core/decoder.py

"""
Intelligent Tokenizer v6.2.0 - 6-Layer Decoder with Multi-Level Cross-Attention
Incorporates GPT-5 suggestions for KV cache optimization
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Dict, List, Optional, Tuple
import math


class KVCacheOptimizedAttention(nn.Module):
    """
    KV Cache Optimized Attention - GPT-5 suggestion
    16Q → 2K/V for 8x memory reduction
    """

    def __init__(self, hidden_dim: int = 1280, num_heads: int = 16, kv_compression: int = 8):
        super().__init__()
        self.hidden_dim = hidden_dim
        self.num_heads = num_heads
        self.kv_heads = max(2, num_heads // kv_compression)  # 16/8 = 2 KV heads
        self.head_dim = hidden_dim // num_heads  # 80

        # Query uses all heads
        self.q_proj = nn.Linear(hidden_dim, hidden_dim)  # 16 heads

        # Key/Value use fewer heads (GPT-5 suggestion)
        self.k_proj = nn.Linear(hidden_dim, self.kv_heads * self.head_dim)  # 2 heads
        self.v_proj = nn.Linear(hidden_dim, self.kv_heads * self.head_dim)  # 2 heads

        # Output projection
        self.o_proj = nn.Linear(hidden_dim, hidden_dim)

        # KV cache for inference
        self.register_buffer('cached_keys', None)
        self.register_buffer('cached_values', None)

    def forward(self,
                hidden_states: torch.Tensor,
                encoder_hidden: Optional[torch.Tensor] = None,
                attention_mask: Optional[torch.Tensor] = None,
                use_cache: bool = False) -> Tuple[torch.Tensor, Optional[Tuple]]:
        """
        Forward pass with KV cache optimization
        """
        batch_size, seq_len = hidden_states.shape[:2]

        # Query projection (all heads)
        Q = self.q_proj(hidden_states).view(batch_size, seq_len, self.num_heads, self.head_dim)
        Q = Q.transpose(1, 2)  # [batch, heads, seq, dim]

        # Key/Value source (self or cross)
        kv_source = encoder_hidden if encoder_hidden is not None else hidden_states

        # Key/Value projection (fewer heads)
        K = self.k_proj(kv_source).view(batch_size, -1, self.kv_heads, self.head_dim)
        V = self.v_proj(kv_source).view(batch_size, -1, self.kv_heads, self.head_dim)
        K = K.transpose(1, 2)  # [batch, kv_heads, seq, dim]
        V = V.transpose(1, 2)

        # Repeat KV heads to match Q heads (broadcast)
        K = K.repeat_interleave(self.num_heads // self.kv_heads, dim=1)
        V = V.repeat_interleave(self.num_heads // self.kv_heads, dim=1)

        # Cache management for incremental generation (GPT suggestion)
        if use_cache:
            # For incremental generation, only process new token
            if self.cached_keys is not None and hidden_states.size(1) == 1:
                # Append new K/V to cache
                K = torch.cat([self.cached_keys, K], dim=2)
                V = torch.cat([self.cached_values, V], dim=2)
            # Update cache
            self.cached_keys = K
            self.cached_values = V

        # Scaled dot-product attention
        scores = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(self.head_dim)

        # Use additive mask (GPT suggestion)
        if attention_mask is not None:
            scores = scores + attention_mask  # additive mask: -inf where masked, 0 elsewhere

        attn_weights = F.softmax(scores, dim=-1)
        attn_output = torch.matmul(attn_weights, V)

        # Reshape and project
        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.view(batch_size, seq_len, self.hidden_dim)
        output = self.o_proj(attn_output)

        return output, (K, V) if use_cache else None


class SelectiveCrossAttention(nn.Module):
    """
    Selective cross-attention - only attend to relevant encoder layers
    Reduces 24 → 8 cross-attentions for efficiency
    """

    def __init__(self, hidden_dim: int = 1280, layer_id: int = 0):
        super().__init__()
        self.hidden_dim = hidden_dim
        self.layer_id = layer_id

        # Define which encoder layers this decoder layer should attend to
        self.encoder_connections = {
            0: [0],      # Decoder L0 → Encoder L0 (byte info)
            1: [0],      # Decoder L1 → Encoder L0 (byte info)
            2: [1, 2],   # Decoder L2 → Encoder L1,2 (language info)
            3: [1, 2],   # Decoder L3 → Encoder L1,2 (language info)
            4: [3],      # Decoder L4 → Encoder L3 (semantic info)
            5: [3],      # Decoder L5 → Encoder L3 (semantic info)
        }

        # Get connections for this layer
        self.connected_layers = self.encoder_connections.get(layer_id, [0])

        # Create attention modules only for connected layers
        self.cross_attentions = nn.ModuleList([
            KVCacheOptimizedAttention(hidden_dim, num_heads=16, kv_compression=8)
            for _ in self.connected_layers
        ])

        # Lightweight fusion with weighted sum (GPT suggestion)
        self.fusion = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim),
            nn.LayerNorm(hidden_dim),
            nn.SiLU(),
            nn.Dropout(0.1)
        )

        # Learnable weights for connected layers only
        self.layer_weights = nn.Parameter(torch.ones(len(self.connected_layers)) / len(self.connected_layers))

    def forward(self,
                decoder_hidden: torch.Tensor,
                encoder_all_hidden: List[torch.Tensor],
                attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        """
        Selectively attend to relevant encoder layers only
        """
        # Only attend to connected encoder layers
        cross_outputs = []
        for i, layer_idx in enumerate(self.connected_layers):
            if layer_idx < len(encoder_all_hidden):
                encoder_hidden = encoder_all_hidden[layer_idx]
                cross_out, _ = self.cross_attentions[i](
                    hidden_states=decoder_hidden,
                    encoder_hidden=encoder_hidden,
                    attention_mask=attention_mask
                )
                cross_outputs.append(cross_out)

        # Weighted sum fusion for connected layers only
        if len(cross_outputs) > 1:
            weighted_outputs = torch.stack(cross_outputs, dim=0)  # [N, batch, seq, hidden]
            weights = F.softmax(self.layer_weights, dim=0).view(-1, 1, 1, 1)
            fused = (weighted_outputs * weights).sum(dim=0)  # [batch, seq, hidden]
        else:
            # Single connection - no fusion needed
            fused = cross_outputs[0] if cross_outputs else decoder_hidden

        # Apply lightweight fusion layer
        fused = self.fusion(fused)

        return fused


class SwiGLU(nn.Module):
    """SwiGLU activation for better convergence (GPT suggestion)"""
    def __init__(self, dim: int, mult: float = 2.66):
        super().__init__()
        inner = int(round(dim * mult / 2)) * 2  # Even alignment
        self.w1 = nn.Linear(dim, inner // 2)
        self.w2 = nn.Linear(dim, inner // 2)
        self.w3 = nn.Linear(inner // 2, dim)

    def forward(self, x):
        return self.w3(F.silu(self.w1(x)) * self.w2(x))


class DecoderLayer(nn.Module):
    """
    Single decoder layer with self-attention and selective cross-attention
    """

    def __init__(self, hidden_dim: int = 1280, num_heads: int = 16, layer_id: int = 0):
        super().__init__()
        self.hidden_dim = hidden_dim
        self.layer_id = layer_id

        # Self-attention (with KV cache optimization)
        self.self_attn = KVCacheOptimizedAttention(hidden_dim, num_heads, kv_compression=8)
        self.self_attn_norm = nn.LayerNorm(hidden_dim)

        # Selective cross-attention to specific encoder layers
        self.cross_attn = SelectiveCrossAttention(hidden_dim, layer_id=layer_id)
        self.cross_attn_norm = nn.LayerNorm(hidden_dim)

        # Feed-forward network with SwiGLU (GPT suggestion)
        self.ffn = SwiGLU(hidden_dim, mult=2.66)
        self.ffn_norm = nn.LayerNorm(hidden_dim)

        # Dropout for residual connections
        self.dropout = nn.Dropout(0.1)

    def forward(self,
                hidden_states: torch.Tensor,
                encoder_all_hidden: List[torch.Tensor],
                self_attention_mask: Optional[torch.Tensor] = None,
                cross_attention_mask: Optional[torch.Tensor] = None,
                use_cache: bool = False) -> Tuple[torch.Tensor, Optional[Tuple]]:
        """
        Forward pass through decoder layer
        """
        # Self-attention with residual
        residual = hidden_states
        hidden_states = self.self_attn_norm(hidden_states)
        self_attn_out, cache = self.self_attn(
            hidden_states,
            attention_mask=self_attention_mask,
            use_cache=use_cache
        )
        hidden_states = residual + self.dropout(self_attn_out)

        # Cross-attention with residual
        residual = hidden_states
        hidden_states = self.cross_attn_norm(hidden_states)
        cross_attn_out = self.cross_attn(
            hidden_states,
            encoder_all_hidden,
            attention_mask=cross_attention_mask
        )
        hidden_states = residual + self.dropout(cross_attn_out)

        # FFN with residual
        residual = hidden_states
        hidden_states = self.ffn_norm(hidden_states)
        ffn_out = self.ffn(hidden_states)
        hidden_states = residual + self.dropout(ffn_out)

        return hidden_states, cache


class DecoderV62(nn.Module):
    """
    6-Layer Decoder with Multi-Level Cross-Attention
    Reduced from 8 layers but compensated with better cross-attention
    """

    def __init__(self, config: Optional[Dict] = None):
        super().__init__()

        # Configuration
        self.hidden_dim = 1280
        self.num_heads = 16
        self.num_layers = 6  # Reduced from 8
        self.vocab_size = 260  # 256 bytes + special tokens
        self.max_seq_len = 48

        # Token constants (GPT suggestion - explicit constants)
        self.PAD = 256
        self.BOS = 257
        self.EOS = 258
        self.MASK = 259

        # Token embedding and position encoding
        self.token_embedding = nn.Embedding(self.vocab_size, self.hidden_dim)
        self.position_embedding = nn.Embedding(self.max_seq_len, self.hidden_dim)

        # 6 decoder layers with layer-specific cross-attention
        self.layers = nn.ModuleList([
            DecoderLayer(self.hidden_dim, self.num_heads, layer_id=i)
            for i in range(self.num_layers)
        ])

        # Output projection
        self.output_norm = nn.LayerNorm(self.hidden_dim)
        self.output_projection = nn.Linear(self.hidden_dim, self.vocab_size)

        # Monitoring (GPT-5 suggestion)
        # Track importance of ENCODER layers (4) used by decoder
        self.register_buffer('layer_importance', torch.zeros(4))  # Track importance of 4 encoder layers

    def forward(self,
                encoder_all_hidden: List[torch.Tensor],
                decoder_input_ids: Optional[torch.Tensor] = None,
                attention_mask: Optional[torch.Tensor] = None,
                use_cache: bool = False,
                past_key_values: Optional[List] = None) -> Dict[str, torch.Tensor]:
        """
        Forward pass through decoder

        Args:
            encoder_all_hidden: All encoder layer outputs (4 layers)
            decoder_input_ids: Input token IDs for teacher forcing
            attention_mask: Attention mask
            use_cache: Whether to cache KV for inference
            past_key_values: Cached KV from previous steps
        """
        batch_size = encoder_all_hidden[0].size(0)
        device = encoder_all_hidden[0].device

        # If no decoder input, start with compressed representation
        if decoder_input_ids is None:
            # Use encoder's final compressed output as starting point
            hidden_states = encoder_all_hidden[-1]  # [batch, M tokens, 1280]
            seq_len = hidden_states.size(1)
        else:
            # Teacher forcing mode: use provided tokens
            seq_len = decoder_input_ids.size(1)

            # Embeddings
            token_embeds = self.token_embedding(decoder_input_ids)
            position_ids = torch.arange(seq_len, device=device).expand(batch_size, -1)
            position_embeds = self.position_embedding(position_ids)

            hidden_states = token_embeds + position_embeds

        # Create causal mask for self-attention (additive mask - GPT suggestion)
        causal_mask = torch.full((1, 1, seq_len, seq_len), float('-inf'), device=device)
        causal_mask = torch.triu(causal_mask, diagonal=1)  # [1, 1, seq, seq]

        # Pass through decoder layers
        all_hidden_states = []
        all_caches = [] if use_cache else None

        for i, layer in enumerate(self.layers):
            # GPT final check: Create proper cross-attention mask for encoder hidden states
            if encoder_all_hidden is not None and len(encoder_all_hidden) > 0:
                S_enc = encoder_all_hidden[0].size(1)  # Encoder sequence length
                # Create additive mask (0 = attend, -inf = mask)
                cross_mask = torch.zeros((batch_size, 1, 1, S_enc), device=hidden_states.device)
            else:
                cross_mask = None

            hidden_states, cache = layer(
                hidden_states,
                encoder_all_hidden,
                self_attention_mask=causal_mask,
                cross_attention_mask=cross_mask,  # Use proper cross mask
                use_cache=use_cache
            )

            all_hidden_states.append(hidden_states)
            if use_cache:
                all_caches.append(cache)

        # Final output projection
        hidden_states = self.output_norm(hidden_states)
        logits = self.output_projection(hidden_states)

        # Update monitoring: track encoder layer importance
        # (This would be computed based on cross-attention weights in practice)
        with torch.no_grad():
            # Simplified: assume equal importance for now
            self.layer_importance = torch.tensor([0.25, 0.25, 0.25, 0.25])

        outputs = {
            'logits': logits,
            'last_hidden_state': hidden_states,
            'all_hidden_states': all_hidden_states,
            'encoder_layer_importance': self.layer_importance
        }

        if use_cache:
            outputs['past_key_values'] = all_caches

        return outputs

    def generate(self,
                 encoder_all_hidden: List[torch.Tensor],
                 max_length: int = 48,
                 temperature: float = 1.0,
                 top_k: int = 50,
                 top_p: float = 0.95) -> torch.Tensor:
        """
        Autoregressive generation
        """
        batch_size = encoder_all_hidden[0].size(0)
        device = encoder_all_hidden[0].device

        # Start with BOS token
        generated = torch.full((batch_size, 1), self.BOS, device=device)

        # Generate tokens one by one
        past_key_values = None
        for _ in range(max_length - 1):
            # GPT optimization: Only pass last token for O(T) complexity
            if past_key_values is not None:
                decoder_input = generated[:, -1:]  # Last token only
            else:
                decoder_input = generated  # Full sequence for first step

            outputs = self.forward(
                encoder_all_hidden,
                decoder_input_ids=decoder_input,
                use_cache=True,
                past_key_values=past_key_values
            )

            logits = outputs['logits'][:, -1, :] / temperature

            # Top-k filtering
            if top_k > 0:
                indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
                logits[indices_to_remove] = float('-inf')

            # Top-p (nucleus) filtering
            if top_p < 1.0:
                sorted_logits, sorted_indices = torch.sort(logits, descending=True)
                cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)

                # Remove tokens with cumulative probability above threshold
                sorted_indices_to_remove = cumulative_probs > top_p
                sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
                sorted_indices_to_remove[..., 0] = 0

                indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
                logits[indices_to_remove] = float('-inf')

            # Sample
            probs = F.softmax(logits, dim=-1)
            next_token = torch.multinomial(probs, num_samples=1)

            # Append to generated sequence
            generated = torch.cat([generated, next_token], dim=1)

            # Check for EOS
            if (next_token == self.EOS).all():
                break

            past_key_values = outputs.get('past_key_values')

        return generated

    def get_memory_usage(self) -> Dict[str, float]:
        """
        Calculate memory usage with KV cache optimization (GPT-5 metric)
        """
        # Standard attention: 16 heads for K and V
        standard_kv_memory = 2 * 16 * self.max_seq_len * 80 * 4  # bytes

        # Optimized: 2 heads for K and V
        optimized_kv_memory = 2 * 2 * self.max_seq_len * 80 * 4  # bytes

        return {
            'standard_kv_mb': standard_kv_memory / (1024 * 1024),
            'optimized_kv_mb': optimized_kv_memory / (1024 * 1024),
            'reduction_ratio': standard_kv_memory / optimized_kv_memory,
            'total_params_m': sum(p.numel() for p in self.parameters()) / 1e6
        }


if __name__ == "__main__":
    # Test the decoder
    decoder = DecoderV62()

    # Simulate encoder outputs (4 layers, 6 tokens each)
    batch_size = 2
    num_tokens = 6  # After progressive splitting
    hidden_dim = 1280

    encoder_outputs = [
        torch.randn(batch_size, num_tokens, hidden_dim)
        for _ in range(4)
    ]

    # Test with teacher forcing
    decoder_input = torch.randint(0, 256, (batch_size, 48))
    output = decoder(encoder_outputs, decoder_input_ids=decoder_input)

    print(f"Decoder output shape: {output['logits'].shape}")
    print(f"Encoder layer importance: {output['encoder_layer_importance']}")

    # Test generation
    generated = decoder.generate(encoder_outputs, max_length=48)
    print(f"Generated shape: {generated.shape}")

    # Memory usage
    memory_stats = decoder.get_memory_usage()
    print(f"Memory optimization: {memory_stats['reduction_ratio']:.1f}x reduction")
    print(f"Total parameters: {memory_stats['total_params_m']:.1f}M")