"""v51: v48 + SiLU-gated FFN (true SwiGLU in binary form).

v47/v48 FFN: `down(sign(gate(x)) * sign(up(x)))` — the gate*up is XNOR of two
±1 vectors. That throws away 1 bit of gate information per channel.

v51 FFN: SwiGLU-style. `down(silu(gate_raw(x)) * sign(up(x)))` where:
  - gate_raw returns the pre-sign float (α·popcount - threshold)
  - silu of that is a float
  - up returns ±1
  - product is float
  - down is a DoubledScaled... wait no, keep it single ±1 per weight.

Keeps weights strictly ±1 per stored parameter. The FFN's forward path now
produces float activations through the gate branch, matching standard
SwiGLU. This is how BitNet-1.58b (and BitNet v1) actually structure FFN.
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F

from model import sign_ste, sign_ste_clipped, BinaryEmbedding
from model_v16 import gumbel_hard_attention
from model_v47 import RMSNorm, BitLinearScaled, BitLinearScaledRaw, IntBinaryAttentionScaled


class BitFFNSwiGLU(nn.Module):
    """SwiGLU: silu(gate_raw) * sign(up) → down. gate has float output; up is ±1.
    down's input is float; it still uses ±1 weights (XNOR-popcount on int8-ish input).
    """
    def __init__(self, d_model, d_ff):
        super().__init__()
        # gate returns raw float (no final sign). up returns ±1.
        self.gate = BitLinearScaledRaw(d_model, d_ff, binarize_input=True)
        self.up = BitLinearScaled(d_model, d_ff, binarize_input=True)
        # down takes float input; still binarizes internally.
        self.down = BitLinearScaledRaw(d_ff, d_model, binarize_input=True)

    def forward(self, x):
        g = F.silu(self.gate(x))       # float
        u = self.up(x)                 # ±1
        return self.down(g * u)        # float, returned as raw (into residual)


class BitBlockV51(nn.Module):
    def __init__(self, d_model, n_heads, d_ff):
        super().__init__()
        self.norm1 = RMSNorm(d_model)
        self.attn = IntBinaryAttentionScaled(d_model, n_heads)
        self.norm2 = RMSNorm(d_model)
        self.ffn = BitFFNSwiGLU(d_model, d_ff)

    def forward(self, x):
        x = x + self.attn(self.norm1(x))
        x = x + self.ffn(self.norm2(x))
        return x


class BitLMv51(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([
            BitBlockV51(d_model, n_heads, d_ff) for _ in range(n_layers)
        ])
        self.norm_out = RMSNorm(d_model)
        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)
        x = self.norm_out(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 = BitLMv51(d_model=512, n_layers=4, d_ff=192)
    n = sum(p.numel() for p in m.parameters())
    print(f'v51 SwiGLU: {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')