File size: 32,062 Bytes
5762a7c | 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 | #!/usr/bin/env python3
"""SymbolicLight V1 model implementation."""
import math
from dataclasses import dataclass, field
from typing import Optional, Tuple, List
import torch
import torch.nn as nn
import torch.nn.functional as F
@dataclass
class SymbolicLightConfig:
"""Default configuration for SymbolicLight V1."""
vocab_size: int = 57344
embed_dim: int = 1536
n_layers: int = 22
n_heads: int = 24
head_dim: int = 64
intermediate_dim: int = 6144
max_seq_len: int = 512
spike_chunk_size: int = 64
dropout: float = 0.1
spike_threshold: float = 1.0
leak_factor: float = 0.95
stdp_lr: float = 0.01
enable_stdp: bool = False
rope_theta: float = 10000.0
frontend_mode: str = "text"
sparse_attn_window: int = 512
n_global_anchors: int = 4
enable_sparse_attn: bool = True
enable_dynamic_prior: bool = True
use_topk_mask: bool = False
topk_sparsity: float = 0.89
class ATanSurrogate(torch.autograd.Function):
"""ATan surrogate-gradient spike function."""
@staticmethod
def forward(ctx, membrane_potential, threshold):
ctx.save_for_backward(membrane_potential, torch.tensor(threshold,
device=membrane_potential.device,
dtype=membrane_potential.dtype))
return (membrane_potential >= threshold).float()
@staticmethod
def backward(ctx, grad_output):
membrane_potential, threshold = ctx.saved_tensors
alpha = 2.0
grad_v = 1.0 / (1.0 + (alpha * (membrane_potential - threshold)) ** 2)
return grad_output * grad_v, None
def surrogate_spike(membrane_potential: torch.Tensor, threshold: float = 1.0) -> torch.Tensor:
"""Apply the surrogate spike function."""
return ATanSurrogate.apply(membrane_potential, threshold)
class RotaryPositionEncoding(nn.Module):
"""Rotary position embedding."""
def __init__(self, dim: int, 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)
def forward(self, x: torch.Tensor, offset: int = 0) -> torch.Tensor:
"""
Apply RoPE rotation to the input tensor.
Args:
x: [B, S, D] input continuous representation
offset: position offset for incremental decoding
Returns:
rotated: [B, S, D] rotated representation
"""
B, S, D = x.shape
t = torch.arange(offset, offset + S, device=x.device, dtype=torch.float32)
freqs = torch.outer(t, self.inv_freq.to(x.device))
emb = torch.cat([freqs, freqs], dim=-1)
cos_emb = emb.cos().unsqueeze(0)
sin_emb = emb.sin().unsqueeze(0)
x_rotated = torch.cat([
-x[..., D // 2:],
x[..., :D // 2],
], dim=-1)
return x * cos_emb + x_rotated * sin_emb
class FrontendRouter(nn.Module):
"""Text embedding frontend."""
def __init__(self, config: SymbolicLightConfig):
super().__init__()
self.config = config
self.text_embedding = nn.Embedding(config.vocab_size, config.embed_dim)
def forward(self, token_ids: torch.Tensor, modality: str = "text") -> torch.Tensor:
if modality == "text":
return self.text_embedding(token_ids)
elif modality == "vision":
raise NotImplementedError("Vision frontend is not included in this release.")
elif modality == "audio":
raise NotImplementedError("Audio frontend is not included in this release.")
else:
raise ValueError(f"Unknown modality: {modality}")
def _lif_scan_forward(x: torch.Tensor, v_mem: torch.Tensor,
leak: float, threshold: float) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
JIT forward pass for temporal LIF neuron scanning.
Returns: (spikes [B,S,D], final_v_mem [B,D], all_v_mem [B,S,D] for backward)
"""
B, S, D = x.shape
spikes = torch.empty_like(x)
all_v = torch.empty_like(x)
for t in range(S):
v_mem = v_mem * leak + x[:, t, :]
v_mem = torch.clamp(v_mem, -3.0, 3.0)
all_v[:, t, :] = v_mem
spike = (v_mem >= threshold).float()
v_mem = v_mem * (1.0 - spike)
spikes[:, t, :] = spike
return spikes, v_mem, all_v
class LIFScan(torch.autograd.Function):
"""LIF scan with ATan surrogate gradient for backward."""
@staticmethod
def forward(ctx, x, v_mem, leak, threshold):
spikes, final_v, all_v = _lif_scan_forward(x, v_mem, leak, threshold)
ctx.save_for_backward(all_v)
ctx.threshold = threshold
return spikes, final_v
@staticmethod
def backward(ctx, grad_spikes, grad_v_mem):
all_v, = ctx.saved_tensors
alpha = 2.0
surrogate_grad = 1.0 / (1.0 + (alpha * (all_v - ctx.threshold)) ** 2)
grad_x = grad_spikes * surrogate_grad
return grad_x, None, None, None
class SpikeEncoder(nn.Module):
"""
Convert discrete token IDs into spatiotemporal spike tensors.
Main design updates:
- remove learned positional embeddings and use RoPE in SparseTCAM
- use chunk-parallel LIF spike conversion to reduce Python loops
- route token embeddings through FrontendRouter instead of a hard-coded embedding
Flow: token_id -> FrontendRouter -> LayerNorm -> parallel LIF spike conversion
"""
def __init__(self, config: SymbolicLightConfig):
super().__init__()
self.config = config
self.frontend = FrontendRouter(config)
self.norm = nn.LayerNorm(config.embed_dim)
self.threshold = config.spike_threshold
self.leak = config.leak_factor
self.v_mem = None
def _init_membrane(self, shape: torch.Size, device: torch.device):
"""Initialize or reset the membrane potential."""
self.v_mem = torch.zeros(shape, device=device)
def forward(self, token_ids: torch.Tensor, use_cache: bool = False,
cache: dict = None) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Args:
token_ids: [batch, seq_len]
use_cache: if True, run O(1) incremental decoding
cache: cache dictionary
Returns:
spikes: [batch, seq_len, embed_dim] sparse 0/1 spikes
continuous: [batch, seq_len, embed_dim] continuous residual stream
"""
B, S = token_ids.shape
if use_cache and cache is not None:
if 'v_mem' not in cache:
cache['v_mem'] = torch.zeros(B, self.config.embed_dim, device=token_ids.device)
if 'seq_len' not in cache:
cache['seq_len'] = 0
self.v_mem = cache['v_mem']
cache['seq_len'] += S
else:
self._init_membrane((B, self.config.embed_dim), token_ids.device)
x = self.frontend(token_ids)
x = self.norm(x)
if getattr(self.config, 'use_topk_mask', False):
k = max(1, int((1.0 - self.config.topk_sparsity) * self.config.embed_dim))
_, topk_indices = torch.topk(x.abs(), k, dim=-1)
spikes = torch.zeros_like(x)
spikes.scatter_(-1, topk_indices, 1.0)
if self.training:
spikes = spikes + (surrogate_spike(x, self.threshold) - spikes).detach()
else:
chunk_size = self.config.spike_chunk_size
spikes_list = []
for chunk_start in range(0, S, chunk_size):
chunk_end = min(chunk_start + chunk_size, S)
x_chunk = x[:, chunk_start:chunk_end, :]
chunk_spikes, self.v_mem = LIFScan.apply(
x_chunk, self.v_mem, self.leak, self.threshold
)
spikes_list.append(chunk_spikes)
if self.training:
self.v_mem = self.v_mem.detach()
spikes = torch.cat(spikes_list, dim=1)
if use_cache and cache is not None:
cache['v_mem'] = self.v_mem.detach()
return spikes, x
class SparseLocalAttention(nn.Module):
"""
Compute attention only among active spike positions with a local window and global anchors.
Key idea:
- dense attention attends across all S positions -> O(S^2)
- this path only attends over active positions inside a local window -> O(S * k * w)
where k is the active fraction and w is the window size
- global anchors let the first few tokens interact broadly and stabilize global context
Relation to the decay path:
- the decay path compresses history into a fixed-size hidden state for coarse long-range memory
- the attention path focuses precisely on recent informative positions for local reasoning
- a learned gate blends both paths
"""
def __init__(self, config: SymbolicLightConfig):
super().__init__()
self.window_size = max(1, int(config.sparse_attn_window))
self.n_global_anchors = config.n_global_anchors
self.n_heads = config.n_heads
self.head_dim = config.head_dim
self.scale = config.head_dim ** -0.5
self._use_sdpa = hasattr(F, "scaled_dot_product_attention")
self.q_proj = nn.Linear(config.embed_dim, config.embed_dim, bias=False)
self.k_proj = nn.Linear(config.embed_dim, config.embed_dim, bias=False)
self.v_proj = nn.Linear(config.embed_dim, config.embed_dim, bias=False)
self.rope = RotaryPositionEncoding(config.head_dim, theta=config.rope_theta)
S = config.max_seq_len
q_pos = torch.arange(S).unsqueeze(1)
k_pos = torch.arange(S).unsqueeze(0)
distance = q_pos - k_pos
causal = distance >= 0
window = (q_pos - k_pos) <= self.window_size
anchors = k_pos < self.n_global_anchors
self.register_buffer('_cached_mask', causal & (window | anchors))
def forward(self, x: torch.Tensor, spike_mask: torch.Tensor,
offset: int = 0, use_cache: bool = False, cache: dict = None) -> torch.Tensor:
"""
Args:
x: [B, S_q, D] continuous representation; RoPE is applied internally to Q/K
spike_mask: [B, S_q] boolean mask, True means the position fired a spike
offset: RoPE position offset for incremental decoding
use_cache: whether to use the KV cache for incremental decoding
cache: KV cache dictionary
Returns:
attn_out: [B, S_q, D] sparse attention output with zeros on inactive positions
"""
B, S_q, D = x.shape
Q = self.q_proj(x).view(B, S_q, self.n_heads, self.head_dim).transpose(1, 2)
K = self.k_proj(x).view(B, S_q, self.n_heads, self.head_dim).transpose(1, 2)
V = self.v_proj(x).view(B, S_q, self.n_heads, self.head_dim).transpose(1, 2)
Q = self.rope(Q.contiguous().view(B * self.n_heads, S_q, self.head_dim), offset=offset)
Q = Q.view(B, self.n_heads, S_q, self.head_dim).to(V.dtype)
K = self.rope(K.contiguous().view(B * self.n_heads, S_q, self.head_dim), offset=offset)
K = K.view(B, self.n_heads, S_q, self.head_dim).to(V.dtype)
if use_cache and cache is not None:
if 'K' in cache:
K = torch.cat([cache['K'], K], dim=2)
V = torch.cat([cache['V'], V], dim=2)
spike_mask_kv = torch.cat([cache['spike_mask'], spike_mask], dim=1)
else:
spike_mask_kv = spike_mask
cache['K'] = K.detach()
cache['V'] = V.detach()
cache['spike_mask'] = spike_mask_kv.detach()
else:
spike_mask_kv = spike_mask
S_kv = K.size(2)
if offset == 0 and S_q == S_kv and S_q == self._cached_mask.size(0):
attn_mask = self._cached_mask
else:
q_pos = torch.arange(offset, offset + S_q, device=x.device).unsqueeze(1)
k_pos = torch.arange(0, S_kv, device=x.device).unsqueeze(0)
distance = q_pos - k_pos
causal = distance >= 0
window = distance <= self.window_size
anchors = k_pos < self.n_global_anchors
attn_mask = causal & (window | anchors)
spike_key_mask = spike_mask_kv.unsqueeze(1).unsqueeze(2)
full_mask = attn_mask.unsqueeze(0).unsqueeze(0) & spike_key_mask
query_has_any_key = full_mask.any(dim=-1, keepdim=True)
if self._use_sdpa:
safe_mask = full_mask | ~query_has_any_key
attn_out = F.scaled_dot_product_attention(Q, K, V, attn_mask=safe_mask, dropout_p=0.0)
attn_out = attn_out.masked_fill(~query_has_any_key, 0.0)
else:
scores = torch.matmul(Q, K.transpose(-2, -1)) * self.scale
scores = scores.masked_fill(~full_mask, float('-inf'))
scores = scores.masked_fill(~query_has_any_key, 0.0)
attn_weights = F.softmax(scores, dim=-1).to(V.dtype)
attn_weights = attn_weights.masked_fill(~query_has_any_key, 0.0)
attn_out = torch.matmul(attn_weights, V)
attn_out = attn_out.transpose(1, 2).contiguous().view(B, S_q, D)
attn_out = attn_out * spike_mask.unsqueeze(-1).to(dtype=attn_out.dtype)
return attn_out
class SparseTCAM(nn.Module):
"""Dual-path spike-gated sequence mixer."""
def __init__(self, config: SymbolicLightConfig):
super().__init__()
self.n_heads = config.n_heads
self.head_dim = config.head_dim
self.embed_dim = config.embed_dim
self.threshold = config.spike_threshold
self.leak = config.leak_factor
self.enable_sparse_attn = config.enable_sparse_attn
self.tcam_proj = nn.Linear(config.embed_dim, config.embed_dim, bias=False)
self.out_proj = nn.Linear(config.embed_dim, config.embed_dim, bias=False)
self.norm = nn.LayerNorm(config.embed_dim)
self.dropout = nn.Dropout(config.dropout)
self.decay_raw = nn.Parameter(torch.full((config.n_heads,), 3.0))
if self.enable_sparse_attn:
self.sparse_attn = SparseLocalAttention(config)
self.attn_gate = nn.Parameter(torch.zeros(1))
def forward(self, spikes: torch.Tensor, continuous: torch.Tensor,
use_cache: bool = False, cache: dict = None) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Main updates:
1. apply RoPE inside SparseLocalAttention after Q/K projection
2. allow training-time cache reads and writes for hidden state h across chunks
"""
B, S, D = spikes.shape
compute_dtype = continuous.dtype
if spikes.dtype != compute_dtype:
spikes = spikes.to(compute_dtype)
offset = 0
if use_cache and cache is not None:
offset = cache.get('rope_offset', 0)
cache['rope_offset'] = offset + S
spike_energy = spikes.sum(dim=-1)
active_mask = (spike_energy > 0).unsqueeze(-1).to(dtype=compute_dtype)
tcam_out = self.tcam_proj(spikes * active_mask)
tcam_out = tcam_out.view(B, S, self.n_heads, self.head_dim)
decay = torch.sigmoid(self.decay_raw)
if cache is not None:
if 'h' not in cache:
cache['h'] = torch.zeros(B, self.n_heads, self.head_dim, device=spikes.device, dtype=compute_dtype)
h = cache['h']
else:
h = torch.zeros(B, self.n_heads, self.head_dim, device=spikes.device, dtype=compute_dtype)
if use_cache and cache is not None and S == 1:
h = decay.view(1, self.n_heads, 1) * h + (1 - decay.view(1, self.n_heads, 1)) * tcam_out[:, 0]
cache['h'] = h.detach()
context = h.unsqueeze(1)
else:
powers = torch.arange(S - 1, -1, -1, dtype=compute_dtype, device=spikes.device)
kernel = ((decay.view(-1, 1) ** powers.view(1, -1)) * (1 - decay).view(-1, 1)).unsqueeze(1)
tcam_out_trans = tcam_out.permute(0, 3, 2, 1).reshape(-1, self.n_heads, S)
tcam_out_pad = F.pad(tcam_out_trans, (S - 1, 0))
out = F.conv1d(tcam_out_pad, kernel, groups=self.n_heads)
context = out.view(-1, self.head_dim, self.n_heads, S).permute(0, 3, 2, 1)
powers_fwd = torch.arange(1, S + 1, dtype=compute_dtype, device=spikes.device).view(1, S, 1, 1)
decay_t = decay.view(1, 1, self.n_heads, 1) ** powers_fwd
context = context + h.unsqueeze(1) * decay_t
if cache is not None:
cache['h'] = context[:, -1, :, :].detach()
decay_output = context.reshape(B, S, D)
if self.enable_sparse_attn:
spike_mask = (spikes.sum(dim=-1) > 0)
attn_cache = cache.setdefault('attn', {}) if cache is not None else None
attn_output = self.sparse_attn(
continuous, spike_mask, offset=offset,
use_cache=use_cache, cache=attn_cache
)
gate = torch.sigmoid(self.attn_gate)
output = gate * attn_output + (1 - gate) * decay_output
else:
output = decay_output
output = self.out_proj(self.dropout(output))
out_continuous = self.norm(continuous + output)
out_spikes = surrogate_spike(out_continuous, self.threshold).to(out_continuous.dtype)
return out_spikes, out_continuous
class SpikingFeedForward(nn.Module):
"""
Two-layer feed-forward block used in place of the standard Transformer MLP.
The main difference is the LIF-style spike activation in the hidden layer.
"""
def __init__(self, config: SymbolicLightConfig):
super().__init__()
self.up = nn.Linear(config.embed_dim, config.intermediate_dim, bias=False)
self.down = nn.Linear(config.intermediate_dim, config.embed_dim, bias=False)
self.norm = nn.LayerNorm(config.embed_dim)
self.threshold = config.spike_threshold
self.dropout = nn.Dropout(config.dropout)
def forward(self, x: torch.Tensor) -> torch.Tensor:
residual = x
h = self.up(x)
h = surrogate_spike(h, self.threshold).to(x.dtype)
h = self.down(self.dropout(h))
return self.norm(residual + h)
class SymbolicLightBlock(nn.Module):
def __init__(self, config: SymbolicLightConfig):
super().__init__()
self.tcam = SparseTCAM(config)
self.ffn = SpikingFeedForward(config)
def forward(self, spikes, continuous, use_cache=False, cache=None):
spikes, continuous = self.tcam(spikes, continuous, use_cache=use_cache, cache=cache)
continuous = self.ffn(continuous)
spikes = surrogate_spike(continuous, self.tcam.threshold).to(continuous.dtype)
return spikes, continuous
class BayesianHead(nn.Module):
"""
Dynamic context-conditioned prior head.
Earlier versions used a static learned log_prior vector.
This version predicts log_prior from the current context with a lightweight network.
Bayesian form:
log P(word|context) = log P(context|word) + log P(word|context_summary)
likelihood term dynamic prior term
Intuition:
- when the context is about cooking, the prior can upweight tokens such as salt or pan
- when the context is about programming, the prior can upweight tokens such as function or loop
- this is more targeted than a static frequency bias
"""
def __init__(self, config: SymbolicLightConfig):
super().__init__()
self.output_proj = nn.Linear(config.embed_dim, config.vocab_size, bias=False)
self.use_dynamic_prior = getattr(config, 'enable_dynamic_prior', True)
self.prior_weight = nn.Parameter(torch.tensor(0.1))
if self.use_dynamic_prior:
bottleneck_dim = config.embed_dim // 4
self.prior_net = nn.Sequential(
nn.Linear(config.embed_dim, bottleneck_dim, bias=False),
nn.GELU(),
nn.Linear(bottleneck_dim, config.vocab_size, bias=False),
)
else:
self.log_prior = nn.Parameter(torch.zeros(config.vocab_size))
def forward(self, continuous: torch.Tensor) -> torch.Tensor:
"""
Args:
continuous: [B, S, D]
Returns:
logits: [B, S, vocab_size]
"""
log_likelihood = self.output_proj(continuous)
if self.use_dynamic_prior:
dynamic_prior = self.prior_net(continuous)
logits = log_likelihood + self.prior_weight * dynamic_prior
else:
logits = log_likelihood + self.prior_weight * self.log_prior
return logits
class STDPUpdater:
"""Optional local spike-timing update rule."""
def __init__(self, config: SymbolicLightConfig):
self.lr = config.stdp_lr
self.enabled = config.enable_stdp
@torch.no_grad()
def update(self, model: nn.Module, pre_spikes: torch.Tensor, post_spikes: torch.Tensor):
if not self.enabled:
return
causal = (pre_spikes.sum(dim=1, keepdim=True) > 0) & (post_spikes.sum(dim=1, keepdim=True) > 0)
if causal.any():
for block in model.blocks:
w = block.tcam.tcam_proj.weight
pre_active = (pre_spikes > 0).float()
post_active = (post_spikes > 0).float()
co_firing = torch.einsum('bsd,bse->de', post_active, pre_active)
delta = self.lr * co_firing / (pre_spikes.size(0) * pre_spikes.size(1))
mask = (co_firing > 0).float()
w.data += delta * mask * 0.05
w.data.clamp_(-5, 5)
class SymbolicLightModel(nn.Module):
"""SymbolicLight language model."""
def __init__(self, config: SymbolicLightConfig):
super().__init__()
self.config = config
self.spike_encoder = SpikeEncoder(config)
self.blocks = nn.ModuleList([
SymbolicLightBlock(config) for _ in range(config.n_layers)
])
self.output_head = BayesianHead(config)
self.stdp = STDPUpdater(config)
self.gradient_checkpointing = False
self.apply(self._init_weights)
n_params = sum(p.numel() for p in self.parameters())
print(f"[SymbolicLight V1] model initialized | parameters: {n_params/1e6:.1f}M ({n_params/1e9:.3f}B)")
def _init_weights(self, module):
if isinstance(module, nn.Linear):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
def compile_for_inference(self):
"""Apply torch.compile to selected inference-critical submodules."""
self.spike_encoder = torch.compile(self.spike_encoder, mode='reduce-overhead')
for block in self.blocks:
block.tcam = torch.compile(block.tcam, mode='reduce-overhead')
block.ffn = torch.compile(block.ffn, mode='reduce-overhead')
print("[SymbolicLight V1] torch.compile applied for inference acceleration")
def gradient_checkpointing_enable(self):
self.gradient_checkpointing = True
def gradient_checkpointing_disable(self):
self.gradient_checkpointing = False
def forward(self, token_ids: torch.Tensor, use_cache: bool = False,
past_key_values: list = None,
streaming_state: list = None):
"""
Shared forward pass for training and inference.
Args:
token_ids: [B, S] input token IDs
use_cache: whether to use the KV cache for inference
past_key_values: list of inference caches
streaming_state: list of streaming caches used to carry hidden state across chunks
during training; format matches past_key_values:
[encoder_cache, block0_cache, block1_cache, ...]
Returns:
logits: [B, S, vocab_size]
"""
if use_cache and past_key_values is None:
past_key_values = [{} for _ in range(len(self.blocks) + 1)]
if not use_cache and streaming_state is not None:
caches = streaming_state
elif use_cache:
caches = past_key_values
else:
caches = [None] * (len(self.blocks) + 1)
encoder_cache = caches[0] if caches[0] is not None else (
past_key_values[0] if use_cache else None
)
spikes, continuous = self.spike_encoder(token_ids, use_cache=use_cache, cache=encoder_cache)
model_dtype = self.output_head.output_proj.weight.dtype
if continuous.dtype != model_dtype:
continuous = continuous.to(model_dtype)
if spikes.dtype != model_dtype:
spikes = spikes.to(model_dtype)
initial_spikes = spikes
for i, block in enumerate(self.blocks):
block_cache = caches[i + 1] if caches[i + 1] is not None else (
past_key_values[i + 1] if use_cache else None
)
if self.training and self.gradient_checkpointing and not use_cache and block_cache is None:
def _checkpointed_block(spk, cont, current_block=block):
out_spikes, out_continuous = current_block(
spk, cont, use_cache=False, cache=None,
)
return out_spikes, out_continuous
spikes, continuous = torch.utils.checkpoint.checkpoint(
_checkpointed_block, spikes, continuous,
use_reentrant=False,
)
else:
spikes, continuous = block(
spikes, continuous,
use_cache=use_cache, cache=block_cache,
)
logits = self.output_head(continuous)
if not self.training and self.config.enable_stdp and initial_spikes.size(1) > 1:
self.stdp.update(self, initial_spikes, spikes)
return logits
@torch.no_grad()
def generate(self, prompt_ids: torch.Tensor, max_new_tokens: int = 100,
temperature: float = 0.8, top_k: int = 50,
adaptive_temperature: bool = True) -> torch.Tensor:
"""
Autoregressive text generation with O(1) cached incremental decoding.
Adaptive temperature:
- lower entropy -> lower temperature for more deterministic outputs
- higher entropy -> higher temperature for more exploratory outputs
- effective range is approximately [0.3, 1.5]
"""
self.eval()
generated = prompt_ids.clone()
past_key_values = [{} for _ in range(len(self.blocks) + 1)]
logits = self.forward(prompt_ids, use_cache=True, past_key_values=past_key_values)
def _adaptive_temp(raw_logits, base_temp):
"""Adjust temperature dynamically from the logits entropy."""
if not adaptive_temperature:
return base_temp
probs = F.softmax(raw_logits, dim=-1)
p = probs.clamp(1e-7, 1.0)
entropy = -(p * p.log()).sum(dim=-1).mean()
max_entropy = math.log(self.config.vocab_size)
norm_entropy = (entropy / max_entropy).clamp(0, 1)
temp = max(0.1, base_temp - norm_entropy.item() * (base_temp - 0.1))
return temp
raw_logits = logits[:, -1, :]
temp = _adaptive_temp(raw_logits, temperature)
next_logits = raw_logits / temp
if top_k > 0:
top_k_vals, _ = torch.topk(next_logits, top_k)
min_top_k = top_k_vals[:, -1].unsqueeze(-1)
next_logits[next_logits < min_top_k] = float('-inf')
probs = F.softmax(next_logits, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
generated = torch.cat([generated, next_token], dim=1)
for _ in range(1, max_new_tokens):
if next_token.item() == 2:
break
logits = self.forward(next_token, use_cache=True, past_key_values=past_key_values)
raw_logits = logits[:, -1, :]
temp = _adaptive_temp(raw_logits, temperature)
next_logits = raw_logits / temp
if top_k > 0:
top_k_vals, _ = torch.topk(next_logits, top_k)
min_top_k = top_k_vals[:, -1].unsqueeze(-1)
next_logits[next_logits < min_top_k] = float('-inf')
probs = F.softmax(next_logits, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
generated = torch.cat([generated, next_token], dim=1)
return generated
def get_sparsity_stats(self) -> dict:
"""Return sparsity statistics for debugging and reporting."""
stats = {}
with torch.no_grad():
dummy = torch.randint(0, 100, (1, 32))
spikes, _ = self.spike_encoder(dummy)
stats['encoder_sparsity'] = 1.0 - spikes.mean().item()
for i, block in enumerate(self.blocks):
spikes, _ = block(spikes, spikes)
stats[f'block_{i}_sparsity'] = 1.0 - spikes.mean().item()
return stats
if __name__ == "__main__":
print("=" * 60)
print(" SymbolicLight V1 model smoke test")
print("=" * 60)
config = SymbolicLightConfig(
vocab_size=57344,
embed_dim=768,
n_layers=12,
n_heads=12,
head_dim=64,
)
model = SymbolicLightModel(config)
dummy_input = torch.randint(0, 57344, (2, 128))
print(f"\nInput: batch=2, seq_len=128")
logits = model(dummy_input)
print(f"Output logits: {logits.shape}")
print(f"\nStreaming context test (2 chunks x 128 tokens)...")
chunk1 = torch.randint(0, 57344, (2, 128))
chunk2 = torch.randint(0, 57344, (2, 128))
streaming_state = [{} for _ in range(len(model.blocks) + 1)]
logits1 = model(chunk1, streaming_state=streaming_state)
print(f" Chunk 1 logits: {logits1.shape}, streaming state saved [OK]")
logits2 = model(chunk2, streaming_state=streaming_state)
print(f" Chunk 2 logits: {logits2.shape}, cross-chunk memory passed [OK]")
stats = model.get_sparsity_stats()
print(f"\nSparsity stats:")
for k, v in stats.items():
print(f" {k}: {v*100:.1f}% silent")
prompt = torch.randint(0, 57344, (1, 10))
print(f"\nAutoregressive generation test (prompt=10, gen=20)...")
output = model.generate(prompt, max_new_tokens=20)
print(f"Generated sequence length: {output.shape[1]}")
print("\n[PASS] SymbolicLight V1 smoke checks completed.")
print(" [1] RoPE rotary position encoding [OK]")
print(" [2] Cross-chunk state passing [OK]")
print(" [3] BayesianHead dynamic prior [OK]")
print(" [4] SpikeEncoder parallel scan [OK]")
print(" [5] FrontendRouter multimodal stub [OK]")
|