File size: 9,176 Bytes
b5d4048 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 | """
model.py — MechanismBase
========================
The transformer decoder implementing P / G → Q.
Two configurations:
SmallConfig (~10M params) — appropriate for ~200K tokens.
Generalizes. Recommended for current corpus.
FullConfig (~235M params) — appropriate for ~2M+ tokens.
Use after expanding the training corpus.
Architecture maps to PL terminology:
wte — token embedding: seeds patterns P with initial loaded history
wpe — position encoding: adds positional loaded history
PropagationBlock — one complete P / G → Q step:
attention = gradient family G applied to P
residual = loaded history H_P accumulating
pre-norm = coherence check before each propagation
MLP = reconfiguration toward coherent state
ln_f — final coherence check
lm_head — output: weight-tied to wte (same carrier in and out)
Parameter counts (approximate):
SmallConfig: 10.5M params
FullConfig: 235.0M params
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from dataclasses import dataclass
# =============================================================================
# CONFIGURATIONS
# =============================================================================
@dataclass
class SmallConfig:
"""
~10M params. Appropriate for 100K–500K tokens.
This is the working configuration for the current corpus (~200K tokens).
Trains in ~30 minutes on RTX 4060 Ti.
Will generalize, not just memorize.
"""
vocab_size: int = 16384 # Carrier V — BPE tokenizer
n_embd: int = 256 # Loaded history vector dimension
n_layer: int = 8 # Propagation steps
n_head: int = 8 # Gradient families per step
block_size: int = 256 # Context window
dropout: float = 0.1
name: str = "SmallBase"
@dataclass
class MediumConfig:
"""
~50M params. Appropriate for 500K–2M tokens.
Use after expanding generate_data.py to produce more derivation traces.
Trains in ~2-3 hours on RTX 4060 Ti.
"""
vocab_size: int = 16384
n_embd: int = 512
n_layer: int = 12
n_head: int = 8
block_size: int = 256
dropout: float = 0.1
name: str = "MediumBase"
@dataclass
class FullConfig:
"""
~235M params. The full AGI Base V1.
Appropriate for 2M+ tokens.
Requires expanding generate_data.py significantly (see comments there).
Trains in ~6 hours on RTX 4060 Ti when data is sufficient.
"""
vocab_size: int = 16384
n_embd: int = 1024
n_layer: int = 16
n_head: int = 16
block_size: int = 256
dropout: float = 0.1
name: str = "FullBase"
# Default: SmallConfig for the current corpus
MechanismConfig = SmallConfig
# =============================================================================
# PROPAGATION BLOCK
# =============================================================================
class PropagationBlock(nn.Module):
"""
One complete P / G → Q propagation step.
Attention : gradient family G applied to pattern P
Residual : loaded history H_P accumulating
LayerNorm : coherence threshold check (pre-norm: check BEFORE propagating)
MLP : reconfiguration toward coherent state
"""
def __init__(self, config):
super().__init__()
self.ln1 = nn.LayerNorm(config.n_embd)
self.attn = nn.MultiheadAttention(
config.n_embd,
config.n_head,
dropout=config.dropout,
batch_first=True,
)
self.ln2 = nn.LayerNorm(config.n_embd)
self.mlp = nn.Sequential(
nn.Linear(config.n_embd, 4 * config.n_embd),
nn.GELU(),
nn.Linear(4 * config.n_embd, config.n_embd),
nn.Dropout(config.dropout),
)
self.drop = nn.Dropout(config.dropout)
def forward(self, x, attn_mask=None):
# Pre-norm: coherence check before gradient application
normed = self.ln1(x)
attn_out, _ = self.attn(
normed, normed, normed,
attn_mask=attn_mask,
need_weights=False,
)
# Residual accumulates loaded history
x = x + self.drop(attn_out)
x = x + self.mlp(self.ln2(x))
return x
# =============================================================================
# MECHANISMBASE
# =============================================================================
class MechanismBase(nn.Module):
"""
The mechanism instantiated in the weight carrier.
wte : token embedding — seeds patterns
wpe : position encoding — adds positional loaded history
h : propagation blocks
ln_f : final coherence check
lm_head : output (weight-tied to wte)
"""
def __init__(self, config):
super().__init__()
self.config = config
self.wte = nn.Embedding(config.vocab_size, config.n_embd)
self.wpe = nn.Embedding(config.block_size, config.n_embd)
self.drop = nn.Dropout(config.dropout)
self.h = nn.ModuleList(
[PropagationBlock(config) for _ in range(config.n_layer)]
)
self.ln_f = nn.LayerNorm(config.n_embd)
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
# Weight tying: input and output in the same carrier
self.lm_head.weight = self.wte.weight
self.apply(self._init_weights)
def _init_weights(self, module):
if isinstance(module, nn.Linear):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
def forward(self, idx, targets=None):
B, T = idx.shape
assert T <= self.config.block_size, \
f"Sequence length {T} exceeds block_size {self.config.block_size}"
positions = torch.arange(T, device=idx.device)
x = self.drop(self.wte(idx) + self.wpe(positions))
# Causal mask: patterns attend only to prior loaded history
causal_mask = nn.Transformer.generate_square_subsequent_mask(
T, device=idx.device
)
for block in self.h:
x = block(x, attn_mask=causal_mask)
x = self.ln_f(x)
logits = self.lm_head(x)
loss = None
if targets is not None:
loss = F.cross_entropy(
logits.view(-1, logits.size(-1)),
targets.view(-1),
)
return logits, loss
@torch.no_grad()
def generate(
self,
idx,
max_new_tokens: int = 200,
temperature: float = 0.8,
top_k: int = 50,
top_p: float = 0.9,
):
"""
Autoregressive generation with temperature + top-k + top-p sampling.
"""
self.eval()
for _ in range(max_new_tokens):
x = idx[:, -self.config.block_size:]
logits, _ = self(x, None)
next_logits = logits[0, -1, :] / temperature
# Top-k
if top_k > 0:
k = min(top_k, next_logits.size(-1))
topk_vals, _ = torch.topk(next_logits, k)
next_logits[next_logits < topk_vals[-1]] = float("-inf")
# Top-p
if top_p < 1.0:
sorted_logits, sorted_idx = torch.sort(next_logits, descending=True)
cumprobs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
remove = (cumprobs - F.softmax(sorted_logits, dim=-1)) > top_p
sorted_logits[remove] = float("-inf")
next_logits = torch.zeros_like(next_logits).scatter_(
0, sorted_idx, sorted_logits
)
probs = F.softmax(next_logits, dim=-1)
next_id = torch.multinomial(probs, num_samples=1)
idx = torch.cat([idx, next_id.unsqueeze(0)], dim=1)
return idx
def count_parameters(self) -> int:
return sum(p.numel() for p in self.parameters())
def parameter_summary(self) -> str:
total = self.count_parameters()
embed = self.wte.weight.numel()
lines = [
f" Configuration: {self.config.name}",
f" Total params: {total:,}",
f" Embed params: {embed:,} ({embed/total:.1%} of total)",
f" n_embd={self.config.n_embd}, "
f"n_layer={self.config.n_layer}, "
f"n_head={self.config.n_head}",
]
return "\n".join(lines)
if __name__ == "__main__":
for ConfigClass in [SmallConfig, MediumConfig, FullConfig]:
config = ConfigClass()
model = MechanismBase(config)
print(model.parameter_summary())
print()
|