # flat_topk_indices = topk_indices.flatten()
            # one_hot_assignments = F.one_hot(flat_topk_indices, num_classes=self.num_experts).float()
            # tokens_per_expert_fraction = one_hot_assignments.sum(0) / num_tokens
            # print(tokens_per_expert_fraction)
            # router_prob_per_expert = torch.mean(probs, dim=0)
    
            # Load Balancing for K > 1
            one_hot = F.one_hot(topk_indices, num_classes=self.num_experts).float()  
            tokens_per_expert = torch.sum(one_hot * topk_probs.unsqueeze(-1), dim=(0, 1)) / num_tokens
            router_prob_per_expert = torch.mean(probs, dim=0)
        
            aux_loss = self.num_experts * torch.mean(tokens_per_expert * router_prob_per_expert)
    
            # --- DEBUG: print expert utilization for this batch ---
            # print("Expert utilization (fraction of tokens per expert):", tokens_per_expert.detach().cpu().numpy())
            # print(aux_loss)
    
            return topk_indices, topk_probs_normalized, aux_loss

# --- 2. The MoE Layer (Minor cleanup for debug prints) ---
class LoraMoELayerTopK(nn.Module):
    """
    Implements the MoL layer with expert merging.
    Allows for efficient dense computation by collapsing experts 
    into a single adapter based on router weights.
    """
    def __init__(self, config):
        super().__init__()
        self.config = config
        
        dim = config.hidden_size
        expert_intermediate_dim = config.expert_intermediate_size
        num_experts = config.num_experts
        lora_rank = config.lora_rank
        
        self.k = config.top_k
        self.num_experts = num_experts
        
        self.norm = nn.LayerNorm(dim, eps=config.layer_norm_eps)
        
        self.router = SwitchRouterTopK(config)
        self.shared_ffn = SharedLoraFFN(config)
        
        # The pool of Expert LoRA weights {\Delta_1, ..., \Delta_E}
        self.lora_A1 = nn.Parameter(torch.randn(num_experts, lora_rank, dim))
        self.lora_B1 = nn.Parameter(torch.zeros(num_experts, expert_intermediate_dim, lora_rank))
        self.lora_A2 = nn.Parameter(torch.randn(num_experts, lora_rank, expert_intermediate_dim))
        self.lora_B2 = nn.Parameter(torch.zeros(num_experts, dim, lora_rank))
        
        # Initialization (Kaiming Uniform)
        for i in range(num_experts):
            nn.init.kaiming_uniform_(self.lora_A1[i], a=math.sqrt(5))
            nn.init.kaiming_uniform_(self.lora_A2[i], a=math.sqrt(5))

    def forward(self, hidden_states: torch.Tensor):
        if self.config.routing_strategy == "ema":
            residual = hidden_states
            hidden_states_norm = self.norm(hidden_states)
            
            # 1. Get the global merging weights (w_t)
            # Returns shape: (num_experts,)
            merge_weights = self.router(hidden_states_norm)
            
            # 2. Weighted Merge of all LoRA parameters
            # Formula: \Delta_{merged} = \sum_{j=1}^E w_j * \Delta_j
            # We reshape weights to [Experts, 1, 1] for broadcasting against [Experts, Rank, Dim]
            w = merge_weights.view(-1, 1, 1)
            
            merged_A1 = torch.sum(w * self.lora_A1, dim=0)
            merged_B1 = torch.sum(w * self.lora_B1, dim=0)
            merged_A2 = torch.sum(w * self.lora_A2, dim=0)
            merged_B2 = torch.sum(w * self.lora_B2, dim=0)
    
            # 3. Dense Forward Pass
            # Pass the merged adapter to the FFN. 
            output = self.shared_ffn(
                hidden_states_norm, 
                merged_A1, merged_B1, 
                merged_A2, merged_B2
            )
    
            # We return 0.0 for aux_loss because we are not doing load balancing in this mode
            return residual + output, torch.tensor(0.0, device=hidden_states.device)
        elif self.config.routing_strategy == "uniform":
            residual = hidden_states
            hidden_states_norm = self.norm(hidden_states)
            
            # 1. Get the global merging weights (w_t)
            # Returns shape: (num_experts,)
            merge_weights = torch.ones(self.config.num_experts, dtype=hidden_states_norm.dtype, device=hidden_states_norm.device) / (self.config.num_experts)
            
            # 2. Weighted Merge of all LoRA parameters
            # Formula: \Delta_{merged} = \sum_{j=1}^E w_j * \Delta_j
            # We reshape weights to [Experts, 1, 1] for broadcasting against [Experts, Rank, Dim]
            w = merge_weights.view(-1, 1, 1)
            
            merged_A1 = torch.sum(w * self.lora_A1, dim=0)
            merged_B1 = torch.sum(w * self.lora_B1, dim=0)
            merged_A2 = torch.sum(w * self.lora_A2, dim=0)
            merged_B2 = torch.sum(w * self.lora_B2, dim=0)
    
            # 3. Dense Forward Pass
            # Pass the merged adapter to the FFN. 
            output = self.shared_ffn(
                hidden_states_norm, 
                merged_A1, merged_B1, 
                merged_A2, merged_B2
            )
    
            # We return 0.0 for aux_loss because we are not doing load balancing in this mode
            return residual + output, torch.tensor(0.0, device=hidden_states.device)
        else:
            residual = hidden_states
            hidden_states_norm = self.norm(hidden_states)

            # The permutation-based expert dispatch below expects a flat
            # (total_tokens, dim) layout, matching FlashAttention's unpadded
            # format. Under the SDPA fallback, tokens stay padded as
            # (batch, seq, dim), so flatten here and restore the shape below.
            orig_shape = hidden_states_norm.shape
            if hidden_states_norm.dim() == 3:
                hidden_states_norm = hidden_states_norm.reshape(-1, orig_shape[-1])

            num_tokens, dim = hidden_states_norm.shape
    
            topk_indices, topk_probs, aux_loss = self.router(hidden_states_norm)
            
            # Efficient permutation-based dispatch
            flat_token_indices = torch.arange(num_tokens, device=hidden_states.device).repeat_interleave(self.k)
            flat_expert_indices = topk_indices.flatten()
            
            perm_indices = torch.argsort(flat_expert_indices)
            sorted_token_indices = flat_token_indices[perm_indices]
            sorted_expert_indices = flat_expert_indices[perm_indices]
            
            permuted_tokens = hidden_states_norm[sorted_token_indices]
            permuted_probs = topk_probs.flatten()[perm_indices]
    
            tokens_per_expert = F.one_hot(sorted_expert_indices, self.num_experts).sum(dim=0)
            split_tokens = torch.split(permuted_tokens, tokens_per_expert.tolist(), dim=0)
            split_probs = torch.split(permuted_probs, tokens_per_expert.tolist(), dim=0)
    
            # Batched processing loop over experts
            expert_outputs = []
            for i in range(self.num_experts):
                if tokens_per_expert[i] > 0:
                    output = self.shared_ffn(
                        split_tokens[i],
                        self.lora_A1[i], self.lora_B1[i],
                        self.lora_A2[i], self.lora_B2[i]
                    )
                    expert_outputs.append(output * split_probs[i].unsqueeze(1))
                else:
                    expert_outputs.append(torch.empty(0, dim, device=hidden_states.device))
    
            # Un-permute and combine results
            concatenated_outputs = torch.cat(expert_outputs, dim=0)
            inverse_perm_indices = torch.argsort(perm_indices)
            unpermuted_outputs = concatenated_outputs[inverse_perm_indices]
            
            final_output = unpermuted_outputs.view(num_tokens, self.k, dim).sum(dim=1)
            if len(orig_shape) == 3:
                final_output = final_output.view(orig_shape)
            
            # Final residual connection
            output = residual + final_output
            return output, aux_loss

        
def rotate_half(x):
    """Rotates half the hidden dims of the input (standard HF RoPE convention)."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
    """Applies rotary position embeddings to q and k (standard HF RoPE convention)."""
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


class ModernAlbertMLP(nn.Module):
    def __init__(self, config: ModernALBERTConfig):
        super().__init__()
        self.config = config
        self.Wi = nn.Linear(config.hidden_size, int(config.intermediate_size) * 2, bias=False)
        # self.act = ACT2FN[config.hidden_activation]
        self.act = ACT2FN["gelu"]
        self.drop = nn.Dropout(config.hidden_dropout_prob)
        self.Wo = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        input, gate = self.Wi(hidden_states).chunk(2, dim=-1)
        return self.Wo(self.drop(self.act(input) * gate))

if FLASH_ATTN_AVAILABLE:
    #Flash Attention Rotatory Embedding
    class ApplyRotaryEmbUnpad(torch.autograd.Function):
        @staticmethod
        def forward(
            ctx,
            qkv,
            cos,
            sin,
            cu_seqlens: Optional[torch.Tensor] = None,
            max_seqlen: Optional[int] = None,
        ):
            # (total_nnz, 3, nheads, headdim)
            qkv = qkv.contiguous()
            total_nnz, _three, _nheads, headdim = qkv.shape
            # We need qkv to be contiguous so that when we reshape to combine (3, nheads) dimensions,
            # we get the same tensor
            # qk = rearrange(qkv[:, :2], "b_s t h d -> b_s (t h) d")
            qk = qkv[:, :2].view(total_nnz, -1, headdim)
            apply_rotary(
                qk,
                cos,
                sin,
                seqlen_offsets=0,
                cu_seqlens=cu_seqlens,
                max_seqlen=max_seqlen,
                interleaved=False,
                inplace=True,
            )

            ctx.save_for_backward(cos, sin, cu_seqlens)
            ctx.max_seqlen = max_seqlen
            return qkv

        @staticmethod
        def backward(ctx, do):
            cos, sin, cu_seqlens = ctx.saved_tensors
            do = do.contiguous()
            total_nnz, _three, _nheads, headdim = do.shape
            # We need dqkv to be contiguous so that when we reshape to combine (3, nheads) dimensions,
            # we get the same tensor
            dqk = do[:, :2].view(total_nnz, -1, headdim)
            apply_rotary(
                dqk,
                cos,
                sin,
                seqlen_offsets=0,
                cu_seqlens=cu_seqlens,
                max_seqlen=ctx.max_seqlen,
                interleaved=False,
                inplace=True,
                conjugate=True,
            )

            return do, None, None, None, None, None, None


    def apply_rotary_unpadded(
        qkv,
        cos,
        sin,
        cu_seqlens: Optional[torch.Tensor] = None,
        max_seqlen: Optional[int] = None,
    ):
        """
        Arguments:
            qkv: (total_nnz, 3, nheads, headdim) - input tensor for packed QKV.
            cos, sin: (seqlen_rotary, rotary_dim / 2)
            interleaved: if True, rotate pairs of even and odd dimensions (GPT-J style) instead
                of 1st half and 2nd half (GPT-NeoX style).
            inplace: if True, apply rotary embedding in-place.
            seqlen_offsets: (batch_size,) or int. Each sequence in x is shifted by this amount.
                Most commonly used in inference when we have KV cache.
            cu_seqlens: (batch + 1,) or None
            max_seqlen: int
        Return:
            out: (total_nnz, dim)
        rotary_dim must be <= headdim
        Apply rotary embedding to the first rotary_dim of x.
        """
        return ApplyRotaryEmbUnpad.apply(qkv, cos, sin, cu_seqlens, max_seqlen)


    class ModernAlbertUnpaddedRotaryEmbedding(RotaryEmbedding):
        """
        The rotary position embeddings applied directly to unpadded sequences.
        """

        def __init__(
            self,
            dim: int,
            base: float = 10000.0,
            max_seqlen: Optional[int] = None,
            device: Optional[torch.device] = None,
            dtype: Optional[torch.dtype] = None,
        ):
            """
            max_seqlen: if max_seqlen, device, and dtype are provided, we precompute the cos_sin_cache
                up to max_seqlen. If the max_seqlen, device, or dtype during training/inference differ,
                the cos_sin_cache will be recomputed during the forward pass.
            """
            super().__init__(dim=dim, base=base, device=device, interleaved=False)
            self.max_seqlen = max_seqlen

            if max_seqlen is not None and device is not None and dtype is not None:
                self._update_cos_sin_cache(max_seqlen, device=device, dtype=dtype)

        def forward(
            self,
            qkv: torch.Tensor,
            cu_seqlens: torch.Tensor,
            max_seqlen: Optional[int] = None,
        ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
            """
            Apply rotary embedding *inplace* to qkv.
            qkv: (total_nnz, 3, nheads, headdim)
            cu_seqlens: (batch + 1,) cumulative sequence lengths
            max_seqlen: int max seq length in the batch
            """
            if max_seqlen is not None:
                self._update_cos_sin_cache(max_seqlen, device=qkv.device, dtype=qkv.dtype)

            qkv = apply_rotary_unpadded(
                qkv,
                self._cos_cached,
                self._sin_cached,
                cu_seqlens=cu_seqlens,
                max_seqlen=max_seqlen,
            )

            return qkv

        def extra_repr(self) -> str:
            return f"dim={self.dim}, base={self.base}, scale_base={self.scale_base}"


class ModernAlbertRotaryEmbedding(nn.Module):
    def __init__(self, config: ModernALBERTConfig, device=None):
        super().__init__()
        # Only standard (unscaled) RoPE is used in ModernALBERT, so the frequencies
        # are computed directly instead of via transformers' ROPE_INIT_FUNCTIONS
        # registry, whose available keys have changed across transformers versions.
        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

        base = getattr(config, "rope_theta", 10000.0)
        head_dim = config.hidden_size // config.num_attention_heads
        inv_freq = 1.0 / (
            base ** (torch.arange(0, head_dim, 2, dtype=torch.int64, device=device).float() / head_dim)
        )
        self.attention_scaling = 1.0
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    def compute_default_rope_parameters(self, config, device=None, **kwargs):
        """Matches the signature transformers' generic weight-init expects for
        any submodule whose class name contains "RotaryEmbedding" (see
        PreTrainedModel._init_weights): returns (inv_freq, attention_scaling)."""
        base = getattr(config, "rope_theta", 10000.0)
        head_dim = config.hidden_size // config.num_attention_heads
        inv_freq = 1.0 / (
            base ** (torch.arange(0, head_dim, 2, dtype=torch.int64, device=device).float() / head_dim)
        )
        return inv_freq, 1.0

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()

        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

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

# GeGLU unchanged
class GeGLU(nn.Module):
    def __init__(self, dim_in, dim_out):
        super().__init__()
        self.w1 = nn.Linear(dim_in, dim_out)
        self.w2 = nn.Linear(dim_in, dim_out)
    def forward(self, x):
        return F.gelu(self.w1(x)) * self.w2(x)


#Flash Attention
def _unpad_modernbert_input(
    inputs: torch.Tensor,
    attention_mask: torch.Tensor,
    position_ids: Optional[torch.Tensor] = None,
    labels: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, int, Optional[torch.Tensor], Optional[torch.Tensor]]:
    """
    Remove padding from input sequences.

    Args:
        inputs: (batch, seqlen, ...) or (batch, seqlen)
        attention_mask: (batch, seqlen), bool / int, 1 means valid and 0 means not valid.
        position_ids: (batch, seqlen), int, position ids
        labels: (batch, seqlen), int, labels

    Returns:
        unpadded_inputs: (total_nnz, ...), where total_nnz = number of tokens selected in attention_mask.
        indices: (total_nnz)
        cu_seqlens: (batch + 1), the cumulative sequence lengths
        max_seqlen_in_batch: int
        unpadded_position_ids: (total_nnz) or None
        unpadded_labels: (total_nnz) or None
    """
    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
    max_seqlen_in_batch = int(seqlens_in_batch.max().item())
    cu_seqlens = torch.nn.functional.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))

    if inputs.dim() == 2:
        unpadded_inputs = inputs.flatten()[indices]
    else:
        batch, seqlen, *rest = inputs.shape
        shape = batch * seqlen
        unpadded_inputs = inputs.view(shape, *rest)[indices]

    unpadded_position_ids = position_ids.flatten()[indices] if position_ids is not None else None
    unpadded_labels = labels.flatten()[indices] if labels is not None else None

    return unpadded_inputs, indices, cu_seqlens, max_seqlen_in_batch, unpadded_position_ids, unpadded_labels


def _pad_modernbert_output(
    inputs: torch.Tensor,
    indices: torch.Tensor,
    batch: int,
    seqlen: int,
) -> torch.Tensor:
    """
    Add padding to sequences.

    Args:
        inputs: (total_nnz, ...) or (total_nnz,), where total_nnz = number of tokens selected in attention_mask.
        indices: (total_nnz)
        batch: int, batch size
        seqlen: int, max sequence length

    Returns:
        padded_inputs: (batch, seqlen, ...) or (batch, seqlen)
    """
    if inputs.dim() == 1:
        output = torch.zeros(batch * seqlen, dtype=inputs.dtype, device=inputs.device)
        output[indices] = inputs
        padded_inputs = output.view(batch, seqlen)
    else:
        _, *rest = inputs.shape
        output = torch.zeros(batch * seqlen, *rest, dtype=inputs.dtype, device=inputs.device)
        output[indices] = inputs
        padded_inputs = output.view(batch, seqlen, *rest)

    return padded_inputs


def flash_attention_forward(
    module: "SharedGroup",
    qkv: torch.Tensor,
    rotary_emb: "ModernAlbertUnpaddedRotaryEmbedding",
    cu_seqlens: torch.Tensor,
    max_seqlen: int,
    local_attention: tuple[int, int],
    bs: int,
    dim: int,
    target_dtype: torch.dtype = torch.bfloat16,
    **_kwargs,
) -> tuple[torch.Tensor]:
    # (total_seqlen, 3, nheads, headdim)
    qkv = rotary_emb(qkv, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen)

    convert_dtype = qkv.dtype not in (torch.float16, torch.bfloat16)
    if convert_dtype:
        # FA2 implementation only supports fp16 and bf16. If FA2 is supported,
        # bfloat16 must be supported as of FA2 2.5.7. (Turing GPUs not supported)
        orig_dtype = qkv.dtype
        qkv = qkv.to(target_dtype)

        attn = flash_attn_varlen_qkvpacked_func(
            qkv,
            cu_seqlens=cu_seqlens,
            max_seqlen=max_seqlen,
            dropout_p=module.att_dropout.p if module.training else 0.0,
            # deterministic=module.deterministic_flash_attn,
            # deterministic=False,
            window_size=local_attention,
        )
        attn = attn.to(orig_dtype)  # type: ignore
    else:
        attn = flash_attn_varlen_qkvpacked_func(
            qkv,
            cu_seqlens=cu_seqlens,
            max_seqlen=max_seqlen,
            dropout_p=module.att_dropout.p if module.training else 0.0,
            # deterministic=module.deterministic_flash_attn,
            window_size=local_attention,
        )
    return (attn.view(bs, dim),)

def sdpa_attention_forward(
    module: "SharedGroup",
    qkv: torch.Tensor,
    attention_mask: torch.Tensor,
    sliding_window_mask: torch.Tensor,
    position_ids: Optional[torch.LongTensor],
    local_attention: tuple[int, int],
    bs: int,
    dim: int,
    **_kwargs,
) -> tuple[torch.Tensor]:
    # qkv: [batch_size, seqlen, 3, nheads, headdim]
    cos, sin = module.rotary_emb(qkv, position_ids=position_ids)
    query, key, value = qkv.transpose(3, 1).unbind(dim=2)
    # query, key, value: [batch_size, heads, seq_len, head_dim]
    query, key = apply_rotary_pos_emb(query, key, cos, sin)

    if local_attention != (-1, -1):
        attention_mask = sliding_window_mask

    attn_output = (
        F.scaled_dot_product_attention(
            query,
            key,
            value,
            dropout_p=module.att_dropout.p if module.training else 0.0,
            attn_mask=attention_mask,
        )
        .transpose(1, 2)
        .contiguous()
    )
    attn_output = attn_output.view(bs, -1, dim)
    return (attn_output,)

class SharedGroup(nn.Module):
    def __init__(self, config): # config: ModernALBERTConfig
        super().__init__()
        self.config = config
    
        hs, nh = config.hidden_size, config.num_attention_heads
        self.head_dim = hs // nh
        self.num_heads = nh
        self.use_adapter = config.use_adapter
        eps = config.layer_norm_eps

        rope_theta = 10000
        
        # Norms
        self.att_pre_norm = nn.LayerNorm(hs, eps=eps)
        self.ffn_pre_norm = nn.LayerNorm(hs, eps=eps)
        
        # Attention
        self.qkv = nn.Linear(hs, 3 * hs)
        self.out_proj = nn.Linear(hs, hs)
        self.att_dropout = nn.Dropout(config.attention_probs_dropout_prob)
        self.local_attention = (-1, -1)
        
        if FLASH_ATTN_AVAILABLE:
            self.rotary_emb = ModernAlbertUnpaddedRotaryEmbedding(
                dim=self.head_dim, max_seqlen=config.max_position_embeddings, base=rope_theta
            )
        else:
            config_copy = copy.deepcopy(config)
            config_copy.rope_theta = rope_theta
            self.rotary_emb = ModernAlbertRotaryEmbedding(config=config_copy)
            
        # FFN
        self.mlp = ModernAlbertMLP(config)

    def forward(self, inputs, mask, config, start_idx=0, use_moa=False, **kwargs):        
        outputs = [] if config.output_hidden_states else None
        attn_maps = [] if config.output_attentions else None

        x = inputs

        for i in range(config.group_depth):
            h = x
            h_norm = self.att_pre_norm(h)
            
            qkv_proj = self.qkv(h_norm)           
            bs = h.shape[0]

            # --- Attention Calculation ---
            if FLASH_ATTN_AVAILABLE:
                qkv = qkv_proj.view(-1, 3, self.num_heads, self.head_dim)
        
                attn_outputs = flash_attention_forward(
                    self,
                    qkv=qkv,
                    rotary_emb=self.rotary_emb,
                    local_attention=self.local_attention,
                    bs=bs,
                    dim=self.head_dim * self.num_heads,
                    **kwargs,
                )
                
                attn_out = attn_outputs[0]
            else: # Fallback to PyTorch Scaled Dot Product Attention
                qkv = qkv_proj.view(bs, -1, 3, self.num_heads, self.head_dim)
                # `mask` is already reshaped to (batch, 1, 1, seq) by
                # ModernALBERTModel.forward's SDPA branch - reuse it as-is.
                attn_mask = mask
                
                attn_outputs = sdpa_attention_forward(
                    self,
                    qkv=qkv,
                    attention_mask=attn_mask,
                    rotary_emb=self.rotary_emb,
                    local_attention=self.local_attention,
                    bs=bs,
                    dim=self.head_dim * self.num_heads,
                    **{k: v for k, v in kwargs.items() if k != "attention_mask"},
                )
                
                attn_out = attn_outputs[0]

            x = self.att_dropout(self.out_proj(attn_out)) + h

            if use_moa == True and i == config.group_depth - 1:
                return x, outputs, attn_maps
            else:
                # FFN block
                h2 = x
                h2_norm = self.ffn_pre_norm(h2)
                    
                x = self.mlp(h2_norm) + h2
                
                # Collect hidden state if needed
                if config.output_hidden_states:
                    outputs.append(x)
                
        return x, outputs, attn_maps


class ModernAlbertEmbeddings(nn.Module):
    """
    Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
    """

    def __init__(self, config: ModernALBERTConfig):
        super().__init__()
        self.config = config
        
        self.tok_embeddings = nn.Embedding(config.vocab_size, config.embedding_size, padding_idx=config.pad_token_id)
        self.embed_proj = nn.Linear(config.embedding_size, config.hidden_size)
        
        # self.norm = nn.LayerNorm(config.hidden_size, eps=config.norm_eps, bias=config.norm_bias)
        self.norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, bias=False)
        self.drop = nn.Dropout(0.0)

    def forward(
        self, input_ids: Optional[torch.LongTensor] = None, inputs_embeds: Optional[torch.Tensor] = None
    ) -> torch.Tensor:
        if inputs_embeds is not None:
            hidden_states = self.drop(self.norm(self.embed_proj(inputs_embeds)))
        else:
            hidden_states = self.drop(self.norm(self.embed_proj(self.tok_embeddings(input_ids))))
        
        return hidden_states

@dataclass
class MoABaseModelOutput(BaseModelOutput):
    load_balancing_loss: Optional[torch.FloatTensor] = None

class ModernALBERTModel(ts.PreTrainedModel):
    config_class = ModernALBERTConfig
    base_model_prefix = "modernAlbert"
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True

    def __init__(self, config: ModernALBERTConfig):
        super().__init__(config)
        self.config = config
        
        # Factorized embeddings
        self.embeddings = ModernAlbertEmbeddings(config)
        self.final_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, bias=False)
        self.num_groups = config.num_hidden_layers // config.group_depth
        self.groups = nn.ModuleList([SharedGroup(config) for _ in range(self.num_groups)])
        
        if config.use_moa:
            self.moa_layers = nn.ModuleList([
                # LoraMoELayerTopK(config) for _ in range(self.num_groups - 1) 
                LoraMoELayerTopK(config) for _ in range(config.num_expert_modules) 
            ])
            
        self.pooler = nn.Linear(config.hidden_size, config.hidden_size)
        self.post_init()

    def forward(self, 
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        sliding_window_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        indices: Optional[torch.Tensor] = None,
        cu_seqlens: Optional[torch.Tensor] = None,
        max_seqlen: Optional[int] = None,
        batch_size: Optional[int] = None,
        seq_len: Optional[int] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ):
        all_hidden_states = []
        
        if batch_size is None and seq_len is None:
            if inputs_embeds is not None:
                batch_size, seq_len = inputs_embeds.shape[:2]
            else:
                batch_size, seq_len = input_ids.shape[:2]
        
        output_hidden_states = output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions

        if output_hidden_states:
            self.config.output_hidden_states = True

        hs, atts = ([] if output_hidden_states else None), ([] if output_attentions else None)
        all_aux_losses = []
        
        repad = False
        if FLASH_ATTN_AVAILABLE:
            if indices is None and cu_seqlens is None and max_seqlen is None:
                repad = True
                if inputs_embeds is None:
                    with torch.no_grad():
                        input_ids, indices, cu_seqlens, max_seqlen, *_ = _unpad_modernbert_input(
                            inputs=input_ids, attention_mask=attention_mask
                        )
                else:
                    inputs_embeds, indices, cu_seqlens, max_seqlen, *_ = _unpad_modernbert_input(
                        inputs=inputs_embeds, attention_mask=attention_mask
                    )
        else:
            _device = input_ids.device if input_ids is not None else inputs_embeds.device
            if position_ids is None:
                position_ids = torch.arange(seq_len, device=_device).unsqueeze(0)

            # Build a boolean attention mask broadcastable to
            # (batch, heads, q_len, kv_len) for F.scaled_dot_product_attention.
            # (Local/sliding-window attention isn't used by any released
            # ModernALBERT variant, so sliding_window_mask just mirrors attention_mask.)
            if attention_mask is not None:
                attention_mask = attention_mask.to(torch.bool)[:, None, None, :].to(_device)
            sliding_window_mask = attention_mask

        hidden_states = self.embeddings(input_ids=input_ids, inputs_embeds=inputs_embeds)        
        x = hidden_states

        if output_hidden_states:
            hs.append(x)

        # Mask
        mask = None
        if attention_mask is not None:
            mask = attention_mask.to(torch.bool)
        
        for i, group in enumerate(self.groups):

            is_moa = self.config.use_moa and (i > len(self.groups) - len(self.moa_layers) - 1)
            moa_idx = i - (len(self.groups) - len(self.moa_layers))
            
            x, layer_hs, layer_atts = group(x, 
                                            mask, 
                                            self.config,             
                                            sliding_window_mask=sliding_window_mask,
                                            position_ids=position_ids,
                                            cu_seqlens=cu_seqlens,
                                            max_seqlen=max_seqlen,
                                            use_moa=is_moa,
                                            output_attentions=output_attentions,)
            
            if output_hidden_states and layer_hs:
                hs.extend(layer_hs)
            if output_attentions and layer_atts:
                atts.extend(layer_atts)

            # After each group (except the last), apply the MoA layer
            if self.config.use_moa and is_moa:
                x, aux_loss = self.moa_layers[moa_idx](x)
                if output_hidden_states:
                    hs.append(x)
                all_aux_losses.append(aux_loss)


        hidden_states = self.final_norm(x)

        # hidden_states = _pad_modernbert_output(
        #         inputs=hidden_states, indices=indices, batch=batch_size, seqlen=seq_len
        # )
        
        if repad:
            hidden_states = _pad_modernbert_output(
                inputs=hidden_states, indices=indices, batch=batch_size, seqlen=seq_len
            )
            if all_hidden_states is not None:
                all_hidden_states = tuple(
                    _pad_modernbert_output(inputs=hs, indices=indices, batch=batch_size, seqlen=seq_len)
                    for hs in all_hidden_states
                )


        load_balancing_loss = None
        if all_aux_losses != []:
            load_balancing_loss = torch.stack(all_aux_losses).mean() * self.config.load_balancing_loss_coef

        return MoABaseModelOutput(last_hidden_state=hidden_states, hidden_states=hs, attentions=atts, load_balancing_loss=load_balancing_loss)

class ModernAlbertPredictionHead(nn.Module):
    def __init__(self, config: ModernALBERTConfig):
        super().__init__()
        self.config = config
        self.dense = nn.Linear(config.hidden_size, config.embedding_size, bias=False)
        self.act = ACT2FN["gelu"]
        self.norm = nn.LayerNorm(config.embedding_size, eps=config.layer_norm_eps, bias=False)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        return self.norm(self.act(self.dense(hidden_states)))

class ModernALBERTForMaskedLM(ts.PreTrainedModel):
    """
    Modern ALBERT model with a Masked Language Modeling (MLM) head,
    optimized to mirror the HuggingFace `AlbertForMaskedLM` API.
    """
    _tied_weights_keys = {"decoder.weight": "albert.embeddings.tok_embeddings.weight"}
    config_class = ModernALBERTConfig
    base_model_prefix = "modernAlbert"
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_cache_class = True

    def __init__(self, config: ModernALBERTConfig):
        super().__init__(config)
        self.config = config

        # Base encoder without pooling
        self.albert = ModernALBERTModel(config)

        # MLM head
        self.head = ModernAlbertPredictionHead(config)
        self.decoder = nn.Linear(config.embedding_size, config.vocab_size, bias=False)

        self.post_init()

    def get_input_embeddings(self):
        return self.albert.embeddings.tok_embeddings

    def get_output_embeddings(self):
        return self.decoder

    def set_output_embeddings(self, new_embeddings: nn.Linear):
        self.decoder = new_embeddings

    @torch.compile(dynamic=True)
    def compiled_head(self, output: torch.Tensor) -> torch.Tensor:
        return self.decoder(self.head(output))

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        sliding_window_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        indices: Optional[torch.Tensor] = None,
        cu_seqlens: Optional[torch.Tensor] = None,
        max_seqlen: Optional[int] = None,
        batch_size: Optional[int] = None,
        seq_len: Optional[int] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        **kwargs,
    ):
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        
        if FLASH_ATTN_AVAILABLE:
            if indices is None and cu_seqlens is None and max_seqlen is None:
                if batch_size is None and seq_len is None:
                    if inputs_embeds is not None:
                        batch_size, seq_len = inputs_embeds.shape[:2]
                    else:
                        batch_size, seq_len = input_ids.shape[:2]
                device = input_ids.device if input_ids is not None else inputs_embeds.device

                if attention_mask is None:
                    attention_mask = torch.ones((batch_size, seq_len), device=device, dtype=torch.bool)

                if inputs_embeds is None:
                    with torch.no_grad():
                        input_ids, indices, cu_seqlens, max_seqlen, position_ids, labels = _unpad_modernbert_input(
                            inputs=input_ids, attention_mask=attention_mask, position_ids=position_ids, labels=labels
                        )
                else:
                    inputs_embeds, indices, cu_seqlens, max_seqlen, position_ids, labels = _unpad_modernbert_input(
                        inputs=inputs_embeds, attention_mask=attention_mask, position_ids=position_ids, labels=labels
                    )
                    
        # Encode
        outputs = self.albert(
            input_ids=input_ids,
            attention_mask=attention_mask,
            sliding_window_mask=sliding_window_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            indices=indices,
            cu_seqlens=cu_seqlens,
            max_seqlen=max_seqlen,
            batch_size=batch_size,
            seq_len=seq_len,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        last_hidden_state = outputs[0]

        if FLASH_ATTN_AVAILABLE:
            last_hidden_state_unpaded = _pad_modernbert_output(inputs=last_hidden_state, indices=indices, batch=batch_size, seqlen=seq_len)
            if outputs.hidden_states != None:
                outputs.hidden_states.append(last_hidden_state_unpaded)
        
        logits = self.decoder(self.head(last_hidden_state))
        
        loss = None
        
        if labels is not None:
            loss = self.loss_function(logits, labels, vocab_size=self.config.vocab_size, **kwargs)

            if outputs.load_balancing_loss != None and self.training:
                # print(outputs.load_balancing_loss)
                loss += outputs.load_balancing_loss
            
        if FLASH_ATTN_AVAILABLE:
            logits = _pad_modernbert_output(inputs=logits, indices=indices, batch=batch_size, seqlen=seq_len)

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

        return MaskedLMOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

class ModernALBERTForSequenceClassification(ts.PreTrainedModel):
    config_class = ModernALBERTConfig
    
    def __init__(self, config: ModernALBERTConfig):
        super().__init__(config)
        self.num_labels = config.num_labels
        self.config = config

        self.albert = ModernALBERTModel(config)
        self.head = ModernAlbertPredictionHead(config)
        self.drop = torch.nn.Dropout(0.0)
        self.classifier = nn.Linear(config.embedding_size, config.num_labels)

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

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        sliding_window_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        indices: Optional[torch.Tensor] = None,
        cu_seqlens: Optional[torch.Tensor] = None,
        max_seqlen: Optional[int] = None,
        batch_size: Optional[int] = None,
        seq_len: Optional[int] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        **kwargs,
    ) -> Union[tuple[torch.Tensor], SequenceClassifierOutput]:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        # self._maybe_set_compile()

        if FLASH_ATTN_AVAILABLE:
            if indices is None and cu_seqlens is None and max_seqlen is None:
                if batch_size is None and seq_len is None:
                    if inputs_embeds is not None:
                        batch_size, seq_len = inputs_embeds.shape[:2]
                    else:
                        batch_size, seq_len = input_ids.shape[:2]
                device = input_ids.device if input_ids is not None else inputs_embeds.device

                if attention_mask is None:
                    attention_mask = torch.ones((batch_size, seq_len), device=device, dtype=torch.bool)

                if inputs_embeds is None:
                    with torch.no_grad():
                        input_ids, indices, cu_seqlens, max_seqlen, position_ids, _ = _unpad_modernbert_input(
                            inputs=input_ids, attention_mask=attention_mask, position_ids=position_ids, labels=None
                        )
                else:
                    inputs_embeds, indices, cu_seqlens, max_seqlen, position_ids, _ = _unpad_modernbert_input(
                        inputs=inputs_embeds, attention_mask=attention_mask, position_ids=position_ids, labels=None
                    )
        
        outputs = self.albert(
            input_ids=input_ids,
            attention_mask=attention_mask,
            # sliding_window_mask=sliding_window_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            indices=indices,
            cu_seqlens=cu_seqlens,
            max_seqlen=max_seqlen,
            batch_size=batch_size,
            seq_len=seq_len,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        
        last_hidden_state = outputs[0]
        if FLASH_ATTN_AVAILABLE:
            last_hidden_state = _pad_modernbert_output(inputs=last_hidden_state, indices=indices, batch=batch_size, seqlen=seq_len)

        # if self.config.classifier_pooling == "cls":
        # last_hidden_state = last_hidden_state[:, 0]
        # elif self.config.classifier_pooling == "mean":
        last_hidden_state = (last_hidden_state * attention_mask.unsqueeze(-1)).sum(dim=1) / attention_mask.sum(
            dim=1, keepdim=True
        )

        pooled_output = self.head(last_hidden_state)
        pooled_output = self.drop(pooled_output)
        logits = self.classifier(pooled_output)

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

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.num_labels == 1:
                    loss = loss_fct(logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)

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

        return SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

class ModernALBERTForQuestionAnswering(ts.PreTrainedModel):
    config_class = ModernALBERTConfig

    def __init__(self, config: ModernALBERTConfig):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.albert = ModernALBERTModel(config)
        self.head = ModernAlbertPredictionHead(config)
        self.drop = torch.nn.Dropout(0.0)
        self.classifier_head = nn.Linear(config.embedding_size, config.num_labels)

        self.post_init()

    def forward(
        self,
        input_ids: Optional[torch.Tensor],
        attention_mask: Optional[torch.Tensor] = None,
        sliding_window_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.Tensor] = None,
        start_positions: Optional[torch.Tensor] = None,
        end_positions: Optional[torch.Tensor] = None,
        indices: Optional[torch.Tensor] = None,
        cu_seqlens: Optional[torch.Tensor] = None,
        max_seqlen: Optional[int] = None,
        batch_size: Optional[int] = None,
        seq_len: Optional[int] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        **kwargs,
    ) -> Union[tuple[torch.Tensor], QuestionAnsweringModelOutput]:

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        # self._maybe_set_compile()

        outputs = self.albert(
            input_ids,
            attention_mask=attention_mask,
            sliding_window_mask=sliding_window_mask,
            position_ids=position_ids,
            indices=indices,
            cu_seqlens=cu_seqlens,
            max_seqlen=max_seqlen,
            batch_size=batch_size,
            seq_len=seq_len,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        last_hidden_state = outputs[0]

        last_hidden_state = self.head(last_hidden_state)
        last_hidden_state = self.drop(last_hidden_state)
        logits = self.classifier_head(last_hidden_state)

        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1).contiguous()
        end_logits = end_logits.squeeze(-1).contiguous()

        loss = None
        if start_positions is not None and end_positions is not None:
            loss = self.loss_function(start_logits, end_logits, start_positions, end_positions, **kwargs)

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

        return QuestionAnsweringModelOutput(
            loss=loss,
            start_logits=start_logits,
            end_logits=end_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )