| """v46: Float residual stream. Every WEIGHT is ±1; the residual stream between |
| blocks is not sign()'d. |
| |
| v18/v17 force x = sign(x + a + f) at every block. That means the entire model |
| lives in a ±1 vector space between layers — a crushing constraint. Each |
| BitLinear binarizes its input anyway via sign_ste_clipped, so removing the |
| outer sign on the residual costs nothing in "1-bit compute": every matmul is |
| still XNOR-popcount, every weight is still ±1. |
| |
| What changes: residual stream accumulates as int/float across depth. The |
| model now has float-valued activations (small dynamic range: sum of L ±1 |
| vectors). Storage is still strictly 1-bit per parameter. |
| |
| This is what BitNet/1-bit-LLMs actually do. Our "maximalist" framing was |
| pessimistic about this tradeoff. |
| """ |
| import math |
| import torch |
| import torch.nn as nn |
| import torch.nn.functional as F |
|
|
| from model import sign_ste, BitLinear, BitFFN, BinaryEmbedding |
| from model_v18 import IntBinaryAttention |
|
|
|
|
| class BitBlockV46(nn.Module): |
| def __init__(self, d_model, n_heads, d_ff): |
| super().__init__() |
| self.attn = IntBinaryAttention(d_model, n_heads) |
| self.ffn = BitFFN(d_model, d_ff) |
|
|
| def forward(self, x): |
| a = self.attn(x) |
| f = self.ffn(x) |
| return x + a + f |
|
|
|
|
| class BitLMv46(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([ |
| BitBlockV46(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) |
| |
| x = sign_ste(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, L, d_ff) in ((256, 8, 512), (512, 4, 192), (336, 4, 192)): |
| m = BitLMv46(d_model=D, n_layers=L, d_ff=d_ff) |
| n = sum(p.numel() for p in m.parameters()) |
| print(f'D={D} L={L} d_ff={d_ff}: {n:,} ({n/1e6:.3f}M)') |
| m = BitLMv46() |
| 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') |
|
|
