"""v43: Doubled-Binary — each BitLinear has TWO ±1 weight matrices summed. Effective weights take values in {-2, 0, +2}: ternary with a neutral/zero state. This is still pure 1-bit-per-parameter (every stored weight is ±1 via sign STE). Motivation: analysis on v29 showed 25–30% of latent weights have |w| < 0.01 — the training signal wants them near zero, but sign() forces ±1 regardless. The model is being forced to commit weights that "don't want to be committed," creating noise. Doubled binary lets two opposing ±1 values cancel (sum=0), so the effective weight can be zero. Same 1-bit storage, more expressive. v17 shape with 2x weight count: d_model=336 (from 512), n_layers=4, d_ff=192. Target: 5.26M ≈ 5.52M v17 baseline. """ 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 class DoubledBitLinearRaw(nn.Module): """Two ±1 weight matrices summed: effective W_eff in {-2, 0, +2}.""" def __init__(self, in_features, out_features, binarize_input=True): super().__init__() self.in_features = in_features self.out_features = out_features self.binarize_input = binarize_input self.weight_a = nn.Parameter(torch.randn(out_features, in_features) * 0.02) self.weight_b = nn.Parameter(torch.randn(out_features, in_features) * 0.02) def forward(self, x): W = sign_ste(self.weight_a) + sign_ste(self.weight_b) # {-2, 0, 2} if self.binarize_input: x = sign_ste_clipped(x) return F.linear(x, W) class DoubledBitLinear(nn.Module): """DoubledBitLinearRaw + learned threshold + sign. Returns ±1. Sum of two ±1 matrices has effective values in {-2, 0, 2}. The raw popcount output variance is ~2x standard BitLinear, so we scale by 1/(2·sqrt(in)). """ def __init__(self, in_features, out_features, binarize_input=True): super().__init__() self.raw = DoubledBitLinearRaw(in_features, out_features, binarize_input=binarize_input) self.threshold = nn.Parameter(torch.zeros(out_features)) # Scale by 1/(2·sqrt(in)) since effective |w| can be 2 and sum over in_features. self.scale = 1.0 / (2.0 * math.sqrt(in_features)) def forward(self, x): s = self.raw(x) * self.scale - self.threshold return sign_ste_clipped(s) class DoubledBitFFN(nn.Module): def __init__(self, d_model, d_ff): super().__init__() self.gate = DoubledBitLinear(d_model, d_ff, binarize_input=True) self.up = DoubledBitLinear(d_model, d_ff, binarize_input=True) self.down = DoubledBitLinear(d_ff, d_model, binarize_input=True) def forward(self, x): return self.down(self.gate(x) * self.up(x)) class DoubledIntBinaryAttention(nn.Module): """v18 attention with DoubledBitLinear Q/K/V/O.""" def __init__(self, d_model, n_heads): 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.q_proj = DoubledBitLinear(d_model, d_model) self.k_proj = DoubledBitLinear(d_model, d_model) self.v_proj = DoubledBitLinear(d_model, d_model) self.o_proj = DoubledBitLinear(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 = self.n_heads, self.head_dim 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) A = gumbel_hard_attention(scores, mask=mask) O = torch.matmul(A, V) O = O.transpose(1, 2).contiguous().view(B, T, D) return self.o_proj(O) class BitBlockV43(nn.Module): def __init__(self, d_model, n_heads, d_ff): super().__init__() self.attn = DoubledIntBinaryAttention(d_model, n_heads) self.ffn = DoubledBitFFN(d_model, d_ff) def forward(self, x): a = self.attn(x) f = self.ffn(x) return sign_ste(x + a + f) class BitLMv43(nn.Module): def __init__(self, vocab_size=128, d_model=336, 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([ BitBlockV43(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) for (D, d_ff) in ((320, 240), (336, 192), (336, 208)): m = BitLMv43(d_model=D, d_ff=d_ff) n = sum(p.numel() for p in m.parameters()) print(f'D={D}, d_ff={d_ff}: {n:,} ({n/1e6:.3f}M)') m = BitLMv43() 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')