vm_backup/code/model_v47.py

"""v47: BitNet-style — per-channel float scale on every BitLinear + RMSNorm
between blocks + float residual stream.

Storage-wise still strict 1-bit per weight: the weight matrices are ±1. What we
add is float *auxiliary* parameters:
  - per-output-channel scale α ∈ R^{d_out} per BitLinear
  - RMSNorm γ ∈ R^{d_model} per block

These are the standard components of every "1-bit LLM" paper in the literature
(BitNet, OneBit). v17's maximalist design stripped them out. If the intern's
gain comes from restoring magnitude/normalization information (the only thing
strict ±1 maximalism destroys), this matches.

Float aux params: ~d_model floats per BitLinear + d_model per RMSNorm. For
v17-shape (d=512, L=4) that's ~20K floats, vs 5.5M ±1 weights. <0.4% overhead.
"""
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 BitLinearScaled(nn.Module):
    """±1 weights, XNOR-popcount matmul, per-channel float scale α.

    forward: sign_ste_clipped(alpha * sign(W) @ sign(x) - threshold).
    Every stored weight is ±1. α and threshold are float (trainable scalars
    per output channel)."""
    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 = nn.Parameter(torch.randn(out_features, in_features) * 0.02)
        self.alpha = nn.Parameter(torch.full((out_features,), 1.0 / math.sqrt(in_features)))
        self.threshold = nn.Parameter(torch.zeros(out_features))

    def forward(self, x):
        W = sign_ste(self.weight)
        if self.binarize_input:
            x = sign_ste_clipped(x)
        s = F.linear(x, W) * self.alpha - self.threshold
        return sign_ste_clipped(s)


class BitLinearScaledRaw(nn.Module):
    """Same as BitLinearScaled but returns the pre-sign (float/int) score.
    Used where we want to sum raw values into the residual stream."""
    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 = nn.Parameter(torch.randn(out_features, in_features) * 0.02)
        self.alpha = nn.Parameter(torch.full((out_features,), 1.0 / math.sqrt(in_features)))
        self.bias = nn.Parameter(torch.zeros(out_features))

    def forward(self, x):
        W = sign_ste(self.weight)
        if self.binarize_input:
            x = sign_ste_clipped(x)
        return F.linear(x, W) * self.alpha + self.bias


class RMSNorm(nn.Module):
    def __init__(self, d_model, eps=1e-6):
        super().__init__()
        self.gamma = nn.Parameter(torch.ones(d_model))
        self.eps = eps

    def forward(self, x):
        rms = x.pow(2).mean(dim=-1, keepdim=True).add(self.eps).rsqrt()
        return x * rms * self.gamma


class BitFFNScaled(nn.Module):
    """SwiGLU-ish: gate * up then down. Intermediate kept float (via scale)."""
    def __init__(self, d_model, d_ff):
        super().__init__()
        self.gate = BitLinearScaled(d_model, d_ff, binarize_input=True)
        self.up = BitLinearScaled(d_model, d_ff, binarize_input=True)
        # down returns raw float into residual stream
        self.down = BitLinearScaledRaw(d_ff, d_model, binarize_input=True)

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


class IntBinaryAttentionScaled(nn.Module):
    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 = BitLinearScaled(d_model, d_model)
        self.k_proj = BitLinearScaled(d_model, d_model)
        self.v_proj = BitLinearScaled(d_model, d_model)
        # o_proj returns raw float for residual
        self.o_proj = BitLinearScaledRaw(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 BitBlockV47(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 = BitFFNScaled(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 BitLMv47(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([
            BitBlockV47(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 = BitLMv47(d_model=512, n_layers=4, d_ff=192)
    n = sum(p.numel() for p in m.parameters())
    float_p = sum(p.numel() for n_, p in m.named_parameters() if 'alpha' in n_ or 'gamma' in n_)
    print(f'total: {n:,} ({n/1e6:.3f}M); float-aux: {float_p:,} ({float_p/n*100:.2f}%)')
    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')