Spaces:
Running
on
Zero
Running
on
Zero
File size: 12,541 Bytes
853528a |
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 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the Apache License, Version 2.0
# found in the LICENSE file in the root directory of this source tree.
# References:
# https://github.com/facebookresearch/dino/blob/master/vision_transformer.py
# https://github.com/rwightman/pytorch-image-models/tree/master/timm/models/vision_transformer.py
import logging
import os
import warnings
from torch import Tensor
from torch import nn
import torch
from torch.nn.functional import scaled_dot_product_attention
from torch.nn.attention import SDPBackend
XFORMERS_ENABLED = os.environ.get("XFORMERS_DISABLED") is None
try:
if XFORMERS_ENABLED:
from xformers.ops import memory_efficient_attention, unbind
XFORMERS_AVAILABLE = True
# warnings.warn("xFormers is available (Attention)")
else:
# warnings.warn("xFormers is disabled (Attention)")
raise ImportError
except ImportError:
XFORMERS_AVAILABLE = False
# warnings.warn("xFormers is not available (Attention)")
class Attention(nn.Module):
def __init__(
self,
dim: int,
num_heads: int = 8,
qkv_bias: bool = False,
proj_bias: bool = True,
attn_drop: float = 0.0,
proj_drop: float = 0.0,
) -> None:
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim**-0.5
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim, bias=proj_bias)
self.proj_drop = nn.Dropout(proj_drop)
def forward(self, x: Tensor, attn_bias=None) -> Tensor:
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0] * self.scale, qkv[1], qkv[2]
attn = q @ k.transpose(-2, -1)
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class MemEffAttention(Attention):
def forward(self, x: Tensor, attn_bias=None) -> Tensor:
if not XFORMERS_AVAILABLE:
if attn_bias is not None:
raise AssertionError("xFormers is required for using nested tensors")
return super().forward(x)
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
# q, k, v = unbind(qkv, 2)
q, k, v = [qkv[:,:,i] for i in range(3)]
x = memory_efficient_attention(q, k, v, attn_bias=attn_bias)
x = x.reshape([B, N, C])
x = self.proj(x)
x = self.proj_drop(x)
return x
class FlashAttention(Attention):
def forward(self, x: Tensor, attn_bias=None) -> Tensor:
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).transpose(1, 3)
# q, k, v = unbind(qkv, 2)
q, k, v = [qkv[:,:,i] for i in range(3)]
if q.dtype == torch.bfloat16:
with nn.attention.sdpa_kernel(SDPBackend.FLASH_ATTENTION):
x = scaled_dot_product_attention(q, k, v)
else:
with nn.attention.sdpa_kernel([SDPBackend.MATH, SDPBackend.EFFICIENT_ATTENTION]):
x = scaled_dot_product_attention(q, k, v)
x = x.transpose(1, 2).reshape([B, N, C])
x = self.proj(x)
x = self.proj_drop(x)
return x
"""
Following is written by GPT-4o
"""
class CrossAttentionRope(nn.Module):
def __init__(
self,
dim: int,
num_heads: int = 8,
qkv_bias: bool = False,
proj_bias: bool = True,
attn_drop: float = 0.0,
proj_drop: float = 0.0,
qk_norm: bool = False,
norm_layer: nn.Module = nn.LayerNorm,
rope=None,
) -> None:
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim**-0.5
# Separate projection layers for query, key, and value
self.q_proj = nn.Linear(dim, dim, bias=qkv_bias)
self.k_proj = nn.Linear(dim, dim, bias=qkv_bias)
self.v_proj = nn.Linear(dim, dim, bias=qkv_bias)
self.q_norm = norm_layer(head_dim) if qk_norm else nn.Identity()
self.k_norm = norm_layer(head_dim) if qk_norm else nn.Identity()
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim, bias=proj_bias)
self.proj_drop = nn.Dropout(proj_drop)
self.rope = rope
def forward(self, query: Tensor, key: Tensor, value: Tensor, attn_bias=None, qpos=None, kpos=None) -> Tensor:
"""
Args:
query: Tensor of shape (B, N, C), input query
key: Tensor of shape (B, M, C), input key
value: Tensor of shape (B, M, C), input value
attn_bias: Optional tensor for attention bias
Returns:
Tensor of shape (B, N, C), output of cross-attention
"""
B, N, C = query.shape
_, M, _ = key.shape
# Project query, key, and value
q = self.q_proj(query).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
k = self.k_proj(key).reshape(B, M, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
v = self.v_proj(value).reshape(B, M, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
q, k = self.q_norm(q).to(v.dtype), self.k_norm(k).to(v.dtype)
if self.rope is not None:
q = self.rope(q, qpos)
k = self.rope(k, kpos)
# Scale query
q = q * self.scale
# Compute attention scores
attn = q @ k.transpose(-2, -1) # (B, num_heads, N, M)
if attn_bias is not None:
attn = attn + attn_bias
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
# Compute attention output
x = (attn @ v).transpose(1, 2).reshape(B, N, C) # (B, N, C)
# Final projection
x = self.proj(x)
x = self.proj_drop(x)
return x
class MemEffCrossAttentionRope(CrossAttentionRope):
def forward(self, query: Tensor, key: Tensor, value: Tensor, attn_bias=None, qpos=None, kpos=None) -> Tensor:
"""
Args:
query: Tensor of shape (B, N, C), input query
key: Tensor of shape (B, M, C), input key
value: Tensor of shape (B, M, C), input value
attn_bias: Optional tensor for attention bias
Returns:
Tensor of shape (B, N, C), output of cross-attention
"""
if not XFORMERS_AVAILABLE:
if attn_bias is not None:
raise AssertionError("xFormers is required for using nested tensors")
return super().forward(query, key, value, attn_bias)
B, N, C = query.shape
_, M, _ = key.shape
# Project query, key, and value
q = self.q_proj(query).reshape(B, N, self.num_heads, C // self.num_heads)
k = self.k_proj(key).reshape(B, M, self.num_heads, C // self.num_heads)
v = self.v_proj(value).reshape(B, M, self.num_heads, C // self.num_heads)
q = q.transpose(1, 2)
k = k.transpose(1, 2)
q, k = self.q_norm(q).to(v.dtype), self.k_norm(k).to(v.dtype)
if self.rope is not None:
q = self.rope(q, qpos)
k = self.rope(k, kpos)
q = q.transpose(1, 2)
k = k.transpose(1, 2)
# Compute memory-efficient attention
x = memory_efficient_attention(q, k, v, attn_bias=attn_bias)
x = x.reshape(B, N, C)
# Final projection
x = self.proj(x)
x = self.proj_drop(x)
return x
class AttentionRope(nn.Module):
def __init__(
self,
dim: int,
num_heads: int = 8,
qkv_bias: bool = False,
proj_bias: bool = True,
attn_drop: float = 0.0,
proj_drop: float = 0.0,
qk_norm: bool = False,
norm_layer: nn.Module = nn.LayerNorm,
rope=None
) -> None:
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = head_dim**-0.5
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim, bias=proj_bias)
self.proj_drop = nn.Dropout(proj_drop)
self.q_norm = norm_layer(head_dim) if qk_norm else nn.Identity()
self.k_norm = norm_layer(head_dim) if qk_norm else nn.Identity()
self.rope = rope
def forward(self, x: Tensor, attn_bias=None, xpos=None) -> Tensor:
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
q, k = self.q_norm(q).to(v.dtype), self.k_norm(k).to(v.dtype)
if self.rope is not None:
q = self.rope(q, xpos)
k = self.rope(k, xpos)
q = q * self.scale
attn = q @ k.transpose(-2, -1)
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class MemEffAttentionRope(AttentionRope):
def forward(self, x: Tensor, attn_bias=None, xpos=None) -> Tensor:
if not XFORMERS_AVAILABLE:
if attn_bias is not None:
raise AssertionError("xFormers is required for using nested tensors")
return super().forward(x)
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
qkv = qkv.transpose(1, 3)
# q, k, v = unbind(qkv, 2)
q, k, v = [qkv[:,:,i] for i in range(3)]
q, k = self.q_norm(q).to(v.dtype), self.k_norm(k).to(v.dtype)
if self.rope is not None:
q = self.rope(q, xpos)
k = self.rope(k, xpos)
q = q.transpose(1, 2)
k = k.transpose(1, 2)
v = v.transpose(1, 2)
x = memory_efficient_attention(q, k, v, attn_bias=attn_bias)
x = x.reshape([B, N, C])
# score_matrix = (q.permute(0, 2, 1, 3) * self.scale @ k.permute(0, 2, 1, 3).transpose(-2, -1)).sum(dim=1).reshape(frame_num, 261, frame_num, 261).mean(dim=[1, 3]).sum(1) # for frame attention matrix
# global_valid_id = torch.where(score_matrix > 0)
# score_matrix = (q.permute(0, 2, 1, 3) * self.scale @ k.permute(0, 2, 1, 3).transpose(-2, -1)).sum(dim=1)
x = self.proj(x)
x = self.proj_drop(x)
return x
class FlashAttentionRope(AttentionRope):
def forward(self, x: Tensor, attn_bias=None, xpos=None) -> Tensor:
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).transpose(1, 3)
# q, k, v = unbind(qkv, 2)
q, k, v = [qkv[:,:,i] for i in range(3)]
q, k = self.q_norm(q).to(v.dtype), self.k_norm(k).to(v.dtype)
if self.rope is not None:
q = self.rope(q, xpos)
k = self.rope(k, xpos)
if q.dtype == torch.bfloat16:
with nn.attention.sdpa_kernel(SDPBackend.FLASH_ATTENTION):
x = scaled_dot_product_attention(q, k, v)
else:
with nn.attention.sdpa_kernel([SDPBackend.MATH, SDPBackend.EFFICIENT_ATTENTION]):
x = scaled_dot_product_attention(q, k, v)
x = x.transpose(1, 2).reshape([B, N, C])
x = self.proj(x)
x = self.proj_drop(x)
return x
def get_attn_score(blk_class, x, frame_num, token_length, xpos=None):
x = blk_class.norm1(x)
B, N, C = x.shape
qkv = blk_class.attn.qkv(x).reshape(B, N, 3, blk_class.attn.num_heads, C // blk_class.attn.num_heads)
qkv = qkv.transpose(1, 3)
# q, k, v = unbind(qkv, 2)
q, k, v = [qkv[:,:,i] for i in range(3)]
q, k = blk_class.attn.q_norm(q).to(v.dtype), blk_class.attn.k_norm(k).to(v.dtype)
if blk_class.attn.rope is not None:
q = blk_class.attn.rope(q, xpos)
k = blk_class.attn.rope(k, xpos)
q = q.transpose(1, 2)
k = k.transpose(1, 2)
score = (q.permute(0, 2, 1, 3) * blk_class.attn.scale @ k.permute(0, 2, 1, 3).transpose(-2, -1)).sum(dim=1).reshape(B, frame_num, token_length, frame_num, token_length).mean(dim=[2, 4]).sum(-1)
return score |