kernel/attention.py

"""
GLADIUS v2.0 — Hybrid Attention (SLA2-inspired) + SDPA

The core attention mechanism: α-blended softmax + linear attention.

Linear path: O(n) — cheap background awareness of all tokens.
Softmax path: O(n·k) — precise attention for important token pairs.
α: Per-token, learned. High when precision matters, low for routine.

This is the SLA2 principle applied to GLADIUS:
    O = α ⊙ softmax_attention(Q, K_important, V) + (1-α) ⊙ linear_attention(Q, K_all, V)

v2.0.1 (Day 45): Softmax path upgraded from manual matmul→softmax→matmul to
    F.scaled_dot_product_attention → FlashAttention2 on CUDA (2-3x faster).
    Linear path unchanged (inherently O(n), already fast).
    All weight shapes identical — checkpoint-compatible, no architecture change.

Reference: Ali's SLA2 attention pipeline diagram (ali-ref/img-09.jpg)
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
import math

from .config import KernelConfig


class RoPE(nn.Module):
    """Rotary Position Embeddings (Su et al., 2021).

    Applied to Q and K before attention computation.
    Does not interfere with our additive temporal encoding
    (they operate in different subspaces — RoPE is rotational,
    time encoding is additive).
    """

    def __init__(self, head_dim: int, max_seq_len: int = 2048):
        super().__init__()
        # Precompute frequency bands
        inv_freq = 1.0 / (10000 ** (torch.arange(0, head_dim, 2).float() / head_dim))
        self.register_buffer('inv_freq', inv_freq)

        # Precompute rotation matrices for max_seq_len
        t = torch.arange(max_seq_len).float()
        freqs = torch.einsum('i,j->ij', t, inv_freq)
        emb = torch.cat([freqs, freqs], dim=-1)
        self.register_buffer('cos_cached', emb.cos())
        self.register_buffer('sin_cached', emb.sin())

    def forward(self, x: torch.Tensor, seq_len: int) -> torch.Tensor:
        """Apply rotary embeddings. x: (batch, heads, seq, head_dim)"""
        cos = self.cos_cached[:seq_len].unsqueeze(0).unsqueeze(0)
        sin = self.sin_cached[:seq_len].unsqueeze(0).unsqueeze(0)
        return (x * cos) + (self._rotate_half(x) * sin)

    @staticmethod
    def _rotate_half(x: torch.Tensor) -> torch.Tensor:
        x1, x2 = x.chunk(2, dim=-1)
        return torch.cat([-x2, x1], dim=-1)


class HybridAttention(nn.Module):
    """
    SLA2-inspired hybrid attention.

    Every token gets linear attention (cheap, global context).
    Important tokens ALSO get softmax attention (expensive, precise).
    The blend ratio α is learned per-token.

    argmax_attention: α = argmax_{blend} S(blend | hidden_state)
    """

    def __init__(self, config: KernelConfig, layer_idx: int = 0):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.num_heads = config.num_heads
        self.head_dim = config.head_dim
        self.hidden_dim = config.hidden_dim

        # Projections
        self.q_proj = nn.Linear(config.hidden_dim, config.hidden_dim, bias=False)
        self.k_proj = nn.Linear(config.hidden_dim, config.hidden_dim, bias=False)
        self.v_proj = nn.Linear(config.hidden_dim, config.hidden_dim, bias=False)
        self.o_proj = nn.Linear(config.hidden_dim, config.hidden_dim, bias=False)

        # Learned blend ratio: per-head α router
        # Input: hidden state → Output: per-head scalar in [0, 1]
        self.alpha_router = nn.Sequential(
            nn.Linear(config.hidden_dim, config.num_heads),
            nn.Sigmoid()
        )

        # RoPE
        self.rope = RoPE(config.head_dim, config.max_seq_len)

        # QK-Clip: softcap for attention logit stability (Gemma 2 / Kimi K2)
        # NOTE: When qk_softcap is set, we fall back to manual attention
        # because SDPA doesn't support logit transformation.
        # When qk_softcap is None (default in Omega), SDPA → FlashAttention2.
        self.qk_softcap = getattr(config, 'qk_softcap', None)

        self._init_weights()

    def _init_weights(self):
        for proj in [self.q_proj, self.k_proj, self.v_proj, self.o_proj]:
            nn.init.normal_(proj.weight, std=0.02)
        # Initialize alpha toward 0.5 (balanced blend)
        nn.init.zeros_(self.alpha_router[0].bias)

    def _softmax_attention_sdpa(
        self,
        Q: torch.Tensor,
        K: torch.Tensor,
        V: torch.Tensor,
        mask: torch.Tensor | None = None,
    ) -> torch.Tensor:
        """
        Softmax attention via F.scaled_dot_product_attention.
        
        Uses FlashAttention2 backend on CUDA with bfloat16/float16.
        Falls back to efficient math backend on CPU or with masks.
        
        The kernel always passes a standard lower-triangular causal mask.
        We use is_causal=True to hit the FlashAttention fast path (no
        explicit mask materialization, much faster and more memory efficient).
        
        Args:
            Q, K, V: (B, H, S, D_h)
            mask: (B, 1, S, S) causal mask (ignored — is_causal=True is equivalent)
        Returns:
            O: (B, H, S, D_h)
        """
        # is_causal=True is equivalent to the tril mask the kernel passes,
        # but enables FlashAttention fast path (no mask materialization).
        # Q and K are already contiguous (B, H, S, D_h).
        return F.scaled_dot_product_attention(
            Q, K, V,
            dropout_p=0.0,
            is_causal=True,
        )

    def _softmax_attention_manual(
        self,
        Q: torch.Tensor,
        K: torch.Tensor,
        V: torch.Tensor,
        mask: torch.Tensor | None = None,
    ) -> torch.Tensor:
        """
        Manual softmax attention with QK-softcap support.
        
        Fallback when qk_softcap is enabled (SDPA doesn't support logit transforms).
        """
        scores = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(self.head_dim)

        if self.qk_softcap is not None and self.qk_softcap > 0:
            scores = self.qk_softcap * torch.tanh(scores / self.qk_softcap)

        if mask is not None:
            scores = scores.masked_fill(mask == 0, float('-inf'))

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

    def forward(
        self,
        x: torch.Tensor,
        mask: torch.Tensor | None = None,
        memory_keys: torch.Tensor | None = None,
        memory_values: torch.Tensor | None = None,
    ) -> torch.Tensor:
        """
        Args:
            x: (batch, seq_len, hidden_dim)
            mask: (batch, 1, seq_len, seq_len) causal mask
            memory_keys: Optional hot memory keys to attend over
            memory_values: Optional hot memory values
        Returns:
            (batch, seq_len, hidden_dim)
        """
        B, S, D = x.shape

        # Project to multi-head
        Q = self.q_proj(x).view(B, S, self.num_heads, self.head_dim).transpose(1, 2)
        K = self.k_proj(x).view(B, S, self.num_heads, self.head_dim).transpose(1, 2)
        V = self.v_proj(x).view(B, S, self.num_heads, self.head_dim).transpose(1, 2)

        # Apply RoPE to Q and K
        Q = self.rope(Q, S)
        K = self.rope(K, S)

        # === Linear Attention Path (O(n)) ===
        # Feature map: elu(x) + 1 → non-negative
        Q_lin = F.elu(Q) + 1  # (B, H, S, D)
        K_lin = F.elu(K) + 1

        # Causal linear attention via cumulative sum
        KV_lin = torch.matmul(K_lin.transpose(-2, -1), V)  # (B, H, D, D)
        Z_lin = K_lin.transpose(-2, -1).sum(dim=-1, keepdim=True)  # normalizer
        O_linear = torch.matmul(Q_lin, KV_lin) / (torch.matmul(Q_lin, Z_lin) + 1e-6)

        # === Softmax Attention Path ===
        # Use SDPA (→ FlashAttention2) when no softcap, manual otherwise
        if self.qk_softcap is not None and self.qk_softcap > 0:
            O_softmax = self._softmax_attention_manual(Q, K, V, mask)
        else:
            O_softmax = self._softmax_attention_sdpa(Q, K, V, mask)

        # === Blend ===
        # α per-token, per-head: (B, S, num_heads) → (B, num_heads, S, 1)
        alpha = self.alpha_router(x)  # (B, S, H)
        alpha = alpha.permute(0, 2, 1).unsqueeze(-1)  # (B, H, S, 1)

        O = alpha * O_softmax + (1 - alpha) * O_linear

        # Reshape back
        O = O.transpose(1, 2).contiguous().view(B, S, D)
        return self.o_proj(O)


class SwiGLU(nn.Module):
    """SwiGLU FFN block (Shazeer, 2020). Used in LLaMA, Mistral, etc."""

    def __init__(self, config: KernelConfig):
        super().__init__()
        self.gate_proj = nn.Linear(config.hidden_dim, config.ffn_dim, bias=False)
        self.up_proj = nn.Linear(config.hidden_dim, config.ffn_dim, bias=False)
        self.down_proj = nn.Linear(config.ffn_dim, config.hidden_dim, bias=False)
        self._init_weights()

    def _init_weights(self):
        for proj in [self.gate_proj, self.up_proj, self.down_proj]:
            nn.init.normal_(proj.weight, std=0.02)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x))


class RMSNorm(nn.Module):
    """Root Mean Square Layer Normalization (Zhang & Sennrich, 2019)."""

    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(dim))
        self.eps = eps

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        norm = torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
        return x * norm * self.weight


class TransformerLayer(nn.Module):
    """Single transformer layer: RMSNorm → HybridAttn → RMSNorm → SwiGLU."""

    def __init__(self, config: KernelConfig, layer_idx: int = 0):
        super().__init__()
        self.attention = HybridAttention(config, layer_idx)
        self.ffn = SwiGLU(config)
        self.attn_norm = RMSNorm(config.hidden_dim)
        self.ffn_norm = RMSNorm(config.hidden_dim)

    def forward(
        self,
        x: torch.Tensor,
        mask: torch.Tensor | None = None,
    ) -> torch.Tensor:
        # Pre-norm residual connections
        x = x + self.attention(self.attn_norm(x), mask=mask)
        x = x + self.ffn(self.ffn_norm(x))
        return x