File size: 7,832 Bytes
9e81e45 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 |
import math
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
# ---------------------
# Utility Layers
# ---------------------
class RMSNorm(nn.Module):
def __init__(self, d: int, eps: float = 1e-6):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(d))
def forward(self, x):
norm = x.pow(2).mean(-1, keepdim=True)
x = x * torch.rsqrt(norm + self.eps)
return self.weight * x
class FeedForward(nn.Module):
def __init__(self, d_model: int, mult: int = 4, dropout: float = 0.0):
super().__init__()
inner = d_model * mult
self.net = nn.Sequential(
nn.Linear(d_model, inner * 2), # GEGLU
nn.GLU(dim=-1),
nn.Linear(inner, d_model),
nn.Dropout(dropout),
)
def forward(self, x):
return self.net(x)
# ---------------------
# SSW Components
# ---------------------
class LocalTextureConv(nn.Module):
"""Depthwise 1D conv + GLU gate. Causal padding. O(n * d * k) with small k."""
def __init__(self, d_model: int, kernel_size: int = 7):
super().__init__()
assert kernel_size % 2 == 1, "kernel_size should be odd for simple causal pad"
self.dw = nn.Conv1d(d_model, d_model, kernel_size, groups=d_model, padding=kernel_size-1)
self.pw = nn.Conv1d(d_model, 2 * d_model, 1)
def forward(self, x):
# x: (B, T, C)
x_c = x.transpose(1, 2) # (B, C, T)
y = self.dw(x_c)
T = x.size(1)
y = y[..., :T] # causal crop
y = self.pw(y).transpose(1, 2) # (B, T, 2C)
y = F.glu(y, dim=-1) # (B, T, C)
return y
class GlobalStatePropagation(nn.Module):
"""Simplified selective SSM-like recurrence (toy, readable)."""
def __init__(self, d_model: int, state_size: int = 128):
super().__init__()
self.state_size = state_size
self.inp = nn.Linear(d_model, state_size * 3)
self.out = nn.Linear(state_size, d_model)
def forward(self, x):
B, T, _ = x.size()
u, f, r = self.inp(x).chunk(3, dim=-1)
f = torch.sigmoid(f)
r = torch.sigmoid(r)
u = torch.tanh(u)
h = torch.zeros(B, self.state_size, device=x.device, dtype=x.dtype)
outs = []
for t in range(T):
h = f[:, t] * h + (1 - f[:, t]) * u[:, t]
outs.append(r[:, t] * h)
y = torch.stack(outs, dim=1) # (B, T, S)
return self.out(y) # (B, T, C)
class ContentBasedSummarizer(nn.Module):
"""Top-k sparse attention over history (causal)."""
def __init__(self, d_model: int, top_k: int = 8):
super().__init__()
self.k = top_k
self.q = nn.Linear(d_model, d_model, bias=False)
self.kv = nn.Linear(d_model, 2 * d_model, bias=False)
self.scale = 1.0 / math.sqrt(d_model)
self.scorer = nn.Linear(d_model, 1, bias=False)
def forward(self, x):
B, T, C = x.size()
q = self.q(x)
k, v = self.kv(x).chunk(2, dim=-1)
imp = self.scorer(x).squeeze(-1) # (B, T)
out = torch.zeros_like(x)
for t in range(T):
topk = min(self.k, t + 1)
vals, idx = torch.topk(imp[:, :t+1], k=topk, dim=-1)
k_sel = torch.gather(k[:, :t+1, :], 1, idx.unsqueeze(-1).expand(-1, -1, C))
v_sel = torch.gather(v[:, :t+1, :], 1, idx.unsqueeze(-1).expand(-1, -1, C))
q_t = q[:, t:t+1, :]
att = torch.matmul(q_t, k_sel.transpose(1, 2)) * self.scale
att = F.softmax(att, dim=-1)
out[:, t:t+1, :] = torch.matmul(att, v_sel)
return out
class WeaverBlock(nn.Module):
def __init__(self, d_model: int, ltc_kernel: int, gsp_state: int, cbs_topk: int, dropout: float):
super().__init__()
self.norm1 = RMSNorm(d_model)
self.ltc = LocalTextureConv(d_model, kernel_size=ltc_kernel)
self.gsp = GlobalStatePropagation(d_model, state_size=gsp_state)
self.cbs = ContentBasedSummarizer(d_model, top_k=cbs_topk)
self.mix = nn.Linear(d_model * 3, d_model)
self.dropout = nn.Dropout(dropout)
self.norm2 = RMSNorm(d_model)
self.ff = FeedForward(d_model, mult=4, dropout=dropout)
def forward(self, x):
h = self.norm1(x)
a = self.ltc(h)
b = self.gsp(h)
c = self.cbs(h)
h = self.mix(torch.cat([a, b, c], dim=-1))
x = x + self.dropout(h)
x = x + self.ff(self.norm2(x))
return x
class SSWLM(nn.Module):
def __init__(self, vocab_size: int, d_model: int = 512, n_layers: int = 8,
ltc_kernel: int = 7, gsp_state: int = 128, cbs_topk: int = 8,
dropout: float = 0.1, max_seq_len: int = 1024):
super().__init__()
self.tok_emb = nn.Embedding(vocab_size, d_model)
self.pos_emb = nn.Embedding(max_seq_len, d_model)
self.layers = nn.ModuleList([
WeaverBlock(d_model, ltc_kernel, gsp_state, cbs_topk, dropout)
for _ in range(n_layers)
])
self.norm = RMSNorm(d_model)
self.head = nn.Linear(d_model, vocab_size, bias=False)
self.max_seq_len = max_seq_len
def forward(self, input_ids: torch.Tensor):
B, T = input_ids.size()
assert T <= self.max_seq_len, "sequence too long"
pos = torch.arange(T, device=input_ids.device)
x = self.tok_emb(input_ids) + self.pos_emb(pos)[None, :, :]
for blk in self.layers:
x = blk(x)
x = self.norm(x)
return self.head(x)
@torch.no_grad()
def generate(
self,
input_ids: torch.Tensor,
max_new_tokens: int = 100,
temperature: float = 1.0,
top_p: float = 0.9,
top_k: int = 50,
repetition_penalty: float = 1.1,
eos_token_id: Optional[int] = None,
):
self.eval()
for _ in range(max_new_tokens):
inp = input_ids[:, -self.max_seq_len:]
logits = self.forward(inp)[:, -1, :] / max(1e-6, temperature)
# repetition penalty (simple): downweight already seen token logits
if repetition_penalty and repetition_penalty > 1.0:
for b in range(input_ids.size(0)):
seen = torch.bincount(input_ids[b], minlength=logits.size(-1)).bool()
logits[b, seen] /= repetition_penalty
# top-k filter
if top_k and top_k > 0:
k = min(top_k, logits.size(-1))
topk_vals, topk_idx = torch.topk(logits, k=k, dim=-1)
mask = torch.full_like(logits, float("-inf"))
logits = mask.scatter(1, topk_idx, topk_vals)
# nucleus (top-p) filter
if top_p < 1.0:
sorted_logits, sorted_idx = torch.sort(logits, descending=True)
probs = torch.softmax(sorted_logits, dim=-1)
cumsum = torch.cumsum(probs, dim=-1)
cutoff = cumsum > top_p
cutoff[..., 0] = False # keep at least one
sorted_logits[cutoff] = float("-inf")
# unsort back
inv_idx = torch.argsort(sorted_idx, dim=-1)
logits = torch.gather(sorted_logits, 1, inv_idx)
probs = torch.softmax(logits, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
input_ids = torch.cat([input_ids, next_token], dim=1)
if eos_token_id is not None and (next_token == eos_token_id).all():
break
return input_ids |