File size: 68,209 Bytes
8022b83 c475135 8022b83 c475135 8022b83 c475135 8022b83 c475135 8022b83 c475135 8022b83 c475135 8022b83 c475135 8022b83 c475135 8022b83 c475135 8022b83 c475135 8022b83 c475135 8022b83 c475135 8022b83 c475135 8022b83 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 | """PostSemClawModel β full-architecture model assembly.
Extracted from the monolithic train.py (W1 modularization). Semantics
unchanged. Imports `GPUEngram` from `hydra.engram` and `MuonAdamW` from
`hydra.optimizer`.
Triton kernel integration status (Phase 2):
HYDRA_FUSED_BCNORM β DEFERRED. The bcnorm_fused Triton kernel fuses
LayerNorm + RoPE on B/C projections. However, mamba-ssm's Mamba3 block
uses RMSNormGated (not LayerNorm) for B/C, and RoPE is applied inside
the mamba3_siso_combined CUDA kernel via the Angles parameter. Replacing
would require either (a) monkey-patching RMSNormGated + intercepting the
fused CUDA scan β invasive, 50+ lines, high breakage risk β or (b) a
full custom Mamba3Block reimplementation. Both are out of scope for
Phase 2. The kernel is validated standalone; integration deferred to
Phase 3 when HYDRA moves to a custom SSM block.
HYDRA_FUSED_SSD β DEFERRED. The ssd_exp_trap Triton kernel implements
exponential-trapezoidal discretization as a sequential scan. mamba-ssm's
Mamba3 block delegates the entire scan + gating + output projection to
mamba3_siso_combined (a compiled CUDA kernel with tilelang). Replacing
it would require decomposing the combined kernel into constituent ops
and substituting only the scan β not feasible without a custom block.
Same Phase 3 gate as above.
Both env vars are accepted but currently no-ops (gates read, logged, but
the code path is unchanged). This avoids silent regression if someone
sets them expecting a speedup.
"""
from __future__ import annotations
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
try:
from mamba_ssm import Mamba3
except ModuleNotFoundError: # local CPU tests may run outside the HF image wheel stack
Mamba3 = None
from subsystems.hestia_mini import HestiaQAT
from subsystems.htm import HTMLayer
from subsystems.mhc_mini import ManifoldHyperConnection
from subsystems.sdr_semantic import SemanticFoldingSDR
from hydra.engram import GPUEngram
from hydra.htm_cache import htm_cache_key, htm_cache_matches
from hydra.hyena_block import HyenaBlock
from hydra.reality_bridge import RealityPoincareBridge
# GDNBlock is imported lazily inside __init__ so the `fla` dependency is
# only required when HYDRA_GDN_LAYERS is actually non-empty. Baseline
# pure-Mamba3 runs continue to work without flash-linear-attention installed.
from hydra.optimizer import MuonAdamW
from hydra.sampled_softmax import UnigramSampler, sampled_softmax_loss
try:
from subsystems.cantor_router import CantorRouter
except ModuleNotFoundError:
from archive.cantor_router import CantorRouter
def norm(x: torch.Tensor) -> torch.Tensor:
"""RMSNorm over the last dim β stateless, autocast-friendly."""
return F.rms_norm(x, (x.size(-1),))
def paired_slow_fast_orthogonality(w: torch.Tensor) -> torch.Tensor:
"""Penalty for aligned adjacent slow/fast vector pairs."""
n = (w.shape[0] // 2) * 2
if n == 0:
return w.new_zeros(())
slow = F.normalize(w[:n:2].float(), dim=-1, eps=1e-8)
fast = F.normalize(w[1:n:2].float(), dim=-1, eps=1e-8)
return (slow * fast).sum(dim=-1).square().mean().to(dtype=w.dtype)
def semantic_gaussian_mollify(
x: torch.Tensor,
std: float = 0.0,
training: bool = True,
eval_enabled: bool = False,
) -> torch.Tensor:
"""Optionally add train-time semantic Gaussian noise; disabled is identity."""
if std <= 0.0 or (not training and not eval_enabled):
return x
return x + torch.randn_like(x) * float(std)
class _LocalMamba3Fallback(nn.Identity):
"""Shape-preserving local fallback used only when mamba_ssm is absent."""
pass
class PostSemClawModel(nn.Module):
"""Full Post-SEM-Claw model assembly.
Architecture:
Token Embedding -> [Mamba3 + residual] x n_layer
-> SDR + Engram (at configured layer) -> norm -> LM head
Interface (must match prepare.py evaluate_bpb):
model(x, y, reduction='none').view(-1) -> per-token losses
model(x, y, reduction='mean') -> scalar loss
"""
def __init__(self, config):
super().__init__()
self.config = config
# Audit 2026-05-09 issue #18: HTM output caching is only safe when the
# current SDR pattern matches the cached pattern. Under MDLM each
# forward can resample a different mask pattern with identical (B, T)
# shape, so shape-only reuse silently feeds stale HTM features into
# the next forward. We gate the cache on (a) MDLM-off and (b) an
# active-index match (htm_cache_matches) below.
self._mdlm_active = os.environ.get("HYDRA_USE_MDLM", "0") == "1"
self._htm_cache_key = None # populated by forward()
# Token embedding
self.wte = nn.Embedding(config.vocab_size, config.d_model)
# Mamba-3 blocks β official mamba-ssm fused CUDA kernel. No fallbacks.
# RoPE is applied internally by the Mamba3 CUDA kernel via the Angles
# parameter; external cos/sin buffers are not needed.
#
# Hyena supplement: layers whose index appears in `config.hyena_layers`
# are instantiated as HyenaBlock instead of Mamba3. The config field
# is populated from HYDRA_HYENA_LAYERS at construction time and then
# persisted to checkpoints, so resume is safe even when the env var
# is unset. Empty tuple β all-Mamba3, byte-identical to pre-port.
_hyena_layer_set = set(getattr(config, "hyena_layers", ()) or ())
_gdn_layer_set = set(getattr(config, "gdn_layers", ()) or ())
# Hyena wins on overlap; conflict is logged at construction time.
_both = _hyena_layer_set & _gdn_layer_set
if _both:
print(f"[WARN] layers in both hyena_layers and gdn_layers; using Hyena: {sorted(_both)}", flush=True)
_gdn_layer_set -= _hyena_layer_set
if os.environ.get("HYDRA_STRICT_OPTIMAL_COMPONENTS", "0") == "1":
if _hyena_layer_set or _gdn_layer_set:
raise RuntimeError(
"HYDRA_STRICT_OPTIMAL_COMPONENTS=1 requires all layers to use Mamba3; "
f"got hyena_layers={sorted(_hyena_layer_set)} gdn_layers={sorted(_gdn_layer_set)}."
)
if Mamba3 is None:
raise RuntimeError("HYDRA_STRICT_OPTIMAL_COMPONENTS=1 requires mamba_ssm/Mamba3 to be importable.")
if _gdn_layer_set:
from hydra.gdn_block import GDNBlock # requires `fla` package
def _build_block(i: int) -> nn.Module:
if i in _hyena_layer_set:
return HyenaBlock(
d_model=config.d_model,
seq_len=config.sequence_len,
order=int(os.environ.get("HYDRA_HYENA_ORDER", "2")),
filter_order=int(os.environ.get("HYDRA_HYENA_FILTER_DIM", "64")),
)
if i in _gdn_layer_set:
return GDNBlock(
d_model=config.d_model,
n_heads=config.n_heads,
)
if Mamba3 is None:
return _LocalMamba3Fallback()
block = Mamba3(
d_model=config.d_model,
d_state=config.d_state,
expand=config.expand,
headdim=config.headdim,
is_mimo=False, # SISO path uses stable mamba3_siso_combined kernel
chunk_size=int(os.environ.get("HYDRA_MAMBA3_CHUNK", "64")), # 64 is the validated default; 128 tripped a Triton autotune hang (>8min, no progress)
is_outproj_norm=False,
dtype=torch.bfloat16,
)
return block
self.blocks = nn.ModuleList([_build_block(i) for i in range(config.n_layer)])
# Full-architecture SDR: offline semantic retina + STE (no-bypass).
self.sdr_semantic = SemanticFoldingSDR(
vocab_size=config.vocab_size,
n_bits=config.sdr_n_bits,
target_active=config.sdr_target_active,
delta_rank=config.sdr_delta_rank,
som_warmup_steps=config.sdr_som_warmup,
som_update_interval=config.sdr_som_interval,
)
# HTM spatial pooler + temporal memory (Rust, Hebbian).
self.htm = HTMLayer(
input_bits=config.sdr_n_bits,
n_columns=config.htm_n_columns,
cells_per_column=config.htm_cells_per_column,
batch_size=1, # grows lazily to actual B on first forward
seed=42,
learn=True,
reset_each_forward=True,
)
# Gradient bridge: (n_columns + anomaly) -> d_model.
self.htm_proj = nn.Linear(config.htm_n_columns + 1, config.d_model, bias=False)
# GPU Engram with Hebbian writes β runs EVERY step.
self.engram = GPUEngram(
d_model=config.d_model,
n_columns=config.engram_n_columns,
max_ngram=3,
)
self.reality_bridge = None
self.cantor = None
if os.environ.get("HYDRA_REALITY_BRIDGE", "0") == "1":
d_reality = int(os.environ.get("HYDRA_REALITY_D", "133"))
self.reality_bridge = RealityPoincareBridge(
d_model=config.d_model,
d_reality=d_reality,
l0_k=int(os.environ.get("HYDRA_REALITY_L0_K", "64")),
)
if os.environ.get("HYDRA_CANTOR_DISABLE", "0") != "1":
self.cantor = CantorRouter(
depth=int(os.environ.get("HYDRA_CANTOR_DEPTH", "7")),
d_query=d_reality,
seed=int(os.environ.get("HYDRA_CANTOR_SEED", "42")),
device=self.wte.weight.device,
)
self.engram_layer_idx = config.engram_layer_idx
# Manifold-Constrained Hyper-Connections (one per Mamba-3 block).
self.mhc = nn.ModuleList([
ManifoldHyperConnection(d_model=config.d_model, n_streams=2, sinkhorn_iters=3)
for _ in range(config.n_layer)
])
# Hestia QAT β ternary weight quantization applied post-optimizer-step.
self.hestia = HestiaQAT(enabled=True, bits=1.58)
# LM head
self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
# Learnability knob 1: Multi-Token Prediction (Llama-3 style).
# MTP_K=1 -> standard next-token. MTP_K>1 -> extra heads predict
# tokens at positions t+1, t+2, ..., t+K. Heads are weight-tied to
# lm_head (we share Parameters), so the only extra compute is
# additional CE losses; no new params. Activated via HYDRA_MTP_K.
self._mtp_k = max(1, int(os.environ.get("HYDRA_MTP_K", "1")))
# Audit 2026-05-09 issue #22 - Cluster E. UnigramSampler attached
# post-init by training setup. Stays None in environments that
# cannot build the unigram cache (CI, unit tests); the legacy
# uniform-negative path covers that case.
self._unigram_sampler = None
# Learnability knob 3: gradient checkpointing on Mamba3 blocks.
self._grad_ckpt = os.environ.get("HYDRA_GRAD_CKPT", "0") == "1"
# Learnability knob 4: doc-separator BOS masking in packed sequences.
self._doc_sep_mask = os.environ.get("HYDRA_DOC_SEP_MASK", "0") == "1"
# BOS token id is looked up lazily on first forward (requires tokenizer
# load); -1 means uninitialized.
self._bos_token_id = -1
# Learnability knob 5: explicit stop-grad on HTM tensor (htm_rust
# outputs already have requires_grad=False; this is defense-in-depth).
self._htm_stop_grad = os.environ.get("HYDRA_HTM_STOP_GRAD", "0") == "1"
# Learnability knob 6: entropy penalty coefficient on LM logits.
self._entropy_penalty = float(os.environ.get("HYDRA_ENTROPY_PENALTY", "0.0"))
# Residual dropout. Audit 2026-05-09 issue #14(a): default lowered from
# 0.2 -> 0.05. The 0.2 default combined with 1/sqrt(2L) out_proj rescale
# (audit issue #14b) and frozen dt_bias (#19) was producing zero
# block_out_rms after ~200 steps (Mamba inertness, see
# docs/sweeps/2026-04-20_inert_mamba_finding.md). 0.05 is GPT-2 style
# and leaves room for the residual stream to actually carry block
# output. Override via HYDRA_DROPOUT env var.
self.drop = nn.Dropout(float(os.environ.get("HYDRA_DROPOUT", "0.05")))
# Logits soft-capping. Default raised 15 -> 30 (audit #20): the lower
# cap was actively flattening the high end of the logit distribution
# at modern eff-vocab. Override via HYDRA_LOGIT_SOFTCAP env var.
self.softcap = float(os.environ.get("HYDRA_LOGIT_SOFTCAP", "30.0"))
# Secondary metrics storage
self._metrics = {}
# Per-layer diagnostic panel. Env-gated; zero overhead when off.
# Emits residual-contribution (delta_ratio), feature std, effective rank,
# gradient norm per layer; used to identify minimum viable n_layer + find
# entropy leakage / dead layers. See docs/depth-sweep.md.
#
# Audit 2026-05-09 issue #14(c): the inertness tripwire (training.py)
# needs `layer_*_block_out_rms` and `layer_*_block_in_rms` metrics
# available at step <= TRIPWIRE_STEP. Force the forward hooks on
# whenever the tripwire is enabled (default) so the check is not a
# no-op on production runs that don't set HYDRA_LAYER_DIAGNOSTICS=1.
# Backward (grad-norm) hooks remain gated by the original env var.
self._diag_enabled = os.environ.get("HYDRA_LAYER_DIAGNOSTICS", "0") == "1"
_tripwire_ratio = float(os.environ.get("HYDRA_INERTNESS_TRIPWIRE_RATIO", "0.01"))
_tripwire_active = _tripwire_ratio > 0.0
self._fwd_hooks_enabled = self._diag_enabled or _tripwire_active
self._diag_step = 0
self._diag_svd_every = int(os.environ.get("HYDRA_LAYER_DIAG_SVD_EVERY", "100"))
if self._diag_enabled:
# Gradient-norm backward hooks on each Mamba3 block output.
for _i, _block in enumerate(self.blocks):
def _mk_grad_hook(_layer_idx):
def _hook(module, grad_input, grad_output):
if grad_output and grad_output[0] is not None:
g = grad_output[0].detach()
self._metrics[f'layer_{_layer_idx}_grad_norm'] = float(
g.pow(2).mean().sqrt().item()
)
return _hook
_block.register_full_backward_hook(_mk_grad_hook(_i))
# Forward hooks on each Mamba3 block capture the block's OUTPUT
# directly. This is the clean measurement: unlike merge_streams()
# sampling which sees (streams + M*block_output) in bf16 β where
# small block contributions round to zero against unit-norm
# residuals β this captures `block_output` itself as produced.
# Reports both its absolute RMS norm and its ratio to the block
# INPUT's RMS norm (contribution magnitude relative to the
# residual it's added to).
# Auto-enabled when the inertness tripwire is active; otherwise
# gated by HYDRA_LAYER_DIAGNOSTICS=1 (see _fwd_hooks_enabled above).
if self._fwd_hooks_enabled:
for _i, _block in enumerate(self.blocks):
def _mk_fwd_hook(_layer_idx):
def _hook(module, inputs, output):
with torch.no_grad():
inp = inputs[0].detach().float() if inputs else None
out = output.detach().float() if isinstance(output, torch.Tensor) else None
if out is not None:
out_rms = out.pow(2).mean().sqrt().item()
self._metrics[f'layer_{_layer_idx}_block_out_rms'] = float(out_rms)
if inp is not None:
in_rms = inp.pow(2).mean().sqrt().item()
self._metrics[f'layer_{_layer_idx}_block_in_rms'] = float(in_rms)
self._metrics[f'layer_{_layer_idx}_contrib_ratio'] = float(
out_rms / (in_rms + 1e-8)
)
return _hook
_block.register_forward_hook(_mk_fwd_hook(_i))
# Triton kernel integration gates (Phase 2 β deferred, see module docstring).
self._fused_bcnorm = os.environ.get("HYDRA_FUSED_BCNORM", "0") == "1"
self._fused_ssd = os.environ.get("HYDRA_FUSED_SSD", "0") == "1"
if self._fused_bcnorm or self._fused_ssd:
import sys
_active = []
if self._fused_bcnorm:
_active.append("HYDRA_FUSED_BCNORM")
if self._fused_ssd:
_active.append("HYDRA_FUSED_SSD")
print(
f"[HYDRA] Triton kernel gates set: {', '.join(_active)}. "
f"NOTE: Both are DEFERRED (mamba-ssm Mamba3 uses internal "
f"CUDA kernels). Gates accepted but currently no-ops.",
file=sys.stderr,
)
# R6 optional torch.compile on the impl forward. Gated (default OFF).
if os.environ.get("HYDRA_MODEL_COMPILE", "0") == "1":
self._forward_impl = torch.compile(
self._forward_impl,
fullgraph=False,
dynamic=True,
mode="default",
)
@torch.no_grad()
def init_weights(self) -> None:
s = 3 ** 0.5 * self.config.d_model ** -0.5
# Move SDR retina indices (plain attribute, not buffer) to same device as params.
# Required because to_empty() only moves params/buffers, and _retina_indices
# is loaded from numpy (always CPU) by SemanticFoldingSDR.__init__.
device = self.wte.weight.device
if hasattr(self.sdr_semantic, '_retina_indices'):
self.sdr_semantic._retina_indices = self.sdr_semantic._retina_indices.to(device)
# Embedding init: GPT-2 / LLaMA convention. std=1.0 was chosen for
# vocab=8192; at larger vocabs, smaller std prevents logit blowup.
# Use std = 1/sqrt(d_model) which scales sensibly with model width.
import math as _math
_d_model = self.wte.weight.shape[1]
wte_std = float(os.environ.get("HYDRA_WTE_STD", str(1.0 / _math.sqrt(_d_model))))
nn.init.normal_(self.wte.weight, mean=0.0, std=wte_std)
# LM head init: was std=0.001 β PATHOLOGICAL at vocab>=32k because
# logits collapse to zero, loss locks at log(V)~=11, gradient through
# head β 1/V is too small to escape. GPT-2 uses std=0.02; LLaMA uses
# std=1/sqrt(d_model). Pick 0.02 as robust default, env-overridable.
lm_head_std = float(os.environ.get("HYDRA_LM_HEAD_STD", "0.02"))
nn.init.normal_(self.lm_head.weight, mean=0.0, std=lm_head_std)
# F8 (NOT APPLIED): Weight tying would save V*D params but current LR
# groups have embedding_lr=1.0 and unembedding_lr=0.005 Γ d_model_scale
# β tying forces the shared tensor under a single LR group and either
# the embeddings learn 200x too slow (under unembed LR) or the LM head
# becomes unstable (under embed LR). Short 15-step smoke with tying +
# embed-group update showed initial loss jump 9 -> 20. Deferred until
# LR groups are re-tuned; see docs/OPTIMIZATION_PLAN.md Post-plan.
for li, block in enumerate(self.blocks):
if hasattr(block, 'in_proj') and hasattr(block.in_proj, 'weight'):
nn.init.uniform_(block.in_proj.weight, -s, s)
if hasattr(block, 'out_proj') and hasattr(block.out_proj, 'weight'):
# Audit 2026-05-09 issue #14(b): out_proj init std is now FLAT
# 0.02 (GPT-2 default for non-residual heads). The previous
# default of 0.02 / sqrt(2 * n_layer) β designed for L>=24
# GPT-2-class stacks β was throwing away gradient signal at
# the L=1-4 configs feather actually trains. Combined with
# dropout=0.2 and frozen dt_bias this drove block_out_rms to
# zero after ~200 steps (Mamba inertness).
#
# Default behaviour: flat 0.02. Legacy 1/sqrt(2L) divisor is
# available via HYDRA_OUT_PROJ_DIVISOR=1 for reproducibility
# of pre-2026-05-09 runs. HYDRA_OUT_PROJ_STD overrides the
# absolute std value (applied AFTER the optional divisor).
n_layer = self.config.n_layer
_legacy_divisor = os.environ.get("HYDRA_OUT_PROJ_DIVISOR", "0") == "1"
_base_std = 0.02 / (2 * n_layer) ** 0.5 if _legacy_divisor else 0.02
out_std = float(os.environ.get("HYDRA_OUT_PROJ_STD", str(_base_std)))
nn.init.normal_(block.out_proj.weight, mean=0.0, std=out_std)
# Audit 2026-05-09 issue #19: explicit per-head dt_bias init.
# Mamba-2 paper convention: softplus(dt_bias) ~ U(dt_min, dt_max)
# so each head learns a different discretisation timescale. Mamba3
# CUDA kernel exposes `dt_bias` of shape (n_heads,). The default
# init in mamba_ssm already follows this distribution, but we
# re-apply it explicitly here both for documentation and for
# env-tunable bounds (HYDRA_DT_MIN, HYDRA_DT_MAX).
if hasattr(block, 'dt_bias') and isinstance(block.dt_bias, nn.Parameter):
_dt_min = float(os.environ.get("HYDRA_DT_MIN", "0.001"))
_dt_max = float(os.environ.get("HYDRA_DT_MAX", "0.1"))
_log_dt_min = _math.log(max(_dt_min, 1e-6))
_log_dt_max = _math.log(max(_dt_max, _dt_min * 1.0001))
# softplus^{-1}(t) = log(exp(t) - 1). For small t (< ~0.5),
# softplus^{-1}(t) ~= log(t), so we initialise log_dt and let
# the kernel's softplus produce dt ~ exp(log_dt) ~ U(dt_min, dt_max).
with torch.no_grad():
block.dt_bias.uniform_(_log_dt_min, _log_dt_max)
nn.init.normal_(self.htm_proj.weight, mean=0.0, std=s)
if hasattr(self.engram, "memory"):
nn.init.normal_(self.engram.memory, mean=0.0, std=0.01)
if hasattr(self.engram, "gate"):
nn.init.zeros_(self.engram.gate.weight)
nn.init.zeros_(self.engram.gate.bias)
if self.reality_bridge is not None:
nn.init.normal_(self.reality_bridge.to_reality.weight, mean=0.0, std=0.02)
nn.init.normal_(self.reality_bridge.to_tangent2.weight, mean=0.0, std=0.02)
if self.cantor is not None and hasattr(self.cantor, "branch"):
bound = (3.0 / float(self.cantor.d_query)) ** 0.5
nn.init.uniform_(self.cantor.branch, -bound, bound)
# Cast to bf16 to match Mamba3 dtype; Muon groups by shape so mixed
# dtypes in the same shape group would break lerp_ dtype checks.
self.wte.to(dtype=torch.bfloat16)
self.blocks.to(dtype=torch.bfloat16)
self.htm_proj.to(dtype=torch.bfloat16)
self.engram.to(dtype=torch.bfloat16)
if self.reality_bridge is not None:
self.reality_bridge.to(dtype=torch.bfloat16)
if self.cantor is not None:
self.cantor.to(dtype=torch.bfloat16)
def set_bos_token_id(self, bos_id: int) -> None:
"""Inform the model of the tokenizer's BOS id so doc-separator
masking (learnability #4) knows which positions to skip. Called from
training setup once the tokenizer is loaded."""
self._bos_token_id = int(bos_id)
def set_unigram_sampler(self, sampler) -> None:
"""Attach a UnigramSampler used by the sampled-softmax LM-head loss.
Audit 2026-05-09 issue #22 - Cluster E. With a sampler attached the
LM-head loss draws negatives from the empirical unigram distribution
and applies the per-id ``log p_unigram`` correction (Jean et al.
2015, importance-sampled NCE). Without a sampler the legacy
uniform-negative path runs (uniform proposal + ``log(V/K)`` constant
correction) - kept as a fallback for environments that cannot build
the unigram cache (no tokenizer, CI, etc.).
Eval always uses full softmax regardless of this setting.
"""
if sampler is not None and not isinstance(sampler, UnigramSampler):
raise TypeError(
f"set_unigram_sampler expects UnigramSampler or None, got {type(sampler)}"
)
self._unigram_sampler = sampler
def invalidate_hyena_caches(self) -> None:
"""Invalidate filter-rfft caches on all Hyena blocks.
MUST be called after each `optimizer.step()` when
`HYDRA_HYENA_FILTER_CACHE=1` is set, otherwise cached rfft values
will be reused with stale filter parameters.
No-op for blocks that are not HyenaBlock (Mamba3, etc.).
"""
for block in self.blocks:
if hasattr(block, "operator") and hasattr(block.operator, "invalidate_filter_cache"):
block.operator.invalidate_filter_cache()
def flush_hyena_pending_grads(self) -> None:
"""Push pending train-cache filter gradients into filter params.
Used ONLY when HYDRA_HYENA_TRAIN_CACHE=1. Must be called exactly once
per optimizer step, BEFORE `optimizer.step()` and BEFORE
`invalidate_hyena_caches()`. The lightning_module wires this in
`optimizer_step` around the existing optimizer.step() call.
No-op if:
* No HyenaBlocks are in the model, OR
* No micro-batch ever ran with grad enabled (e.g. all-eval step).
"""
for block in self.blocks:
if hasattr(block, "operator") and hasattr(block.operator, "flush_pending_filter_grads"):
block.operator.flush_pending_filter_grads()
def estimate_flops(self) -> int:
nparams = sum(p.numel() for p in self.parameters())
embed_params = self.wte.weight.numel()
return 6 * (nparams - embed_params)
def num_scaling_params(self) -> dict:
wte = sum(p.numel() for p in self.wte.parameters())
lm_head = sum(p.numel() for p in self.lm_head.parameters())
blocks = sum(p.numel() for p in self.blocks.parameters())
sdr = sum(p.numel() for p in self.sdr_semantic.parameters())
htm_proj = sum(p.numel() for p in self.htm_proj.parameters())
engram = sum(p.numel() for p in self.engram.parameters())
total = sum(p.numel() for p in self.parameters())
return {
'wte': wte, 'lm_head': lm_head, 'blocks': blocks,
'sdr_semantic': sdr, 'htm_proj': htm_proj,
'engram': engram, 'total': total,
}
def get_secondary_metrics(self) -> dict:
"""Flush any lingering CUDA tensors to host (single sync)."""
flushed = {}
for k, v in self._metrics.items():
if hasattr(v, 'item'):
try:
flushed[k] = float(v.item())
except Exception:
flushed[k] = v
else:
flushed[k] = v
return flushed
def setup_optimizer(self, unembedding_lr=0.004, embedding_lr=0.6, matrix_lr=0.04,
weight_decay=0.2, adam_betas=(0.8, 0.95), scalar_lr=0.5):
"""Setup MuonAdamW optimizer with per-component LR groups."""
model_dim = self.config.d_model
embedding_params = list(self.wte.parameters())
lm_head_params = list(self.lm_head.parameters())
# Muon routing guard: 2D parameters are NOT automatically matrices.
# Exclude:
# (a) params whose name ends in `.freq` β Sin frequency vectors used
# by Hyena's implicit filter MLP. Shape (1, dim) is nominally 2D
# but semantically a per-dim scalar. Muon's polar-express
# orthogonalization would force it toward an orthogonal matrix,
# destroying the learned modulation frequencies.
# (b) 2-D params with min(shape) < MUON_MIN_DIM. Tiny projections
# (e.g. HyenaFilter.implicit_filter.0.weight of shape (64, 3))
# get collapsed toward near-identity by orthogonalization on the
# narrow axis, damaging expressivity. These belong in AdamW.
# These exclusions route the params into the AdamW scalar/vector group.
MUON_MIN_DIM = 8
def _muon_eligible(name: str, p: torch.Tensor) -> bool:
if p.dim() != 2:
return False
if name.endswith(".freq"):
return False
if min(p.shape) < MUON_MIN_DIM:
return False
return True
# Matrix params -> Muon (2D weight matrices passing the routing guard).
matrix_params = []
for name, p in self.blocks.named_parameters():
if _muon_eligible(name, p):
matrix_params.append(p)
# NOTE (W1 audit REG-2): SemanticFoldingSDR.delta_u / delta_v are
# currently GRADIENT-DEAD. The forward path uses `binary_only(idx)` for
# HTM and stores it as `self._last_sdr`, but does NOT route the STE
# output through any downstream op. Including them in the Muon group
# burns compute (stack + orthogonalize + lerp) on zero-grad params
# every step. Excluded here; a later W5 pass can reconnect STE via a
# gated residual if the SDR signal is wanted back in-graph. The
# parameters still exist, so no state_dict break.
# for p in self.sdr_semantic.parameters():
# if p.dim() == 2:
# matrix_params.append(p)
for name, p in self.htm_proj.named_parameters():
if _muon_eligible(name, p):
matrix_params.append(p)
for name, p in self.engram.named_parameters():
if _muon_eligible(name, p):
matrix_params.append(p)
# ------------------------------------------------------------------
# Contrastive retina (Retina-D): retina_contrastive gets its OWN AdamW
# group with a dedicated LR so it can learn word2vec-like similarity
# structure without interfering with the main model LR schedule.
#
# Audit 2026-05-09 issue #21: pre-2026-05-09 this group bound to
# `retina_logits`, which was the contrastive embedding (the shadowing
# parameter β see subsystems/sdr_semantic.py). After the rename, the
# group binds to `retina_contrastive`. The learnable-mode binary
# retina_logits (shape [V, n_bits]) is now bound to the SDR delta
# group at HYDRA_SDR_LR (1e-3 default) so it actually learns instead
# of sitting at lr=0.
#
# delta_u / delta_v are still gradient-dead (REG-2), but the audit
# now routes them to a real lr group so subsequent reconnection of
# the STE through HTM (deferred work) trains them automatically.
# retina_contrastive is in-graph via contrastive_loss() every N steps.
#
# LR (contrastive): 1e-3 β higher than embedding (0.6 * scale) to
# compensate for the sparse gradient signal (only pairs from current
# batch update each row). weight_decay=0.01 provides light row-norm
# regularisation without collapsing logits toward zero.
# LR (sdr delta + binary retina_logits): HYDRA_SDR_LR=1e-3.
# ------------------------------------------------------------------
_contrastive_enabled = os.environ.get("HYDRA_CONTRASTIVE_RETINA", "0") == "1"
_contrastive_lr = float(os.environ.get("HYDRA_CONTRASTIVE_LR", "1e-3"))
_sdr_lr = float(os.environ.get("HYDRA_SDR_LR", "1e-3"))
# Separate delta params (gradient-dead) and binary retina_logits
# (learnable mode only) from retina_contrastive (in-graph when enabled).
sdr_delta_ids = {id(self.sdr_semantic.delta_u), id(self.sdr_semantic.delta_v)}
sdr_contrastive_params = []
if self.sdr_semantic.retina_contrastive is not None:
sdr_contrastive_params.append(self.sdr_semantic.retina_contrastive)
sdr_delta_params = [self.sdr_semantic.delta_u, self.sdr_semantic.delta_v]
# Binary retina_logits exists only in learnable mode. When present,
# route it through the SDR delta group with HYDRA_SDR_LR so it can
# learn alongside delta_u/delta_v rather than sit at lr=0.
sdr_binary_logits = []
_binary_logits_param = getattr(self.sdr_semantic, 'retina_logits', None)
if isinstance(_binary_logits_param, torch.nn.Parameter):
sdr_binary_logits.append(_binary_logits_param)
sdr_delta_params.append(_binary_logits_param)
# All SDR param ids β excluded from scalar group regardless.
sdr_param_ids = {id(p) for p in self.sdr_semantic.parameters()}
# Audit 2026-05-09 issue #19: dt_bias gets its own optimizer group.
# Previously rolled into the scalar AdamW group at lr=scalar_lr*scale
# (~0.5 * dmodel_lr_scale), which is far too low for the per-head
# discretization-step bias to escape its initialisation. Mamba-2 paper
# uses lr=1e-2 for dt_proj/dt_bias; replicate that here. wd=0 because
# this is a per-head bias, not a regulariseable matrix.
_dt_bias_lr = float(os.environ.get("HYDRA_DT_BIAS_LR", "1e-2"))
dt_bias_params = []
for name, p in self.blocks.named_parameters():
if name.endswith("dt_bias") or name == "dt_bias":
dt_bias_params.append(p)
dt_bias_ids = {id(p) for p in dt_bias_params}
assigned = set(id(p) for p in embedding_params + lm_head_params + matrix_params)
scalar_params = [
p for p in self.parameters()
if id(p) not in assigned
and id(p) not in sdr_param_ids
and id(p) not in dt_bias_ids
]
total_assigned = (
len(embedding_params) + len(lm_head_params) + len(matrix_params)
+ len(scalar_params) + len(sdr_delta_params) + len(sdr_contrastive_params)
+ len(dt_bias_params)
)
total_params = len(list(self.parameters()))
assert total_assigned == total_params, (
f"Parameter count mismatch: assigned {total_assigned} vs total {total_params}"
)
dmodel_lr_scale = (model_dim / 768) ** -0.5
print(f"Scaling AdamW LRs by 1/sqrt({model_dim}/768) = {dmodel_lr_scale:.6f}")
if _contrastive_enabled:
print(f"[retina-d] Contrastive retina ENABLED. retina_contrastive LR={_contrastive_lr:.2e}")
else:
print("[retina-d] Contrastive retina DISABLED (set HYDRA_CONTRASTIVE_RETINA=1 to enable)")
print(
f"[sdr] delta_u/delta_v + binary retina_logits routed to SDR group "
f"(lr={_sdr_lr:.2e}) β audit issue #21."
)
if dt_bias_params:
print(
f"[dt-bias] {len(dt_bias_params)} dt_bias parameter(s) routed to "
f"dedicated AdamW group (lr={_dt_bias_lr:.2e}, betas=(0.9, 0.95)) β "
f"audit issue #19."
)
else:
print("[dt-bias] no dt_bias parameters detected (Hyena/GDN-only stack?)")
param_groups = [
dict(kind='adamw', params=lm_head_params,
lr=unembedding_lr * dmodel_lr_scale, betas=adam_betas,
eps=1e-10, weight_decay=0.0),
dict(kind='adamw', params=embedding_params,
lr=embedding_lr * dmodel_lr_scale, betas=adam_betas,
eps=1e-10, weight_decay=0.0),
]
# Contrastive retina embedding: dedicated group, always present
# if the parameter exists, so optimizer state is consistent.
if sdr_contrastive_params:
# kind='retina_contrastive' marks this group so
# training.py skips cosine-LR scaling.
param_groups.append(
dict(kind='retina_contrastive', params=sdr_contrastive_params,
lr=_contrastive_lr, betas=adam_betas,
eps=1e-10, weight_decay=0.01)
)
# Audit 2026-05-09 issue #21: SDR delta_u/delta_v (and the
# learnable-mode binary retina_logits when present) get a real
# lr=HYDRA_SDR_LR (1e-3 default). Pre-2026-05-09 this group sat
# at lr=0 β even when STE was reconnected through HTM the
# parameters could not learn. retina_logits is gradient-active
# only in learnable mode where contrastive_loss / binary_only
# use it; in offline mode binary_only goes through _retina_data
# and these params don't accumulate updates.
param_groups.append(
dict(kind='adamw', params=sdr_delta_params,
lr=_sdr_lr, betas=adam_betas,
eps=1e-10, weight_decay=0.0)
)
# Audit 2026-05-09 issue #19: dedicated AdamW group for dt_bias. Held
# out of the cosine LR scaling applied to model groups by setting
# kind='dt_bias' (training.py treats this kind like 'retina_contrastive'
# and pins lr to its initial value). betas=(0.9, 0.95) follows the
# Mamba-2 paper convention for dt-related parameters.
if dt_bias_params:
param_groups.append(dict(
kind='dt_bias', params=dt_bias_params,
lr=_dt_bias_lr, betas=(0.9, 0.95),
eps=1e-8, weight_decay=0.0,
))
if scalar_params:
param_groups.append(
dict(kind='adamw', params=scalar_params,
lr=scalar_lr * dmodel_lr_scale, betas=adam_betas,
eps=1e-10, weight_decay=0.0)
)
for shape in sorted({p.shape for p in matrix_params}):
group_params = [p for p in matrix_params if p.shape == shape]
# ns_steps: Muon polar-express inner iterations. Default 5 (paper),
# but 3 converges on small matrices (d_model ~ 384) with ~40% lower
# optimizer step cost. Env-tunable for experimentation.
_ns_steps = int(os.environ.get("HYDRA_MUON_NS_STEPS", "3"))
param_groups.append(dict(
kind='muon', params=group_params, lr=matrix_lr,
momentum=0.95, ns_steps=_ns_steps, beta2=0.95, weight_decay=weight_decay,
))
optimizer = MuonAdamW(param_groups)
for group in optimizer.param_groups:
group["initial_lr"] = group["lr"]
return optimizer
def forward(self, idx, targets=None, reduction='mean'):
"""idx: (B, T) int64. Returns loss if targets given, else logits.
Nested bf16 autocast is a no-op when ambient autocast is already on;
when it's off (e.g. integration tests) we establish the dtype contract.
"""
if torch.is_autocast_enabled():
return self._forward_impl(idx, targets=targets, reduction=reduction)
with torch.amp.autocast(device_type="cuda", dtype=torch.bfloat16):
return self._forward_impl(idx, targets=targets, reduction=reduction)
def _forward_impl(self, idx, targets=None, reduction='mean'):
B, T = idx.shape
# Diagnostic: per-subsystem CUDA event timing. Env-gated; zero overhead
# when disabled. Logs one timing line per forward call. Used to isolate
# which subsystem is the tps bottleneck on paid hardware.
_profile = os.environ.get("HYDRA_PROFILE_FORWARD", "0") == "1"
if _profile:
def _ev():
e = torch.cuda.Event(enable_timing=True)
e.record()
return e
_t0 = _ev()
else:
_t0 = None
# Compute SDR binary ONCE and reuse for both HTM input and the stash.
sdr_binary = self.sdr_semantic.binary_only(idx)
self._last_sdr = sdr_binary # uint8 stash (not bf16 β 256MB avoidance)
# HTM subsampling: run HTM on 1 of every N micro-batches within a
# gradient accumulation step, reuse the cached result for the other
# N-1 micro-batches. Cooperative launch monopolizes all SMs (grid.sync
# requires full-grid residency), so HTM and mamba can't overlap via
# streams. Subsampling removes HTM from most micro-batches' critical
# path instead.
#
# Math: N=8, 64 accum steps β 8 HTM calls (10.6ms each) + 56 fast
# calls (4ms each). Total = 84.8 + 224 = 309ms β 106k tps.
#
# HYDRA_HTM_SUBSAMPLE=N (default 8). Set =1 for every-microbatch HTM.
_htm_sub = int(os.environ.get("HYDRA_HTM_SUBSAMPLE", "8"))
if not hasattr(self, '_htm_call_idx'):
self._htm_call_idx = int(os.environ.get("HYDRA_HTM_INITIAL_OFFSET", "0"))
_run_htm = (self._htm_call_idx % _htm_sub == 0)
self._htm_call_idx += 1
if _run_htm:
htm_handle = self.htm.forward_async(sdr_binary, output_dtype=self.wte.weight.dtype)
else:
htm_handle = None
if _profile: _t_htm_async = _ev()
dense_emb = self.wte(idx) # (B, T, d_model) bf16
dense_emb = semantic_gaussian_mollify(
dense_emb,
std=float(os.environ.get("HYDRA_SEMANTIC_SMOOTH_STD", "0.0")),
training=self.training,
eval_enabled=os.environ.get("HYDRA_SEMANTIC_SMOOTH_EVAL", "0") == "1",
)
if _profile: _t_wte = _ev()
# Audit 2026-05-09 issue #18: cache reuse must verify the current
# SDR active-index pattern matches the one the cache was built from.
# Under MDLM, mask resampling produces forwards with identical shape
# but different content; shape-only reuse fed stale HTM features.
# nonzero() is computed lazily β only when we are about to write the
# cache or test reuse β so the runtime hit on the no-cache path is
# zero.
if _run_htm:
htm_out = self.htm.forward_await(htm_handle)
self._htm_cache = htm_out.detach() # cache for non-HTM micro-batches
self._htm_cache_key = htm_cache_key(sdr_binary.nonzero())
self._htm_cache_shape = (B, T)
elif (
# Throughput mode for ordinary CE training: real GPU HTM still runs
# every HYDRA_HTM_SUBSAMPLE forwards and learns there. Between real
# HTM passes, reuse the previous same-shape HTM features instead of
# exact SDR-key matching, because streaming Nemotron changes SDR
# content every batch and exact matching degenerates to HTM every
# microbatch. Keep exact matching for MDLM/masked forwards.
self.training
and not self._mdlm_active
and os.environ.get("HYDRA_HTM_CACHE_MODE", "exact").lower() == "shape"
and hasattr(self, '_htm_cache') and self._htm_cache is not None
and getattr(self, '_htm_cache_shape', None) == (B, T)
):
htm_out = self._htm_cache
elif (
hasattr(self, '_htm_cache') and self._htm_cache is not None
and self._htm_cache.shape[0] == B and self._htm_cache.shape[1] == T
and not self._mdlm_active
and htm_cache_matches(self._htm_cache_key, sdr_binary.nonzero())
):
htm_out = self._htm_cache
elif (
os.environ.get("HYDRA_HTM_ZERO_CACHE_ON_MISS", "0") == "1"
and self.training
and not self._mdlm_active
):
htm_out = torch.zeros((B, T, self.config.htm_n_columns + 1), device=dense_emb.device, dtype=dense_emb.dtype)
self._htm_cache = htm_out.detach()
self._htm_cache_key = None
self._htm_cache_shape = (B, T)
else:
# Very first call with subsample > 1, OR MDLM is on, OR the SDR
# pattern has changed from the cached one under exact mode: run HTM.
htm_handle = self.htm.forward_async(sdr_binary, output_dtype=self.wte.weight.dtype)
htm_out = self.htm.forward_await(htm_handle)
self._htm_cache = htm_out.detach()
self._htm_cache_key = htm_cache_key(sdr_binary.nonzero())
self._htm_cache_shape = (B, T)
if _profile: _t_htm_await = _ev()
with torch.no_grad():
sdr_active_bits = float(self.sdr_semantic.target_active)
htm_anomaly = htm_out[..., -1].mean()
# Learnability #5: explicit stop-grad on HTM output. htm_rust already
# produces a detached tensor, but making it explicit here hardens the
# contract against future refactors that might route HTM through a
# grad-enabled op.
if self._htm_stop_grad:
htm_out = htm_out.detach()
# Gradient bridge: HTM columns+anomaly -> d_model.
htm_proj_out = self.htm_proj(htm_out.to(dense_emb.dtype))
x = dense_emb + htm_proj_out
x = norm(x)
if _profile: _t_htm_proj = _ev()
# mHC-routed Mamba-3 stack with Engram injection at configured layer.
streams = self.mhc[0].init_streams(x)
_engram_ev = None
# Per-layer diagnostic panel. The pre-layer merged state h_pre lets us
# measure residual contribution of each layer: delta_N = h_post - h_pre.
# All reads are detached no-grad to avoid autograd graph pollution.
_diag = self._diag_enabled
if _diag:
# Cast to float32 for the diagnostic arithmetic: the layer's
# residual contribution is small (~0.5 Γ rms-normed block output),
# which underflows in bf16 subtraction (3-digit mantissa) and
# reports delta_ratio=0 at the boundaries. float32 snapshot is
# ~3.8 MB extra memory per diag sample (B=1, T=2048, d=96) β
# negligible vs peak VRAM.
with torch.no_grad():
h_pre = self.mhc[0].merge_streams(streams).detach().float()
_run_svd = (self._diag_step % self._diag_svd_every) == 0
for i, (block, mhc_layer) in enumerate(zip(self.blocks, self.mhc)):
def _block_fn(h, _block=block):
return self.drop(_block(norm(h)))
# Learnability #3: gradient checkpointing. Wrap the block-fn so
# the mhc layer's internal uses of it re-run the block in backward
# (trading compute for activation memory). use_reentrant=False is
# the modern API and works cleanly under autocast.
if self._grad_ckpt and self.training:
import torch.utils.checkpoint as _ckpt
_raw_fn = _block_fn
def _block_fn(h, _raw=_raw_fn): # noqa: E731
return _ckpt.checkpoint(_raw, h, use_reentrant=False)
streams = mhc_layer(streams, _block_fn)
if i == self.engram_layer_idx:
if _profile: _t_pre_engram = _ev()
# mHC stream-2 preservation at the engram boundary.
# Old behavior (HYDRA_ENGRAM_RESET_STREAMS=1) re-initialized
# both streams from the merged tensor, discarding stream-1's
# accumulated state from layers 0..engram_layer_idx and
# collapsing mHC topology to n_streams=1 below the engram.
# New default: treat the engram as a residual update on
# stream-0 only, so stream-1 carries its prior state into
# the post-engram half of the network. `streams` here is a
# tensor of shape (n_streams, B, T, d_model) β see
# subsystems/mhc_mini.ManifoldHyperConnection.
x_mid = mhc_layer.merge_streams(streams)
if self.reality_bridge is not None and self.cantor is not None:
rb = self.reality_bridge(x_mid)
cantor_leaf_ids, _ = self.cantor(rb.reality, return_scores=False)
x_after_engram, hit_rate = self.engram(
x_mid,
idx,
sdr_active_indices=rb.l0_indices,
cantor_leaf_ids=cantor_leaf_ids,
cantor_n_leaves=self.cantor.n_leaves,
)
else:
x_after_engram, hit_rate = self.engram(x_mid, idx)
if os.environ.get("HYDRA_ENGRAM_RESET_STREAMS", "0") == "1":
streams = mhc_layer.init_streams(x_after_engram)
else:
engram_delta = x_after_engram - x_mid
streams = streams.clone()
streams[0] = streams[0] + engram_delta
self._metrics['engram_hit_rate'] = hit_rate
if _diag:
with torch.no_grad():
s0 = streams[0].detach().float()
s1 = streams[1].detach().float() if streams.shape[0] > 1 else None
self._metrics['engram_stream0_rms'] = float(s0.pow(2).mean().sqrt().item())
if s1 is not None:
self._metrics['engram_stream1_rms'] = float(s1.pow(2).mean().sqrt().item())
self._metrics['engram_stream_divergence_rms'] = float(
(s0 - s1).pow(2).mean().sqrt().item()
)
if _profile: _engram_ev = _ev()
if _diag:
with torch.no_grad():
h_post = mhc_layer.merge_streams(streams).detach().float()
in_n = h_pre.pow(2).mean().sqrt()
out_n = h_post.pow(2).mean().sqrt()
d_n = (h_post - h_pre).pow(2).mean().sqrt()
self._metrics[f'layer_{i}_in_norm'] = float(in_n.item())
self._metrics[f'layer_{i}_out_norm'] = float(out_n.item())
self._metrics[f'layer_{i}_delta_ratio'] = float((d_n / (in_n + 1e-6)).item())
self._metrics[f'layer_{i}_feat_std'] = float(h_post.std(dim=-1).mean().item())
if _run_svd:
# Effective rank via participation ratio of singular values.
# eff_rank = (Ξ£Ο)^2 / Ξ£ΟΒ² β smooth rank proxy, bounded by d_model.
# Sampled to keep overhead low (SVD is O(min(B*T, D)^2Β·D)).
flat = h_post.reshape(-1, h_post.shape[-1])[:512].float()
try:
s = torch.linalg.svdvals(flat)
eff_rank = float(((s.sum() ** 2) / (s.pow(2).sum() + 1e-6)).item())
self._metrics[f'layer_{i}_eff_rank'] = eff_rank
except Exception:
pass
h_pre = h_post
if _diag:
self._diag_step += 1
if _profile: _t_blocks = _ev()
self._metrics['sdr_active_bits'] = sdr_active_bits
self._metrics['htm_anomaly'] = htm_anomaly
x = self.mhc[-1].merge_streams(streams)
x = norm(x)
if _profile: _t_merge = _ev()
softcap = self.softcap
_softcap_clamp = os.environ.get("HYDRA_SOFTCAP_CLAMP", "0") == "1"
if targets is not None:
smoothing = self.config.label_smoothing
V = self.config.vocab_size
# Learnability #4: doc-separator masking. In packed rows,
# tokenizer.encode(..., prepend=bos_token) places a BOS at every
# document boundary. Without masking, the model is penalized for
# failing to predict "doc B's BOS" from the last tokens of doc A
# β pure noise. We set targets==bos to -1 (ignore_index). Done
# BEFORE MTP/entropy/sampled-softmax branches so all downstream
# losses inherit the mask.
if self._doc_sep_mask and self._bos_token_id >= 0:
targets = torch.where(
targets == self._bos_token_id,
torch.full_like(targets, -1),
targets,
)
# Sampled softmax: instead of computing logits for ALL V tokens,
# compute only for the target + K negatives drawn from a proposal
# distribution. Reduces the lm_head matmul from (B*T, d) x (d, V)
# to per-row (1, d) x (d, K) matmuls fused as einsum. At V=65536
# and K=4096: 16x less compute, ~1.3-1.8x tps improvement once
# the LM-head matmul is the bottleneck.
#
# Audit 2026-05-09 issue #22 - Cluster E. Two paths:
# * UnigramSampler attached -> per-row independent negatives
# drawn from the unigram distribution; per-id `log p_unigram`
# correction (Jean et al. 2015 importance-sampled NCE). This
# is the new default once training.py builds the cache.
# * No sampler -> legacy uniform-negative path, single shared
# batch of K negatives, constant `log(V/K)` correction.
# Retained as fallback when the unigram cache is unavailable
# (CI, unit tests, missing tokenizer).
# Eval (self.training=False) always uses full softmax.
#
# Set HYDRA_SAMPLED_SOFTMAX=0 to force full softmax even in train.
K_neg = int(os.environ.get("HYDRA_SAMPLED_SOFTMAX", "4096"))
use_sampled = K_neg > 0 and K_neg < V and self.training
unigram_sampler = getattr(self, "_unigram_sampler", None)
# neg_logits / log_correction are populated by the legacy uniform
# path so MTP heads can reuse them. The unigram path leaves them
# at None and each MTP head draws its own per-row negatives.
neg_logits = None
log_correction = None
if use_sampled and unigram_sampler is not None:
# ----- Unigram-sampler path (importance-weighted NCE) -----
h_flat = x.reshape(-1, x.shape[-1])
t_flat = targets.reshape(-1)
valid_mask_flat = (t_flat >= 0)
if reduction == 'none':
per_tok = sampled_softmax_loss(
h_flat, t_flat, self.lm_head.weight, unigram_sampler, K_neg,
label_smoothing=0.0, softcap=softcap,
softcap_clamp=_softcap_clamp,
valid_mask=valid_mask_flat, reduction='none',
)
return per_tok
out = sampled_softmax_loss(
h_flat, t_flat, self.lm_head.weight, unigram_sampler, K_neg,
label_smoothing=smoothing, softcap=softcap,
softcap_clamp=_softcap_clamp,
valid_mask=valid_mask_flat, reduction='mean',
)
elif use_sampled:
# ----- Legacy uniform-negative path (fallback) -----
# Flatten hidden states + targets
h_flat = x.reshape(-1, x.shape[-1]) # (B*T, d)
t_flat = targets.reshape(-1) # (B*T,)
n = h_flat.shape[0]
# Learnability #4 hardening: sampled-softmax gather crashes on
# negative ids (-1 from doc-sep mask). Replace -1 with 0 for
# gather; the actual loss is masked below.
valid_mask_flat = (t_flat >= 0)
t_flat_safe = torch.where(valid_mask_flat, t_flat, torch.zeros_like(t_flat))
# Sample K negatives uniformly from [0, V)
neg_ids = torch.randint(0, V, (K_neg,), device=x.device)
# Gather lm_head weights for target + negatives
all_ids = torch.cat([t_flat_safe, neg_ids]) # (B*T + K,)
sampled_w = self.lm_head.weight[all_ids] # (B*T + K, d)
# Compute sampled logits: for each position, dot with its
# target weight and all K negative weights.
# Target logit: dot product of h[i] with w[target[i]]
target_w = sampled_w[:n] # (B*T, d)
neg_w = sampled_w[n:] # (K, d)
target_logit = (h_flat * target_w).sum(-1) # (B*T,)
neg_logits = h_flat @ neg_w.t() # (B*T, K)
if not _softcap_clamp:
target_logit = softcap * torch.tanh(target_logit / softcap)
neg_logits = softcap * torch.tanh(neg_logits / softcap)
# Sampled softmax loss: -log(exp(target) / (exp(target) + sum(exp(neg))))
# With log-sum-exp correction for sampling K of V negatives.
# Correction: add log(V/K) to negative logits to account for
# the fact that we're only seeing K of V possible negatives.
log_correction = torch.tensor(V / K_neg, device=x.device).log()
all_logits = torch.cat([
target_logit.unsqueeze(-1), # (B*T, 1)
neg_logits + log_correction, # (B*T, K)
], dim=-1).float() # (B*T, K+1)
# CE with target always at index 0
ce_targets = torch.zeros(n, dtype=torch.long, device=x.device)
if reduction == 'none':
per_tok = F.cross_entropy(all_logits, ce_targets, reduction='none')
if self._doc_sep_mask and self._bos_token_id >= 0:
per_tok = torch.where(valid_mask_flat, per_tok, torch.zeros_like(per_tok))
return per_tok
per_tok_ce = F.cross_entropy(
all_logits, ce_targets, reduction='none',
label_smoothing=smoothing,
)
# Mask doc-separator positions. valid_mask_flat is always
# computed; when doc_sep_mask is off every token is valid so
# this reduces to a plain mean.
valid_f = valid_mask_flat.float()
valid_n = valid_f.sum().clamp(min=1)
out = (per_tok_ce * valid_f).sum() / valid_n
else:
# Full softmax path (eval or HYDRA_SAMPLED_SOFTMAX=0)
chunk_size = int(os.environ.get("HYDRA_CE_CHUNK", "1024"))
if chunk_size <= 0:
MAX_LOGITS_BYTES = 256 * 1024 * 1024
tokens_per_chunk = max(V, MAX_LOGITS_BYTES // (V * 4))
chunk_size = max(1, tokens_per_chunk // max(1, B))
chunk_size = min(chunk_size, T)
if reduction == 'none':
loss_parts = []
for start in range(0, T, chunk_size):
end = min(start + chunk_size, T)
chunk_logits = self.lm_head(x[:, start:end, :]).float()
if _softcap_clamp:
chunk_logits = torch.clamp(chunk_logits, -softcap, softcap)
else:
chunk_logits = softcap * torch.tanh(chunk_logits / softcap)
chunk_targets = targets[:, start:end].reshape(-1)
chunk_loss = F.cross_entropy(
chunk_logits.view(-1, chunk_logits.size(-1)),
chunk_targets, ignore_index=-1, reduction='none',
)
loss_parts.append(chunk_loss)
return torch.cat(loss_parts)
total_loss = 0.0
total_tokens = 0
for start in range(0, T, chunk_size):
end = min(start + chunk_size, T)
chunk_logits = self.lm_head(x[:, start:end, :]).float()
if _softcap_clamp:
chunk_logits = torch.clamp(chunk_logits, -softcap, softcap)
else:
chunk_logits = softcap * torch.tanh(chunk_logits / softcap)
chunk_targets = targets[:, start:end].reshape(-1)
chunk_loss = F.cross_entropy(
chunk_logits.view(-1, chunk_logits.size(-1)),
chunk_targets, ignore_index=-1, reduction='sum',
label_smoothing=smoothing,
)
total_loss = total_loss + chunk_loss
total_tokens += (chunk_targets != -1).sum()
out = total_loss / total_tokens
# -----------------------------------------------------------
# Learnability #1: Multi-Token Prediction.
# For k in {2..K}, add a CE loss at position (t) predicting
# the token at position (t+k), using the SAME lm_head weights
# (weight-tied). Cost: K-1 extra CEs on a subset of positions.
# Only triggered in reduction='mean' path, training only.
# -----------------------------------------------------------
if reduction == 'mean' and self._mtp_k > 1 and self.training and use_sampled:
# Audit 2026-05-09 issue #22 - Cluster E. When a UnigramSampler
# is attached each MTP head draws its OWN per-row negatives
# (independent of the primary head). This matches the math of
# importance-sampled NCE (each loss term has its own negative
# pool drawn from the same proposal q) and removes the bias
# from sharing one batch of negatives across K heads. Cost:
# one extra alias-method sample call per head - cheap relative
# to the einsum.
# Legacy (uniform) path retains the original neg_logits reuse
# for backward-compat with environments that have no sampler.
mtp_loss_sum = out.new_tensor(0.0)
mtp_terms = 0
neg_logits_bt = None
if unigram_sampler is None and neg_logits is not None:
neg_logits_bt = neg_logits.view(B, T, K_neg)
for k in range(2, self._mtp_k + 1):
shift = k - 1
if T <= shift:
continue
n_k = B * (T - shift)
h_k_flat = x[:, :T - shift, :].reshape(n_k, -1) # (n_k, d)
t_k = targets[:, shift:].reshape(-1) # (n_k,)
mask_k = (t_k >= 0)
if unigram_sampler is not None:
# Independent unigram-distributed negatives per head.
per_tok_ce_k = sampled_softmax_loss(
h_k_flat, t_k, self.lm_head.weight,
unigram_sampler, K_neg,
label_smoothing=smoothing, softcap=softcap,
softcap_clamp=_softcap_clamp,
valid_mask=mask_k, reduction='none',
)
# sampled_softmax_loss already zeros invalid rows.
n_valid_k = mask_k.float().sum().clamp(min=1)
mtp_loss_sum = mtp_loss_sum + per_tok_ce_k.sum() / n_valid_k
else:
# Legacy path: reuse primary's negatives. neg_logits_bt
# is non-None here because the legacy branch runs only
# when use_sampled and unigram_sampler is None.
t_k_safe = torch.where(mask_k, t_k, torch.zeros_like(t_k))
tgt_w_k = self.lm_head.weight[t_k_safe] # (n_k, d)
tgt_logit_k = (h_k_flat * tgt_w_k).sum(-1) # (n_k,)
if not _softcap_clamp:
tgt_logit_k = softcap * torch.tanh(tgt_logit_k / softcap)
neg_logits_k = neg_logits_bt[:, :T - shift, :].reshape(n_k, K_neg)
all_logits_k = torch.cat([
tgt_logit_k.unsqueeze(-1),
neg_logits_k + log_correction,
], dim=-1).float()
ce_targets_k = torch.zeros(n_k, dtype=torch.long, device=x.device)
per_tok_ce_k = F.cross_entropy(
all_logits_k, ce_targets_k, reduction='none',
label_smoothing=smoothing,
)
per_tok_ce_k = torch.where(mask_k, per_tok_ce_k, torch.zeros_like(per_tok_ce_k))
n_valid_k = mask_k.sum().clamp(min=1)
mtp_loss_sum = mtp_loss_sum + per_tok_ce_k.sum() / n_valid_k
mtp_terms += 1
if mtp_terms > 0:
out = (out + mtp_loss_sum) / float(mtp_terms + 1)
# -----------------------------------------------------------
# Learnability #6: output entropy penalty.
# L += -lambda * H(softmax(logits)). Negative entropy penalizes
# peaked distributions; encourages diverse predictions and
# breaks repetition loops. Computed on a small subset of
# positions to keep V-sized logits cost bounded.
# -----------------------------------------------------------
if reduction == 'mean' and self._entropy_penalty > 0.0 and self.training:
# Sample up to 64 random positions. V-sized logits on 64
# positions = 64 * V * 4 bytes (~50 MB at V=200k) β fits
# on the 3060 and adds ~2 ms.
h_flat = x.reshape(-1, x.shape[-1])
n_pos = h_flat.shape[0]
n_sample = min(64, n_pos)
idx_sample = torch.randint(0, n_pos, (n_sample,), device=x.device)
h_sample = h_flat[idx_sample]
logits_s = F.linear(h_sample, self.lm_head.weight).float()
if _softcap_clamp:
logits_s = torch.clamp(logits_s, -softcap, softcap)
else:
logits_s = softcap * torch.tanh(logits_s / softcap)
log_probs = F.log_softmax(logits_s, dim=-1)
probs = log_probs.exp()
entropy = -(probs * log_probs).sum(-1).mean() # scalar, nats
out = out - self._entropy_penalty * entropy
if _profile:
_t_end = _ev()
torch.cuda.synchronize()
def _ms(a, b): return a.elapsed_time(b)
print(
f"[PROFILE B={B} T={T}] "
f"htm_launch={_ms(_t0, _t_htm_async):.2f} "
f"wte={_ms(_t_htm_async, _t_wte):.2f} "
f"htm_await={_ms(_t_wte, _t_htm_await):.2f} "
f"htm_proj={_ms(_t_htm_await, _t_htm_proj):.2f} "
f"mamba_mhc_engram={_ms(_t_htm_proj, _t_blocks):.2f} "
f"merge={_ms(_t_blocks, _t_merge):.2f} "
f"lm_head_loss={_ms(_t_merge, _t_end):.2f} "
f"total={_ms(_t0, _t_end):.2f} ms",
flush=True,
)
return out
logits = self.lm_head(x).float()
if _softcap_clamp:
logits = torch.clamp(logits, -softcap, softcap)
else:
logits = softcap * torch.tanh(logits / softcap)
return logits
|