Spaces:
Running on Zero
Running on Zero
File size: 46,300 Bytes
45e7dfb | 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 | """
model.py -- SpikeWhaleLM: combined architecture from SpikeTransformer (My Project) + NanoWhale.
Architecture flow:
Embedding
-> Engram delta (N-gram memory, My Project)
-> [expand to hc_mult copies if HC enabled]
-> N x TransformerBlock:
HC pre-op (NanoWhale) -> RMSNorm -> MLA+DERF+XSA Attention (combined)
-> HC post-op
HC pre-op -> RMSNorm -> MoE FFN w/ shared expert (NanoWhale)
-> HC post-op
-> [mean-pool hc_mult copies if HC enabled]
-> RMSNorm
-> LM head + MTP heads (NanoWhale)
Component origins:
RMSNorm, RotaryEmbedding -- both (standard)
Engram / DERFContextGate -- My Project
MLADerfXSAAttention -- MLA from NanoWhale + DERF+XSA from My Project
SparseMoEFFN w/ shared expert -- NanoWhale MoE structure + My Project aux loss
HyperConnectionLayer -- NanoWhale
SpikeWhaleLM + MTP heads -- NanoWhale
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Tuple, List
from transformers import PreTrainedModel
from transformers.modeling_outputs import CausalLMOutputWithPast
from torch.utils.checkpoint import checkpoint as gradient_checkpoint
from config import SpikeWhaleConfig
# ---------------------------------------------------------------------------
# Primitives
# ---------------------------------------------------------------------------
class RMSNorm(nn.Module):
def __init__(self, dim: int, eps: float = 1e-6):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(dim))
def forward(self, x: torch.Tensor) -> torch.Tensor:
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) * self.weight
class RotaryEmbedding(nn.Module):
"""RoPE for the rope partition of Q and K (qk_rope_head_dim dims only)."""
def __init__(self, dim: int, max_positions: int = 4096, theta: float = 10000.0):
super().__init__()
inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2).float() / dim))
self.register_buffer("inv_freq", inv_freq)
t = torch.arange(max_positions).float()
freqs = torch.outer(t, inv_freq)
self.register_buffer("cos_cache", freqs.cos())
self.register_buffer("sin_cache", freqs.sin())
def forward(self, x: torch.Tensor, position_ids: torch.Tensor) -> torch.Tensor:
"""
x: [B, H, S, rope_dim]
position_ids: [B, S]
"""
cos = self.cos_cache[position_ids].unsqueeze(1) # [B, 1, S, rope_dim//2]
sin = self.sin_cache[position_ids].unsqueeze(1)
d = cos.shape[-1]
x1, x2 = x[..., :d], x[..., d:]
return torch.cat([x1 * cos - x2 * sin, x1 * sin + x2 * cos], dim=-1)
# ---------------------------------------------------------------------------
# Engram: N-gram hash lookup + DERF gate (My Project, preserved)
# ---------------------------------------------------------------------------
class TokenCompressor(nn.Module):
def __init__(self, embed_dim: int, compress_dim: int):
super().__init__()
self.proj = nn.Linear(embed_dim, compress_dim, bias=False)
nn.init.normal_(self.proj.weight, std=0.02)
# BUGFIX: this projection feeds ONLY the integer hash index
# (idx = h.abs().long() % table_size) in MultiHeadHashLookup. The .long()
# cast is non-differentiable, so no gradient ever reaches this weight --
# it can never learn. Worse, _classify_params put it in the weight-decay
# group, so AdamW was steadily shrinking it toward zero and degrading the
# hash projection over a long run. Freeze it: a fixed random projection is
# exactly the right behavior for an LSH-style hash, and freezing drops it
# from the optimizer (saves state) and from weight decay. Checkpoint-safe:
# the parameter still exists and is still saved/loaded in state_dict.
self.proj.weight.requires_grad_(False)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.proj(x)
class MultiHeadHashLookup(nn.Module):
def __init__(self, num_heads: int, table_size: int,
compress_dim: int, out_dim: int, max_ngram: int = 3):
super().__init__()
self.num_heads = num_heads
self.table_size = table_size
self.max_ngram = max_ngram
self.out_dim = out_dim
self.tables = nn.ModuleList([
nn.Embedding(table_size, out_dim) for _ in range(num_heads)
])
for t in self.tables:
nn.init.normal_(t.weight, std=0.01)
for n in range(1, max_ngram + 1):
for k in range(n):
proj = torch.randn(num_heads, compress_dim)
proj = proj / (proj.norm(dim=1, keepdim=True) + 1e-8)
self.register_buffer(f"hash_proj_n{n}_p{k}", proj)
def forward(self, compressed: torch.Tensor) -> torch.Tensor:
"""
compressed: [B, S, compress_dim]
returns: [B, S, out_dim]
All positions are processed in parallel. The outer loop runs max_ngram
times (≤3), not S times (≤2048). Each iteration is a single matmul +
embedding lookup across the whole sequence, making this GPU-friendly
and compatible with torch.compile.
"""
B, S, _ = compressed.shape
device = compressed.device
out = torch.zeros(B, S, self.out_dim, device=device, dtype=compressed.dtype)
# Per-position normalization: tracks how many (n-gram × head) contributions
# each position receives. Positions near the start get fewer contributions
# because shorter n-grams don't exist yet (matches original causal behavior).
norm = torch.zeros(S, device=device)
for n in range(1, self.max_ngram + 1):
if S < n:
continue
valid_len = S - n + 1 # positions [n-1 .. S-1] are valid for order-n
start = n - 1
# Accumulate position-k contribution to the order-n hash.
# compressed[:, k : k+valid_len, :] is the k-th token of every n-gram
# window simultaneously → [B, valid_len, num_heads] after projection.
h = torch.zeros(B, valid_len, self.num_heads, device=device)
for k in range(n):
proj = getattr(self, f"hash_proj_n{n}_p{k}") # [num_heads, compress_dim]
h = h + torch.matmul(compressed[:, k:k + valid_len, :].float(), proj.t())
idx = h.abs().long() % self.table_size # [B, valid_len, num_heads]
for head_idx, table in enumerate(self.tables):
out[:, start:, :] = out[:, start:, :] + table(idx[:, :, head_idx])
norm[start:] += self.num_heads
# Cast back to input dtype: the norm division promotes bf16→float32 under autocast.
# Keeping the output in the same dtype as the input avoids a silent dtype mismatch
# when EngramModule adds this result back onto the (bf16) embedding tensor.
return (out / norm.view(1, -1, 1).clamp(min=1)).to(compressed.dtype)
class DERFContextGate(nn.Module):
"""
DERF gate: gate = gamma * erf(alpha * proj([retrieved, x]) + bias)
Positive probability = (gate + 1) / 2 applied to retrieved embedding.
Large negative init_bias keeps gate closed at start of training.
"""
def __init__(self, obs_size: int, init_bias: float = -4.0):
super().__init__()
self.proj = nn.Linear(obs_size * 2, obs_size)
self.alpha = nn.Parameter(torch.ones(obs_size))
self.bias = nn.Parameter(torch.full((obs_size,), init_bias))
self.gamma = nn.Parameter(torch.ones(obs_size))
def forward(self, retrieved: torch.Tensor, x: torch.Tensor) -> torch.Tensor:
logits = self.proj(torch.cat([retrieved, x], dim=-1))
gate = self.gamma * ((torch.erf(self.alpha * logits + self.bias) + 1.0) / 2.0)
return retrieved * gate
class EngramModule(nn.Module):
"""
N-gram hash lookup with DERF gate (My Project), fully vectorized.
All S positions are processed in parallel — the sequential Python loop
over sequence positions has been eliminated. The lookup now accepts the
full [B, S, compress_dim] compressed tensor and returns [B, S, H] in one pass.
"""
def __init__(self, cfg: SpikeWhaleConfig):
super().__init__()
self.compressor = TokenCompressor(cfg.hidden_size, cfg.engram_compress_dim)
self.lookup = MultiHeadHashLookup(
cfg.engram_num_heads, cfg.engram_table_size,
cfg.engram_compress_dim, cfg.hidden_size, cfg.engram_max_ngram,
)
self.gate = DERFContextGate(cfg.hidden_size, cfg.engram_gate_init_bias)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""x: [B, S, H] -> engram_delta: [B, S, H]"""
compressed = self.compressor(x.detach()) # [B, S, compress_dim]
retrieved = self.lookup(compressed) # [B, S, H]
return self.gate(retrieved, x) # [B, S, H]
# ---------------------------------------------------------------------------
# Hyper-Connections (NanoWhale, simplified)
# ---------------------------------------------------------------------------
class HyperConnectionLayer(nn.Module):
"""
Simplified Hyper-Connections for one sublayer (attention or FFN).
Maintains hc_mult parallel residual streams.
Pre-op: learned weighted average of hc_mult copies -> single hidden state for sublayer.
Post-op: sublayer output added to each copy with learned per-stream weights.
Full HC uses Sinkhorn-normalized 2D routing matrices; this uses softmax-normalized
1D weights for pre/post routing -- captures the same multi-stream routing spirit.
"""
def __init__(self, hidden_size: int, hc_mult: int,
sinkhorn_iters: int = 20, eps: float = 1e-6):
super().__init__()
self.hc_mult = hc_mult
# pre_weight: how to mix hc_mult copies into one sublayer input
# post_weight: how to distribute the sublayer delta to each copy
#
# BUGFIX: these must NOT be initialized identically across streams.
# The model expands the hidden state into hc_mult *identical* copies.
# With uniform pre/post weights, pre_op produces sum_i copy_i * w_i =
# copy * sum(softmax)=copy (all copies equal), and post_op adds the same
# delta to every copy -- so the streams stay byte-for-byte identical at
# every layer. When all streams are equal, the softmax Jacobian applied
# to the (equal) per-stream gradients is exactly zero, so pre_weight and
# post_weight receive ZERO gradient and never move off 1/hc_mult. The HC
# routing then learns nothing and just burns hc_mult x memory/compute.
#
# Breaking the post_weight symmetry at init makes the streams diverge
# after the first sublayer, which restores gradient flow to all HC
# weights. We center post_weight so softmax starts near-uniform (keeps
# the residual baseline ~unchanged) but with a distinct value per stream.
self.pre_weight = nn.Parameter(
torch.linspace(0.5, -0.5, hc_mult) / max(hc_mult, 1)
)
self.post_weight = nn.Parameter(
torch.linspace(-0.5, 0.5, hc_mult) / max(hc_mult, 1)
)
def pre_op(self, copies: torch.Tensor) -> torch.Tensor:
"""copies: [B, hc_mult, S, H] -> [B, S, H]"""
w = F.softmax(self.pre_weight, dim=0) # [hc_mult]
return (copies * w.view(1, -1, 1, 1)).sum(dim=1)
def post_op(self, copies: torch.Tensor, delta: torch.Tensor) -> torch.Tensor:
"""
copies: [B, hc_mult, S, H]
delta: [B, S, H]
Returns updated copies: [B, hc_mult, S, H]
"""
w = F.softmax(self.post_weight, dim=0) # [hc_mult]
return copies + delta.unsqueeze(1) * w.view(1, -1, 1, 1)
# ---------------------------------------------------------------------------
# MLA + DERF + XSA Attention (combined)
# ---------------------------------------------------------------------------
class MLADerfXSAAttention(nn.Module):
"""
Multi-Head Latent Attention (NanoWhale) with DERF scores + XSA correction (My Project).
MLA (from NanoWhale):
Q: hidden -> q_lora_rank (RMSNorm) -> num_heads * head_dim (low-rank projection)
K, V: hidden -> num_kv_heads * head_dim (direct, MQA by default with num_kv_heads=1)
Output: num_heads * head_dim -> o_lora_rank -> hidden (low-rank output)
Partial RoPE: applied only to the last qk_rope_head_dim dims of Q and K
DERF (from My Project):
Replaces softmax: erf(alpha * scores + bias) * gamma, shifted to [0,1] then normalized.
Per-head learnable alpha, bias, gamma.
XSA (from My Project):
After computing the weighted value sum y, subtract the component of y that
projects onto each position's own value vector. Forces the output to carry
only cross-position information, not echo the current token back.
"""
def __init__(self, cfg: SpikeWhaleConfig):
super().__init__()
self.num_heads = cfg.num_attention_heads
self.num_kv_heads = cfg.num_key_value_heads
self.head_dim = cfg.head_dim
self.qk_rope_head_dim = cfg.qk_rope_head_dim
self.nope_head_dim = cfg.nope_head_dim
self.hidden_size = cfg.hidden_size
self.use_derf = cfg.use_derf
self.use_xsa = cfg.use_xsa
self.dropout_p = cfg.attention_dropout
self.kv_groups = self.num_heads // self.num_kv_heads
# Low-rank Q projection (MLA)
self.q_a_proj = nn.Linear(cfg.hidden_size, cfg.q_lora_rank, bias=False)
self.q_a_norm = RMSNorm(cfg.q_lora_rank, cfg.rms_norm_eps)
self.q_b_proj = nn.Linear(cfg.q_lora_rank, self.num_heads * self.head_dim, bias=False)
# Direct K, V projections (MQA/GQA)
self.k_proj = nn.Linear(cfg.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
self.v_proj = nn.Linear(cfg.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
# Low-rank output projection (MLA)
self.o_a_proj = nn.Linear(self.num_heads * self.head_dim, cfg.o_lora_rank, bias=False)
self.o_b_proj = nn.Linear(cfg.o_lora_rank, cfg.hidden_size, bias=False)
# Partial RoPE: applied to qk_rope_head_dim dims only
self.rope = RotaryEmbedding(
self.qk_rope_head_dim,
max_positions=cfg.max_position_embeddings,
theta=cfg.rope_theta,
)
# DERF parameters: one per query head (My Project)
if self.use_derf:
self.derf_alpha = nn.Parameter(torch.ones(self.num_heads))
self.derf_bias = nn.Parameter(torch.zeros(self.num_heads))
self.derf_gamma = nn.Parameter(torch.ones(self.num_heads))
nn.init.normal_(self.q_a_proj.weight, std=cfg.initializer_range)
nn.init.normal_(self.q_b_proj.weight, std=cfg.initializer_range)
nn.init.normal_(self.k_proj.weight, std=cfg.initializer_range)
nn.init.normal_(self.v_proj.weight, std=cfg.initializer_range)
nn.init.normal_(self.o_a_proj.weight, std=cfg.initializer_range)
nn.init.normal_(self.o_b_proj.weight, std=cfg.initializer_range)
def forward(
self,
x: torch.Tensor,
position_ids: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
use_cache: bool = False,
) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
B, S, _ = x.shape
# Q via low-rank projection with intermediate norm (MLA)
q = self.q_a_norm(self.q_a_proj(x))
q = self.q_b_proj(q).view(B, S, self.num_heads, self.head_dim).transpose(1, 2)
# [B, num_heads, S, head_dim]
# K, V direct projections
k = self.k_proj(x).view(B, S, self.num_kv_heads, self.head_dim).transpose(1, 2)
v = self.v_proj(x).view(B, S, self.num_kv_heads, self.head_dim).transpose(1, 2)
# Partial RoPE: split into nope and rope partitions, rotate only the rope part
q_nope = q[..., :self.nope_head_dim]
q_rope = q[..., self.nope_head_dim:] # qk_rope_head_dim dims
k_nope = k[..., :self.nope_head_dim]
k_rope = k[..., self.nope_head_dim:]
q_rope = self.rope(q_rope, position_ids)
k_rope = self.rope(k_rope, position_ids)
q = torch.cat([q_nope, q_rope], dim=-1)
k = torch.cat([k_nope, k_rope], dim=-1)
# KV cache for inference
if past_key_value is not None:
k = torch.cat([past_key_value[0], k], dim=2)
v = torch.cat([past_key_value[1], v], dim=2)
present = (k, v) if use_cache else None
N = k.shape[2] # total key positions (past + current)
# Expand KV heads for MQA/GQA
if self.kv_groups > 1:
k = k.unsqueeze(2).expand(-1, -1, self.kv_groups, -1, -1).reshape(
B, self.num_heads, N, self.head_dim)
v = v.unsqueeze(2).expand(-1, -1, self.kv_groups, -1, -1).reshape(
B, self.num_heads, N, self.head_dim)
# Scaled dot-product attention.
if self.use_derf:
# DERF replaces softmax with a custom erf nonlinearity, so it cannot
# use the fused kernel and must materialize scores explicitly.
scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
# Build boolean mask for causality (this avoids the -inf math errors)
if attention_mask is None and past_key_value is None:
is_masked = torch.triu(torch.ones(S, N, dtype=torch.bool, device=scores.device), diagonal=N - S + 1).unsqueeze(0).unsqueeze(0)
else:
is_masked = (attention_mask < -1.0) if attention_mask is not None else torch.zeros_like(scores, dtype=torch.bool)
# FIX 2: Do NOT use float('-inf'). If alpha ever hits 0.0, 0.0 * -inf = NaN.
# Use a safe negative scalar (-10000.0) for masked positions.
safe_scores = scores.masked_fill(is_masked, -10000.0)
a = self.derf_alpha.view(1, -1, 1, 1)
b = self.derf_bias.view(1, -1, 1, 1)
g = self.derf_gamma.view(1, -1, 1, 1)
attn_weights = g * torch.erf(a * safe_scores + b) # [-gamma, gamma]
attn_weights = (attn_weights + g) / 2.0 # shift to [0, gamma]
attn_weights = attn_weights.masked_fill(is_masked, 0.0) # enforce causal mask safely
attn_weights = attn_weights / (attn_weights.sum(dim=-1, keepdim=True) + 1e-8)
if self.dropout_p > 0 and self.training:
attn_weights = F.dropout(attn_weights, p=self.dropout_p)
y = torch.matmul(attn_weights, v) # [B, num_heads, S, head_dim]
else:
# OPTIMIZATION: standard (softmax) attention goes through the fused
# scaled_dot_product_attention kernel (FlashAttention / mem-efficient
# backends). This is the hot path during pretraining (use_derf=False)
# and is much faster + lower memory than materializing [B,H,S,N]
# scores and a softmax. SDPA already scales by 1/sqrt(head_dim).
#
# CONTIGUITY FIX: with MQA/GQA, k and v above are built via
# .unsqueeze(2).expand(...).reshape(...). Under torch.compile, inductor
# can trace the broadcasted (zero-stride) view through to the fused
# flash-attention BACKWARD kernel, whose meta-kernel then asserts on the
# mismatched stride (e.g. "stride 120==245760 at dim=1") and aborts.
# Forcing contiguity guarantees standard strides into the fused kernel.
q = q.contiguous()
k = k.contiguous()
v = v.contiguous()
drop = self.dropout_p if self.training else 0.0
if past_key_value is None and attention_mask is None:
# Prefill / training: pure causal mask, no materialization needed.
y = F.scaled_dot_product_attention(q, k, v, is_causal=True, dropout_p=drop)
else:
# Incremental decode or a provided mask: pass an explicit boolean
# keep-mask (True = attend). SDPA fills masked positions with -inf.
if attention_mask is not None:
is_masked = (attention_mask < -1.0)
else:
is_masked = torch.triu(
torch.ones(S, N, dtype=torch.bool, device=q.device),
diagonal=N - S + 1,
).unsqueeze(0).unsqueeze(0)
y = F.scaled_dot_product_attention(
q, k, v, attn_mask=~is_masked, dropout_p=drop)
# XSA: remove self-projection from output (My Project)
# For each query position s, subtract the component of y[:,:,s,:] that
# projects onto the normalized value vector at the same position.
if self.use_xsa:
past_len = N - S
v_self = v[:, :, past_len:past_len + S, :] # [B, H, S, D]
vn = v_self / (v_self.norm(dim=-1, keepdim=True) + 1e-8)
projection = (y * vn).sum(dim=-1, keepdim=True) * vn
y = y - projection
# Low-rank output projection (MLA)
y = y.transpose(1, 2).contiguous().view(B, S, self.num_heads * self.head_dim)
y = self.o_b_proj(self.o_a_proj(y))
return y, present
# ---------------------------------------------------------------------------
# MoE FFN: shared expert + sqrtsoftplus + hash routing (NanoWhale) + aux loss (My Project)
# ---------------------------------------------------------------------------
class ExpertFFN(nn.Module):
"""Single SwiGLU expert."""
def __init__(self, hidden_size: int, intermediate_size: int):
super().__init__()
self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x))
def sqrtsoftplus(x: torch.Tensor) -> torch.Tensor:
"""sqrt(softplus(x)) = sqrt(log(1+exp(x))). NanoWhale expert scoring."""
# FIX 1: Added 1e-8. If F.softplus(x) evaluates to 0.0, torch.sqrt(0) produces NaN gradients on backward pass.
return torch.sqrt(F.softplus(x) + 1e-8)
class SparseMoEFFN(nn.Module):
"""
Combines NanoWhale MoE structure with My Project aux loss:
- n_shared_experts always-active experts (NanoWhale)
- n_routed_experts sparse routed experts, top-k activation
- sqrtsoftplus scoring (NanoWhale) vs softmax
- hash routing for early layers (NanoWhale)
- norm_topk_prob + routed_scaling_factor (NanoWhale)
- load-balancing aux loss (My Project)
"""
def __init__(self, cfg: SpikeWhaleConfig, layer_idx: int = 0):
super().__init__()
self.n_routed_experts = cfg.n_routed_experts
self.n_shared_experts = cfg.n_shared_experts
self.num_experts_per_tok = cfg.num_experts_per_tok
self.norm_topk_prob = cfg.norm_topk_prob
self.scoring_func = cfg.scoring_func
self.routed_scaling_factor = cfg.routed_scaling_factor
self.use_hash_routing = layer_idx < cfg.num_hash_layers
self.aux_loss_coef = cfg.moe_aux_loss_coef
self.router = nn.Linear(cfg.hidden_size, cfg.n_routed_experts, bias=False)
self.experts = nn.ModuleList([
ExpertFFN(cfg.hidden_size, cfg.moe_intermediate_size)
for _ in range(cfg.n_routed_experts)
])
self.shared_experts = nn.ModuleList([
ExpertFFN(cfg.hidden_size, cfg.moe_intermediate_size)
for _ in range(cfg.n_shared_experts)
]) if cfg.n_shared_experts > 0 else None
self._last_aux_loss: Optional[torch.Tensor] = None
def forward(self, x: torch.Tensor, position_ids: Optional[torch.Tensor] = None) -> torch.Tensor:
B, S, H = x.shape
x_flat = x.view(B * S, H)
T = B * S
# Shared experts: always active (NanoWhale)
shared_out = torch.zeros_like(x_flat)
if self.shared_experts:
for expert in self.shared_experts:
shared_out = shared_out + expert(x_flat)
if len(self.shared_experts) > 1:
shared_out = shared_out / len(self.shared_experts)
# Router
if self.use_hash_routing:
# Hash routing: deterministic assignment without learned router (NanoWhale).
# Assign each of the num_experts_per_tok slots a DISTINCT expert by cycling:
# token at absolute position p -> experts [p%n, (p+1)%n, ..., (p+k-1)%n].
#
# BUGFIX: the assignment must key off the token's ABSOLUTE sequence
# position, not torch.arange(T) (its index in the current flattened
# batch). With arange(T), incremental KV-cache decoding (S=1) always
# sees index 0 and routes every token to expert 0, so generation used
# a different expert assignment than training and silently diverged.
# Using position_ids makes prefill, full-sequence training, and
# step-by-step generation all agree. (For S divisible by n_experts,
# this matches the previous training-time behavior exactly, so existing
# checkpoints stay valid.)
if position_ids is not None:
base = (position_ids.reshape(T, 1) % self.n_routed_experts).long()
else:
base = (torch.arange(T, device=x.device) % self.n_routed_experts).unsqueeze(1)
offsets = torch.arange(self.num_experts_per_tok, device=x.device) # [k]
top_k_indices = (base + offsets.unsqueeze(0)) % self.n_routed_experts # [T, k]
top_k_weights = torch.ones(T, self.num_experts_per_tok, device=x.device) / self.num_experts_per_tok
self._last_aux_loss = None
else:
router_logits = self.router(x_flat)
if self.scoring_func == "sqrtsoftplus":
routing_scores = sqrtsoftplus(router_logits)
else:
routing_scores = F.softmax(router_logits, dim=-1)
top_k_scores, top_k_indices = torch.topk(routing_scores, self.num_experts_per_tok, dim=-1)
if self.norm_topk_prob:
top_k_weights = top_k_scores / (top_k_scores.sum(dim=-1, keepdim=True) + 1e-8)
else:
top_k_weights = top_k_scores
top_k_weights = top_k_weights * self.routed_scaling_factor
# Load-balancing aux loss (My Project)
softmax_probs = F.softmax(router_logits, dim=-1)
expert_mask = torch.zeros_like(softmax_probs)
expert_mask.scatter_(1, top_k_indices, 1.0)
f_e = expert_mask.mean(0)
p_e = softmax_probs.mean(0)
self._last_aux_loss = self.n_routed_experts * (f_e * p_e).sum() * self.aux_loss_coef
# Dispatch tokens to routed experts
out = torch.zeros_like(x_flat)
for expert_idx, expert in enumerate(self.experts):
token_mask = (top_k_indices == expert_idx).any(dim=-1)
if not token_mask.any():
continue
expert_input = x_flat[token_mask]
expert_output = expert(expert_input)
k_pos = (top_k_indices[token_mask] == expert_idx).nonzero(as_tuple=False)
weights = top_k_weights[token_mask][k_pos[:, 0], k_pos[:, 1]].unsqueeze(-1)
out[token_mask] = out[token_mask] + expert_output * weights
out = out + shared_out
return out.view(B, S, H)
def get_aux_loss(self) -> Optional[torch.Tensor]:
# Return None when hash routing (no aux loss) or when forward hasn't run yet.
# Returning torch.tensor(0.0) here would be a CPU tensor and cause a device
# mismatch when added to the CUDA total_aux_loss in SpikeWhaleModel.
return self._last_aux_loss
class DenseFFN(nn.Module):
"""Dense SwiGLU FFN for non-MoE layers."""
def __init__(self, cfg: SpikeWhaleConfig):
super().__init__()
self.gate_proj = nn.Linear(cfg.hidden_size, cfg.moe_intermediate_size, bias=False)
self.up_proj = nn.Linear(cfg.hidden_size, cfg.moe_intermediate_size, bias=False)
self.down_proj = nn.Linear(cfg.moe_intermediate_size, cfg.hidden_size, bias=False)
def forward(self, x: torch.Tensor, position_ids: Optional[torch.Tensor] = None) -> torch.Tensor:
return self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x))
def get_aux_loss(self) -> Optional[torch.Tensor]:
return None # dense layers have no aux loss; None avoids CPU-tensor device mismatch
# ---------------------------------------------------------------------------
# Transformer block with Hyper-Connections
# ---------------------------------------------------------------------------
class TransformerBlock(nn.Module):
"""
Transformer block combining all features:
- Hyper-Connections: pre/post routing through hc_mult streams (NanoWhale)
- MLA + DERF + XSA attention (combined)
- MoE FFN with shared expert (NanoWhale) + aux loss (My Project)
"""
def __init__(self, cfg: SpikeWhaleConfig, layer_idx: int):
super().__init__()
self.use_hc = cfg.use_hyper_connections
self.hidden_dropout = cfg.hidden_dropout
self.attn_norm = RMSNorm(cfg.hidden_size, cfg.rms_norm_eps)
self.attn = MLADerfXSAAttention(cfg)
self.ffn_norm = RMSNorm(cfg.hidden_size, cfg.rms_norm_eps)
if cfg.use_moe and layer_idx in cfg.moe_layers:
self.ffn = SparseMoEFFN(cfg, layer_idx)
self.is_moe = True
else:
self.ffn = DenseFFN(cfg)
self.is_moe = False
if self.use_hc:
self.hc_attn = HyperConnectionLayer(cfg.hidden_size, cfg.hc_mult,
cfg.hc_sinkhorn_iters, cfg.hc_eps)
self.hc_ffn = HyperConnectionLayer(cfg.hidden_size, cfg.hc_mult,
cfg.hc_sinkhorn_iters, cfg.hc_eps)
def forward(
self,
x: torch.Tensor, # [B, hc_mult, S, H] if HC else [B, S, H]
position_ids: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
past_key_value: Optional[Tuple] = None,
use_cache: bool = False,
) -> Tuple[torch.Tensor, Optional[Tuple], Optional[torch.Tensor]]:
# --- Attention sub-layer ---
if self.use_hc:
h = self.hc_attn.pre_op(x) # [B, S, H]
else:
h = x
attn_out, present = self.attn(
self.attn_norm(h), position_ids, attention_mask, past_key_value, use_cache
)
attn_out = F.dropout(attn_out, p=self.hidden_dropout, training=self.training)
if self.use_hc:
x = self.hc_attn.post_op(x, attn_out)
h = self.hc_ffn.pre_op(x) # [B, S, H]
else:
h = h + attn_out
# --- FFN sub-layer ---
ffn_out = self.ffn(self.ffn_norm(h), position_ids)
ffn_out = F.dropout(ffn_out, p=self.hidden_dropout, training=self.training)
if self.use_hc:
x = self.hc_ffn.post_op(x, ffn_out)
else:
x = h + ffn_out
return x, present, self.ffn.get_aux_loss()
# ---------------------------------------------------------------------------
# Full model
# ---------------------------------------------------------------------------
class HRMRefinementBlock(nn.Module):
"""
HRM-INSPIRED iterative refinement (EXPERIMENTAL, off by default). NOT the full
Hierarchical Reasoning Model -- only the iterative-refinement mechanism that the
independent ARC-Prize ablation found carried most of HRM's benefit, adapted to a
causal LM's final hidden state.
Runs N inner steps; each computes a small gated update conditioned on the current
state AND the original ('anchor') input. Per-step gate inits at 0 and up.weight is
zero-init -> the block is an EXACT identity at init, so enabling it cannot hurt a
fresh model; it only contributes if training opens the gate. Pointwise over
positions -> causal-safe (no future-token leakage). In/out [B,S,H].
"""
def __init__(self, hidden_size: int, refine_dim: int, steps: int, eps: float = 1e-6):
super().__init__()
self.steps = steps
self.norm = RMSNorm(hidden_size, eps)
self.down = nn.Linear(hidden_size * 2, refine_dim, bias=False)
self.up = nn.Linear(refine_dim, hidden_size, bias=False)
self.gate = nn.Parameter(torch.zeros(steps))
nn.init.normal_(self.down.weight, std=0.02)
nn.init.zeros_(self.up.weight)
def forward(self, x: torch.Tensor) -> torch.Tensor:
anchor = x
h = x
for t in range(self.steps):
inp = torch.cat([self.norm(h), anchor], dim=-1)
update = self.up(F.silu(self.down(inp)))
h = h + torch.tanh(self.gate[t]) * update
return h
class LatentProjection(nn.Module):
"""ModularMind-on-V2: pool final hidden state -> d_latent output vector.
Mirrors ModularMind's contract: mean-pool over sequence, ReLU^2 activation
(sparse latent codes), Xavier init (NOT zero) so the latent carries signal
from step 1 — zero-init would make the chain unable to bootstrap."""
def __init__(self, hidden_size: int, d_latent: int, eps: float = 1e-6):
super().__init__()
self.proj1 = nn.Linear(hidden_size, hidden_size, bias=False)
self.proj2 = nn.Linear(hidden_size, d_latent, bias=False)
self.norm = RMSNorm(d_latent, eps)
nn.init.xavier_uniform_(self.proj1.weight)
nn.init.xavier_uniform_(self.proj2.weight)
def forward(self, x: torch.Tensor) -> torch.Tensor:
pooled = x.mean(dim=1) # [B, S, H] -> [B, H]
h = torch.relu(self.proj1(pooled)) ** 2
return self.norm(self.proj2(h)) # [B, d_latent]
class LatentInjection(nn.Module):
"""ModularMind-on-V2: fold an incoming d_latent vector into embeddings.
Broadcast across positions, ReGLU-gated add. Gate starts SMALL (not exactly
zero): the injection is near-identity at init (stable) while still passing a
little gradient, so the upstream RecursiveLink + specialist can bootstrap from
step 1. (Exact-zero gate would block all gradient to the link -- the
bootstrapping problem ModularMind's LatentProjection docstring warns about.)
This is the INPUT side of RecursiveLink (the prev specialist's latent)."""
def __init__(self, hidden_size: int, d_latent: int, eps: float = 1e-6,
gate_init: float = 1e-3):
super().__init__()
self.up = nn.Linear(d_latent, hidden_size, bias=False)
self.norm = RMSNorm(hidden_size, eps)
self.value_proj = nn.Linear(hidden_size, hidden_size, bias=False)
self.gate_proj = nn.Linear(hidden_size, hidden_size, bias=False)
self.gate_init = gate_init
nn.init.xavier_uniform_(self.up.weight)
nn.init.xavier_uniform_(self.value_proj.weight)
nn.init.normal_(self.gate_proj.weight, std=gate_init) # small, not zero
def forward(self, x: torch.Tensor, latent: torch.Tensor) -> torch.Tensor:
# x: [B, S, H], latent: [B, d_latent]
inj = self.norm(self.up(latent)).unsqueeze(1) # [B, 1, H] broadcast over S
value = self.value_proj(inj)
gate = torch.relu(self.gate_proj(inj))
return x + value * gate
class RecursiveLink(nn.Module):
"""ModularMind cross-specialist bridge, V2 build. Converts one specialist's
output latent into the next specialist's input latent. ReGLU + residual,
single shared module reused for every hop. Fully differentiable."""
def __init__(self, d_latent: int = 256, expansion: float = 2.0):
super().__init__()
d_hidden = int(d_latent * expansion)
self.norm = nn.LayerNorm(d_latent)
self.value_proj = nn.Linear(d_latent, d_hidden, bias=False)
self.gate_proj = nn.Linear(d_latent, d_hidden, bias=False)
self.down = nn.Linear(d_hidden, d_latent, bias=False)
self.residual_gate = nn.Parameter(torch.ones(1))
nn.init.xavier_uniform_(self.value_proj.weight)
nn.init.xavier_uniform_(self.gate_proj.weight)
nn.init.xavier_uniform_(self.down.weight)
def forward(self, z: torch.Tensor) -> torch.Tensor:
n = self.norm(z)
h = self.value_proj(n) * torch.relu(self.gate_proj(n))
return z + self.residual_gate * self.down(h)
class SpikeWhaleModel(nn.Module):
"""Decoder stack without LM head."""
def __init__(self, cfg: SpikeWhaleConfig):
super().__init__()
self.cfg = cfg
self.embed_tokens = nn.Embedding(cfg.vocab_size, cfg.hidden_size)
nn.init.normal_(self.embed_tokens.weight, std=cfg.initializer_range)
self.engram = EngramModule(cfg) if cfg.use_engram else None
self.layers = nn.ModuleList([
TransformerBlock(cfg, layer_idx=i)
for i in range(cfg.num_hidden_layers)
])
self.norm = RMSNorm(cfg.hidden_size, cfg.rms_norm_eps)
self.hrm_refine = (
HRMRefinementBlock(cfg.hidden_size, cfg.hrm_refine_dim,
cfg.hrm_refine_steps, cfg.rms_norm_eps)
if getattr(cfg, "use_hrm_refine", False) else None
)
# ModularMind-on-V2: latent input/output (off unless use_latent_io)
if getattr(cfg, "use_latent_io", False):
self.latent_inject = LatentInjection(cfg.hidden_size, cfg.d_latent, cfg.rms_norm_eps)
self.latent_out = LatentProjection(cfg.hidden_size, cfg.d_latent, cfg.rms_norm_eps)
else:
self.latent_inject = None
self.latent_out = None
self.gradient_checkpointing = False
def reset_latent_gate(self):
"""Re-init the injection gate SMALL (not zero). Must be called AFTER any HF
post_init/_init_weights pass, which otherwise re-randomizes the gate to full
scale. Small-but-nonzero keeps injection near-identity at start while letting
gradient reach the upstream RecursiveLink (so the chain can bootstrap)."""
if self.latent_inject is not None:
nn.init.normal_(self.latent_inject.gate_proj.weight,
std=self.latent_inject.gate_init)
def forward(
self,
input_ids: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
past_key_values: Optional[List[Tuple]] = None,
use_cache: bool = False,
inject_latent: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, Optional[List[Tuple]], torch.Tensor]:
B, S = input_ids.shape
device = input_ids.device
if position_ids is None:
past_len = past_key_values[0][0].shape[2] if past_key_values else 0
position_ids = torch.arange(
past_len, past_len + S, device=device
).unsqueeze(0).expand(B, -1)
# Token embedding
x = self.embed_tokens(input_ids) # [B, S, H]
# Engram N-gram delta (My Project)
if self.engram is not None:
x = x + self.engram(x)
# ModularMind-on-V2: inject the previous specialist's latent (broadcast
# across positions, ReGLU-gated). No-op at init (gate zero) and skipped
# entirely if no latent is passed.
if self.latent_inject is not None and inject_latent is not None:
x = self.latent_inject(x, inject_latent)
# Expand to hc_mult streams for Hyper-Connections (NanoWhale)
if self.cfg.use_hyper_connections:
x = x.unsqueeze(1).expand(-1, self.cfg.hc_mult, -1, -1).clone()
# [B, hc_mult, S, H]
present_key_values = [] if use_cache else None
total_aux_loss = torch.tensor(0.0, device=device)
for layer_idx, layer in enumerate(self.layers):
pkv = past_key_values[layer_idx] if past_key_values else None
if self.gradient_checkpointing and self.training:
# Gradient checkpointing with use_reentrant=False (NanoWhale)
x, present, aux_loss = gradient_checkpoint(
layer, x, position_ids, attention_mask, None, False,
use_reentrant=False,
)
else:
x, present, aux_loss = layer(x, position_ids, attention_mask, pkv, use_cache)
if use_cache:
present_key_values.append(present)
if aux_loss is not None:
total_aux_loss = total_aux_loss + aux_loss
# Reduce HC streams to single hidden state
if self.cfg.use_hyper_connections:
x = x.mean(dim=1) # [B, S, H]
if self.hrm_refine is not None:
x = self.hrm_refine(x)
x = self.norm(x)
# ModularMind-on-V2: emit this specialist's output latent (for RecursiveLink).
out_latent = self.latent_out(x) if self.latent_out is not None else None
return x, present_key_values, total_aux_loss, out_latent
class SpikeWhaleLM(PreTrainedModel):
"""
Full causal LM combining all SpikeTransformer + NanoWhale features.
Training (forward with labels):
out = model(input_ids=ids, labels=ids)
loss = out.loss # CE + MTP loss + MoE aux loss
Generation:
out = model(input_ids=ids, use_cache=True)
past = out.past_key_values
out2 = model(input_ids=next_id, past_key_values=past, use_cache=True)
"""
config_class = SpikeWhaleConfig
base_model_prefix = "model"
supports_gradient_checkpointing = True
_no_split_modules = ["TransformerBlock"]
def __init__(self, cfg: SpikeWhaleConfig):
super().__init__(cfg)
self.model = SpikeWhaleModel(cfg)
self.lm_head = nn.Linear(cfg.hidden_size, cfg.vocab_size, bias=False)
nn.init.normal_(self.lm_head.weight, std=cfg.initializer_range)
if cfg.tie_word_embeddings:
self.lm_head.weight = self.model.embed_tokens.weight
# Multi-Token Prediction heads (NanoWhale): predict token at position+k
self.mtp_heads = nn.ModuleList([
nn.Linear(cfg.hidden_size, cfg.vocab_size, bias=False)
for _ in range(cfg.num_nextn_predict_layers)
]) if cfg.num_nextn_predict_layers > 0 else None
self.post_init()
# HF post_init re-randomizes Linear weights, clobbering the zero-init
# injection gate. Restore it so the latent injection is identity-at-start.
self.model.reset_latent_gate()
def get_input_embeddings(self):
return self.model.embed_tokens
def set_input_embeddings(self, value):
self.model.embed_tokens = value
def get_output_embeddings(self):
return self.lm_head
def set_output_embeddings(self, new_embeddings):
self.lm_head = new_embeddings
def _set_gradient_checkpointing(self, module, value=False):
if isinstance(module, SpikeWhaleModel):
module.gradient_checkpointing = value
def forward(
self,
input_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
past_key_values: Optional[List[Tuple]] = None,
labels: Optional[torch.Tensor] = None,
use_cache: bool = False,
inject_latent: Optional[torch.Tensor] = None,
**kwargs,
) -> CausalLMOutputWithPast:
hidden, present_kvs, aux_loss, out_latent = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
use_cache=use_cache,
inject_latent=inject_latent,
)
logits = self.lm_head(hidden)
loss = None
if labels is not None:
# Standard next-token CE loss (shifted by 1)
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
loss = F.cross_entropy(
shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1),
ignore_index=-100,
)
# Multi-Token Prediction loss (NanoWhale)
# Each MTP head k predicts token at position + k+1 (beyond the standard +1)
if self.mtp_heads is not None:
mtp_total = torch.tensor(0.0, device=loss.device)
for k, head in enumerate(self.mtp_heads, start=1):
offset = k + 1 # predicts position + offset
if hidden.size(1) > offset:
mtp_logits = head(hidden[..., :-offset, :].contiguous())
mtp_labels = labels[..., offset:].contiguous()
mtp_total = mtp_total + F.cross_entropy(
mtp_logits.view(-1, mtp_logits.size(-1)),
mtp_labels.view(-1),
ignore_index=-100,
)
loss = loss + mtp_total / max(len(self.mtp_heads), 1)
# MoE load-balancing aux loss (My Project)
loss = loss + aux_loss
out = CausalLMOutputWithPast(
loss=loss,
logits=logits,
past_key_values=present_kvs,
)
out.latent = out_latent # ModularMind-on-V2: this specialist's output latent
return out
def count_parameters(self) -> int:
return sum(p.numel() for p in self.parameters()) |