modeling_neuroblast.py

import torch
import math
from torch import nn
import torch.nn.functional as F
from transformers import PreTrainedModel, GenerationMixin
from transformers.cache_utils import DynamicCache, Cache
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
from transformers.cache_utils import Cache, DynamicCache, StaticCache
from transformers.utils import logging
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
)
from transformers.activations import ACT2FN
from typing import Optional, Tuple, Union, List
from .configuration_neuroblast import NeuroBLASTConfig

CLAMP_VALUE = 1e5

logger = logging.get_logger(__name__)

def apply_gradient_scaling(
    tensor: torch.Tensor, scale: float, enabled: bool = True
) -> torch.Tensor:
    """
    Apply gradient scaling to a tensor.
    This scales the gradients during backward pass while keeping forward pass unchanged.
    """
    if not enabled or scale == 1.0 or not tensor.requires_grad:
        return tensor

    # Use a custom autograd function for gradient scaling
    class GradientScale(torch.autograd.Function):
        @staticmethod
        def forward(ctx, input_tensor, scale_factor):
            ctx.scale = scale_factor
            return input_tensor.clone()

        @staticmethod
        def backward(ctx, grad_output):
            if grad_output is None:
                return None, None
            return grad_output * ctx.scale, None

    return GradientScale.apply(tensor, scale)


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,
    min_dtype: float,
    cache_position: torch.Tensor,
    batch_size: int,
):
    if attention_mask is not None and attention_mask.dim() == 4:
        causal_mask = attention_mask
    else:
        causal_mask = torch.full(
            (sequence_length, target_length),
            fill_value=min_dtype,
            dtype=dtype,
            device=device,
        )
        if sequence_length != 1:
            causal_mask = torch.triu(causal_mask, diagonal=1)
        causal_mask *= torch.arange(
            target_length, device=device
        ) > cache_position.reshape(-1, 1)
        causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
        if attention_mask is not None:
            causal_mask = (
                causal_mask.clone()
            )
            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


# --- RoPE Implementation (using HF LlamaRotaryEmbedding) ---


class LlamaRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        LlamaRMSNorm 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}"


class NeuroBLASTRotaryEmbedding(nn.Module):
    """
        Rotary Positional Embedding for NeuroBLAST model.        
        Source: LlamaRotaryEmbedding
    """
    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, dtype=torch.float32).to(device) / self.dim)
        )
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self._set_cos_sin_cache(
            seq_len=max_position_embeddings, device="cpu", dtype=torch.float32
        )

    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.einsum("i,j->ij", t, self.inv_freq)
        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):
        if seq_len > self.max_seq_len_cached:
            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
        elif self.cos_cached.device != x.device or self.cos_cached.dtype != x.dtype:
            self.cos_cached = self.cos_cached.to(device=x.device, dtype=x.dtype)
            self.sin_cached = self.sin_cached.to(device=x.device, dtype=x.dtype)

        return (
            self.cos_cached[:seq_len],
            self.sin_cached[:seq_len],
        )


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


def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
    """ Applies rotary positional embeddings to query and key tensors."""
    cos = cos[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
    sin = sin[position_ids].unsqueeze(1)  # [bs, 1, seq_len, dim]
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


# Overload for Cross Attention where only query is rotated
def apply_rotary_pos_emb_single(q, cos, sin, position_ids):
    """ Applies rotary positional embeddings to query tensor. """
    cos = cos[position_ids].unsqueeze(1)  # [1, 1, seq_len, dim]
    sin = sin[position_ids].unsqueeze(1)  # [1, 1, seq_len, dim]
    q_embed = (q * cos) + (rotate_half(q) * sin)
    return q_embed


class SwiGLUMLP(nn.Module):
    """SwiGLU MLP block"""

    def __init__(self, hidden_size, config: NeuroBLASTConfig, dropout):
        super().__init__()
        intermediate_size = getattr(config, "intermediate_size", int(hidden_size * 2.5))
        self.init_std = getattr(config, "initializer_range", 0.02)
        self.clamp_value = config.clamp_value

        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=True)
        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
        self.act_fn = nn.SiLU()
        self.dropout = nn.Dropout(dropout)

        if config.num_experts is not None:
            self.experts = NeuroBLASTSparseMoeBlock(config)

        # Re-enable scaled initialization
        with torch.no_grad():
            # Scale down initial weights in the up/gate projections
            self.gate_proj.weight.data.normal_(
                mean=0.0,
                std=self.init_std / math.sqrt(hidden_size),  # Scale by input dim
            )
            if self.gate_proj.bias is not None:
                self.gate_proj.bias.data.zero_()

            self.up_proj.weight.data.normal_(
                mean=0.0, std=self.init_std / math.sqrt(hidden_size)
            )  # Scale by input dim
            # Scale down initial weights in the down projection even further
            self.down_proj.weight.data.normal_(
                mean=0.0,
                std=self.init_std
                / math.sqrt(intermediate_size),  # Scale by intermediate dim
            )

    def forward(self, x):
        gated_x = self.gate_proj(x)
        activated_x = self.act_fn(gated_x)
        up_projected_x = self.up_proj(x)

        intermediate_activation = activated_x * up_projected_x

        # Clamp the intermediate activation before down_proj
        clamp_value = self.clamp_value

        intermediate_activation = torch.clamp(
            intermediate_activation, min=-clamp_value, max=clamp_value
        )
        intermediate_activation = torch.nan_to_num(
            intermediate_activation
        )  # Safeguard against NaNs

        y = self.down_proj(intermediate_activation)
        y = self.dropout(y)

        if hasattr(self, "experts"):
            z = self.experts(y)

            y = y + z

        return y


class NeuroBLASTMoeMLP(nn.Module):
    """ Source: Qwen3MoeMLP """
    def __init__(self, config, intermediate_size=None):
        super().__init__()
        self.config = config
        self.clamp_value = config.clamp_value
        self.hidden_size = config.hidden_size
        self.intermediate_size = (
            intermediate_size
            if intermediate_size is not None
            else 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]

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

        down_proj = torch.clamp(down_proj, min=-self.clamp_value, max=self.clamp_value)
        down_proj = torch.nan_to_num(down_proj)  # Safeguard against NaNs

        return down_proj


class NeuroBLASTSparseMoeBlock(nn.Module):
    """ Source: Qwen3SparseMoeBlock """
    def __init__(self, config):
        super().__init__()
        self.num_experts = config.num_experts
        self.top_k = config.num_experts_per_tok
        self.norm_topk_prob = config.norm_topk_prob

        # gating
        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
        self.experts = nn.ModuleList(
            [NeuroBLASTMoeMLP(config) for _ in range(self.num_experts)]
        )

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

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

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

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

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

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

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


class NeuroBLASTRouterBlock(nn.Module):
    """ Memory router; overcomplicated due to backward compatibility """
    def __init__(
        self,
        config,
        hidden_size,
    ):
        super().__init__()
        self.num_experts = 2
        self.top_k = 1
        self.norm_topk_prob = config.norm_topk_prob

        # gating
        self.gate = nn.Linear(hidden_size, self.num_experts, bias=False)

    def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        """ """
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
        # router_logits: (batch * sequence_length, n_experts)
        router_logits = self.gate(hidden_states)

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

        return routing_weights, selected_experts


class SelfAttention(torch.nn.Module):
    def __init__(
        self,
        config: NeuroBLASTConfig,
        hidden_size: int,
        is_causal: bool = False,
        layer_idx: Optional[int] = None,
    ):
        super().__init__()
        self.is_causal = is_causal
        self.dropout_p = config.dropout  # Will apply based on self.training
        self.layer_idx = layer_idx
        self.hidden_size = hidden_size
        self.intermediate_size = config.kv_dim
        self.use_flash_attn = config.use_flash_attn
        # Allow overriding num_heads, default to config.num_attention_heads
        self.num_heads = getattr(
            config, f"num_heads_{layer_idx}", config.num_attention_heads
        )
        self.head_dim = self.intermediate_size // self.num_heads

        if (self.head_dim * self.num_heads) != self.intermediate_size:
            raise ValueError(
                f"Layer {self.layer_idx}: hidden_size ({self.intermediate_size}) must be divisible by num_heads ({self.num_heads})"
            )

        self.qkv_proj = nn.Linear(
            self.hidden_size, self.intermediate_size * 3, bias=True
        )
        self.o_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.dropout = nn.Dropout(config.dropout)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Tuple[torch.Tensor, torch.Tensor],  # (cos, sin)
        attention_mask: Optional[
            torch.Tensor
        ],  # Not directly used by flash_attn causal
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        position_ids: Optional[torch.LongTensor] = None,
    ):
        batch_size, seq_len, _ = hidden_states.shape
        dropout_p = self.dropout_p if self.training else 0.0

        qkv = self.qkv_proj(hidden_states)
        query_states, key_states, value_states = qkv.chunk(3, dim=-1)

        query_states = query_states.view(
            batch_size, seq_len, self.num_heads, self.head_dim
        ).transpose(1, 2)
        key_states = key_states.view(
            batch_size, seq_len, self.num_heads, self.head_dim
        ).transpose(1, 2)
        value_states = value_states.view(
            batch_size, seq_len, self.num_heads, self.head_dim
        ).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(
            query_states, key_states, cos, sin, position_ids
        )

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            # When using Cache object, updating happens in-place.
            key_states, value_states = past_key_value.update(
                key_states, value_states, self.layer_idx, cache_kwargs
            )

        if self.use_flash_attn:
            causal_mask = attention_mask
            if attention_mask is not None:
                if attention_mask.dim() == 4:
                    causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
                elif attention_mask.dim() == 2:
                    causal_mask = attention_mask

                if causal_mask.dtype not in [
                    torch.bool,
                    torch.float16,
                    torch.float32,
                    torch.bfloat16,
                ]:
                    causal_mask = causal_mask.to(query_states.dtype)

            is_causal = (
                True if causal_mask is None and query_states.shape[-2] > 1 else False
            )

            attn_output = F.scaled_dot_product_attention(
                query_states,
                key_states,
                value_states,
                attn_mask=causal_mask,
                dropout_p=dropout_p,
                enable_gqa=False,
                scale=self.head_dim**-0.5,
                is_causal=is_causal,
            )
            attn_weights = None
        else:
            attn_weights = torch.matmul(query_states, key_states.transpose(-1, -2)) / (
                self.head_dim**0.5
            )

            if attention_mask is not None:
                attn_weights = attn_weights + attention_mask
            elif self.is_causal and seq_len > 1:
                causal_mask = torch.triu(
                    torch.ones(
                        (seq_len, key_states.shape[2]),
                        dtype=torch.bool,
                        device=query_states.device,
                    ),
                    diagonal=1,
                )
                attn_weights = attn_weights.masked_fill(
                    causal_mask.unsqueeze(0).unsqueeze(0), float("-inf")
                )

            attn_weights = F.softmax(attn_weights, dim=-1)
            attn_weights_dropped = F.dropout(
                attn_weights, p=dropout_p, training=self.training
            )
            attn_output = torch.matmul(attn_weights_dropped, value_states)

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.view(batch_size, seq_len, -1)
        attn_output = self.o_proj(attn_output)
        attn_output = self.dropout(attn_output)

        if output_attentions:
            logger.warning(
                f"Layer {self.layer_idx}: Flash Attention does not return attention weights."
            )

        outputs = (attn_output,)
        if output_attentions:
            outputs += (attn_weights,)
        if use_cache:
            outputs += (past_key_value,)  # Return the cache object

        return outputs


class CrossAttention(torch.nn.Module):
    def __init__(
        self,
        config: NeuroBLASTConfig,
        query_dim: int,
        kv_dim: int,
        layer_idx: int,
        is_causal: bool = True,
    ):
        super().__init__()
        self.dropout_p = config.dropout
        self.layer_idx = layer_idx
        self.query_dim = query_dim
        self.kv_dim = kv_dim
        self.is_causal = is_causal

        self.num_heads = config.num_attention_heads
        self.head_dim = self.kv_dim // self.num_heads
        self.kv_head_dim = (
            self.kv_dim // self.num_heads
        )

        if (self.head_dim * self.num_heads) != self.kv_dim:
            raise ValueError(
                f"CrossAttn {layer_idx}: query_dim ({self.kv_dim}) must be divisible by num_heads ({self.num_heads})"
            )
        if (self.kv_head_dim * self.num_heads) != self.kv_dim:
            raise ValueError(
                f"CrossAttn {layer_idx}: kv_dim ({kv_dim}) must be divisible by num_heads ({self.num_heads})"
            )

        self.q_proj = nn.Linear(self.query_dim, self.kv_dim, bias=True)
        self.k_proj = nn.Linear(self.query_dim, self.kv_dim, bias=True)
        self.v_proj = nn.Linear(self.query_dim, self.kv_dim, bias=True)
        self.o_proj = nn.Linear(self.kv_dim, self.query_dim, bias=False)
        self.dropout = nn.Dropout(config.dropout)

        self.use_flash_attn = hasattr(
            F, "scaled_dot_product_attention"
        )

    def forward(
        self,
        query_states: torch.Tensor,
        kv_states: torch.Tensor,
        position_embeddings: Tuple[
            torch.Tensor, torch.Tensor
        ],
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = False,
        position_ids: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = False,
    ):
        batch_size, q_seq_len, _ = query_states.shape
        kv_seq_len = kv_states.shape[1]
        dropout_p = self.dropout_p if self.training else 0.0

        query = self.q_proj(query_states)
        key = self.k_proj(kv_states)
        value = self.v_proj(kv_states)

        query = query.view(
            batch_size, q_seq_len, self.num_heads, self.head_dim
        ).transpose(1, 2)
        
        cos, sin = position_embeddings
        query = apply_rotary_pos_emb_single(query, cos, sin, position_ids)

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key, value = past_key_value.update(
                key, value, self.layer_idx or 0, cache_kwargs
            )
            kv_seq_len = key.shape[1]

        key = key.view(
            batch_size, kv_seq_len, self.num_heads, self.kv_head_dim
        ).transpose(1, 2)
        value = value.view(
            batch_size, kv_seq_len, self.num_heads, self.kv_head_dim
        ).transpose(1, 2)

        sdpa_attn_mask = attention_mask

        if self.use_flash_attn:
            is_causal = True if sdpa_attn_mask is None and q_seq_len > 1 else False
            attn_output = F.scaled_dot_product_attention(
                query,
                key,
                value,
                attn_mask=sdpa_attn_mask if not is_causal else None,
                dropout_p=dropout_p,
                is_causal=is_causal,
                enable_gqa=False,
                scale=self.head_dim**-0.5,
            )
            attn_weights = None
        else:
            attn_weights = torch.matmul(query, key.transpose(-1, -2)) / (
                self.head_dim**0.5
            )

            if sdpa_attn_mask is not None:
                attn_weights = attn_weights + sdpa_attn_mask
            elif self.is_causal and q_seq_len > 1:
                causal_mask = torch.triu(
                    torch.ones(
                        (q_seq_len, kv_seq_len), dtype=torch.bool, device=query.device
                    ),
                    diagonal=1,
                )
                attn_weights = attn_weights.masked_fill(
                    causal_mask.unsqueeze(0).unsqueeze(0), float("-inf")
                )

            attn_weights = F.softmax(attn_weights, dim=-1)
            attn_weights_dropped = F.dropout(
                attn_weights, p=dropout_p, training=self.training
            )
            attn_output = torch.matmul(attn_weights_dropped, value)

        attn_output = attn_output.transpose(
            1, 2
        ).contiguous()
        attn_output = attn_output.reshape(batch_size, q_seq_len, self.kv_dim)

        attn_output = self.o_proj(attn_output)
        attn_output = self.dropout(attn_output)

        outputs = (attn_output,)
        if output_attentions:
            outputs += (attn_weights,)

        if use_cache:
            outputs += (past_key_value,)

        return outputs


class AttentionBlock(torch.nn.Module):
    """Modified Attention Block with Pre-Norm and choice of Self/Cross Attention & MLP"""

    def __init__(
        self,
        config: NeuroBLASTConfig,
        hidden_size: int,
        attention_module: nn.Module,
        mlp_module: nn.Module,
        is_cross_attention: bool = False,
        layer_idx: int = 0,
        precomputed_total_layers: Optional[int] = None,
    ):
        super().__init__()
        self.hidden_size = hidden_size
        self.config = config
        self.layer_idx = layer_idx
        self.is_cross_attention = is_cross_attention

        self.input_layernorm = nn.LayerNorm(
            hidden_size, eps=getattr(config, "rms_norm_eps", 1e-5)
        )
        self.attention = attention_module

        self.post_attention_layernorm = nn.LayerNorm(
            hidden_size, eps=getattr(config, "rms_norm_eps", 1e-5)
        )
        self.mlp = mlp_module

        if self.config.use_zero_memory and (
            self.config.zero_memory_layers is None
            or self.layer_idx in self.config.zero_memory_layers
        ):
            self.router = NeuroBLASTRouterBlock(config, hidden_size)
            self.memory = NeuroBLASTMemory(
                config,
                hidden_size=hidden_size,
                layer_idx=layer_idx,
                precomputed_total_layers=precomputed_total_layers,
            )

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        position_ids: Optional[torch.LongTensor],
        kv_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
        previous_states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
    ):
        residual = hidden_states
        hidden_states = torch.nan_to_num(hidden_states)
        normed_hidden_states = self.input_layernorm(hidden_states)

        if self.is_cross_attention:
            if kv_states is None:
                raise ValueError("kv_states must be provided for CrossAttention")
            attn_outputs = self.attention(
                query_states=normed_hidden_states,
                kv_states=kv_states,
                past_key_value=past_key_value,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                attention_mask=attention_mask,
                output_attentions=output_attentions,
                position_ids=position_ids,
                use_cache=use_cache,
            )
        else:
            attn_outputs = self.attention(
                normed_hidden_states,
                position_embeddings=position_embeddings,
                attention_mask=attention_mask,
                past_key_value=past_key_value,
                cache_position=cache_position,
                output_attentions=output_attentions,
                use_cache=use_cache,
                position_ids=position_ids,
            )
        attn_output = attn_outputs[0]
        past_key_value = attn_outputs[-1] if use_cache else None

        hidden_states = residual + attn_output
        hidden_states = torch.nan_to_num(hidden_states)

        residual = hidden_states

        normed_hidden_states = self.post_attention_layernorm(hidden_states)

        mlp_output = self.mlp(normed_hidden_states)

        hidden_states = residual + mlp_output

        hidden_states = torch.nan_to_num(hidden_states)

        if self.config.use_zero_memory and (
            self.config.zero_memory_layers is None
            or self.layer_idx in self.config.zero_memory_layers
        ):
            routing_weights, selected_experts = self.router(hidden_states)

            residual = hidden_states

            hidden_states, (hx, cx), past_key_value = self.memory(
                hidden_states,
                previous_states,
                past_key_value=past_key_value,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                attention_mask=attention_mask,
                output_attentions=output_attentions,
                position_ids=position_ids,
                use_cache=use_cache,
            )

            hidden_states = torch.nan_to_num(hidden_states)
            hx = torch.nan_to_num(hx)
            cx = torch.nan_to_num(cx)

            hidden_states = hidden_states * routing_weights.reshape(
                hidden_states.shape[:-1]
            ).unsqueeze(-1)

            hidden_states = residual + self.config.zero_memory_alpha * hidden_states
            hidden_states = torch.nan_to_num(hidden_states)

        outputs = (hidden_states,) + attn_outputs[1:]

        if self.config.use_zero_memory and (
            self.config.zero_memory_layers is None
            or self.layer_idx in self.config.zero_memory_layers
        ):
            outputs += ((hx, cx),)
        else:
            outputs += (previous_states,)

        if use_cache:
            outputs += (past_key_value,)

        return outputs


class NeuroBLASTMemory(nn.Module):
    def __init__(
        self,
        config: NeuroBLASTConfig,
        hidden_size: int = 256,
        num_heads: int = 4,
        scale_factor: int = 4,
        layer_idx: int = 0,
        with_hx: bool = True,
        precomputed_total_layers: Optional[int] = None,
        *args,
        **kwargs,
    ):
        super().__init__(*args, **kwargs)

        self.hidden_size = hidden_size
        self.num_heads = num_heads
        self.scale_factor = scale_factor
        self.clamp_value = config.clamp_value
        # Use precomputed_total_layers instead of hardcoded 100 for layer index shift
        layer_shift = (
            precomputed_total_layers if precomputed_total_layers is not None else 100
        )
        self.layer_idx = layer_idx + layer_shift
        self.with_hx = with_hx
        self.kv_dim = (
            config.kv_dim
        )

        self.scaled_dim = hidden_size * scale_factor
        self.head_dim = self.kv_dim // config.num_attention_heads
        self.num_heads = self.hidden_size // self.head_dim

        self.norm1 = nn.LayerNorm(hidden_size)

        if self.with_hx:
            self.lin1 = nn.Linear(self.hidden_size, self.scaled_dim)
        self.lin2 = nn.Linear(self.hidden_size, self.scaled_dim)
        self.lin3 = nn.Linear(self.scaled_dim, self.hidden_size)

        self.lin4 = nn.Linear(self.hidden_size, self.scaled_dim)
        self.lin5 = nn.Linear(self.scaled_dim, self.hidden_size)

        if self.with_hx:
            self.lin6 = nn.Linear(self.scaled_dim, self.hidden_size)

        self.gate1 = nn.Linear(self.scaled_dim, self.scaled_dim)
        self.act1 = nn.SiLU()

        self.gate2 = nn.Linear(self.scaled_dim, self.scaled_dim)
        self.act2 = nn.SiLU()

        self.last_token_reg = nn.Linear(self.hidden_size, self.hidden_size, bias=True)
        self.prev_tokens_reg = nn.Linear(self.hidden_size, self.hidden_size, bias=True)

        self.norm2 = nn.LayerNorm(hidden_size)
        self.dropout = nn.Dropout(config.dropout)

    def forward(
        self,
        x: torch.Tensor,
        previous_state: tuple[torch.Tensor, torch.Tensor],
        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        position_ids: Optional[torch.LongTensor],
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = False,
    ):
        hx, cx = previous_state

        b, s, d = x.size()

        x = torch.nan_to_num(x)
        norm_x = self.norm1(x)

        norm_x = norm_x.view(b, s, self.num_heads, self.head_dim).transpose(1, 2)

        cos, sin = position_embeddings
        norm_x = apply_rotary_pos_emb_single(norm_x, cos, sin, position_ids)

        kv_seq_len = s

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            norm_x, _ = past_key_value.update(
                norm_x,
                torch.zeros((b, 1, kv_seq_len, d)),
                self.layer_idx,
                cache_kwargs,
            )
            kv_seq_len = norm_x.shape[2]

        norm_x = norm_x.transpose(1, 2).contiguous()
        norm_x = norm_x.view(b, kv_seq_len, d)

        norm_x = torch.nan_to_num(norm_x)

        expanded_x = None

        shifted_x = torch.cat(
            [
                torch.zeros((b, 1, d), device=x.device, dtype=x.dtype),
                (norm_x[:, :-1].contiguous()),
            ],
            dim=1,
        ).contiguous()

        shifted_x = torch.nan_to_num(shifted_x)  # Replace NaNs with zeros

        prev_tokens_x = norm_x.cumsum(dim=1)

        prev_tokens_x = prev_tokens_x - shifted_x
        prev_tokens_x = torch.nan_to_num(prev_tokens_x)[:, -s:].contiguous()

        if self.with_hx:
            expanded_x = self.lin1(norm_x[:, -s:].contiguous())
            expanded_x = torch.nan_to_num(expanded_x)

        expanded_shifted_x = self.lin2(shifted_x[:, -s:].contiguous())

        expanded_shifted_x = torch.nan_to_num(expanded_shifted_x)

        gated_shifted_x = self.gate1(expanded_shifted_x)
        gated_shifted_x = self.act1(gated_shifted_x)
        gated_shifted_x = torch.clamp(
            gated_shifted_x, min=-self.clamp_value, max=self.clamp_value
        )

        gated_shifted_x = torch.nan_to_num(gated_shifted_x)

        collapsed_shifted_x = self.lin3(gated_shifted_x)
        collapsed_shifted_x = torch.nan_to_num(collapsed_shifted_x)

        prev_tokens_x = torch.nan_to_num(prev_tokens_x)

        expanded_prev_tokens_x = self.lin4(prev_tokens_x)
        expanded_prev_tokens_x = torch.nan_to_num(expanded_prev_tokens_x)

        gated_prev_tokens_x = self.gate2(expanded_prev_tokens_x)
        gated_prev_tokens_x = self.act2(gated_prev_tokens_x)
        gated_prev_tokens_x = torch.clamp(
            gated_prev_tokens_x, min=-self.clamp_value, max=self.clamp_value
        )

        gated_prev_tokens_x = torch.nan_to_num(gated_prev_tokens_x)

        collapsed_prev_tokens_x = self.lin5(gated_prev_tokens_x)
        collapsed_prev_tokens_x = torch.nan_to_num(collapsed_prev_tokens_x)

        if self.with_hx:
            weights = torch.softmax(expanded_x * expanded_shifted_x, dim=-1)

            expanded_x_attn = weights * expanded_x

            expanded_x_attn = torch.nan_to_num(expanded_x_attn)
            hx = hx + self.lin6(expanded_x_attn)
            hx = torch.nan_to_num(hx)

        if self.with_hx:
            x = torch.nan_to_num(x)
            hx = torch.nan_to_num(hx)
            collapsed_shifted_x = torch.nan_to_num(collapsed_shifted_x)
            collapsed_prev_tokens_x = torch.nan_to_num(collapsed_prev_tokens_x)
            output = x + (
                hx
                * (
                    self.last_token_reg(collapsed_shifted_x)
                    + self.prev_tokens_reg(collapsed_prev_tokens_x)
                )
            )
            output = torch.nan_to_num(output)

        else:
            output = (
                x
                + (
                    self.last_token_reg(collapsed_shifted_x)
                    + self.prev_tokens_reg(collapsed_prev_tokens_x)
                )[:, -s:].contiguous()
            )

        output = self.norm2(output)
        output = self.dropout(output)
        output = torch.nan_to_num(output)

        return (
            output,
            (hx, cx),
            past_key_value,
        )


class NeuroBLASTPreTrainedModel(PreTrainedModel):
    config_class = NeuroBLASTConfig
    base_model_prefix = "brain"
    supports_gradient_checkpointing = True
    _no_split_modules = []
    _keys_to_ignore_on_load_unexpected = [r"decoder\.version"]
    _supports_flash_attn_2 = False
    _supports_sdpa = True

    def _init_weights(self, module):
        std = getattr(self.config, "initializer_range", 0.02)
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm):
            if module.bias is not None:
                module.bias.data.zero_()
            module.weight.data.fill_(1.0)


class NeuroBLASTModel(NeuroBLASTPreTrainedModel):
    def __init__(self, config: NeuroBLASTConfig):
        super(NeuroBLASTModel, self).__init__(config)
        self.config = config
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(
            config.vocab_size,
            config.hidden_size,  # Using main hidden_size for embeddings now
            padding_idx=self.padding_idx,
        )
        self.rotary_emb = NeuroBLASTRotaryEmbedding(
            config.kv_dim // config.num_attention_heads,
            max_position_embeddings=config.max_position_embeddings,
            base=config.rope_theta,
        )
        self.dropout = nn.Dropout(config.dropout)

        self.assoc_to_sensory_pooler = nn.Sequential(
            nn.Linear(config.hidden_size, config.hidden_size),
            nn.Identity(), # Backward compatibility - previously LayerNorm, but we found that removing it improve generalization
            nn.GELU(),
            nn.LayerNorm(config.hidden_size),
        )
        self.assoc_to_motor_pooler = nn.Sequential(
            nn.Linear(config.hidden_size, config.hidden_size),
            nn.Identity(), # Backward compatibility
            nn.GELU(),
            nn.LayerNorm(config.hidden_size),
        )
        self.sensory_to_motor_pooler = nn.Sequential(
            nn.Linear(config.hidden_size, config.hidden_size),
            nn.Identity(),  # Backward compatibility
            nn.GELU(),
            nn.LayerNorm(config.hidden_size),
        )

        # --- Cortex Layers ---
        # Using generic AttentionBlock with specific Self/Cross Attention modules passed in
        total_layers = 0

        # Precompute total layers for Memory layer indexing before creating any layers
        precomputed_total_layers = (
            config.num_association_cortex_layers
            + config.num_sensory_cortex_layers * 2
            + config.num_motor_cortex_layers
            * 3
        )
        self.precomputed_total_layers = precomputed_total_layers
        config.precomputed_total_layers = precomputed_total_layers

        # 1. Association Cortex (Self-Attention)
        self.association_cortex = nn.ModuleList()
        for i in range(config.num_association_cortex_layers):
            layer_idx = total_layers + i
            print(f"Adding layer {layer_idx} to association cortex")
            self.association_cortex.append(
                AttentionBlock(
                    config,
                    config.hidden_size,  # Use main hidden_size
                    attention_module=SelfAttention(
                        config, config.hidden_size, is_causal=True, layer_idx=layer_idx
                    ),
                    mlp_module=(
                        NeuroBLASTSparseMoeBlock(
                            config,
                        )
                        if config.num_experts
                        else NeuroBLASTMoeMLP(config)
                    ),
                    is_cross_attention=False,
                    layer_idx=layer_idx,
                    precomputed_total_layers=precomputed_total_layers,
                )
            )

        total_layers += config.num_association_cortex_layers
        # 2. Sensory Cortex (Self-Attention + Cross-Attention to Association)
        self.sensory_self_attn_layers = nn.ModuleList()
        self.sensory_cross_attn_layers = (
            nn.ModuleList()
        )  # One cross-attn per self-attn layer
        for i in range(config.num_sensory_cortex_layers):
            layer_idx = total_layers + i
            print(f"Adding layer {layer_idx} to sensory cortex")
            self.sensory_self_attn_layers.append(
                AttentionBlock(
                    config,
                    config.hidden_size,
                    attention_module=SelfAttention(
                        config,
                        config.hidden_size,
                        is_causal=True,
                        layer_idx=layer_idx,
                    ),
                    mlp_module=(
                        NeuroBLASTSparseMoeBlock(
                            config,
                        )
                        if config.num_experts
                        else NeuroBLASTMoeMLP(config)
                    ),
                    is_cross_attention=False,
                    layer_idx=layer_idx,
                    precomputed_total_layers=precomputed_total_layers,
                )
            )

        total_layers += config.num_sensory_cortex_layers
        for i in range(config.num_sensory_cortex_layers):
            layer_idx = total_layers + i
            print(f"Adding layer {layer_idx} to sensory cross-attention")
            # Add Cross-Attention layer: Sensory queries, Association is K/V source
            self.sensory_cross_attn_layers.append(
                AttentionBlock(
                    config,
                    config.hidden_size,  # Query Dim
                    attention_module=CrossAttention(
                        config,
                        query_dim=config.hidden_size,
                        kv_dim=config.kv_dim,  # Assoc output dim
                        layer_idx=layer_idx,
                    ),
                    mlp_module=(
                        NeuroBLASTSparseMoeBlock(
                            config,
                        )
                        if config.num_experts
                        else NeuroBLASTMoeMLP(config)
                    ),
                    is_cross_attention=True,
                    layer_idx=layer_idx,
                    precomputed_total_layers=precomputed_total_layers,
                )
            )

        total_layers += config.num_sensory_cortex_layers

        # 3. Motor Cortex (Self-Attention + Cross-Attention to Sensory + Cross-Attention to Association)
        self.motor_self_attn_layers = nn.ModuleList()
        self.motor_cross_sensory_layers = nn.ModuleList()
        self.motor_cross_assoc_layers = nn.ModuleList()
        for i in range(config.num_motor_cortex_layers):
            layer_idx = total_layers + i
            print(f"Adding layer {layer_idx} to motor cortex")
            self.motor_self_attn_layers.append(
                AttentionBlock(
                    config,
                    config.hidden_size,
                    attention_module=SelfAttention(
                        config,
                        config.hidden_size,
                        is_causal=True,
                        layer_idx=layer_idx,
                    ),
                    mlp_module=(
                        NeuroBLASTSparseMoeBlock(
                            config,
                        )
                        if config.num_experts
                        else NeuroBLASTMoeMLP(config)
                    ),
                    is_cross_attention=False,
                    layer_idx=layer_idx,
                    precomputed_total_layers=precomputed_total_layers,
                )
            )

        total_layers += config.num_motor_cortex_layers
        for i in range(config.num_motor_cortex_layers):
            layer_idx = total_layers + i
            print(f"Adding layer {layer_idx} to motor cross-sensory")
            # Cross-Attend to Sensory Output
            self.motor_cross_sensory_layers.append(
                AttentionBlock(
                    config,
                    config.hidden_size,  # Query Dim
                    attention_module=CrossAttention(
                        config,
                        query_dim=config.hidden_size,
                        kv_dim=config.kv_dim,  # Sensory output dim
                        layer_idx=layer_idx,
                    ),
                    mlp_module=(
                        NeuroBLASTSparseMoeBlock(
                            config,
                        )
                        if config.num_experts
                        else NeuroBLASTMoeMLP(config)
                    ),
                    is_cross_attention=True,
                    layer_idx=layer_idx,
                    precomputed_total_layers=precomputed_total_layers,
                )
            )

        total_layers += config.num_motor_cortex_layers
        for i in range(config.num_motor_cortex_layers):
            layer_idx = total_layers + i
            print(f"Adding layer {layer_idx} to motor cross-association")
            # Cross-Attend to Association Output
            self.motor_cross_assoc_layers.append(
                AttentionBlock(
                    config,
                    config.hidden_size,  # Query Dim
                    attention_module=CrossAttention(
                        config,
                        query_dim=config.hidden_size,
                        kv_dim=config.kv_dim,  # Assoc output dim
                        layer_idx=layer_idx,
                    ),
                    mlp_module=(
                        NeuroBLASTSparseMoeBlock(
                            config,
                        )
                        if config.num_experts
                        else NeuroBLASTMoeMLP(config)
                    ),
                    is_cross_attention=True,
                    layer_idx=layer_idx,
                    precomputed_total_layers=precomputed_total_layers,
                )
            )

        total_layers += config.num_motor_cortex_layers

        # Initialize more conservatively to prevent strong gradient flow initially
        self.sensory_cross_assoc_gate = NeuroBLASTMoeMLP(
            config,
        )
        self.motor_cross_sensory_gate = NeuroBLASTMoeMLP(
            config,
        )
        self.motor_cross_assoc_gate = NeuroBLASTMoeMLP(
            config,
        )

        # Final normalization before output head
        self.norm = nn.LayerNorm(
            config.hidden_size, eps=getattr(config, "rms_norm_eps", 1e-5)
        )

        self.gradient_checkpointing = False
        self.post_init()

    def forward(
        self,
        input_ids: torch.LongTensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[
            Cache
        ] = 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,
    ):
        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
        # use_cache = False
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("Specify either input_ids or inputs_embeds")
        batch_size, seq_length = (
            input_ids.shape if input_ids is not None else inputs_embeds.shape[:2]
        )

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` incompatible with gradient checkpointing. Setting `use_cache=False`"
                )
                use_cache = False

            if not any(param.requires_grad for param in self.parameters()):
                logger.warning_once(
                    "No parameters require gradients. Disabling gradient checkpointing to avoid warnings."
                )
                self.gradient_checkpointing = False

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

        past_key_values_length = 0
        if use_cache:
            if not isinstance(past_key_values, Cache):
                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
            past_key_values_length = past_key_values.get_seq_length()

        if cache_position is None:
            past_seen_tokens = (
                past_key_values.get_seq_length() if past_key_values is not None else 0
            )
            cache_position = torch.arange(
                past_seen_tokens,
                past_seen_tokens + inputs_embeds.shape[1],
                device=inputs_embeds.device,
            )

        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

        causal_mask = self._update_causal_mask(
            attention_mask,
            inputs_embeds,
            cache_position,
            past_key_values,
            output_attentions,
        )

        hidden_states = inputs_embeds

        cos, sin = self.rotary_emb(
            hidden_states, seq_len=seq_length + past_key_values_length
        )
        position_embeddings = (cos, sin)

        all_hidden_states = () if output_hidden_states else None
        all_attentions = (
            () if output_attentions else None
        )
        next_decoder_cache = (
            past_key_values if use_cache else None
        )

        if self.config.use_zero_memory:
            hx = torch.ones(
                (batch_size, seq_length, hidden_states.size(-1)),
                device=hidden_states.device,
                dtype=hidden_states.dtype,
            )
            cx = torch.ones(
                (batch_size, seq_length, hidden_states.size(-1)),
                device=hidden_states.device,
                dtype=hidden_states.dtype,
            )

            if self.training:
                hx.requires_grad_()
                cx.requires_grad_()
        else:
            hx = None
            cx = None

        # 1. Association Cortex (Self-Attention)
        assoc_output = hidden_states
        for i, layer in enumerate(self.association_cortex):
            if output_hidden_states:
                all_hidden_states += (assoc_output,)
            if self.gradient_checkpointing and self.training:
                outputs = torch.utils.checkpoint.checkpoint(
                    layer,
                    assoc_output,
                    position_embeddings,
                    causal_mask,
                    position_ids,
                    None,
                    next_decoder_cache,
                    cache_position,
                    output_attentions,
                    use_cache,
                    (hx, cx),
                )
            else:
                outputs = layer(
                    assoc_output,
                    position_embeddings,
                    causal_mask,
                    position_ids,
                    kv_states=None,
                    past_key_value=next_decoder_cache,
                    cache_position=cache_position,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    previous_states=(hx, cx),
                )
            assoc_output = outputs[0]

            hx, cx = outputs[-1 if not use_cache else -2]

            if output_attentions:
                all_attentions += (outputs[1],)

            if use_cache:
                next_decoder_cache = outputs[-1]
            else:
                next_decoder_cache = None

        sensory_state = self.assoc_to_sensory_pooler(assoc_output)

        sensory_state = apply_gradient_scaling(
            sensory_state,
            self.config.association_gradient_scale,
            self.config.gradient_scaling_enabled,
        )

        # 2. Sensory Cortex (Self-Attention + Cross-Attention to Association)
        for i in range(self.config.num_sensory_cortex_layers):
            if output_hidden_states:
                all_hidden_states += (sensory_state,)

            self_attn_layer = self.sensory_self_attn_layers[i]
            if self.gradient_checkpointing and self.training:
                outputs_self = torch.utils.checkpoint.checkpoint(
                    self_attn_layer,
                    sensory_state,
                    position_embeddings,
                    causal_mask,
                    position_ids,
                    None,
                    next_decoder_cache,
                    cache_position,
                    output_attentions,
                    use_cache,
                    (hx, cx),
                )
            else:
                outputs_self = self_attn_layer(
                    sensory_state,
                    position_embeddings,
                    causal_mask,
                    position_ids,
                    kv_states=None,
                    past_key_value=next_decoder_cache,
                    cache_position=cache_position,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    previous_states=(hx, cx),
                )
            sensory_state = outputs_self[0]
            hx, cx = outputs_self[-1 if not use_cache else -2]

            if output_attentions:
                all_attentions += (outputs_self[1],)

            if use_cache:
                next_decoder_cache = outputs_self[-1]
            else:
                next_decoder_cache = None

            cross_attn_layer = self.sensory_cross_attn_layers[i]
            
            cross_attn_causal_mask = None

            if causal_mask is not None:
                q_seq_len = sensory_state.size(1)
                kv_seq_len = assoc_output.size(1)

                cross_attn_causal_mask = torch.ones(
                    (batch_size, 1, q_seq_len, kv_seq_len),
                    device=hidden_states.device,
                    dtype=hidden_states.dtype,
                )

                causal_mask_upper = torch.triu(
                    torch.ones((q_seq_len, kv_seq_len), device=hidden_states.device),
                    diagonal=1,
                )

                cross_attn_causal_mask = cross_attn_causal_mask.masked_fill(
                    causal_mask_upper.unsqueeze(0).unsqueeze(0).bool(),
                    torch.finfo(hidden_states.dtype).min,
                )

            if self.gradient_checkpointing and self.training:
                outputs_cross = torch.utils.checkpoint.checkpoint(
                    cross_attn_layer,
                    sensory_state,
                    position_embeddings,
                    cross_attn_causal_mask,
                    position_ids,
                    assoc_output,
                    next_decoder_cache,
                    cache_position,
                    output_attentions,
                    use_cache,
                    (hx, cx),
                )
            else:
                outputs_cross = cross_attn_layer(
                    sensory_state,
                    position_embeddings,
                    past_key_value=next_decoder_cache,
                    cache_position=cache_position,
                    attention_mask=cross_attn_causal_mask,
                    position_ids=position_ids,
                    kv_states=apply_gradient_scaling(
                        assoc_output,
                        self.config.cross_attention_gradient_scale,
                        self.config.gradient_scaling_enabled,
                    ),
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    previous_states=(hx, cx),
                )
                
            cross_contribution = nn.functional.layer_norm(
                outputs_cross[0],
                normalized_shape=(self.config.hidden_size,),
                eps=getattr(self.config, "rms_norm_eps", 1e-5),
            )

            sensory_state = sensory_state + self.sensory_cross_assoc_gate(
                cross_contribution
            )

            hx, cx = outputs_cross[-1 if not use_cache else -2]
            if output_attentions:
                all_attentions += (outputs_cross[1],)

            if use_cache:
                next_decoder_cache = outputs_cross[-1]
            else:
                next_decoder_cache = None

        motor_state = self.sensory_to_motor_pooler(sensory_state)
        
        motor_state = apply_gradient_scaling(
            motor_state,
            self.config.sensory_gradient_scale,
            self.config.gradient_scaling_enabled,
        )

        motor_state_from_assoc = self.assoc_to_motor_pooler(assoc_output)
        
        motor_state_from_assoc = apply_gradient_scaling(
            motor_state_from_assoc,
            self.config.association_gradient_scale,
            self.config.gradient_scaling_enabled,
        )

        motor_state = motor_state + motor_state_from_assoc  # Combine pooled inputs

        # 3. Motor Cortex (Self + Cross-Sensory + Cross-Association)
        for i in range(self.config.num_motor_cortex_layers):
            if output_hidden_states:
                all_hidden_states += (motor_state,)

            self_attn_layer = self.motor_self_attn_layers[i]
            if self.gradient_checkpointing and self.training:
                outputs_self = torch.utils.checkpoint.checkpoint(
                    self_attn_layer,
                    motor_state,
                    position_embeddings,
                    causal_mask,
                    position_ids,
                    None,
                    next_decoder_cache,
                    cache_position,
                    output_attentions,
                    use_cache,
                    (hx, cx),
                )
            else:
                outputs_self = self_attn_layer(
                    motor_state,
                    position_embeddings,
                    causal_mask,
                    position_ids,
                    kv_states=None,
                    past_key_value=next_decoder_cache,
                    cache_position=cache_position,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                    previous_states=(hx, cx),
                )
            motor_state = outputs_self[0]
            hx, cx = outputs_self[-1 if not use_cache else -2]
            if output_attentions:
                all_attentions += (outputs_self[1],)

            if use_cache:
                next_decoder_cache = outputs_self[-1]
            else:
                next_decoder_cache = None

            cross_sensory_layer = self.motor_cross_sensory_layers[i]

            motor_cross_sensory_mask = None

            if causal_mask is not None:
                motor_q_seq_len = motor_state.size(1)
                sensory_kv_seq_len = sensory_state.size(
                    1
                ) 
                
                motor_cross_sensory_mask = torch.ones(
                    (batch_size, 1, motor_q_seq_len, sensory_kv_seq_len),
                    device=hidden_states.device,
                    dtype=hidden_states.dtype,
                )
                
                causal_mask_upper = torch.triu(
                    torch.ones(
                        (motor_q_seq_len, sensory_kv_seq_len),
                        device=hidden_states.device,
                    ),
                    diagonal=1,
                )

                motor_cross_sensory_mask = motor_cross_sensory_mask.masked_fill(
                    causal_mask_upper.unsqueeze(0).unsqueeze(0).bool(),
                    torch.finfo(hidden_states.dtype).min,
                )

            if self.gradient_checkpointing and self.training:
                outputs_cross_sensory = torch.utils.checkpoint.checkpoint(
                    cross_sensory_layer,
                    motor_state,
                    position_embeddings,
                    motor_cross_sensory_mask,
                    position_ids,
                    sensory_state,
                    next_decoder_cache,
                    cache_position,
                    output_attentions,
                    use_cache,
                    (hx, cx),
                )
            else:
                outputs_cross_sensory = cross_sensory_layer(
                    motor_state,
                    position_embeddings,
                    attention_mask=motor_cross_sensory_mask,
                    position_ids=position_ids,
                    kv_states=apply_gradient_scaling(
                        sensory_state,
                        self.config.cross_attention_gradient_scale,
                        self.config.gradient_scaling_enabled,
                    ),
                    output_attentions=output_attentions,
                    past_key_value=next_decoder_cache,
                    cache_position=cache_position,
                    use_cache=use_cache,
                    previous_states=(hx, cx),
                )
            motor_state = motor_state + self.motor_cross_sensory_gate(
                outputs_cross_sensory[0]
            )
            hx, cx = outputs_cross_sensory[-1 if not use_cache else -2]
            if output_attentions:
                all_attentions += (outputs_cross_sensory[1],)

            if use_cache:
                next_decoder_cache = outputs_cross_sensory[-1]
            else:
                next_decoder_cache = None

            cross_assoc_layer = self.motor_cross_assoc_layers[i]
            
            motor_cross_assoc_mask = None
            if causal_mask is not None:
                motor_q_seq_len = motor_state.size(1)
                assoc_kv_seq_len = assoc_output.size(
                    1
                )
                
                motor_cross_assoc_mask = torch.ones(
                    (batch_size, 1, motor_q_seq_len, assoc_kv_seq_len),
                    device=hidden_states.device,
                    dtype=hidden_states.dtype,
                )

                causal_mask_upper = torch.triu(
                    torch.ones(
                        (motor_q_seq_len, assoc_kv_seq_len), device=hidden_states.device
                    ),
                    diagonal=1,
                )

                motor_cross_assoc_mask = motor_cross_assoc_mask.masked_fill(
                    causal_mask_upper.unsqueeze(0).unsqueeze(0).bool(),
                    torch.finfo(hidden_states.dtype).min,
                )

            if self.gradient_checkpointing and self.training:
                outputs_cross_assoc = torch.utils.checkpoint.checkpoint(
                    cross_assoc_layer,
                    motor_state,
                    position_embeddings,
                    motor_cross_assoc_mask,
                    position_ids,
                    assoc_output,
                    next_decoder_cache,
                    cache_position,
                    output_attentions,
                    use_cache,
                    (hx, cx),
                )
            else:
                outputs_cross_assoc = cross_assoc_layer(
                    motor_state,
                    position_embeddings,
                    attention_mask=motor_cross_assoc_mask,
                    position_ids=position_ids,
                    kv_states=apply_gradient_scaling(
                        assoc_output,
                        self.config.cross_attention_gradient_scale,
                        self.config.gradient_scaling_enabled,
                    ),
                    output_attentions=output_attentions,
                    past_key_value=next_decoder_cache,
                    cache_position=cache_position,
                    use_cache=use_cache,
                    previous_states=(hx, cx),
                )
            motor_state = motor_state + self.motor_cross_assoc_gate(
                outputs_cross_assoc[0]
            )
            hx, cx = outputs_cross_assoc[-1 if not use_cache else -2]
            if output_attentions:
                all_attentions += (outputs_cross_assoc[1],)

            if use_cache:
                next_decoder_cache = outputs_cross_assoc[-1]
            else:
                next_decoder_cache = None

        final_output = self.norm(motor_state)

        if output_hidden_states:
            all_hidden_states += (final_output,) 

        if not return_dict:
            outputs_tuple = (final_output,)
            if use_cache:
                outputs_tuple += (next_decoder_cache,)
            if output_hidden_states:
                outputs_tuple += (all_hidden_states,)
            if output_attentions:
                outputs_tuple += (all_attentions,)
            return tuple(v for v in outputs_tuple if v is not None)

        return BaseModelOutputWithPast(
            last_hidden_state=final_output,
            past_key_values=next_decoder_cache,
            hidden_states=all_hidden_states,
            attentions=all_attentions,
        )

    def get_input_embeddings(self):
        return self.embed_tokens

    def set_input_embeddings(self, value):
        self.embed_tokens = value

    def _update_causal_mask(
        self,
        attention_mask: torch.Tensor,
        input_tensor: torch.Tensor,
        cache_position: torch.Tensor,
        past_key_values: Cache,
        output_attentions: bool,
    ):
        if self.config._attn_implementation == "flash_attention_2":
            if attention_mask is not None and 0.0 in attention_mask:
                return attention_mask
            return None

        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)

        if (
            self.config._attn_implementation == "sdpa"
            and not using_static_cache
            and not output_attentions
        ):
            if AttentionMaskConverter._ignore_causal_mask_sdpa(
                attention_mask,
                inputs_embeds=input_tensor,
                past_key_values_length=past_seen_tokens,
                is_training=self.training,
            ):
                return None

        dtype, device = input_tensor.dtype, input_tensor.device
        min_dtype = torch.finfo(dtype).min
        sequence_length = input_tensor.shape[1]
        if using_static_cache:
            target_length = (
                getattr(
                    past_key_values,
                    "get_max_length",
                    lambda: past_key_values.get_seq_length(),
                )()
                if hasattr(past_key_values, "get_seq_length")
                else sequence_length + past_seen_tokens
            )
        else:
            target_length = (
                attention_mask.shape[-1]
                if isinstance(attention_mask, torch.Tensor)
                else past_seen_tokens + sequence_length + 1
            )

        causal_mask = _prepare_4d_causal_attention_mask_with_cache_position(
            attention_mask,
            sequence_length=sequence_length,
            target_length=target_length,
            dtype=dtype,
            device=device,
            min_dtype=min_dtype,
            cache_position=cache_position,
            batch_size=input_tensor.shape[0],
        )

        if (
            self.config._attn_implementation == "sdpa"
            and attention_mask is not None
            and attention_mask.device.type == "cuda"
            and not output_attentions
        ):
            causal_mask = AttentionMaskConverter._unmask_unattended(
                causal_mask, min_dtype
            )

        return causal_mask


class NeuroBLASTForCausalLM(NeuroBLASTPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config: NeuroBLASTConfig):
        super().__init__(config)
        self.config = config
        self.model = NeuroBLASTModel(config)
        self.vocab_size = config.vocab_size  # Ensure vocab_size is accessible
        self.lm_head = torch.nn.Linear(config.hidden_size, self.vocab_size, bias=False)
        self.loss_steps = 0
        self.post_init()  # Initialize weights

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

    def set_input_embeddings(self, value):
        self.model.set_input_embeddings(value)

    def get_output_embeddings(self):
        return self.lm_head

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

    def tie_weights(self):
        if getattr(self.config, "tie_word_embeddings", False):
            output_embeddings = self.get_output_embeddings()
            input_embeddings = self.get_input_embeddings()
            output_embeddings.weight = input_embeddings.weight
            if getattr(output_embeddings, "bias", None) is not None:
                output_embeddings.bias.data.zero_()

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = 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,
        **loss_kwargs,
    ) -> Union[Tuple, CausalLMOutputWithPast]:

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

        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            cache_position=cache_position,
        )

        hidden_states = outputs[0]

        logits = self.lm_head(hidden_states)
        logits = logits.float()

        loss = None
        if labels is not None:
            loss = self.loss_function(
                logits=logits,
                labels=labels,
                vocab_size=self.config.vocab_size,
                **loss_kwargs,
            )

        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=hidden_states,
            attentions=outputs.attentions,
        )

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

    def set_input_embeddings(self, value):
        self.model.set_input_embeddings(value)

    def __str__(self):
        return f"NeuroBLASTForCausalLM(config={self.config})"