"""v3 variant: parallel attention+FFN residual so the 3-way sum is always odd (no ties).

Rationale: the v2 block is
    x = sign(x + attn(x))        # values {-2, 0, 2}, 0 -> +1 (bias)
    x = sign(x + ffn(x))          # values {-2, 0, 2}, 0 -> +1 (bias)
The sign-on-zero bias pushes every residual toward +1, compounds across 8 layers.

v3 block:
    a = attn(x); f = ffn(x)
    x_out = sign(x + a + f)       # values {-3, -1, 1, 3}, never 0, no bias

Same ±1 invariant but strictly unbiased at the residual.
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F

from model import (
    sign_ste, sign_ste_clipped, BitLinearRaw, BitLinear,
    BiAttention, BitFFN, BinaryEmbedding,
)


class BitBlockV3(nn.Module):
    def __init__(self, d_model, n_heads, d_ff):
        super().__init__()
        self.attn = BiAttention(d_model, n_heads)
        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)


class BitLMv3(nn.Module):
    def __init__(self, vocab_size=128, d_model=256, n_layers=8, n_heads=8, d_ff=512, 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([
            BitBlockV3(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

    @torch.no_grad()
    def generate(self, idx, max_new_tokens=200, temperature=1.0, top_k=None):
        self.eval()
        for _ in range(max_new_tokens):
            idx_cond = idx[:, -self.max_seq_len:]
            logits, _ = self(idx_cond)
            logits = logits[:, -1, :] / max(temperature, 1e-5)
            if top_k is not None:
                v, _ = torch.topk(logits, top_k)
                logits[logits < v[:, [-1]]] = -float('inf')
            probs = F.softmax(logits, dim=-1)
            nxt = torch.multinomial(probs, num_samples=1)
            idx = torch.cat([idx, nxt], dim=1)
        return idx


if __name__ == '__main__':
    m = BitLMv3()
    n = sum(p.numel() for p in m.parameters())
    print(f"v3 params: {n:,} ({n/1e6:.2f}M)")
    x = torch.randint(0, 128, (2, 64))
    y = torch.randint(0, 128, (2, 64))
    logits, loss = m(x, y)
    print("logits:", logits.shape, "loss:", loss.item())
    loss.backward()
    print("backward OK")