Spaces:
Sleeping
Sleeping
File size: 10,862 Bytes
78d2329 | 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 | import torch
import torch.nn as nn
from einops import rearrange
import warnings
import torch.nn.functional as F
USE_FLASH_ATTENTION3 = True
try:
from flash_attn_interface import flash_attn_func
FA3_AVAILABLE = True
warnings.warn('flash attention 3 is available (LVSM)')
except ImportError:
FA3_AVAILABLE = False
warnings.warn('flash attention 3 is not available (LVSM)')
try:
import xformers.ops as xops
XFORMERS_AVAILABLE = True
except ImportError:
XFORMERS_AVAILABLE = False
warnings.warn('xformers is not available (LVSM)')
# raise ImportError("Please install xformers to use flashatt v2")
def init_weights(module, std=0.02):
"""Initialize weights for linear and embedding layers.
Args:
module: Module to initialize
std: Standard deviation for normal initialization
"""
if isinstance(module, (nn.Linear, nn.Embedding)):
torch.nn.init.normal_(module.weight, mean=0.0, std=std)
if isinstance(module, nn.Linear) and module.bias is not None:
torch.nn.init.zeros_(module.bias)
# src: https://github.com/pytorch/benchmark/blob/main/torchbenchmark/models/llama/model.py#L28
class RMSNorm(nn.Module):
def __init__(self, dim: int, eps: float = 1e-5):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(dim))
def _norm(self, x):
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
def forward(self, x):
output = self._norm(x.float()).type_as(x)
return output * self.weight.type_as(x)
class MLP(nn.Module):
"""
Multi-Layer Perceptron block.
Reference: https://github.com/facebookresearch/dino/blob/7c446df5b9f45747937fb0d72314eb9f7b66930a/vision_transformer.py#L49-L65
"""
def __init__(
self,
dim,
mlp_ratio=4,
bias=False,
dropout=0.0,
activation=nn.GELU,
mlp_dim=None,
):
"""
Args:
dim: Input dimension
mlp_ratio: Multiplier for hidden dimension
bias: Whether to use bias in linear layers
dropout: Dropout probability
activation: Activation function
mlp_dim: Optional explicit hidden dimension (overrides mlp_ratio)
"""
super().__init__()
hidden_dim = mlp_dim if mlp_dim is not None else int(dim * mlp_ratio)
self.mlp = nn.Sequential(
nn.Linear(dim, hidden_dim, bias=bias),
activation(),
nn.Linear(hidden_dim, dim, bias=bias),
nn.Dropout(dropout),
)
def forward(self, x):
return self.mlp(x)
class QK_Norm_SelfAttention(nn.Module):
"""
Self-attention with optional Q-K normalization.
Reference: https://github.com/facebookresearch/dino/blob/7c446df5b9f45747937fb0d72314eb9f7b66930a/vision_transformer.py#L68-L92
"""
def __init__(
self,
dim,
head_dim,
qkv_bias=False,
fc_bias=True,
attn_dropout=0.0,
fc_dropout=0.0,
use_qk_norm=True,
):
"""
Args:
dim: Input dimension
head_dim: Dimension of each attention head
qkv_bias: Whether to use bias in QKV projection
fc_bias: Whether to use bias in output projection
attn_dropout: Dropout probability for attention weights
fc_dropout: Dropout probability for output projection
use_qk_norm: Whether to use Q-K normalization
We use flash attention V2 for efficiency.
"""
super().__init__()
assert dim % head_dim == 0, f"Token dimension {dim} should be divisible by head dimension {head_dim}"
self.dim = dim
self.head_dim = head_dim
self.num_heads = dim // head_dim
self.attn_dropout = attn_dropout
self.use_qk_norm = use_qk_norm
self.to_qkv = nn.Linear(dim, 3 * dim, bias=qkv_bias)
self.fc = nn.Linear(dim, dim, bias=fc_bias)
self.attn_fc_dropout = nn.Dropout(fc_dropout)
# Optional Q-K normalization
if self.use_qk_norm:
self.q_norm = RMSNorm(head_dim)
self.k_norm = RMSNorm(head_dim)
def forward(self, x, attn_bias=None):
"""
Args:
x: Input tensor of shape (batch, seq_len, dim)
attn_bias: Optional attention bias mask
Returns:
Output tensor of shape (batch, seq_len, dim)
"""
q, k, v = self.to_qkv(x).chunk(3, dim=-1)
q, k, v = (rearrange(t, "b l (nh dh) -> b l nh dh", dh=self.head_dim) for t in (q, k, v))
# Apply qk normalization if enabled
if self.use_qk_norm:
q = self.q_norm(q)
k = self.k_norm(k)
if USE_FLASH_ATTENTION3 and FA3_AVAILABLE:
x = flash_attn_func(q, k, v)[0]
elif XFORMERS_AVAILABLE:
x = xops.memory_efficient_attention(
q, k, v,
attn_bias=attn_bias,
p=self.attn_dropout if self.training else 0.0,
op=(xops.fmha.flash.FwOp, xops.fmha.flash.BwOp),
)
else:
# use pytorch's built-in attention
q = q.permute(0, 2, 1, 3).contiguous() # [B, H, L, C]
k = k.permute(0, 2, 1, 3).contiguous()
v = v.permute(0, 2, 1, 3).contiguous()
x = F.scaled_dot_product_attention(q, k, v)
x = x.permute(0, 2, 1, 3).contiguous() # [B, L, H, C]
x = rearrange(x, "b l nh dh -> b l (nh dh)")
x = self.attn_fc_dropout(self.fc(x))
return x
class SubsetAttention(nn.Module):
"""Attention that can attend to subsets of queries or keys/values."""
def __init__(
self,
dim,
head_dim,
qkv_bias=False,
attn_dropout=0.0,
fc_bias=False,
fc_dropout=0.0,
use_qk_norm=False
):
"""
Args:
dim: Input dimension
head_dim: Dimension of each attention head
qkv_bias: Whether to use bias in QKV projection
attn_dropout: Dropout probability for attention weights
fc_bias: Whether to use bias in output projection
fc_dropout: Dropout probability for output projection
use_qk_norm: Whether to use Q-K normalization
We use flash attention V2 for efficiency.
"""
super().__init__()
assert dim % head_dim == 0, f"Token dimension {dim} should be divisible by head dimension {head_dim}"
self.dim = dim
self.head_dim = head_dim
self.num_heads = dim // head_dim
self.attn_dropout = attn_dropout
self.use_qk_norm = use_qk_norm
# Projections
self.to_qkv = nn.Linear(dim, 3 * dim, bias=qkv_bias)
self.fc = nn.Linear(dim, dim, bias=fc_bias)
self.attn_fc_dropout = nn.Dropout(fc_dropout)
# Optional Q-K normalization
if self.use_qk_norm:
self.q_norm = RMSNorm(head_dim)
self.k_norm = RMSNorm(head_dim)
def forward(self, x, subset_kv_size=None, subset_q_size=None):
"""
Args:
x: Input tensor of shape (batch, seq_len, dim)
subset_kv_size: If provided, only attend to tokens after this index in KV
subset_q_size: If provided, only compute attention for queries up to this index
Returns:
Output tensor of shape (batch, seq_len, dim)
"""
# Only one subset parameter can be provided
assert not (subset_kv_size is not None and subset_q_size is not None), \
"Only one of subset_kv_size or subset_q_size can be provided"
q, k, v = self.to_qkv(x).chunk(3, dim=-1)
q, k, v = (rearrange(t, "b l (nh dh) -> b l nh dh", dh=self.head_dim) for t in (q, k, v))
if self.use_qk_norm:
q = self.q_norm(q)
k = self.k_norm(k)
# Handle subset attention cases
if subset_kv_size is not None and subset_kv_size < k.shape[1]:
# Attend to subset of key/value tokens
k_subset = k[:, subset_kv_size:, :, :].contiguous()
v_subset = v[:, subset_kv_size:, :, :].contiguous()
x = xops.memory_efficient_attention(
q, k_subset, v_subset,
attn_bias=None,
p=self.attn_dropout if self.training else 0.0,
op=(xops.fmha.flash.FwOp, xops.fmha.flash.BwOp),
)
elif subset_q_size is not None and subset_q_size < q.shape[1]:
# Only compute attention for subset of query tokens
q_subset = q[:, :subset_q_size, :, :].contiguous()
x = xops.memory_efficient_attention(
q_subset, k, v,
attn_bias=None,
p=self.attn_dropout if self.training else 0.0,
op=(xops.fmha.flash.FwOp, xops.fmha.flash.BwOp),
)
else:
# Regular attention for all tokens
x = xops.memory_efficient_attention(
q, k, v,
attn_bias=None,
p=self.attn_dropout if self.training else 0.0,
op=(xops.fmha.flash.FwOp, xops.fmha.flash.BwOp),
)
x = rearrange(x, "b l nh dh -> b l (nh dh)")
# Final projection
x = self.attn_fc_dropout(self.fc(x))
return x
class QK_Norm_TransformerBlock(nn.Module):
"""
Standard transformer block with pre-normalization architecture.
Reference: https://github.com/facebookresearch/dino/blob/7c446df5b9f45747937fb0d72314eb9f7b66930a/vision_transformer.py#L95-L113
"""
def __init__(
self,
dim,
head_dim,
ln_bias=False,
attn_qkv_bias=False,
attn_dropout=0.0,
attn_fc_bias=False,
attn_fc_dropout=0.0,
mlp_ratio=4,
mlp_bias=False,
mlp_dropout=0.0,
use_qk_norm=True,
):
super().__init__()
self.norm1 = nn.LayerNorm(dim, bias=ln_bias)
self.attn = QK_Norm_SelfAttention(
dim=dim,
head_dim=head_dim,
qkv_bias=attn_qkv_bias,
fc_bias=attn_fc_bias,
attn_dropout=attn_dropout,
fc_dropout=attn_fc_dropout,
use_qk_norm=use_qk_norm,
)
self.norm2 = nn.LayerNorm(dim, bias=ln_bias)
self.mlp = MLP(
dim=dim,
mlp_ratio=mlp_ratio,
bias=mlp_bias,
dropout=mlp_dropout,
)
def forward(self, x):
x = x + self.attn(self.norm1(x))
x = x + self.mlp(self.norm2(x))
return x
|