File size: 18,092 Bytes
908ea05 bd692df 908ea05 bd692df 908ea05 bd692df 908ea05 bd692df 908ea05 bd692df 908ea05 bd692df 908ea05 bd692df 908ea05 bd692df 908ea05 bd692df 908ea05 bd692df 908ea05 bd692df 908ea05 bd692df 908ea05 bd692df | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 | """GEMEO-CDF v13 — audit-driven Chinchilla-correct architecture.
Per the SOTA audit (May 2026):
- Path B (CLMBR fine-tune) BLOCKED: CLMBR-T-base is HF-gated (manual approval)
- Path A adopted: small from-scratch model + KG adapters + MEDS interop
Architecture:
- 12M backbone params (Chinchilla-respecting for ~20M token corpus)
- d_model=384, n_layers=8, n_heads=6, ffn=1024, ctx=512
- SwiGLU MLP (ffn:d_model = 2.67)
- Tied embeddings (saves ~12M at vocab=32k)
- Dropout 0.1 everywhere (small-data critical)
- Block-causal attention (Diffusion Forcing)
- Per-token sigma noise (independent)
- GATED KG cross-attention (tanh(α)·xattn, α init=0)
- Layers 4, 6, 7 (3 of 8)
- Lets model learn to use KG progressively, doesn't disrupt early loss
- DF objective + LM-aux loss (joint training, paper-grade)
Sources audited:
- CoMET (Aug 2025): tokens-per-param ratio
- CLMBR (Stanford): adapter pattern for cross-site transfer
- MDLM (Sahoo 2024): masked diffusion, matches AR at equal FLOPs
- Genie (DeepMind 2024): gated cross-attention pattern
- SD3 (Esser 2024): AdaLN-Zero zero-init gates
"""
from __future__ import annotations
import math
from dataclasses import dataclass, field
import torch
import torch.nn as nn
import torch.nn.functional as F
@dataclass
class CDFv13Config:
# Vocab + sequence
vocab_size: int = 32768 # MEDS-derived (will be much smaller in practice)
mask_token: int = 32767
max_seq_len: int = 512
block_size: int = 16
# Architecture (Chinchilla-correct for ~20M tokens)
d_model: int = 384
n_heads: int = 6
n_layers: int = 8
ffn: int = 1024 # SwiGLU effective; flag below uses 2 projections
dropout: float = 0.1
emb_dropout: float = 0.1
use_swiglu: bool = True
use_rmsnorm: bool = True
tie_embeddings: bool = True
# SOTA upgrades (opt-in; default off keeps backward-compat with v13 checkpoints)
use_qk_norm: bool = False # RMSNorm on Q,K per head before RoPE (Gemma2/3-style)
use_adaln: bool = False # AdaLN-Zero (DiT/SD3) per-token sigma+cond conditioning
bidirectional: bool = False # full attention (pure masked diffusion); else block-causal
# Diffusion forcing
cond_dropout: float = 0.10
# KG conditioning (GATED adapters)
use_kg: bool = True
kg_dim: int = 3072
kg_attn_layers: list = field(default_factory=lambda: [4, 6, 7])
# Latent action
use_latent_action: bool = False # Dropped per audit (concept shaky)
n_latent_actions: int = 512
# Conditioning
n_conditions: int = 64
class RMSNorm(nn.Module):
"""Root-mean-square LayerNorm (LLaMA/Mistral style)."""
def __init__(self, d: int, eps: float = 1e-6):
super().__init__()
self.weight = nn.Parameter(torch.ones(d))
self.eps = eps
def forward(self, x: torch.Tensor) -> torch.Tensor:
norm = x.float() * torch.rsqrt(x.float().pow(2).mean(-1, keepdim=True) + self.eps)
return (norm * self.weight.float()).to(x.dtype)
class SwiGLU(nn.Module):
"""SwiGLU MLP (used in LLaMA/Gemma/Mistral)."""
def __init__(self, d_in: int, d_hidden: int, dropout: float = 0.1):
super().__init__()
self.w_gate = nn.Linear(d_in, d_hidden, bias=False)
self.w_up = nn.Linear(d_in, d_hidden, bias=False)
self.w_down = nn.Linear(d_hidden, d_in, bias=False)
self.dropout = nn.Dropout(dropout)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.dropout(self.w_down(F.silu(self.w_gate(x)) * self.w_up(x)))
class RotaryEmbedding(nn.Module):
"""RoPE (Su et al. 2021)."""
def __init__(self, dim: int, max_seq: int = 8192, base: float = 10000.0):
super().__init__()
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
t = torch.arange(max_seq).float()
freqs = torch.einsum("i,j->ij", t, inv_freq)
emb = torch.cat([freqs, freqs], dim=-1)
self.register_buffer("cos", emb.cos(), persistent=False)
self.register_buffer("sin", emb.sin(), persistent=False)
def forward(self, q, k, seq_len):
cos = self.cos[:seq_len].to(q.dtype).to(q.device)
sin = self.sin[:seq_len].to(q.dtype).to(q.device)
def rot_half(x):
x1, x2 = x.chunk(2, dim=-1)
return torch.cat([-x2, x1], dim=-1)
return (q * cos) + (rot_half(q) * sin), (k * cos) + (rot_half(k) * sin)
class PerTokenSigmaEmbed(nn.Module):
"""Sinusoidal embedding of per-position diffusion noise sigma in [0,1]."""
def __init__(self, d: int):
super().__init__()
self.d = d
self.proj = nn.Sequential(
nn.Linear(d, d), nn.SiLU(), nn.Linear(d, d),
)
def forward(self, sigma: torch.Tensor) -> torch.Tensor:
half = self.d // 2
freqs = torch.exp(
-math.log(10000.0) * torch.arange(half, device=sigma.device) / half
)
ang = sigma.float().unsqueeze(-1) * freqs
emb = torch.cat([torch.sin(ang), torch.cos(ang)], dim=-1)
return self.proj(emb)
class GatedKGCrossAttention(nn.Module):
"""Cross-attention to KG ego-subgraph, with GATED output.
`tanh(alpha) * cross_attn(x_seq, x_kg)` where alpha is a learnable scalar
initialized to 0. This means at init the cross-attention contributes
NOTHING to the residual stream, so the model trains identically to
no-KG until it discovers KG is useful. Prevents catastrophic loss
spikes on small data.
Pattern from: Genie (DeepMind 2024), Flamingo (DeepMind 2022).
"""
def __init__(self, d_model: int, kg_dim: int, n_heads: int = 8, dropout: float = 0.1):
super().__init__()
self.n_heads = n_heads
self.head_dim = d_model // n_heads
# Project KG to d_model (run inline so we don't need separate KGProjector module)
self.kg_in_proj = nn.Linear(kg_dim, d_model, bias=False)
self.q_proj = nn.Linear(d_model, d_model, bias=False)
self.kv_proj = nn.Linear(d_model, 2 * d_model, bias=False)
self.out_proj = nn.Linear(d_model, d_model, bias=False)
self.norm_q = RMSNorm(d_model)
self.norm_kv = RMSNorm(d_model)
self.dropout = nn.Dropout(dropout)
# Gate (scalar per block, init=0)
self.alpha = nn.Parameter(torch.zeros(1))
def forward(self, x_seq: torch.Tensor, kg_raw: torch.Tensor) -> torch.Tensor:
"""
x_seq: (B, T, d_model)
kg_raw: (B, N_kg, kg_dim) -- raw KG embeddings (e.g. 3072)
"""
B, T, D = x_seq.shape
kg_proj = self.kg_in_proj(kg_raw) # (B, N_kg, D)
N_kg = kg_proj.size(1)
q = self.q_proj(self.norm_q(x_seq))
kv = self.kv_proj(self.norm_kv(kg_proj))
k, v = kv.chunk(2, dim=-1)
q = q.reshape(B, T, self.n_heads, self.head_dim).transpose(1, 2)
k = k.reshape(B, N_kg, self.n_heads, self.head_dim).transpose(1, 2)
v = v.reshape(B, N_kg, self.n_heads, self.head_dim).transpose(1, 2)
out = F.scaled_dot_product_attention(
q, k, v, dropout_p=self.dropout.p if self.training else 0.0)
out = out.transpose(1, 2).reshape(B, T, D)
gate = torch.tanh(self.alpha)
return x_seq + gate * self.dropout(self.out_proj(out))
class CDFv13Block(nn.Module):
"""Pre-norm transformer block + optional gated KG cross-attn."""
def __init__(self, cfg: CDFv13Config, rope: RotaryEmbedding,
layer_idx: int):
super().__init__()
self.cfg = cfg
self.rope = rope
self.layer_idx = layer_idx
norm_cls = RMSNorm if cfg.use_rmsnorm else nn.LayerNorm
self.norm1 = norm_cls(cfg.d_model)
self.norm2 = norm_cls(cfg.d_model)
self.qkv = nn.Linear(cfg.d_model, 3 * cfg.d_model, bias=False)
self.proj = nn.Linear(cfg.d_model, cfg.d_model, bias=False)
self.head_dim = cfg.d_model // cfg.n_heads
# QK-norm: per-head RMSNorm on Q,K before RoPE (stabilises attn logits)
if cfg.use_qk_norm:
self.q_norm = RMSNorm(self.head_dim)
self.k_norm = RMSNorm(self.head_dim)
# AdaLN-Zero: per-token modulation (shift/scale/gate) for MSA + MLP
if cfg.use_adaln:
self.adaln = nn.Sequential(nn.SiLU(), nn.Linear(cfg.d_model, 6 * cfg.d_model, bias=True))
if cfg.use_swiglu:
self.mlp = SwiGLU(cfg.d_model, cfg.ffn, cfg.dropout)
else:
self.mlp = nn.Sequential(
nn.Linear(cfg.d_model, cfg.ffn, bias=False),
nn.GELU(),
nn.Linear(cfg.ffn, cfg.d_model, bias=False),
nn.Dropout(cfg.dropout),
)
self.dropout = nn.Dropout(cfg.dropout)
self.head_dim = cfg.d_model // cfg.n_heads
# Gated KG cross-attention (only in specified layers)
self.use_kg_in_layer = cfg.use_kg and layer_idx in cfg.kg_attn_layers
if self.use_kg_in_layer:
self.kg_xattn = GatedKGCrossAttention(
cfg.d_model, cfg.kg_dim, cfg.n_heads, cfg.dropout)
def forward(self, x, attn_mask, kg_raw=None, cond_vec=None):
B, T, D = x.shape
# AdaLN-Zero modulation (per-token shift/scale/gate) from sigma+cond
if self.cfg.use_adaln and cond_vec is not None:
sh_msa, sc_msa, g_msa, sh_mlp, sc_mlp, g_mlp = self.adaln(cond_vec).chunk(6, dim=-1)
else:
sh_msa = sc_msa = g_msa = sh_mlp = sc_mlp = g_mlp = None
# MSA
h = self.norm1(x)
if sc_msa is not None:
h = h * (1 + sc_msa) + sh_msa
qkv = self.qkv(h).reshape(B, T, 3, self.cfg.n_heads, self.head_dim)
q, k, v = qkv.permute(2, 0, 3, 1, 4).unbind(0)
if self.cfg.use_qk_norm:
q = self.q_norm(q); k = self.k_norm(k)
q, k = self.rope(q, k, T)
out = F.scaled_dot_product_attention(
q, k, v,
attn_mask=(~attn_mask).float().masked_fill(attn_mask, float("-inf"))[None, None],
dropout_p=self.cfg.dropout if self.training else 0.0,
)
out = out.transpose(1, 2).reshape(B, T, D)
attn_out = self.dropout(self.proj(out))
x = x + (g_msa * attn_out if g_msa is not None else attn_out)
# Gated KG cross-attn (if enabled at this layer)
if self.use_kg_in_layer and kg_raw is not None:
x = self.kg_xattn(x, kg_raw)
# MLP
h2 = self.norm2(x)
if sc_mlp is not None:
h2 = h2 * (1 + sc_mlp) + sh_mlp
mlp_out = self.mlp(h2)
x = x + (g_mlp * mlp_out if g_mlp is not None else mlp_out)
return x
class CDFv13Transformer(nn.Module):
"""Audit-compliant CDF v13: 12M backbone + KG adapters + DF objective."""
def __init__(self, cfg: CDFv13Config | None = None):
super().__init__()
self.cfg = cfg or CDFv13Config()
c = self.cfg
norm_cls = RMSNorm if c.use_rmsnorm else nn.LayerNorm
self.tok_emb = nn.Embedding(c.vocab_size, c.d_model)
self.emb_dropout = nn.Dropout(c.emb_dropout)
# Per-token sigma embedding (additive)
self.sigma_emb = PerTokenSigmaEmbed(c.d_model)
# Global condition embedding (additive, broadcast)
self.cond_emb = nn.Embedding(c.n_conditions, c.d_model)
# RoPE
self.rope = RotaryEmbedding(c.d_model // c.n_heads, max_seq=c.max_seq_len * 2)
# Blocks
self.blocks = nn.ModuleList([
CDFv13Block(c, self.rope, layer_idx=i) for i in range(c.n_layers)
])
self.final_norm = norm_cls(c.d_model)
self.head = nn.Linear(c.d_model, c.vocab_size, bias=False)
if c.tie_embeddings:
self.head.weight = self.tok_emb.weight
# Block-causal mask buffer
T = c.max_seq_len
block_id = torch.arange(T) // c.block_size
# Block-causal (Diffusion Forcing): a query may attend to its own block and
# all EARLIER blocks; future blocks are masked. mask[i,j]=True => BLOCKED.
# (Fixes a prior inverted mask that blocked the past instead of the future.)
# Set cfg.bidirectional=True for full bidirectional attention (pure masked
# diffusion / gap-fill), which disables the causal mask entirely.
if getattr(c, "bidirectional", False):
mask = torch.zeros(T, T, dtype=torch.bool)
else:
mask = block_id.unsqueeze(0) > block_id.unsqueeze(1)
self.register_buffer("block_mask", mask, persistent=False)
# Init
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
nn.init.normal_(m.weight, mean=0.0, std=0.02)
if m.bias is not None: nn.init.zeros_(m.bias)
elif isinstance(m, nn.Embedding):
nn.init.normal_(m.weight, mean=0.0, std=0.02)
# AdaLN-Zero: zero the modulation output so each block starts as identity
if self.cfg.use_adaln:
for blk in self.blocks:
nn.init.zeros_(blk.adaln[-1].weight)
nn.init.zeros_(blk.adaln[-1].bias)
def forward(self, x, sigma, cond, kg_raw=None):
B, T = x.shape
cond_vec = None
if self.cfg.use_adaln:
# AdaLN path: conditioning enters via per-token modulation, not additive
cond_vec = self.sigma_emb(sigma) + self.cond_emb(cond).unsqueeze(1)
h = self.tok_emb(x)
else:
h = self.tok_emb(x) + self.sigma_emb(sigma) + self.cond_emb(cond).unsqueeze(1)
h = self.emb_dropout(h)
mask = self.block_mask[:T, :T]
for blk in self.blocks:
h = blk(h, mask, kg_raw=kg_raw, cond_vec=cond_vec)
h = self.final_norm(h)
return self.head(h)
def diffusion_forcing_loss(self, x_clean, cond, kg_raw=None,
mode: str = "uniform") -> torch.Tensor:
"""Standard absorbing-state DF loss with per-token sigma.
mode: 'uniform' (default — safer for discrete than logit-normal per audit)
'logit_normal' (SD3-style — keep as ablation only)
"""
B, T = x_clean.shape
device = x_clean.device
# CFG cond dropout
drop = torch.rand(B, device=device) < self.cfg.cond_dropout
cond = torch.where(drop, torch.zeros_like(cond), cond)
if kg_raw is not None:
drop_kg = (torch.rand(B, device=device) < self.cfg.cond_dropout).float()
kg_raw = kg_raw * (1 - drop_kg).reshape(B, 1, 1)
# Sample per-token sigma
if mode == "logit_normal":
sigma = torch.sigmoid(torch.randn(B, T, device=device)).clamp(0.01, 0.99)
else:
sigma = torch.rand(B, T, device=device).clamp(0.01, 0.99)
# Absorbing-state corruption
corrupt = torch.rand(B, T, device=device) < sigma
x_noisy = torch.where(corrupt, self.cfg.mask_token, x_clean)
logits = self.forward(x_noisy, sigma, cond, kg_raw=kg_raw)
ce = F.cross_entropy(
logits.reshape(-1, self.cfg.vocab_size),
x_clean.reshape(-1),
reduction="none",
).reshape(B, T)
n = corrupt.float().sum().clamp(min=1.0)
return (ce * corrupt.float()).sum() / n
@staticmethod
def recurrence_weights(x_clean, struct_ids, lam: float = 0.25, w_min: float = 0.02):
"""RAVEN recurrence-aware weights (Rajamohan et al., arXiv 2603.24562).
w[i,t] = max(lam ** count, w_min), where `count` is the number of prior
occurrences of token x[i,t] earlier in patient i's sequence. First
occurrences get full weight; repeats decay geometrically toward w_min.
Structural tokens get weight 0. Vectorized (no Python Counter loop).
Returns a (B, T) float tensor on x_clean.device.
"""
B, T = x_clean.shape
device = x_clean.device
# prior-occurrence count per position via equality-with-earlier-positions
eq = (x_clean.unsqueeze(2) == x_clean.unsqueeze(1)) # (B,T,T): eq[b,t,s] = x[b,t]==x[b,s]
earlier = torch.tril(torch.ones(T, T, device=device), diagonal=-1).bool() # [t,s]=True if s<t
count = (eq & earlier.unsqueeze(0)).sum(dim=2).float() # (B,T): #earlier positions s<t with same token
w = torch.clamp(lam ** count, min=w_min)
if struct_ids:
sid = torch.tensor(sorted(struct_ids), device=device)
is_struct = (x_clean.unsqueeze(-1) == sid).any(-1)
w = w.masked_fill(is_struct, 0.0)
return w
def recurrence_aware_loss(self, x_clean, cond, struct_ids, kg_raw=None,
lam: float = 0.25, w_min: float = 0.02,
mode: str = "uniform") -> torch.Tensor:
"""Diffusion-forcing loss reweighted by RAVEN recurrence decay — the
objective that makes GEMEO predict NOVEL events, not repeats. This is the
loss used to train the released `gemeo-sus` flagship."""
B, T = x_clean.shape
device = x_clean.device
drop = torch.rand(B, device=device) < self.cfg.cond_dropout
cond = torch.where(drop, torch.zeros_like(cond), cond)
if kg_raw is not None:
drop_kg = (torch.rand(B, device=device) < self.cfg.cond_dropout).float()
kg_raw = kg_raw * (1 - drop_kg).reshape(B, 1, 1)
if mode == "logit_normal":
sigma = torch.sigmoid(torch.randn(B, T, device=device)).clamp(0.01, 0.99)
else:
sigma = torch.rand(B, T, device=device).clamp(0.01, 0.99)
corrupt = torch.rand(B, T, device=device) < sigma
x_noisy = torch.where(corrupt, self.cfg.mask_token, x_clean)
logits = self.forward(x_noisy, sigma, cond, kg_raw=kg_raw)
ce = F.cross_entropy(
logits.reshape(-1, self.cfg.vocab_size), x_clean.reshape(-1),
reduction="none").reshape(B, T)
w = self.recurrence_weights(x_clean, struct_ids, lam, w_min) * corrupt.float()
return (ce * w).sum() / w.sum().clamp(min=1.0)
|