File size: 14,139 Bytes
5c43f61 | 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 | """
VortexLocalAttention: Local windowed attention with global token support.
Uses a sliding window of 512 tokens for efficiency, with special handling
for global tokens that can attend across the entire sequence.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Tuple
class VortexLocalAttention(nn.Module):
"""
Local windowed attention with window_size=512.
Science documents have strong local coherence — equations reference
nearby text, not distant paragraphs.
Global tokens (special [SCIENCE] tokens) attend to everything.
"""
def __init__(
self,
d_model: int,
num_heads: int,
window_size: int = 512,
use_flash_attention: bool = True,
):
"""
Initialize local windowed attention.
Args:
d_model: Model dimension
num_heads: Number of attention heads
window_size: Size of local attention window
use_flash_attention: Use Flash Attention 2 if available (CUDA only)
"""
super().__init__()
self.d_model = d_model
self.num_heads = num_heads
self.head_dim = d_model // num_heads
self.window_size = window_size
self.use_flash_attention = use_flash_attention
assert d_model % num_heads == 0, "d_model must be divisible by num_heads"
# QKV projection
self.qkv = nn.Linear(d_model, d_model * 3, bias=False)
self.out_proj = nn.Linear(d_model, d_model, bias=False)
# Global token projection (for tokens that attend globally)
self.global_qkv = nn.Linear(d_model, d_model * 3, bias=False)
# Initialize weights
self._initialize_weights()
def _initialize_weights(self):
"""Initialize weights."""
for module in [self.qkv, self.global_qkv, self.out_proj]:
if hasattr(module, 'weight'):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
def forward(
self,
x: torch.Tensor,
global_mask: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
Forward pass with local windowed attention.
Args:
x: Input tensor (batch, seq_len, d_model)
global_mask: Boolean mask indicating which tokens are global (attend everywhere)
Shape: (batch, seq_len) or None
attention_mask: Optional padding mask (batch, seq_len)
Returns:
Output tensor (batch, seq_len, d_model)
"""
batch, seq_len, _ = x.shape
device = x.device
dtype = x.dtype
if global_mask is None:
global_mask = torch.zeros(batch, seq_len, dtype=torch.bool, device=device)
# Compute QKV for all tokens
qkv = self.qkv(x)
q, k, v = qkv.chunk(3, dim=-1)
# Reshape for multi-head attention
q = q.view(batch, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
k = k.view(batch, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
v = v.view(batch, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
# Compute global token QKV separately
if global_mask.any():
global_qkv = self.global_qkv(x)
gq, gk, gv = global_qkv.chunk(3, dim=-1)
gq = gq.view(batch, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
gk = gk.view(batch, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
gv = gv.view(batch, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
# Build output tensor
output = torch.zeros_like(x)
# Process each position
for t in range(seq_len):
# Determine window
window_start = max(0, t - self.window_size // 2)
window_end = min(seq_len, t + self.window_size // 2 + 1)
window_len = window_end - window_start
# Get window indices
window_indices = slice(window_start, window_end)
# Extract window queries (for position t)
q_t = q[:, :, t:t+1, :] # (batch, heads, 1, head_dim)
# Determine which keys/values to use
# Local tokens: only those in window
# Global tokens: all positions (if they are global)
k_window = k[:, :, window_indices, :]
v_window = v[:, :, window_indices, :]
# Build full key/value set including global tokens
# Global tokens attend to all positions
if global_mask.any():
# Find global positions
global_positions = global_mask[0] # (seq_len) - assume same across batch
if global_positions.any():
gk_all = gk[:, :, :, :] # All global keys
gv_all = gv[:, :, :, :]
# Concatenate window keys with global keys
k_full = torch.cat([k_window, gk_all], dim=2)
v_full = torch.cat([v_window, gv_all], dim=2)
else:
k_full = k_window
v_full = v_window
else:
k_full = k_window
v_full = v_window
# Compute attention scores
# q_t: (batch, heads, 1, head_dim)
# k_full: (batch, heads, window_len + num_global, head_dim)
attn_scores = torch.matmul(q_t, k_full.transpose(-2, -1)) / (self.head_dim ** 0.5)
# (batch, heads, 1, k_len)
# Apply attention mask if provided
if attention_mask is not None:
mask_t = attention_mask[:, window_indices].unsqueeze(1).unsqueeze(2)
attn_scores = attn_scores.masked_fill(mask_t == 0, -1e9)
# Softmax
attn_weights = F.softmax(attn_scores, dim=-1)
# Weighted sum
attn_output = torch.matmul(attn_weights, v_full)
# (batch, heads, 1, head_dim)
# Reshape and project
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.view(batch, 1, self.d_model)
attn_output = self.out_proj(attn_output)
# Place in output
output[:, t:t+1, :] = attn_output
return output
def forward_optimized(
self,
x: torch.Tensor,
global_mask: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
Optimized forward pass using Flash Attention or efficient windowed attention.
This is a placeholder for actual Flash Attention integration.
"""
batch, seq_len, _ = x.shape
if self.use_flash_attention and self.window_size >= seq_len:
# For short sequences, can use full attention
return self._flash_attention_forward(x, attention_mask)
else:
# Use windowed attention
return self._windowed_attention_forward(x, global_mask, attention_mask)
def _flash_attention_forward(
self,
x: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
Use Flash Attention 2 if available.
Requires: pip install flash-attn
"""
try:
from flash_attn import flash_attn_func
batch, seq_len, _ = x.shape
qkv = self.qkv(x)
q, k, v = qkv.chunk(3, dim=-1)
# Reshape for flash attention
q = q.view(batch, seq_len, self.num_heads, self.head_dim)
k = k.view(batch, seq_len, self.num_heads, self.head_dim)
v = v.view(batch, seq_len, self.num_heads, self.head_dim)
# Flash attention expects (batch, seq_len, num_heads, head_dim)
# and returns same shape
if attention_mask is not None:
# Flash attention uses causal mask or padding mask
output = flash_attn_func(
q, k, v,
causal=False,
softmax_scale=1.0 / (self.head_dim ** 0.5),
)
else:
output = flash_attn_func(
q, k, v,
causal=False,
)
output = output.view(batch, seq_len, self.d_model)
return self.out_proj(output)
except ImportError:
print("Flash Attention not available, falling back to standard attention")
return self._standard_attention(x, attention_mask)
def _standard_attention(
self,
x: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""Standard full attention (quadratic)."""
batch, seq_len, _ = x.shape
qkv = self.qkv(x)
q, k, v = qkv.chunk(3, dim=-1)
q = q.view(batch, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
k = k.view(batch, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
v = v.view(batch, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
# Compute attention scores
attn_scores = torch.matmul(q, k.transpose(-2, -1)) / (self.head_dim ** 0.5)
if attention_mask is not None:
attn_scores = attn_scores.masked_fill(
attention_mask.unsqueeze(1).unsqueeze(2) == 0,
-1e9
)
attn_weights = F.softmax(attn_scores, dim=-1)
attn_output = torch.matmul(attn_weights, v)
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.view(batch, seq_len, self.d_model)
return self.out_proj(attn_output)
def _windowed_attention_forward(
self,
x: torch.Tensor,
global_mask: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
Efficient windowed attention implementation.
Uses unfold to extract windows and batched matrix multiply.
"""
batch, seq_len, _ = x.shape
device = x.device
if global_mask is None:
global_mask = torch.zeros(batch, seq_len, dtype=torch.bool, device=device)
# Compute QKV
qkv = self.qkv(x)
q, k, v = qkv.chunk(3, dim=-1)
# Reshape: (batch, seq_len, num_heads, head_dim)
q = q.view(batch, seq_len, self.num_heads, self.head_dim)
k = k.view(batch, seq_len, self.num_heads, self.head_dim)
v = v.view(batch, seq_len, self.num_heads, self.head_dim)
# Pad sequence for windowing
pad_len = self.window_size // 2
k_padded = F.pad(k, (0, 0, 0, 0, pad_len, pad_len))
v_padded = F.pad(v, (0, 0, 0, 0, pad_len, pad_len))
# Extract windows using unfold
# (batch, seq_len, num_heads, head_dim) -> (batch, seq_len, window_size, num_heads, head_dim)
k_windows = k_padded.unfold(1, self.window_size, 1)
v_windows = v_padded.unfold(1, self.window_size, 1)
# Permute to (batch, seq_len, num_heads, window_size, head_dim)
k_windows = k_windows.permute(0, 1, 3, 2, 4)
v_windows = v_windows.permute(0, 1, 3, 2, 4)
# Compute attention for each position
# q: (batch, seq_len, num_heads, 1, head_dim)
q_expanded = q.unsqueeze(3)
k_windows = k_windows
# Scores: (batch, seq_len, num_heads, 1, window_size)
attn_scores = torch.matmul(q_expanded, k_windows.transpose(-2, -1)) / (self.head_dim ** 0.5)
attn_scores = attn_scores.squeeze(3) # (batch, seq_len, num_heads, window_size)
# Apply softmax
attn_weights = F.softmax(attn_scores, dim=-1)
# Weighted sum
attn_output = torch.matmul(attn_weights.unsqueeze(3), v_windows).squeeze(3)
# (batch, seq_len, num_heads, head_dim)
# Concatenate heads
attn_output = attn_output.view(batch, seq_len, self.d_model)
# Add global token contribution if any
if global_mask.any():
# Compute full attention for global tokens only
# This is a simplified version - in practice would be optimized
global_indices = global_mask[0].nonzero(as_tuple=True)[0]
if len(global_indices) > 0:
# For positions with global tokens, add full attention
# (simplified: compute full attention for all)
full_attn = self._standard_attention(x, attention_mask)
# Blend: local for most, full for global positions
attn_output = torch.where(
global_mask.unsqueeze(-1),
full_attn,
attn_output
)
return self.out_proj(attn_output)
def test_vortex_local_attention():
"""Test the VortexLocalAttention layer."""
batch_size = 2
seq_len = 256
d_model = 4096
num_heads = 32
window_size = 512
attn = VortexLocalAttention(d_model, num_heads, window_size, use_flash_attention=False)
x = torch.randn(batch_size, seq_len, d_model)
# Forward pass
output = attn(x)
print(f"Input shape: {x.shape}")
print(f"Output shape: {output.shape}")
assert output.shape == x.shape, f"Expected {x.shape}, got {output.shape}"
# With global mask
global_mask = torch.zeros(batch_size, seq_len, dtype=torch.bool)
global_mask[0, 0] = True # First token is global
global_mask[1, -1] = True # Last token is global
output2 = attn(x, global_mask=global_mask)
assert output2.shape == x.shape
print("VortexLocalAttention test passed!")
if __name__ == "__main__":
test_vortex_local_attention()
|