"""v55: XOR-residual binary transformer.

Standard transformer: x_new = sign(x + attn(x) + ffn(x))  — majority vote.
v55:                  x_new = x ⊙ attn(x) ⊙ ffn(x)        — elementwise XOR.

For ±1 vectors, elementwise multiply IS XOR. Unlike majority-vote-sum, XOR
preserves all three bits of information in a different algebra: it's invertible
in a group-theoretic sense (commutative, associative, own inverse).

This is a fundamentally different binary residual operator. Every residual
stream value is strictly ±1 at every step. Attention is still Gumbel hard-argmax.
FFN is still XNOR gated. Weights ±1. Nothing float anywhere.

Config: v17 shape (d=512, L=4, d_ff=192, 5.52M), 10K steps.
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F

from model import sign_ste, BitLinear, BitFFN, BinaryEmbedding
from model_v18 import IntBinaryAttention


class BitBlockV55(nn.Module):
    def __init__(self, d_model, n_heads, d_ff):
        super().__init__()
        self.attn = IntBinaryAttention(d_model, n_heads)
        self.ffn = BitFFN(d_model, d_ff)

    def forward(self, x):
        a = self.attn(x)
        f = self.ffn(x)
        return x * a * f            # ±1 elementwise XOR of three ±1 tensors


class BitLMv55(nn.Module):
    def __init__(self, vocab_size=128, d_model=512, n_layers=4, n_heads=8,
                 d_ff=192, 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.embed = BinaryEmbedding(vocab_size, d_model)
        self.blocks = nn.ModuleList([
            BitBlockV55(d_model, n_heads, d_ff) 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__':
    from model_v16 import set_gumbel_tau
    set_gumbel_tau(0.5)
    m = BitLMv55(d_model=512, n_layers=4, d_ff=192)
    n = sum(p.numel() for p in m.parameters())
    print(f'v55 XOR-res: {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')