File size: 21,460 Bytes
63e99b4 | 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 | """
GLADIUS v2.0 β MoDA: Multi-Head Depth Attention
The insight: Standard transformers compute Q, K, V from the CURRENT layer's hidden
state only. But every previous layer already computed useful representations that get
discarded. MoDA adds a second set of K, V projections that attend over a "depth cache"
β the hidden states from ALL previous layers at each position.
This is NOT cross-attention (fixed external memory). This is SELF-attention through
depth β the model attending to its own computation history.
Architecture per layer l:
Sequence path (standard):
Q_seq = W_q @ x_l (query from current layer)
K_seq = W_k @ x_l (key from current layer)
V_seq = W_v @ x_l (value from current layer)
O_seq = softmax(Q_seq @ K_seq^T / sqrt(d)) @ V_seq
Depth path (NEW):
K_depth = W_k_depth @ stack(x_0, x_1, ..., x_{l-1})
V_depth = W_v_depth @ stack(x_0, x_1, ..., x_{l-1})
O_depth = softmax(Q_seq @ K_depth^T / sqrt(d)) @ V_depth
Combined:
O = gate * O_seq + (1 - gate) * O_depth
The depth KV projections are TINY (hidden_dim β head_dim per group), and the depth
cache grows linearly with layers (not sequence length), so the cost is negligible.
For GLADIUS Wyrm (640d, 14L, 20H):
Depth cache at layer 13: 13 Γ seq_len depth tokens per position
Extra params per layer: 2 Γ hidden_dim Γ (hidden_dim / num_groups) β 40K
Total extra: 14 Γ 40K β 560K params (0.5% of 104.9M)
Reference: MoDA paper (Multi-Head Depth Attention) + Ali's SLA2 hybrid architecture.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from .config import KernelConfig
from .attention import RoPE, RMSNorm, SwiGLU
class DepthKVProjection(nn.Module):
"""
Projects depth cache hidden states into K, V for depth attention.
Uses Grouped Query Attention (GQA) style β fewer KV heads than Q heads
to keep the depth path lightweight.
For Wyrm: 20 Q heads, 4 KV groups β 5 Q heads per KV group.
"""
def __init__(self, hidden_dim: int, num_kv_heads: int, head_dim: int):
super().__init__()
self.num_kv_heads = num_kv_heads
self.head_dim = head_dim
self.kv_dim = num_kv_heads * head_dim
self.k_proj = nn.Linear(hidden_dim, self.kv_dim, bias=False)
self.v_proj = nn.Linear(hidden_dim, self.kv_dim, bias=False)
self._init_weights()
def _init_weights(self):
# Initialize from small noise β depth path starts quiet
nn.init.normal_(self.k_proj.weight, std=0.005)
nn.init.normal_(self.v_proj.weight, std=0.005)
def forward(self, depth_cache: torch.Tensor):
"""
Args:
depth_cache: (batch, depth_len, hidden_dim) β stacked hidden states from previous layers
Returns:
K_depth: (batch, num_kv_heads, depth_len, head_dim)
V_depth: (batch, num_kv_heads, depth_len, head_dim)
"""
B, D_len, _ = depth_cache.shape
K = self.k_proj(depth_cache).view(B, D_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
V = self.v_proj(depth_cache).view(B, D_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
return K, V
class MoDAAttention(nn.Module):
"""
Multi-Head Depth Attention β the core MoDA mechanism.
Combines standard sequence attention (SLA2 hybrid: softmax + linear blend)
with depth attention over previous layers' hidden states.
The depth path uses GQA with fewer KV heads for efficiency.
A learned gate controls the blend between sequence and depth paths.
"""
def __init__(self, config: KernelConfig, layer_idx: int = 0,
num_depth_kv_heads: int = 4):
super().__init__()
self.config = config
self.layer_idx = layer_idx
self.num_heads = config.num_heads
self.head_dim = config.head_dim
self.hidden_dim = config.hidden_dim
self.num_depth_kv_heads = num_depth_kv_heads
# How many Q heads share each depth KV head
assert config.num_heads % num_depth_kv_heads == 0, \
f"num_heads ({config.num_heads}) must be divisible by num_depth_kv_heads ({num_depth_kv_heads})"
self.q_per_kv = config.num_heads // num_depth_kv_heads
# === Sequence path (standard projections) ===
self.q_proj = nn.Linear(config.hidden_dim, config.hidden_dim, bias=False)
self.k_proj = nn.Linear(config.hidden_dim, config.hidden_dim, bias=False)
self.v_proj = nn.Linear(config.hidden_dim, config.hidden_dim, bias=False)
self.o_proj = nn.Linear(config.hidden_dim, config.hidden_dim, bias=False)
# === Depth path (new KV projections over depth cache) ===
self.depth_kv = DepthKVProjection(config.hidden_dim, num_depth_kv_heads, config.head_dim)
# === Depth gate: per-head learned blend between seq and depth ===
# Input: current hidden state β per-head gate in [0, 1]
# 0 = pure sequence attention, 1 = pure depth attention
self.depth_gate = nn.Sequential(
nn.Linear(config.hidden_dim, config.num_heads),
nn.Sigmoid()
)
# === SLA2 alpha: blend between softmax and linear (sequence path only) ===
self.alpha_router = nn.Sequential(
nn.Linear(config.hidden_dim, config.num_heads),
nn.Sigmoid()
)
# === RoPE (sequence positions only β depth has no position) ===
self.rope = RoPE(config.head_dim, config.max_seq_len)
# QK-Clip for stability
self.qk_softcap = getattr(config, 'qk_softcap', None)
self._init_weights()
def _init_weights(self):
for proj in [self.q_proj, self.k_proj, self.v_proj, self.o_proj]:
nn.init.normal_(proj.weight, std=0.02)
# Initialize depth gate bias negative β start with mostly sequence attention
# sigmoid(-2) β 0.12, so depth starts at ~12% influence
nn.init.constant_(self.depth_gate[0].bias, -2.0)
nn.init.zeros_(self.alpha_router[0].bias)
def _expand_kv_heads(self, kv: torch.Tensor) -> torch.Tensor:
"""
Expand GQA KV heads to match Q heads.
(B, num_kv_heads, L, D) β (B, num_heads, L, D)
"""
B, H_kv, L, D = kv.shape
# Repeat each KV head for q_per_kv Q heads
return kv.unsqueeze(2).expand(B, H_kv, self.q_per_kv, L, D).reshape(B, self.num_heads, L, D)
def forward(
self,
x: torch.Tensor,
mask: torch.Tensor | None = None,
depth_cache: torch.Tensor | None = None,
) -> torch.Tensor:
"""
Args:
x: (batch, seq_len, hidden_dim) β current layer input
mask: (batch, 1, seq_len, seq_len) β causal mask
depth_cache: (batch, num_prev_layers * seq_len, hidden_dim) β stacked previous layers
OR None for layer 0 (no depth history yet)
Returns:
(batch, seq_len, hidden_dim)
"""
B, S, D = x.shape
# === Sequence projections ===
Q = self.q_proj(x).view(B, S, self.num_heads, self.head_dim).transpose(1, 2) # (B, H, S, D_h)
K = self.k_proj(x).view(B, S, self.num_heads, self.head_dim).transpose(1, 2)
V = self.v_proj(x).view(B, S, self.num_heads, self.head_dim).transpose(1, 2)
# Apply RoPE to sequence Q, K
Q_rope = self.rope(Q, S)
K_rope = self.rope(K, S)
# === SLA2 Hybrid Sequence Attention ===
# Linear path
Q_lin = F.elu(Q_rope) + 1
K_lin = F.elu(K_rope) + 1
KV_lin = torch.matmul(K_lin.transpose(-2, -1), V)
Z_lin = K_lin.transpose(-2, -1).sum(dim=-1, keepdim=True)
O_linear = torch.matmul(Q_lin, KV_lin) / (torch.matmul(Q_lin, Z_lin) + 1e-6)
# Softmax path
scores = torch.matmul(Q_rope, K_rope.transpose(-2, -1)) / math.sqrt(self.head_dim)
if self.qk_softcap is not None and self.qk_softcap > 0:
scores = self.qk_softcap * torch.tanh(scores / self.qk_softcap)
if mask is not None:
scores = scores.masked_fill(mask == 0, float('-inf'))
attn_weights = F.softmax(scores, dim=-1)
O_softmax = torch.matmul(attn_weights, V)
# SLA2 blend
alpha = self.alpha_router(x).permute(0, 2, 1).unsqueeze(-1) # (B, H, S, 1)
O_seq = alpha * O_softmax + (1 - alpha) * O_linear
# === Depth Attention (if we have depth history) ===
if depth_cache is not None and depth_cache.shape[1] > 0:
# Project depth cache to K, V
K_depth, V_depth = self.depth_kv(depth_cache) # (B, H_kv, D_len, D_h)
# Expand GQA to match Q heads
K_depth = self._expand_kv_heads(K_depth) # (B, H, D_len, D_h)
V_depth = self._expand_kv_heads(V_depth)
# Depth attention scores β Q from current layer, K from depth cache
# No RoPE on depth (depth positions are layer indices, not sequence positions)
# Use un-rotated Q for depth to keep the paths independent
depth_scores = torch.matmul(Q, K_depth.transpose(-2, -1)) / math.sqrt(self.head_dim)
if self.qk_softcap is not None and self.qk_softcap > 0:
depth_scores = self.qk_softcap * torch.tanh(depth_scores / self.qk_softcap)
# No causal mask needed for depth β all previous layers are always available
depth_attn = F.softmax(depth_scores, dim=-1)
O_depth = torch.matmul(depth_attn, V_depth) # (B, H, S, D_h)
# Depth gate: how much to blend in depth attention
gate = self.depth_gate(x).permute(0, 2, 1).unsqueeze(-1) # (B, H, S, 1)
O = (1 - gate) * O_seq + gate * O_depth
else:
# Layer 0: no depth history, pure sequence attention
O = O_seq
# Reshape and project output
O = O.transpose(1, 2).contiguous().view(B, S, D)
return self.o_proj(O)
class MoDATransformerLayer(nn.Module):
"""
Transformer layer with MoDA attention.
Drop-in replacement for TransformerLayer, but forward() now accepts
and returns depth_cache for the depth attention mechanism.
"""
def __init__(self, config: KernelConfig, layer_idx: int = 0,
num_depth_kv_heads: int = 4):
super().__init__()
self.layer_idx = layer_idx
self.attention = MoDAAttention(config, layer_idx, num_depth_kv_heads)
self.ffn = SwiGLU(config)
self.attn_norm = RMSNorm(config.hidden_dim)
self.ffn_norm = RMSNorm(config.hidden_dim)
def forward(
self,
x: torch.Tensor,
mask: torch.Tensor | None = None,
depth_cache: torch.Tensor | None = None,
) -> torch.Tensor:
"""
Args:
x: (batch, seq_len, hidden_dim)
mask: causal mask
depth_cache: stacked previous layer outputs (batch, prev_layers * seq_len, hidden_dim)
Returns:
x: (batch, seq_len, hidden_dim) β output of this layer
"""
x = x + self.attention(self.attn_norm(x), mask=mask, depth_cache=depth_cache)
x = x + self.ffn(self.ffn_norm(x))
return x
def upgrade_kernel_to_moda(kernel, num_depth_kv_heads: int = 4,
init_from_sequence: bool = True):
"""
Surgical upgrade: replace HybridAttention layers with MoDA layers.
Preserves ALL existing weights (Q, K, V, O projections, FFN, norms).
Only adds new depth_kv projections and depth_gate.
Args:
kernel: GladiusKernel instance with existing trained weights
num_depth_kv_heads: number of KV heads for depth attention (GQA)
init_from_sequence: if True, initialize depth KV from sequence KV weights
Returns:
Modified kernel with MoDA layers (in-place)
"""
config = kernel.config
device = next(kernel.parameters()).device
dtype = next(kernel.parameters()).dtype
new_layers = nn.ModuleList()
for i, old_layer in enumerate(kernel.layers):
# Create new MoDA layer
moda_layer = MoDATransformerLayer(config, layer_idx=i,
num_depth_kv_heads=num_depth_kv_heads)
# === Transfer existing weights ===
# Sequence Q, K, V, O projections
moda_layer.attention.q_proj.weight.data.copy_(old_layer.attention.q_proj.weight.data)
moda_layer.attention.k_proj.weight.data.copy_(old_layer.attention.k_proj.weight.data)
moda_layer.attention.v_proj.weight.data.copy_(old_layer.attention.v_proj.weight.data)
moda_layer.attention.o_proj.weight.data.copy_(old_layer.attention.o_proj.weight.data)
# SLA2 alpha router
moda_layer.attention.alpha_router[0].weight.data.copy_(old_layer.attention.alpha_router[0].weight.data)
moda_layer.attention.alpha_router[0].bias.data.copy_(old_layer.attention.alpha_router[0].bias.data)
# RoPE buffers
moda_layer.attention.rope.inv_freq.data.copy_(old_layer.attention.rope.inv_freq.data)
moda_layer.attention.rope.cos_cached.data.copy_(old_layer.attention.rope.cos_cached.data)
moda_layer.attention.rope.sin_cached.data.copy_(old_layer.attention.rope.sin_cached.data)
# FFN (SwiGLU)
moda_layer.ffn.gate_proj.weight.data.copy_(old_layer.ffn.gate_proj.weight.data)
moda_layer.ffn.up_proj.weight.data.copy_(old_layer.ffn.up_proj.weight.data)
moda_layer.ffn.down_proj.weight.data.copy_(old_layer.ffn.down_proj.weight.data)
# Norms
moda_layer.attn_norm.weight.data.copy_(old_layer.attn_norm.weight.data)
moda_layer.ffn_norm.weight.data.copy_(old_layer.ffn_norm.weight.data)
# === Initialize depth KV from sequence KV (warm start) ===
if init_from_sequence:
# Take a subset of the sequence K, V weights for depth initialization
# Map from full hidden_dim projection to GQA-sized projection
seq_k_weight = old_layer.attention.k_proj.weight.data # (hidden_dim, hidden_dim)
seq_v_weight = old_layer.attention.v_proj.weight.data
# Select every q_per_kv-th head's K,V weights for the depth KV heads
head_dim = config.head_dim
q_per_kv = config.num_heads // num_depth_kv_heads
kv_dim = num_depth_kv_heads * head_dim
# Extract weights for the GQA KV heads (take every q_per_kv-th head)
depth_k_weight = torch.zeros(kv_dim, config.hidden_dim, device=device, dtype=dtype)
depth_v_weight = torch.zeros(kv_dim, config.hidden_dim, device=device, dtype=dtype)
for g in range(num_depth_kv_heads):
src_head = g * q_per_kv # Take the first Q head from each group
src_start = src_head * head_dim
src_end = src_start + head_dim
dst_start = g * head_dim
dst_end = dst_start + head_dim
depth_k_weight[dst_start:dst_end] = seq_k_weight[src_start:src_end]
depth_v_weight[dst_start:dst_end] = seq_v_weight[src_start:src_end]
# Scale down to let depth path start gentle
moda_layer.attention.depth_kv.k_proj.weight.data.copy_(depth_k_weight * 0.1)
moda_layer.attention.depth_kv.v_proj.weight.data.copy_(depth_v_weight * 0.1)
new_layers.append(moda_layer)
# Replace layers in kernel
kernel.layers = new_layers.to(device)
# We need to patch the forward to pass depth_cache through layers
# Store reference to original forward
kernel._moda_enabled = True
kernel._num_depth_kv_heads = num_depth_kv_heads
# Report new param count
total = sum(p.numel() for p in kernel.parameters())
trainable = sum(p.numel() for p in kernel.parameters() if p.requires_grad)
depth_params = sum(
sum(p.numel() for p in layer.attention.depth_kv.parameters()) +
sum(p.numel() for p in layer.attention.depth_gate.parameters())
for layer in kernel.layers
)
print(f"\n=== MoDA Upgrade Complete ===")
print(f" Total params: {total:,} (+{depth_params:,} depth params)")
print(f" Trainable: {trainable:,}")
print(f" Depth overhead: {depth_params/total*100:.2f}%")
print(f" Depth KV heads: {num_depth_kv_heads} (GQA ratio: {config.num_heads // num_depth_kv_heads}:1)")
print(f" Memory: {total * 2 / 1024 / 1024:.1f} MB (bfloat16)")
return kernel
class MoDAKernelMixin:
"""
Mixin to patch GladiusKernel.forward() for depth cache propagation.
Usage:
kernel = GladiusKernel.load_checkpoint(path)
kernel = upgrade_kernel_to_moda(kernel)
patch_kernel_forward_for_moda(kernel)
"""
pass
def patch_kernel_forward_for_moda(kernel):
"""
Monkey-patch the kernel's forward to thread depth cache through layers.
The original forward just does:
for layer in self.layers:
x = layer(x, mask=mask)
MoDA needs:
depth_cache = []
for layer in self.layers:
x = layer(x, mask=mask, depth_cache=stack(depth_cache))
depth_cache.append(x)
"""
import types
_original_forward = kernel.forward
def moda_forward(self, input_ids=None, timestamp=None, images=None, audio=None):
"""MoDA-patched forward: threads depth cache through transformer layers."""
# Replicate pre-transformer logic from original forward
text_embeds = None
if input_ids is not None:
B, S = input_ids.shape
text_embeds = self.embeddings.embed(input_ids)
modality_mask = None
if self.has_senses and (images is not None or audio is not None):
x, modality_mask = self.senses(text_embeds=text_embeds, images=images, audio=audio)
B, S = x.shape[0], x.shape[1]
elif text_embeds is not None:
x = text_embeds
B, S = x.shape[0], x.shape[1]
else:
raise ValueError("Must provide input_ids, images, or audio")
# Memory read
x = self.memory.read(x)
# Temporal encoding
time_embed = None
if timestamp is not None:
if isinstance(timestamp, (int, float)):
timestamp = torch.tensor([timestamp] * B, dtype=torch.float32, device=x.device)
time_embed = self.time_engine(timestamp)
x = x + time_embed.unsqueeze(1)
# === MoDA Transformer layers with depth cache ===
if S <= self.config.max_seq_len:
mask = self.causal_mask[:, :, :S, :S]
else:
mask = torch.tril(torch.ones(1, 1, S, S, device=x.device))
depth_states = [] # Collect hidden states for depth attention
for layer in self.layers:
# Build depth cache from all previous layers
if len(depth_states) > 0:
# Stack previous layer summaries: (B, num_prev_layers, D)
# Each layer contributes ONE summary vector per position, not S vectors
# This keeps depth cache at O(L) not O(L*S) β critical for VRAM
depth_cache = torch.stack(depth_states, dim=1) # (B, num_prev, D)
else:
depth_cache = None
x = layer(x, mask=mask, depth_cache=depth_cache)
# Store mean-pooled representation of this layer (detached)
depth_states.append(x.mean(dim=1).detach()) # (B, D)
# NOTE: detach() means depth cache doesn't backprop through previous layers.
# This is intentional β depth attention learns to READ from previous layers,
# not to CHANGE them. Keeps memory O(L*S*D) instead of O(L^2*S*D).
# Final norm
x = self.final_norm(x)
# Tool check
tool_result = self.tool_cortex.check_activation(x)
if tool_result is not None:
x = x + tool_result
# Modulate
logits, silence, pixel_output = self.modulator(x, self.embeddings.output_head, temporal_embedding=time_embed)
# Memory write
importance = self.memory.write(x)
# Cognition heartbeat
mode, cognitive_state, mode_probs = self.cognition.heartbeat(x)
if self.cognition.should_consolidate():
self.memory.consolidate()
self.time_engine.record_event()
return {
'logits': logits,
'silence': silence,
'pixel_output': pixel_output,
'mode': mode,
'importance': importance,
'modality_mask': modality_mask,
'cognitive_state': cognitive_state,
'mode_probs': mode_probs,
}
kernel.forward = types.MethodType(moda_forward, kernel)
print(" Forward pass patched for MoDA depth cache propagation β
")
return kernel
|