Spaces:
Paused
Paused
File size: 21,065 Bytes
ca19627 | 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 | """
AETHER-Net Attention Layers
5 types: GDN, Full, Mamba2, Sliding Window, Cross Attention
Each layer follows the same interface:
forward(hidden_states, attention_mask=None, position_ids=None, **kwargs) -> hidden_states
"""
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Tuple
class RMSNorm(nn.Module):
def __init__(self, hidden_size: int, eps: float = 1e-6):
super().__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.eps = eps
def forward(self, x):
variance = x.float().pow(2).mean(-1, keepdim=True)
x = x * torch.rsqrt(variance + self.eps)
return (self.weight * x).to(x.dtype)
def rotate_half(x):
x1, x2 = x.chunk(2, dim=-1)
return torch.cat((-x2, x1), dim=-1)
def apply_rotary_pos_emb(q, k, cos, sin):
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
class RotaryEmbedding(nn.Module):
def __init__(self, dim: int, max_seq_len: int = 262144, theta: float = 10000000.0):
super().__init__()
inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2).float() / dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
self.max_seq_len = max_seq_len
def forward(self, x, position_ids):
# position_ids: [B, L] β take first batch (all same for standard positions)
pos = position_ids[0] if position_ids.dim() == 2 else position_ids
freqs = torch.outer(pos.float(), self.inv_freq.to(pos.device))
emb = torch.cat((freqs, freqs), dim=-1)
return emb.cos().unsqueeze(0), emb.sin().unsqueeze(0)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 1. FULL ATTENTION (Softmax, GQA, RoPE) β O(nΒ²)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class FullAttention(nn.Module):
"""Standard grouped-query attention with RoPE.
Kept for 5 layers β provides precise token-to-token reasoning.
These layers maintain KV cache."""
def __init__(self, config):
super().__init__()
self.num_heads = config.num_attention_heads
self.num_kv_heads = config.num_kv_heads
self.head_dim = config.head_dim
self.num_kv_groups = self.num_heads // self.num_kv_heads
self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
self.k_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
self.v_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
self.o_proj = nn.Linear(self.num_heads * self.head_dim, config.hidden_size, bias=False)
# Output gate (Qwen3.5 style gated attention)
self.gate = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
self.rotary_emb = RotaryEmbedding(self.head_dim, config.max_position_embeddings, config.rope_theta)
def forward(self, hidden_states, attention_mask=None, position_ids=None, **kwargs):
B, L, _ = hidden_states.shape
q = self.q_proj(hidden_states).view(B, L, self.num_heads, self.head_dim).transpose(1, 2)
k = self.k_proj(hidden_states).view(B, L, self.num_kv_heads, self.head_dim).transpose(1, 2)
v = self.v_proj(hidden_states).view(B, L, self.num_kv_heads, self.head_dim).transpose(1, 2)
# RoPE
cos, sin = self.rotary_emb(hidden_states, position_ids)
cos = cos.unsqueeze(1)
sin = sin.unsqueeze(1)
q, k = apply_rotary_pos_emb(q, k, cos, sin)
# GQA: expand KV heads
if self.num_kv_groups > 1:
k = k.repeat_interleave(self.num_kv_groups, dim=1)
v = v.repeat_interleave(self.num_kv_groups, dim=1)
# Scaled dot-product attention
attn = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
# Causal mask
causal = torch.triu(torch.full((L, L), float('-inf'), device=attn.device), diagonal=1)
attn = attn + causal.unsqueeze(0).unsqueeze(0)
if attention_mask is not None:
attn = attn + attention_mask
attn = F.softmax(attn, dim=-1, dtype=torch.float32).to(q.dtype)
out = torch.matmul(attn, v)
out = out.transpose(1, 2).contiguous().view(B, L, -1)
# Output gating
gate = torch.sigmoid(self.gate(hidden_states))
out = out * gate
return self.o_proj(out)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 2. GATED DELTANET (GDN) β O(n) linear time
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class GatedDeltaNet(nn.Module):
"""Gated DeltaNet: Mamba-style gating + DeltaNet fast-weight update.
Core linear attention mechanism β 10 layers (40% of model).
Implements: M_t = Ξ±_t * M_{t-1} * (I - k_t * q_t^T) + k_t * v_t^T
with SiLU output gating for gradient flow stability.
Weight transplant: Q,K,V projections map directly from Qwen3.5 GDN layers.
"""
def __init__(self, config):
super().__init__()
self.hidden_size = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = config.head_dim
self.state_size = config.gdn_state_size
# Input projections (transplantable from Qwen3.5 GDN)
self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
self.k_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
self.v_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
self.o_proj = nn.Linear(self.num_heads * self.head_dim, config.hidden_size, bias=False)
# Decay gate (Ξ±): controls memory decay speed
self.decay_proj = nn.Linear(config.hidden_size, self.num_heads, bias=True)
# Update gate (Ξ²): controls state update strength
self.beta_proj = nn.Linear(config.hidden_size, self.num_heads, bias=True)
# Output gate (SiLU activation for gradient stability)
self.gate = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
# Short convolution for local context (replaces positional encoding)
self.conv1d = nn.Conv1d(
in_channels=config.hidden_size,
out_channels=config.hidden_size,
kernel_size=4, padding=3, groups=config.hidden_size, bias=True
)
def forward(self, hidden_states, attention_mask=None, position_ids=None, **kwargs):
B, L, D = hidden_states.shape
# Local context mixing via causal conv1d
conv_out = self.conv1d(hidden_states.transpose(1, 2))[..., :L].transpose(1, 2)
q = self.q_proj(conv_out).view(B, L, self.num_heads, self.head_dim)
k = self.k_proj(conv_out).view(B, L, self.num_heads, self.head_dim)
v = self.v_proj(hidden_states).view(B, L, self.num_heads, self.head_dim)
# L2 normalize Q, K (replaces softmax normalization)
q = F.normalize(q, p=2, dim=-1)
k = F.normalize(k, p=2, dim=-1)
# Decay and update gates
alpha = torch.sigmoid(self.decay_proj(hidden_states)).unsqueeze(-1) # [B, L, H, 1]
beta = torch.sigmoid(self.beta_proj(hidden_states)).unsqueeze(-1)
# Recurrent scan with delta rule
# M_t = Ξ± * M_{t-1} * (I - Ξ² * k * q^T) + Ξ² * k * v^T
# For efficiency, compute as: o_t = q^T @ M_t
outputs = []
state = torch.zeros(B, self.num_heads, self.head_dim, self.head_dim,
device=hidden_states.device, dtype=hidden_states.dtype)
for t in range(L):
q_t = q[:, t] # [B, H, D]
k_t = k[:, t]
v_t = v[:, t]
a_t = alpha[:, t] # [B, H, 1]
b_t = beta[:, t]
# Delta rule update
# Erase: state = Ξ± * state * (I - Ξ² * k * q^T)
# Write: state += Ξ² * k * v^T
erase = torch.einsum('bhd,bhe->bhde', k_t * b_t, q_t)
write = torch.einsum('bhd,bhe->bhde', k_t * b_t, v_t)
state = a_t.unsqueeze(-1) * (state - state * erase) + write
# Read: o_t = q^T @ state
o_t = torch.einsum('bhd,bhde->bhe', q_t, state)
outputs.append(o_t)
out = torch.stack(outputs, dim=1) # [B, L, H, D]
out = out.reshape(B, L, -1)
# Output gating with SiLU
gate = F.silu(self.gate(hidden_states))
out = out * gate
return self.o_proj(out)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 3. MAMBA2 β O(n) with SSM state-space duality
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class Mamba2Block(nn.Module):
"""Mamba-2 block with Structured State Space Duality.
5 layers β provides state compression for memory efficiency.
Weight transplant: Via MOHAWK SSD duality from Llama-3.1 Q,K,V β C,B,X.
"""
def __init__(self, config):
super().__init__()
self.hidden_size = config.hidden_size
expand = config.mamba2_expand
self.inner_size = config.hidden_size * expand
self.state_size = config.mamba2_state_size
self.conv_size = config.mamba2_conv_size
self.num_heads = config.num_attention_heads
# Input projection: x β (z, x_ssm) split
self.in_proj = nn.Linear(config.hidden_size, self.inner_size * 2, bias=False)
# Causal conv1d
self.conv1d = nn.Conv1d(
self.inner_size, self.inner_size,
kernel_size=self.conv_size, padding=self.conv_size - 1,
groups=self.inner_size, bias=True
)
# SSM parameters
self.dt_proj = nn.Linear(self.inner_size, self.num_heads, bias=True)
self.A_log = nn.Parameter(torch.log(torch.arange(1, self.num_heads + 1, dtype=torch.float32)))
self.D = nn.Parameter(torch.ones(self.num_heads))
# B, C projections (state-space)
head_dim_ssm = self.inner_size // self.num_heads
self.B_proj = nn.Linear(self.inner_size, self.state_size * self.num_heads, bias=False)
self.C_proj = nn.Linear(self.inner_size, self.state_size * self.num_heads, bias=False)
# Output
self.out_proj = nn.Linear(self.inner_size, config.hidden_size, bias=False)
self.norm = RMSNorm(self.inner_size)
def forward(self, hidden_states, attention_mask=None, position_ids=None, **kwargs):
B, L, _ = hidden_states.shape
# Input split
zx = self.in_proj(hidden_states)
z, x = zx.chunk(2, dim=-1)
# Causal conv
x = self.conv1d(x.transpose(1, 2))[..., :L].transpose(1, 2)
x = F.silu(x)
# SSM parameters
A = -torch.exp(self.A_log) # [H]
dt = F.softplus(self.dt_proj(x)) # [B, L, H]
B_state = self.B_proj(x).view(B, L, self.num_heads, self.state_size)
C_state = self.C_proj(x).view(B, L, self.num_heads, self.state_size)
# Discretize: A_bar = exp(dt * A), B_bar = dt * B
dt_A = dt.unsqueeze(-1) * A.view(1, 1, -1, 1) # [B, L, H, 1]
A_bar = torch.exp(dt_A)
B_bar = dt.unsqueeze(-1) * B_state # [B, L, H, N]
# Selective scan (sequential for correctness; replace with FLA parallel kernel)
head_dim = self.inner_size // self.num_heads
x_heads = x.view(B, L, self.num_heads, head_dim)
outputs = []
state = torch.zeros(B, self.num_heads, self.state_size, device=x.device, dtype=x.dtype)
for t in range(L):
state = A_bar[:, t] * state + B_bar[:, t] * x_heads[:, t, :, :1].expand_as(B_bar[:, t])
y_t = torch.sum(state * C_state[:, t], dim=-1) # [B, H]
outputs.append(y_t)
y = torch.stack(outputs, dim=1) # [B, L, H]
# Skip connection with D
y = y + self.D.view(1, 1, -1) * x.view(B, L, self.num_heads, head_dim).mean(-1)
# Expand back and gate with z
y = y.unsqueeze(-1).expand(-1, -1, -1, head_dim).reshape(B, L, self.inner_size)
y = self.norm(y)
y = y * F.silu(z)
return self.out_proj(y)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 4. SLIDING WINDOW ATTENTION β O(n * w)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class SlidingWindowAttention(nn.Module):
"""Sliding window attention for local pattern capture.
5 layers β complements GDN's global view with fine-grained local context."""
def __init__(self, config):
super().__init__()
self.num_heads = config.num_attention_heads
self.num_kv_heads = config.num_kv_heads
self.head_dim = config.head_dim
self.window_size = config.sliding_window_size
self.num_kv_groups = self.num_heads // self.num_kv_heads
self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
self.k_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
self.v_proj = nn.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
self.o_proj = nn.Linear(self.num_heads * self.head_dim, config.hidden_size, bias=False)
self.gate = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
self.rotary_emb = RotaryEmbedding(self.head_dim, config.max_position_embeddings, config.rope_theta)
def forward(self, hidden_states, attention_mask=None, position_ids=None, **kwargs):
B, L, _ = hidden_states.shape
q = self.q_proj(hidden_states).view(B, L, self.num_heads, self.head_dim).transpose(1, 2)
k = self.k_proj(hidden_states).view(B, L, self.num_kv_heads, self.head_dim).transpose(1, 2)
v = self.v_proj(hidden_states).view(B, L, self.num_kv_heads, self.head_dim).transpose(1, 2)
cos, sin = self.rotary_emb(hidden_states, position_ids)
cos = cos.unsqueeze(1)
sin = sin.unsqueeze(1)
q, k = apply_rotary_pos_emb(q, k, cos, sin)
if self.num_kv_groups > 1:
k = k.repeat_interleave(self.num_kv_groups, dim=1)
v = v.repeat_interleave(self.num_kv_groups, dim=1)
attn = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
# Sliding window + causal mask
mask = torch.ones(L, L, device=attn.device, dtype=torch.bool)
mask = torch.triu(mask, diagonal=1) # causal
mask = mask | torch.tril(torch.ones_like(mask), diagonal=-self.window_size) # window
attn = attn.masked_fill(mask.unsqueeze(0).unsqueeze(0), float('-inf'))
attn = F.softmax(attn, dim=-1, dtype=torch.float32).to(q.dtype)
out = torch.matmul(attn, v)
out = out.transpose(1, 2).contiguous().view(B, L, -1)
gate = torch.sigmoid(self.gate(hidden_states))
out = out * gate
return self.o_proj(out)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# 5. CROSS ATTENTION β for multimodal / tool bridging
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class CrossAttention(nn.Module):
"""Cross attention for PROMETHEUS (world model) and HEPHAESTUS (embodiment) connection.
5 layers β bridges AETHER-Net to external modalities.
When no external context: falls back to self-attention with gating."""
def __init__(self, config):
super().__init__()
self.num_heads = config.num_attention_heads
self.head_dim = config.head_dim
# Self-attention path (default when no external context)
self.q_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
self.k_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
self.v_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
self.o_proj = nn.Linear(self.num_heads * self.head_dim, config.hidden_size, bias=False)
# Cross-attention path (when external context available)
self.cross_k_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
self.cross_v_proj = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
# Modality gate: lerp between self and cross
self.modality_gate = nn.Linear(config.hidden_size, 1, bias=True)
nn.init.constant_(self.modality_gate.bias, -2.0) # default: mostly self-attention
self.gate = nn.Linear(config.hidden_size, self.num_heads * self.head_dim, bias=False)
def forward(self, hidden_states, attention_mask=None, position_ids=None,
encoder_hidden_states=None, **kwargs):
B, L, _ = hidden_states.shape
q = self.q_proj(hidden_states).view(B, L, self.num_heads, self.head_dim).transpose(1, 2)
if encoder_hidden_states is not None:
# Cross-attention mode
k_cross = self.cross_k_proj(encoder_hidden_states).view(
B, -1, self.num_heads, self.head_dim).transpose(1, 2)
v_cross = self.cross_v_proj(encoder_hidden_states).view(
B, -1, self.num_heads, self.head_dim).transpose(1, 2)
attn_cross = torch.matmul(q, k_cross.transpose(-2, -1)) / math.sqrt(self.head_dim)
attn_cross = F.softmax(attn_cross, dim=-1, dtype=torch.float32).to(q.dtype)
out_cross = torch.matmul(attn_cross, v_cross)
out_cross = out_cross.transpose(1, 2).contiguous().view(B, L, -1)
# Self-attention path (always runs)
k_self = self.k_proj(hidden_states).view(B, L, self.num_heads, self.head_dim).transpose(1, 2)
v_self = self.v_proj(hidden_states).view(B, L, self.num_heads, self.head_dim).transpose(1, 2)
attn_self = torch.matmul(q, k_self.transpose(-2, -1)) / math.sqrt(self.head_dim)
causal = torch.triu(torch.full((L, L), float('-inf'), device=attn_self.device), diagonal=1)
attn_self = attn_self + causal.unsqueeze(0).unsqueeze(0)
attn_self = F.softmax(attn_self, dim=-1, dtype=torch.float32).to(q.dtype)
out_self = torch.matmul(attn_self, v_self).transpose(1, 2).contiguous().view(B, L, -1)
# Blend via modality gate
mg = torch.sigmoid(self.modality_gate(hidden_states))
out = mg * out_cross + (1 - mg) * out_self
else:
# Pure self-attention fallback
k = self.k_proj(hidden_states).view(B, L, self.num_heads, self.head_dim).transpose(1, 2)
v = self.v_proj(hidden_states).view(B, L, self.num_heads, self.head_dim).transpose(1, 2)
attn = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
causal = torch.triu(torch.full((L, L), float('-inf'), device=attn.device), diagonal=1)
attn = attn + causal.unsqueeze(0).unsqueeze(0)
attn = F.softmax(attn, dim=-1, dtype=torch.float32).to(q.dtype)
out = torch.matmul(attn, v).transpose(1, 2).contiguous().view(B, L, -1)
gate = torch.sigmoid(self.gate(hidden_states))
out = out * gate
return self.o_proj(out)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Factory
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
ATTENTION_CLASSES = {
"gdn": GatedDeltaNet,
"full": FullAttention,
"mamba2": Mamba2Block,
"slide": SlidingWindowAttention,
"cross": CrossAttention,
}
def build_attention(layer_type: str, config):
cls = ATTENTION_CLASSES.get(layer_type)
if cls is None:
raise ValueError(f"Unknown attention type: {layer_type}. Choose from {list(ATTENTION_CLASSES.keys())}")
return cls(config)
|