| """v57: v17 architecture with 32 heads (d_head=16) — strict ±1. |
| |
| Only change from v17: more heads. Each head is an independent binary voter |
| via Gumbel hard-argmax. Aggregate concat'd into residual. 32 heads = 32× |
| independent ±1 votes per position vs v17's 8. |
| |
| ALiBi slopes are grouped: 8 distinct slopes, 4 heads per group, to avoid |
| int64 overflow for more-than-16 heads. |
| """ |
| 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_v16 import gumbel_hard_attention |
|
|
|
|
| class ManyHeadAttention(nn.Module): |
| def __init__(self, d_model, n_heads, slope_groups=8): |
| 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 = BitLinear(d_model, d_model) |
| self.k_proj = BitLinear(d_model, d_model) |
| self.v_proj = BitLinear(d_model, d_model) |
| self.o_proj = BitLinear(d_model, d_model) |
| |
| |
| heads_per_group = n_heads // slope_groups |
| slopes = [] |
| for g in range(slope_groups): |
| slopes.extend([1 << g] * heads_per_group) |
| while len(slopes) < n_heads: |
| slopes.append(1 << (slope_groups - 1)) |
| self.register_buffer('alibi_slopes_int', torch.tensor(slopes[:n_heads], dtype=torch.long)) |
|
|
| 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 BitBlockV57(nn.Module): |
| def __init__(self, d_model, n_heads, d_ff): |
| super().__init__() |
| self.attn = ManyHeadAttention(d_model, n_heads) |
| self.ffn = BitFFN(d_model, d_ff) |
|
|
| def forward(self, x): |
| a = self.attn(x) |
| f = self.ffn(x) |
| return sign_ste(x + a + f) |
|
|
|
|
| class BitLMv57(nn.Module): |
| def __init__(self, vocab_size=128, d_model=512, n_layers=4, n_heads=32, |
| 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([ |
| BitBlockV57(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 H in (16, 32, 64): |
| m = BitLMv57(d_model=512, n_layers=4, n_heads=H, d_ff=192) |
| n = sum(p.numel() for p in m.parameters()) |
| print(f'n_heads={H}: {n:,} ({n/1e6:.3f}M)') |
| m = BitLMv57(d_model=512, n_layers=4, n_heads=32, d_ff=192) |
| 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') |
|
|
