Spaces:
Running on Zero
Running on Zero
File size: 9,168 Bytes
f37be5a | 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 | """Structured motif-aware transformer closer to the FOG hypothesis.
This variant goes beyond a single repeated motif-aware block and introduces:
1. depth-wise geometry changes (early / middle / late motif stages),
2. explicit stage-specialized feed-forward computations,
3. learnable residual scales to stabilize heterogeneous training.
"""
from __future__ import annotations
from dataclasses import dataclass
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from src.fog.config import FOGConfig
@dataclass(frozen=True)
class LayerGeometry:
stage: str
d_compare: int
d_memory: int
d_expand: int
d_gate: int
residual_scale: float
def _align_to_heads(value: int, n_heads: int) -> int:
aligned = max(n_heads, (value // n_heads) * n_heads)
if aligned < value:
aligned += n_heads
return aligned
def build_layer_geometries(cfg: FOGConfig) -> list[LayerGeometry]:
"""Construct a simple early/middle/late morphology schedule.
Early layers bias toward projection/compression.
Middle layers bias toward memory/compose.
Late layers bias toward expand/transform.
"""
geoms: list[LayerGeometry] = []
n_layers = cfg.n_layers
for idx in range(n_layers):
depth_pos = idx / max(n_layers - 1, 1)
if depth_pos < 0.34:
stage = "early"
d_compare = _align_to_heads(max(cfg.d_compare, cfg.d_model // 4), cfg.n_heads)
d_memory = _align_to_heads(max(cfg.n_heads * 12, int(cfg.d_memory * 0.75)), cfg.n_heads)
d_expand = max(cfg.d_gate * 4, int(cfg.d_expand * 0.75))
d_gate = max(cfg.d_gate, cfg.d_model // 10)
residual_scale = 0.12
elif depth_pos < 0.67:
stage = "middle"
d_compare = _align_to_heads(max(cfg.d_compare, cfg.d_model // 4), cfg.n_heads)
d_memory = _align_to_heads(max(cfg.d_memory, int(cfg.d_model * 0.875)), cfg.n_heads)
d_expand = max(cfg.d_expand, int(cfg.d_model * 2.25))
d_gate = max(cfg.d_gate, cfg.d_model // 8)
residual_scale = 0.16
else:
stage = "late"
d_compare = _align_to_heads(max(cfg.d_compare, cfg.d_model // 3), cfg.n_heads)
d_memory = _align_to_heads(max(cfg.n_heads * 12, int(cfg.d_memory * 0.875)), cfg.n_heads)
d_expand = max(cfg.d_expand, int(cfg.d_expand * 1.25))
d_gate = max(cfg.d_gate, cfg.d_model // 6)
residual_scale = 0.18
geoms.append(
LayerGeometry(
stage=stage,
d_compare=d_compare,
d_memory=d_memory,
d_expand=d_expand,
d_gate=d_gate,
residual_scale=residual_scale,
)
)
return geoms
class StructuredAttention(nn.Module):
def __init__(self, d_model: int, d_compare: int, d_memory: int, n_heads: int) -> None:
super().__init__()
assert d_compare % n_heads == 0
assert d_memory % n_heads == 0
self.n_heads = n_heads
self.compare_head_dim = d_compare // n_heads
self.memory_head_dim = d_memory // n_heads
self.d_compare = d_compare
self.d_memory = d_memory
self.q_proj = nn.Linear(d_model, d_compare)
self.k_proj = nn.Linear(d_model, d_compare)
self.v_proj = nn.Linear(d_model, d_memory)
self.out_proj = nn.Linear(d_memory, d_model)
def forward(self, x: torch.Tensor, mask: torch.Tensor | None = None) -> torch.Tensor:
b, t, _ = x.shape
q = self.q_proj(x).view(b, t, self.n_heads, self.compare_head_dim).transpose(1, 2)
k = self.k_proj(x).view(b, t, self.n_heads, self.compare_head_dim).transpose(1, 2)
v = self.v_proj(x).view(b, t, self.n_heads, self.memory_head_dim).transpose(1, 2)
scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.compare_head_dim)
if mask is not None:
scores = scores.masked_fill(mask == 0, float("-inf"))
attn = F.softmax(scores, dim=-1)
y = torch.matmul(attn, v)
y = y.transpose(1, 2).contiguous().view(b, t, self.d_memory)
return self.out_proj(y)
class StructuredMotifFFN(nn.Module):
def __init__(self, d_model: int, geometry: LayerGeometry, dropout: float) -> None:
super().__init__()
self.stage = geometry.stage
self.expand = nn.Linear(d_model, geometry.d_expand)
self.gate = nn.Linear(d_model, geometry.d_gate)
self.gate_up = nn.Linear(geometry.d_gate, geometry.d_expand)
self.drop = nn.Dropout(dropout)
if self.stage == "middle":
self.compose_proj = nn.Linear(geometry.d_expand, geometry.d_expand)
self.transform_proj = None
elif self.stage == "late":
self.compose_proj = None
self.transform_proj = nn.Linear(geometry.d_expand, geometry.d_expand)
else:
self.compose_proj = None
self.transform_proj = None
self.compress = nn.Linear(geometry.d_expand, d_model)
def forward(self, x: torch.Tensor) -> torch.Tensor:
expanded = F.silu(self.expand(x))
gate = torch.sigmoid(self.gate_up(F.silu(self.gate(x))))
h = expanded * gate
if self.compose_proj is not None:
h = h + 0.5 * F.silu(self.compose_proj(h))
if self.transform_proj is not None:
h = F.silu(self.transform_proj(h))
h = self.drop(h)
return self.compress(h)
class StructuredMotifBlock(nn.Module):
def __init__(self, d_model: int, n_heads: int, geometry: LayerGeometry, dropout: float) -> None:
super().__init__()
self.geometry = geometry
self.ln1 = nn.LayerNorm(d_model)
self.ln2 = nn.LayerNorm(d_model)
self.attn = StructuredAttention(
d_model=d_model,
d_compare=geometry.d_compare,
d_memory=geometry.d_memory,
n_heads=n_heads,
)
self.ffn = StructuredMotifFFN(d_model=d_model, geometry=geometry, dropout=dropout)
self.drop = nn.Dropout(dropout)
self.attn_scale = nn.Parameter(torch.tensor(float(geometry.residual_scale)))
self.ffn_scale = nn.Parameter(torch.tensor(float(geometry.residual_scale)))
def forward(self, x: torch.Tensor, mask: torch.Tensor | None = None) -> torch.Tensor:
x = x + self.attn_scale * self.drop(self.attn(self.ln1(x), mask))
x = x + self.ffn_scale * self.drop(self.ffn(self.ln2(x)))
return x
class StructuredMotifTransformer(nn.Module):
def __init__(self, cfg: FOGConfig) -> None:
super().__init__()
self.cfg = cfg
self.layer_geometries = build_layer_geometries(cfg)
self.tok_emb = nn.Embedding(cfg.vocab_size, cfg.d_model)
self.pos_emb = nn.Embedding(cfg.max_seq_len, cfg.d_model)
self.drop = nn.Dropout(cfg.dropout)
self.blocks = nn.ModuleList(
[
StructuredMotifBlock(
d_model=cfg.d_model,
n_heads=cfg.n_heads,
geometry=geometry,
dropout=cfg.dropout,
)
for geometry in self.layer_geometries
]
)
self.ln_f = nn.LayerNorm(cfg.d_model)
self.head = nn.Linear(cfg.d_model, cfg.vocab_size, bias=False)
self.tok_emb.weight = self.head.weight
self.register_buffer(
"_causal_mask",
torch.tril(torch.ones(cfg.max_seq_len, cfg.max_seq_len, dtype=torch.bool)).unsqueeze(0).unsqueeze(0),
persistent=False,
)
def forward(
self,
input_ids: torch.Tensor,
targets: torch.Tensor | None = None,
loss_mask: torch.Tensor | None = None,
) -> dict[str, torch.Tensor | list[dict[str, int | str]]]:
b, t = input_ids.shape
pos = torch.arange(t, device=input_ids.device).unsqueeze(0)
x = self.drop(self.tok_emb(input_ids) + self.pos_emb(pos))
mask = self._causal_mask[:, :, :t, :t]
for block in self.blocks:
x = block(x, mask)
x = self.ln_f(x)
logits = self.head(x)
loss = None
if targets is not None:
if loss_mask is not None:
flat_logits = logits.view(-1, logits.size(-1))
flat_targets = targets.view(-1)
flat_mask = loss_mask.view(-1).bool()
if flat_mask.any():
loss = F.cross_entropy(flat_logits[flat_mask], flat_targets[flat_mask])
else:
loss = torch.tensor(0.0, device=logits.device)
else:
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
geometry_summary = [
{
"stage": g.stage,
"d_compare": g.d_compare,
"d_memory": g.d_memory,
"d_expand": g.d_expand,
"d_gate": g.d_gate,
}
for g in self.layer_geometries
]
return {"logits": logits, "loss": loss, "geometry": geometry_summary}
|