"""v38: BitMixer — drop attention entirely.

Deep analysis on v29 showed 102 of 120 attention heads collapsed to static
local/positional patterns; only 6 heads were content-sensitive. Binary attention
with ±1 QK + ALiBi cannot form content-addressable routing. Honest response:
replace the attention module with a static binary mix of tokens — a single
causal-masked (T × T) ±1 weight matrix per layer, applied per-channel.

Every signal path remains strictly ±1:
  x (±1) -> W_mix (±1, T×T, causal) @ x -> rowwise rescale -> sign_ste -> ±1

No Q/K/V, no softmax, no ALiBi. Like MLP-Mixer but everything is binary.
If this matches v17 (1.68 BPC) at 5M/10K, binary attention was deadweight.
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F

from model import sign_ste, sign_ste_clipped, BitLinear, BinaryEmbedding


class BitTokenMix(nn.Module):
    """Static causal ±1 token-mix. Shared across channels, independent across layers."""
    def __init__(self, max_seq_len):
        super().__init__()
        self.max_seq_len = max_seq_len
        # Latent float weight; forward uses sign(weight) * causal_mask.
        self.weight = nn.Parameter(torch.randn(max_seq_len, max_seq_len) * 0.02)
        mask = torch.tril(torch.ones(max_seq_len, max_seq_len))
        self.register_buffer('causal_mask', mask)
        # Row t has t+1 non-zero entries; std of signed sum ~ sqrt(t+1).
        row_norm = 1.0 / torch.sqrt(torch.arange(1, max_seq_len + 1).float())
        self.register_buffer('row_norm', row_norm)

    def forward(self, x):
        B, T, D = x.shape
        W = sign_ste(self.weight[:T, :T])
        W = W * self.causal_mask[:T, :T]
        # y[b, t, d] = sum_s W[t, s] x[b, s, d]  =>  (B,D,T) @ W.T -> (B,D,T)
        x_bdt = x.transpose(1, 2)
        y_bdt = x_bdt @ W.t()
        y_bdt = y_bdt * self.row_norm[:T].view(1, 1, T)
        return sign_ste_clipped(y_bdt.transpose(1, 2))


class BitFFNV38(nn.Module):
    def __init__(self, d_model, d_ff):
        super().__init__()
        self.gate = BitLinear(d_model, d_ff, binarize_input=True)
        self.up = BitLinear(d_model, d_ff, binarize_input=True)
        self.down = BitLinear(d_ff, d_model, binarize_input=True)

    def forward(self, x):
        return self.down(self.gate(x) * self.up(x))


class BitBlockV38(nn.Module):
    def __init__(self, d_model, d_ff, max_seq_len):
        super().__init__()
        self.mix = BitTokenMix(max_seq_len)
        self.ffn = BitFFNV38(d_model, d_ff)

    def forward(self, x):
        m = self.mix(x)
        f = self.ffn(x)
        return sign_ste(x + m + f)


class BitLMv38(nn.Module):
    def __init__(self, vocab_size=128, d_model=256, n_layers=8, d_ff=720,
                 max_seq_len=256):
        super().__init__()
        self.vocab_size = vocab_size
        self.d_model = d_model
        self.n_layers = n_layers
        self.d_ff = d_ff
        self.max_seq_len = max_seq_len
        self.embed = BinaryEmbedding(vocab_size, d_model)
        self.blocks = nn.ModuleList([
            BitBlockV38(d_model, d_ff, max_seq_len) 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 = BitLMv38(vocab_size=128, d_model=256, n_layers=8, d_ff=720, max_seq_len=256)
    n = sum(p.numel() for p in m.parameters())
    print(f'v38 BitMixer: {n:,} params ({n/1e6:.2f}M), d_ff=720')
    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')