"""v56: Top-K attention with strict ±1 everywhere.

Gumbel hard-argmax (v16/v17) forces each query to attend to EXACTLY ONE past
position — a crushing expressivity constraint. Top-K relaxes this to K
positions (K > 1) while keeping every activation strictly ±1:

  scores = Q @ K^T  (integer popcount) minus ALiBi
  causal mask
  top-K per query → indices + gather
  O = sign_ste(sum of gathered V's)   # integer sum of K ±1 vectors, signed

Differentiability: top-K is non-differentiable. We use a soft-hard trick: the
forward produces hard top-K gather; the backward flows through the scores with
straight-through on the top-K set. Implemented by adding a soft-softmax surrogate
for gradient + hard top-K for forward.

Config: v17 shape (d=512, L=4, d_ff=192, 5.52M). K=4.
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F

from model import sign_ste, BitLinear, BitFFN, BinaryEmbedding


class TopKBinaryAttention(nn.Module):
    def __init__(self, d_model, n_heads, top_k=4):
        super().__init__()
        assert d_model % n_heads == 0
        self.d_model = d_model
        self.n_heads = n_heads
        self.head_dim = d_model // n_heads
        self.top_k = top_k
        self.q_proj = BitLinear(d_model, d_model)
        self.k_proj = BitLinear(d_model, d_model)
        self.v_proj = BitLinear(d_model, d_model)
        self.o_proj = BitLinear(d_model, d_model)
        slopes = torch.tensor([1 << i for i in range(n_heads)], dtype=torch.long)
        self.register_buffer('alibi_slopes_int', slopes)

    def forward(self, x):
        B, T, D = x.shape
        H, Dh, K = self.n_heads, self.head_dim, self.top_k
        Q = self.q_proj(x).view(B, T, H, Dh).transpose(1, 2)
        K_ = self.k_proj(x).view(B, T, H, Dh).transpose(1, 2)
        V = self.v_proj(x).view(B, T, H, Dh).transpose(1, 2)

        scores = torch.matmul(Q, K_.transpose(-2, -1))
        pos = torch.arange(T, device=x.device)
        dist = (pos.unsqueeze(0) - pos.unsqueeze(1)).abs()
        alibi = self.alibi_slopes_int.view(1, H, 1, 1).to(scores.dtype) \
                * dist.view(1, 1, T, T).to(scores.dtype)
        scores = scores - alibi

        mask = torch.triu(torch.ones(T, T, device=x.device, dtype=torch.bool), diagonal=1)
        scores = scores.masked_fill(mask, -1e9)

        # Soft-hard top-K:
        # - hard: pick top-K, one-hot mask
        # - soft: softmax over scores for gradient
        eff_k = min(K, T)
        top_vals, top_idx = scores.topk(eff_k, dim=-1)    # (B, H, T, K)
        # hard attention mask
        hard_A = torch.zeros_like(scores)
        hard_A.scatter_(-1, top_idx, 1.0)
        # soft softmax (normalized) for backward path
        soft_A = F.softmax(scores, dim=-1)
        A = soft_A + (hard_A - soft_A).detach()            # STE: forward=hard, backward=soft

        # Sum of K ±1 V's per query → sign
        O = torch.matmul(A, V)                              # (B, H, T, Dh) integer in [-K, K]
        O = sign_ste(O)                                     # ±1
        O = O.transpose(1, 2).contiguous().view(B, T, D)
        return self.o_proj(O)


class BitBlockV56(nn.Module):
    def __init__(self, d_model, n_heads, d_ff, top_k=4):
        super().__init__()
        self.attn = TopKBinaryAttention(d_model, n_heads, top_k=top_k)
        self.ffn = BitFFN(d_model, d_ff)

    def forward(self, x):
        a = self.attn(x)
        f = self.ffn(x)
        return sign_ste(x + a + f)      # strict ±1 residual


class BitLMv56(nn.Module):
    def __init__(self, vocab_size=128, d_model=512, n_layers=4, n_heads=8,
                 d_ff=192, top_k=4, max_seq_len=256):
        super().__init__()
        self.vocab_size = vocab_size
        self.d_model = d_model
        self.n_layers = n_layers
        self.max_seq_len = max_seq_len
        self.top_k = top_k
        self.embed = BinaryEmbedding(vocab_size, d_model)
        self.blocks = nn.ModuleList([
            BitBlockV56(d_model, n_heads, d_ff, top_k=top_k) for _ in range(n_layers)
        ])
        self.out_codebook = nn.Parameter(torch.randn(vocab_size, d_model) * 0.02)
        self.logit_scale = nn.Parameter(torch.tensor(1.0 / math.sqrt(d_model)))
        self.out_bias = nn.Parameter(torch.zeros(vocab_size))

    def forward(self, idx, targets=None):
        x = self.embed(idx)
        for blk in self.blocks:
            x = blk(x)
        W_out = sign_ste(self.out_codebook)
        scores = torch.matmul(x, W_out.t())
        logits = scores * self.logit_scale + self.out_bias
        loss = None
        if targets is not None:
            loss = F.cross_entropy(logits.view(-1, self.vocab_size), targets.view(-1))
        return logits, loss


if __name__ == '__main__':
    m = BitLMv56(d_model=512, n_layers=4, d_ff=192, top_k=4)
    n = sum(p.numel() for p in m.parameters())
    print(f'v56 top-4 attn: {n:,} ({n/1e6:.3f}M)')
    x = torch.randint(0, 128, (2, 64))
    y = torch.randint(0, 128, (2, 64))
    logits, loss = m(x, y)
    loss.backward()
    print(f'loss={loss.item():.3f}, backward OK')