File size: 20,104 Bytes
482e70b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 | """G1-FS16 nucleus: M3 byte re-decoding executor + order-sensitive slot
compiler (spec v4.3, G1-FS16 preregistration).
This replaces the falsified G1-DR field. The three registered defects it
repairs, by construction:
* ORDERED WORDS. The compiler scores a word as a sum of per-slot dot
products <h_j, e_{w_j}> (slots encoded by ONE stationary projection of
the spine's own residual, codes as exchangeable embeddings, one shared
blank embedding for empty slots) — s(PQ) != s(QP), permutation-
equivariant, extrapolates to any preregistered depth cap with no
learned length bias.
* TRUE M3 EXECUTION. A word acts on a byte through the chain
u_0 = S(x); v_j = F_{w_j}(u_{j-1});
l_j = b(x_{j-1}) + T_fix(v_j - u_{j-1}); p_j = softmax(l_j);
soft re-seed u_j = sum_z p_j(z) S_z (training)
hard re-seed u_j = S(argmax p_j) (deploy)
where b is the model's OWN context-free byte decode (spine embed ->
rmsnorm -> head, shared tensors, no recurrence): it cannot see the
program, so the ONLY program-dependent path into an answer byte is the
latched word. The empty word contributes exactly b(x) — zero
correction. T_fix is fixed at init and never trained: no learned head
can turn the displacement readout into a lookup.
* EXCHANGEABLE, FULL-RANK ACTIONS. F_k = I + near-zero iid init, one per
code, no code-specific meaning anywhere; full rank by preregistration
so "rank starvation" is not an available excuse.
A0: every trainable tensor here is gradient-traceable from the exact
event codelength; the executor receives only carrier + exchangeable code
index; the compiler reads only the spine's own residual at fixed-stride
program slots; deploy is hard argmax.
"""
from __future__ import annotations
import math
from dataclasses import dataclass
from typing import List, Optional, Tuple
import torch
from torch import Tensor, nn
class _STOneHot(torch.autograd.Function):
"""Exact one-hot forward, identity backward onto the softmax."""
@staticmethod
def forward(ctx, soft: Tensor) -> Tensor:
return torch.nn.functional.one_hot(
soft.argmax(-1), soft.shape[-1]).to(soft.dtype)
@staticmethod
def backward(ctx, grad: Tensor) -> Tensor:
return grad
def _haar(n: int, c: int) -> Tensor:
"""n Haar-distributed orthogonal c x c matrices."""
q, r = torch.linalg.qr(torch.randn(n, c, c))
return q * torch.sign(torch.diagonal(r, dim1=-2, dim2=-1)).unsqueeze(-2)
@dataclass(frozen=True)
class M3Config:
d_model: int # spine residual width (compiler input)
n_codes: int = 2
carrier: int = 256
d_compile: int = 32 # slot/code embedding width
n_slots: int = 16 # program slot count (fixed-stride law)
train_depth: int = 10 # words up to this depth are trained
closure_depth: int = 16 # preregistered global cap (eval closure)
tie_actions: bool = False # DENSE-SHARED control: one shared action
act_init: float = 4.0 # Haar scale on F_k = I + act_init * Q_k
byte_local: bool = True # stationary compiler; False replays the 24k matrix
class M3Field(nn.Module):
def __init__(self, cfg: M3Config):
super().__init__()
self.cfg = cfg
C, K = cfg.carrier, cfg.n_codes
self.S = nn.Parameter(torch.randn(256, C) * (1.0 / math.sqrt(C)))
eye = torch.eye(C)
n_act = 1 if cfg.tie_actions else K
# Actions start SEPARATED, not near-identity. At the old scale
# (0.02/sqrt(C)) the displacement logits of two different codes had
# std 0.0013 against base logits of std 0.16, so no gradient could
# tell the codes apart and the compiler had nothing to select
# between: the field sat idle because it was born degenerate, not
# because compression declined to use it. A Haar-orthogonal
# perturbation keeps every action invertible (the semigroup stays a
# group at init) while putting the codes O(1) apart.
self.F_raw = nn.Parameter(
eye.unsqueeze(0).repeat(n_act, 1, 1)
+ cfg.act_init * _haar(n_act, C))
T = torch.randn(256, C) / math.sqrt(C)
self.register_buffer("T_fix", T, persistent=True) # never trained
# Exactly ONE compiler is allocated. Carrying both would leave a
# trainable tensor in the checkpoint that no term of L_PC can reach,
# which the A0 gradient-traceability audit rejects -- correctly.
# Byte-local compiler. The contextual compiler reads causal spine
# states, so the SAME program byte can select different codes after
# different prefixes and nothing forces w(uv) = w(u)w(v). That let the
# optimizer settle on prefix-dependent partial programs which raise
# exact-match while destroying word purity -- one of the two failure
# signatures of the 24k matrix. Reading the raw byte embedding makes
# the symbol map stationary by construction, so concatenation holds
# mechanically rather than by hope.
#
# A0: this supplies no code meanings and no labels. One embedding and
# one scorer serve all 256 bytes, the code columns stay exchangeable,
# and its gradient arrives only through R_A / the mirror step and the
# shared byte likelihood.
if cfg.byte_local:
self.E_byte = nn.Parameter(torch.randn(256, cfg.d_compile) * 0.2)
self.W_b = nn.Linear(cfg.d_compile, cfg.d_compile, bias=False)
else:
self.W_c = nn.Linear(cfg.d_model, cfg.d_compile, bias=False)
self.e_code = nn.Parameter(torch.randn(K, cfg.d_compile) * 0.2)
self.e_blank = nn.Parameter(torch.randn(cfg.d_compile) * 0.2)
self._plan_cache: dict = {}
@property
def F(self) -> Tensor:
"""Per-code actions; DENSE-SHARED ties every code to one action."""
if self.cfg.tie_actions:
return self.F_raw.expand(self.cfg.n_codes, -1, -1)
return self.F_raw
# ---------------- compiler ----------------
def _symbols(self) -> Tensor:
return torch.cat([self.e_code, self.e_blank.unsqueeze(0)], dim=0)
def slot_logits(self, prog_states: Tensor) -> Tensor:
"""CONTEXTUAL compiler, retained for the registered 24k matrix only.
(B, n_slots, d_model) spine residuals at program slots ->
(B, n_slots, K+1) per-slot symbol scores (last column = blank).
Prefer `slot_logits_bytes`: this reads causal spine state, so the
symbol map is not stationary and concatenation is not guaranteed."""
if self.cfg.byte_local:
raise RuntimeError(
"contextual compiler not allocated under byte_local=True; "
"set M3Config(byte_local=False) to replay the 24k matrix")
h = self.W_c(prog_states) / math.sqrt(self.cfg.d_compile)
return torch.einsum("bsd,kd->bsk", h, self._symbols())
def slot_logits_bytes(self, prog_bytes: Tensor) -> Tensor:
"""Byte-local compiler: (B, n_slots) int64 program bytes ->
(B, n_slots, K+1) per-slot symbol scores.
Depends on the current byte alone, so g(v) = argmax_k s(v, k) is a
fixed symbol map and w(c_1..c_d) = g(c_1)..g(c_d) holds mechanically.
The only residual ambiguity is the legitimate global permutation of
code identities, which the controls already account for."""
if not self.cfg.byte_local:
raise RuntimeError("byte-local compiler not allocated")
r = self.E_byte[prog_bytes]
h = self.W_b(r) / math.sqrt(self.cfg.d_compile)
return torch.einsum("bsd,kd->bsk", h, self._symbols())
def word_scores(self, slot_logits: Tensor,
words: List[Tuple[int, ...]]) -> Tensor:
"""Score every word: sum over its code slots + blanks after."""
B, S, _ = slot_logits.shape
K = self.cfg.n_codes
blank = slot_logits[:, :, K] # (B,S)
blank_suffix = torch.flip(
torch.cumsum(torch.flip(blank, [1]), dim=1), [1])
zero = torch.zeros(B, 1, device=slot_logits.device,
dtype=slot_logits.dtype)
blank_suffix = torch.cat([blank_suffix, zero], dim=1) # (B,S+1)
scores = []
for w in words:
s = blank_suffix[:, len(w)]
for j, c in enumerate(w):
s = s + slot_logits[:, j, c]
scores.append(s)
return torch.stack(scores, dim=1) # (B,W)
def word_scores_indexed(self, slot_logits: Tensor,
onehot: Tensor) -> Tensor:
"""Vectorized `word_scores` for a large fixed alphabet.
`onehot` is (W, n_slots, K+1): slot j of word w selects its code
for j < |w| and the blank symbol for j >= |w|. Mathematically
identical to `word_scores`, which the tests pin.
"""
return torch.einsum("bsk,wsk->bw", slot_logits, onehot)
def alphabet_onehot(self, words: List[Tuple[int, ...]]) -> Tensor:
K, S = self.cfg.n_codes, self.cfg.n_slots
oh = torch.zeros(len(words), S, K + 1)
for w, word in enumerate(words):
for j in range(S):
oh[w, j, word[j] if j < len(word) else K] = 1.0
return oh
def argmax_word(self, slot_logits: Tensor) -> List[Tuple[int, ...]]:
"""Exact hard argmax over the FULL closure alphabet, factorized:
the best word of each length d takes the per-slot best code for
slots < d and blanks after; then argmax over d <= closure cap."""
B, S, _ = slot_logits.shape
K = self.cfg.n_codes
best_code, best_idx = slot_logits[:, :, :K].max(dim=2) # (B,S)
blank = slot_logits[:, :, K]
code_prefix = torch.cumsum(best_code, dim=1)
zero = torch.zeros(B, 1, device=slot_logits.device,
dtype=slot_logits.dtype)
code_prefix = torch.cat([zero, code_prefix], dim=1) # (B,S+1)
blank_suffix = torch.flip(
torch.cumsum(torch.flip(blank, [1]), dim=1), [1])
blank_suffix = torch.cat([blank_suffix, zero], dim=1)
D = self.cfg.closure_depth
totals = torch.stack(
[code_prefix[:, d] + blank_suffix[:, d] for d in range(D + 1)],
dim=1) # (B,D+1)
dbest = totals.argmax(dim=1)
out: List[Tuple[int, ...]] = []
for b in range(B):
d = int(dbest[b])
out.append(tuple(int(best_idx[b, j]) for j in range(d)))
return out
# ---------------- executor ----------------
@staticmethod
def _straight_through(soft: Tensor) -> Tensor:
"""One-hot forward, softmax gradient backward.
The M3 bottleneck is only a BYTE if the training forward pass is the
deploy forward pass. With a soft mixture the carrier is re-seeded
from a convex combination of all 256 byte embeddings, which carries
far more than 8 bits and is strictly more expressive than anything
deploy can do — the relaxation is a continuous scratchpad, and the
entropy charge was the price levied to discourage using it. Forcing
the forward pass onto the one-hot removes the scratchpad by
construction instead of by price, so train and deploy compute the
identical function and the charge has nothing left to buy.
A custom Function rather than the usual `oh + p - p.detach()`: that
idiom evaluates (oh + p) - p in floating point and is NOT exactly oh,
so the train/deploy identity would hold only to ~1e-7. The identity
is the entire justification for dropping the entropy charge, so it
is made exact.
"""
return _STOneHot.apply(soft)
def chain(self, x: Tensor, word: Tuple[int, ...],
base_fn, hard: bool, st: bool = False
) -> Tuple[Tensor, Tensor]:
"""Run the M3 chain for one word on a batch of bytes x (N,).
Returns (final_logits (N,256), total_intermediate_entropy (N,)).
`base_fn(probs_or_ids)` returns the model's own context-free
decode logits either from hard ids (N,) or soft byte probs
(N,256).
"""
N = x.shape[0]
u = self.S[x] # (N,C)
ent = x.new_zeros(N, dtype=torch.float32)
prev_hard: Optional[Tensor] = x
prev_soft: Optional[Tensor] = None
logits = base_fn(x) # empty word
for j, c in enumerate(word):
v = torch.einsum("cd,nd->nc", self.F[c], u)
base = base_fn(prev_hard if prev_soft is None else prev_soft)
logits = base + torch.einsum("zc,nc->nz", self.T_fix, v - u)
p = torch.softmax(logits, dim=-1)
if j < len(word) - 1:
ent = ent + (-(p * (p + 1e-12).log()).sum(-1)
/ math.log(2.0))
if hard:
prev_hard, prev_soft = logits.argmax(-1), None
u = self.S[prev_hard]
elif st:
thru = self._straight_through(p)
prev_soft, prev_hard = thru, None
u = thru @ self.S
else:
prev_soft, prev_hard = p, None
u = p @ self.S
return logits, ent
def _tree_plan(self, words: List[Tuple[int, ...]],
device: torch.device) -> "_TreePlan":
key = (tuple(words), str(device))
plan = self._plan_cache.get(key)
if plan is None:
plan = _TreePlan(words, self.cfg.n_codes, device)
self._plan_cache[key] = plan
return plan
def tree_execute(self, x: Tensor, words: List[Tuple[int, ...]],
base_fn, hard: bool = False, st: bool = False
) -> Tuple[Tensor, Tensor]:
"""Execute EVERY word in `words` on bytes x, sharing prefixes.
`words` must be prefix-closed and contain the empty word. Each
prefix's chain step runs exactly once, so an alphabet of
2^(D+1)-1 words costs 2^(D+1)-1 steps rather than sum |w|.
Returns per-word (FULL M3 logits (W,N,256), intermediate entropy
(W,N)) — the same quantity `chain` returns, so the two are
interchangeable and the oracle's capacity certificate transfers.
This used to return only the final step's DISPLACEMENT, leaving the
caller to supply a base. The probe supplied the spine's contextual
answer logits, which silently replaced the law's own
b(z_{L-1}) term: latent execution then differed from the executor
the oracle certifies, so the certificate guaranteed nothing about
what was actually measured. The base belongs to the law, not to
the caller. A displacement, if one is wanted, is `logits - base_fn(x)`
which is exactly zero for the empty word.
The walk runs one DEPTH LEVEL at a time, not one word at a time:
every node at a level shares the same handful of code actions, so
a level is a constant number of batched kernels regardless of its
width. Word-at-a-time was arithmetically identical but issued
~5 tiny kernels per node, and at the registered alphabet
(2,047 words) that launch overhead measured 3.7 s per training
step on a T4 — a 25-hour run for the preregistered 24k steps.
"""
plan = self._tree_plan(words, x.device)
N = x.shape[0]
dt = self.S.dtype
root_logits = base_fn(x) # (N,256)
u = self.S[x].unsqueeze(0) # (1,N,C)
feed: Optional[Tensor] = None # what base_fn re-reads; None = x
soft: Optional[Tensor] = None # softmax, charged by lam_mid
ent_lvl = x.new_zeros(1, N, dtype=torch.float32)
logit_levels = [root_logits.unsqueeze(0)]
ent_levels = [ent_lvl]
for lvl in range(1, plan.depth + 1):
par, counts = plan.parent[lvl], plan.counts[lvl]
u_par = u.index_select(0, par) # (M,N,C)
outs, start = [], 0
for k in range(self.cfg.n_codes):
m = counts[k]
if m == 0:
continue
outs.append(torch.einsum(
"cd,mnd->mnc", self.F[k], u_par[start:start + m]))
start += m
v = torch.cat(outs, dim=0) if len(outs) > 1 else outs[0]
d = torch.einsum("zc,mnc->mnz", self.T_fix, v - u_par)
M = d.shape[0]
if feed is None: # parents = empty
base = root_logits.unsqueeze(0).expand(M, N, 256)
ent_child = ent_lvl.index_select(0, par)
else:
base = base_fn(feed.index_select(0, par).reshape(M * N, 256)
).reshape(M, N, 256)
# the MDL charge is on the model's UNCERTAINTY at the
# re-seed point, which is the softmax even when the
# carrier is re-seeded from the hard byte
sp = soft.index_select(0, par) # (M,N,256)
h = -(sp * (sp + 1e-12).log()).sum(-1) / math.log(2.0)
ent_child = ent_lvl.index_select(0, par) + h
node_logits = base + d
logit_levels.append(node_logits)
ent_levels.append(ent_child)
if lvl < plan.depth:
logits = node_logits
soft = torch.softmax(logits, dim=-1)
if hard:
hb = logits.argmax(-1)
u = self.S[hb]
feed = torch.nn.functional.one_hot(hb, 256).to(dt)
elif st:
feed = self._straight_through(soft)
u = feed @ self.S
else:
u = soft @ self.S
feed = soft
ent_lvl = ent_child
out = torch.cat(logit_levels, dim=0).index_select(0, plan.order)
ent = torch.cat(ent_levels, dim=0).index_select(0, plan.order)
return out, ent
class _TreePlan:
"""Static level structure of a prefix-closed alphabet.
Nodes at each level are held grouped by last code so a level's actions
are a few contiguous slices instead of a per-node gather of (C,C)
matrices, which would materialize K^depth copies of the action.
"""
def __init__(self, words: List[Tuple[int, ...]], n_codes: int,
device: torch.device):
if () not in words:
raise ValueError("alphabet must contain the empty word")
self.depth = max(len(w) for w in words)
levels: List[List[Tuple[int, ...]]] = [[] for _ in
range(self.depth + 1)]
for w in words:
levels[len(w)].append(w)
for lvl in range(1, self.depth + 1):
levels[lvl].sort(key=lambda w: w[-1])
pos = [{w: i for i, w in enumerate(lv)} for lv in levels]
self.parent: List[Optional[Tensor]] = [None]
self.counts: List[Optional[List[int]]] = [None]
for lvl in range(1, self.depth + 1):
par = []
for w in levels[lvl]:
if w[:-1] not in pos[lvl - 1]:
raise ValueError(f"alphabet is not prefix-closed: {w}")
par.append(pos[lvl - 1][w[:-1]])
self.parent.append(torch.tensor(par, dtype=torch.long,
device=device))
self.counts.append([sum(1 for w in levels[lvl] if w[-1] == k)
for k in range(n_codes)])
offset, flat = 0, {}
for lv in levels:
for i, w in enumerate(lv):
flat[w] = offset + i
offset += len(lv)
self.order = torch.tensor([flat[w] for w in words],
dtype=torch.long, device=device)
|