Spaces:
Paused
Paused
File size: 31,938 Bytes
6497501 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 |
import math
from inspect import isfunction
from typing import Any, Optional
import torch
import torch.nn.functional as F
from einops import rearrange, repeat
from packaging import version
from torch import nn, einsum
if version.parse(torch.__version__) >= version.parse("2.0.0"):
SDP_IS_AVAILABLE = True
from torch.backends.cuda import SDPBackend, sdp_kernel
BACKEND_MAP = {
SDPBackend.MATH: {
"enable_math": True,
"enable_flash": False,
"enable_mem_efficient": False,
},
SDPBackend.FLASH_ATTENTION: {
"enable_math": False,
"enable_flash": True,
"enable_mem_efficient": False,
},
SDPBackend.EFFICIENT_ATTENTION: {
"enable_math": False,
"enable_flash": False,
"enable_mem_efficient": True,
},
None: {"enable_math": True, "enable_flash": True, "enable_mem_efficient": True},
}
else:
from contextlib import nullcontext
SDP_IS_AVAILABLE = False
sdp_kernel = nullcontext
BACKEND_MAP = {}
print(
f"No SDP backend available, likely because you are running in pytorch versions < 2.0. In fact, "
f"you are using PyTorch {torch.__version__}. You might want to consider upgrading."
)
try:
import xformers
import xformers.ops
XFORMERS_IS_AVAILABLE = True
except:
XFORMERS_IS_AVAILABLE = False
print("no module 'xformers'. Processing without...")
from .diffusionmodules.util import checkpoint
def exists(val):
return val is not None
def uniq(arr):
return {el: True for el in arr}.keys()
def default(val, d):
if exists(val):
return val
return d() if isfunction(d) else d
def max_neg_value(t):
return -torch.finfo(t.dtype).max
def init_(tensor):
dim = tensor.shape[-1]
std = 1 / math.sqrt(dim)
tensor.uniform_(-std, std)
return tensor
# feedforward
class GEGLU(nn.Module):
def __init__(self, dim_in, dim_out):
super().__init__()
self.proj = nn.Linear(dim_in, dim_out * 2)
def forward(self, x):
x, gate = self.proj(x).chunk(2, dim=-1)
return x * F.gelu(gate)
class FeedForward(nn.Module):
def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.0):
super().__init__()
inner_dim = int(dim * mult)
dim_out = default(dim_out, dim)
project_in = (
nn.Sequential(nn.Linear(dim, inner_dim), nn.GELU())
if not glu
else GEGLU(dim, inner_dim)
)
self.net = nn.Sequential(
project_in, nn.Dropout(dropout), nn.Linear(inner_dim, dim_out)
)
def forward(self, x):
return self.net(x)
def zero_module(module):
"""
Zero out the parameters of a module and return it.
"""
for p in module.parameters():
p.detach().zero_()
return module
def Normalize(in_channels):
return torch.nn.GroupNorm(
num_groups=32, num_channels=in_channels, eps=1e-6, affine=True
)
class LinearAttention(nn.Module):
def __init__(self, dim, heads=4, dim_head=32):
super().__init__()
self.heads = heads
hidden_dim = dim_head * heads
self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias=False)
self.to_out = nn.Conv2d(hidden_dim, dim, 1)
def forward(self, x):
b, c, h, w = x.shape
qkv = self.to_qkv(x)
q, k, v = rearrange(
qkv, "b (qkv heads c) h w -> qkv b heads c (h w)", heads=self.heads, qkv=3
)
k = k.softmax(dim=-1)
context = torch.einsum("bhdn,bhen->bhde", k, v)
out = torch.einsum("bhde,bhdn->bhen", context, q)
out = rearrange(
out, "b heads c (h w) -> b (heads c) h w", heads=self.heads, h=h, w=w
)
return self.to_out(out)
class SpatialSelfAttention(nn.Module):
def __init__(self, in_channels):
super().__init__()
self.in_channels = in_channels
self.norm = Normalize(in_channels)
self.q = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.k = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.v = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
self.proj_out = torch.nn.Conv2d(
in_channels, in_channels, kernel_size=1, stride=1, padding=0
)
def forward(self, x):
h_ = x
h_ = self.norm(h_)
q = self.q(h_)
k = self.k(h_)
v = self.v(h_)
# compute attention
b, c, h, w = q.shape
q = rearrange(q, "b c h w -> b (h w) c")
k = rearrange(k, "b c h w -> b c (h w)")
w_ = torch.einsum("bij,bjk->bik", q, k)
w_ = w_ * (int(c) ** (-0.5))
w_ = torch.nn.functional.softmax(w_, dim=2)
# attend to values
v = rearrange(v, "b c h w -> b c (h w)")
w_ = rearrange(w_, "b i j -> b j i")
h_ = torch.einsum("bij,bjk->bik", v, w_)
h_ = rearrange(h_, "b c (h w) -> b c h w", h=h)
h_ = self.proj_out(h_)
return x + h_
class CrossAttention(nn.Module):
def __init__(
self,
query_dim,
context_dim=None,
heads=8,
dim_head=64,
dropout=0.0,
backend=None,
):
super().__init__()
inner_dim = dim_head * heads
context_dim = default(context_dim, query_dim)
self.scale = dim_head**-0.5
self.heads = heads
self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
self.to_out = zero_module(nn.Sequential(
nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)
))
self.backend = backend
self.attn_map_cache = None
def forward(
self,
x,
context=None,
mask=None,
additional_tokens=None,
n_times_crossframe_attn_in_self=0,
):
h = self.heads
if additional_tokens is not None:
# get the number of masked tokens at the beginning of the output sequence
n_tokens_to_mask = additional_tokens.shape[1]
# add additional token
x = torch.cat([additional_tokens, x], dim=1)
q = self.to_q(x)
context = default(context, x)
k = self.to_k(context)
v = self.to_v(context)
if n_times_crossframe_attn_in_self:
# reprogramming cross-frame attention as in https://arxiv.org/abs/2303.13439
assert x.shape[0] % n_times_crossframe_attn_in_self == 0
n_cp = x.shape[0] // n_times_crossframe_attn_in_self
k = repeat(
k[::n_times_crossframe_attn_in_self], "b ... -> (b n) ...", n=n_cp
)
v = repeat(
v[::n_times_crossframe_attn_in_self], "b ... -> (b n) ...", n=n_cp
)
q, k, v = map(lambda t: rearrange(t, "b n (h d) -> (b h) n d", h=h), (q, k, v))
## old
sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
del q, k
if exists(mask):
mask = rearrange(mask, 'b ... -> b (...)')
max_neg_value = -torch.finfo(sim.dtype).max
mask = repeat(mask, 'b j -> (b h) () j', h=h)
sim.masked_fill_(~mask, max_neg_value)
# attention, what we cannot get enough of
sim = sim.softmax(dim=-1)
# save attn_map
if self.attn_map_cache is not None:
bh, n, l = sim.shape
size = int(n**0.5)
self.attn_map_cache["size"] = size
self.attn_map_cache["attn_map"] = sim
out = einsum('b i j, b j d -> b i d', sim, v)
out = rearrange(out, "(b h) n d -> b n (h d)", h=h)
## new
# with sdp_kernel(**BACKEND_MAP[self.backend]):
# # print("dispatching into backend", self.backend, "q/k/v shape: ", q.shape, k.shape, v.shape)
# out = F.scaled_dot_product_attention(
# q, k, v, attn_mask=mask
# ) # scale is dim_head ** -0.5 per default
# del q, k, v
# out = rearrange(out, "b h n d -> b n (h d)", h=h)
if additional_tokens is not None:
# remove additional token
out = out[:, n_tokens_to_mask:]
return self.to_out(out)
class MemoryEfficientCrossAttention(nn.Module):
# https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
def __init__(
self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0, **kwargs
):
super().__init__()
# print(
# f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, context_dim is {context_dim} and using "
# f"{heads} heads with a dimension of {dim_head}."
# )
inner_dim = dim_head * heads
context_dim = default(context_dim, query_dim)
self.heads = heads
self.dim_head = dim_head
self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
self.to_out = nn.Sequential(
nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)
)
self.attention_op: Optional[Any] = None
def forward(
self,
x,
context=None,
mask=None,
additional_tokens=None,
n_times_crossframe_attn_in_self=0,
):
if additional_tokens is not None:
# get the number of masked tokens at the beginning of the output sequence
n_tokens_to_mask = additional_tokens.shape[1]
# add additional token
x = torch.cat([additional_tokens, x], dim=1)
q = self.to_q(x)
context = default(context, x)
k = self.to_k(context)
v = self.to_v(context)
if n_times_crossframe_attn_in_self:
# reprogramming cross-frame attention as in https://arxiv.org/abs/2303.13439
assert x.shape[0] % n_times_crossframe_attn_in_self == 0
# n_cp = x.shape[0]//n_times_crossframe_attn_in_self
k = repeat(
k[::n_times_crossframe_attn_in_self],
"b ... -> (b n) ...",
n=n_times_crossframe_attn_in_self,
)
v = repeat(
v[::n_times_crossframe_attn_in_self],
"b ... -> (b n) ...",
n=n_times_crossframe_attn_in_self,
)
b, _, _ = q.shape
q, k, v = map(
lambda t: t.unsqueeze(3)
.reshape(b, t.shape[1], self.heads, self.dim_head)
.permute(0, 2, 1, 3)
.reshape(b * self.heads, t.shape[1], self.dim_head)
.contiguous(),
(q, k, v),
)
# actually compute the attention, what we cannot get enough of
out = xformers.ops.memory_efficient_attention(
q, k, v, attn_bias=None, op=self.attention_op
)
# TODO: Use this directly in the attention operation, as a bias
if exists(mask):
raise NotImplementedError
out = (
out.unsqueeze(0)
.reshape(b, self.heads, out.shape[1], self.dim_head)
.permute(0, 2, 1, 3)
.reshape(b, out.shape[1], self.heads * self.dim_head)
)
if additional_tokens is not None:
# remove additional token
out = out[:, n_tokens_to_mask:]
return self.to_out(out)
class BasicTransformerBlock(nn.Module):
ATTENTION_MODES = {
"softmax": CrossAttention, # vanilla attention
"softmax-xformers": MemoryEfficientCrossAttention, # ampere
}
def __init__(
self,
dim,
n_heads,
d_head,
dropout=0.0,
context_dim=None,
add_context_dim=None,
gated_ff=True,
checkpoint=True,
disable_self_attn=False,
attn_mode="softmax",
sdp_backend=None,
):
super().__init__()
assert attn_mode in self.ATTENTION_MODES
if attn_mode != "softmax" and not XFORMERS_IS_AVAILABLE:
print(
f"Attention mode '{attn_mode}' is not available. Falling back to native attention. "
f"This is not a problem in Pytorch >= 2.0. FYI, you are running with PyTorch version {torch.__version__}"
)
attn_mode = "softmax"
elif attn_mode == "softmax" and not SDP_IS_AVAILABLE:
print(
"We do not support vanilla attention anymore, as it is too expensive. Sorry."
)
if not XFORMERS_IS_AVAILABLE:
assert (
False
), "Please install xformers via e.g. 'pip install xformers==0.0.16'"
else:
print("Falling back to xformers efficient attention.")
attn_mode = "softmax-xformers"
attn_cls = self.ATTENTION_MODES[attn_mode]
if version.parse(torch.__version__) >= version.parse("2.0.0"):
assert sdp_backend is None or isinstance(sdp_backend, SDPBackend)
else:
assert sdp_backend is None
self.disable_self_attn = disable_self_attn
self.attn1 = attn_cls(
query_dim=dim,
heads=n_heads,
dim_head=d_head,
dropout=dropout,
context_dim=context_dim if self.disable_self_attn else None,
backend=sdp_backend,
) # is a self-attention if not self.disable_self_attn
self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
if context_dim is not None and context_dim > 0:
self.attn2 = attn_cls(
query_dim=dim,
context_dim=context_dim,
heads=n_heads,
dim_head=d_head,
dropout=dropout,
backend=sdp_backend,
) # is self-attn if context is none
if add_context_dim is not None and add_context_dim > 0:
self.add_attn = attn_cls(
query_dim=dim,
context_dim=add_context_dim,
heads=n_heads,
dim_head=d_head,
dropout=dropout,
backend=sdp_backend,
) # is self-attn if context is none
self.add_norm = nn.LayerNorm(dim)
self.norm1 = nn.LayerNorm(dim)
self.norm2 = nn.LayerNorm(dim)
self.norm3 = nn.LayerNorm(dim)
self.checkpoint = checkpoint
def forward(
self, x, context=None, add_context=None, additional_tokens=None, n_times_crossframe_attn_in_self=0
):
kwargs = {"x": x}
if context is not None:
kwargs.update({"context": context})
if additional_tokens is not None:
kwargs.update({"additional_tokens": additional_tokens})
if n_times_crossframe_attn_in_self:
kwargs.update(
{"n_times_crossframe_attn_in_self": n_times_crossframe_attn_in_self}
)
return checkpoint(
self._forward, (x, context, add_context), self.parameters(), self.checkpoint
)
def _forward(
self, x, context=None, add_context=None, additional_tokens=None, n_times_crossframe_attn_in_self=0
):
x = (
self.attn1(
self.norm1(x),
context=context if self.disable_self_attn else None,
additional_tokens=additional_tokens,
n_times_crossframe_attn_in_self=n_times_crossframe_attn_in_self
if not self.disable_self_attn
else 0,
)
+ x
)
if hasattr(self, "attn2"):
x = (
self.attn2(
self.norm2(x), context=context, additional_tokens=additional_tokens
)
+ x
)
if hasattr(self, "add_attn"):
x = (
self.add_attn(
self.add_norm(x), context=add_context, additional_tokens=additional_tokens
)
+ x
)
x = self.ff(self.norm3(x)) + x
return x
class BasicTransformerSingleLayerBlock(nn.Module):
ATTENTION_MODES = {
"softmax": CrossAttention, # vanilla attention
"softmax-xformers": MemoryEfficientCrossAttention # on the A100s not quite as fast as the above version
# (todo might depend on head_dim, check, falls back to semi-optimized kernels for dim!=[16,32,64,128])
}
def __init__(
self,
dim,
n_heads,
d_head,
dropout=0.0,
context_dim=None,
gated_ff=True,
checkpoint=True,
attn_mode="softmax",
):
super().__init__()
assert attn_mode in self.ATTENTION_MODES
attn_cls = self.ATTENTION_MODES[attn_mode]
self.attn1 = attn_cls(
query_dim=dim,
heads=n_heads,
dim_head=d_head,
dropout=dropout,
context_dim=context_dim,
)
self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
self.norm1 = nn.LayerNorm(dim)
self.norm2 = nn.LayerNorm(dim)
self.checkpoint = checkpoint
def forward(self, x, context=None):
return checkpoint(
self._forward, (x, context), self.parameters(), self.checkpoint
)
def _forward(self, x, context=None):
x = self.attn1(self.norm1(x), context=context) + x
x = self.ff(self.norm2(x)) + x
return x
class SpatialTransformer(nn.Module):
"""
Transformer block for image-like data.
First, project the input (aka embedding)
and reshape to b, t, d.
Then apply standard transformer action.
Finally, reshape to image
NEW: use_linear for more efficiency instead of the 1x1 convs
"""
def __init__(
self,
in_channels,
n_heads,
d_head,
depth=1,
dropout=0.0,
context_dim=None,
add_context_dim=None,
disable_self_attn=False,
use_linear=False,
attn_type="softmax",
use_checkpoint=True,
# sdp_backend=SDPBackend.FLASH_ATTENTION
sdp_backend=None,
):
super().__init__()
# print(
# f"constructing {self.__class__.__name__} of depth {depth} w/ {in_channels} channels and {n_heads} heads"
# )
from omegaconf import ListConfig
if exists(context_dim) and not isinstance(context_dim, (list, ListConfig)):
context_dim = [context_dim]
if exists(context_dim) and isinstance(context_dim, list):
if depth != len(context_dim):
# print(
# f"WARNING: {self.__class__.__name__}: Found context dims {context_dim} of depth {len(context_dim)}, "
# f"which does not match the specified 'depth' of {depth}. Setting context_dim to {depth * [context_dim[0]]} now."
# )
# depth does not match context dims.
assert all(
map(lambda x: x == context_dim[0], context_dim)
), "need homogenous context_dim to match depth automatically"
context_dim = depth * [context_dim[0]]
elif context_dim is None:
context_dim = [None] * depth
self.in_channels = in_channels
inner_dim = n_heads * d_head
self.norm = Normalize(in_channels)
if not use_linear:
self.proj_in = nn.Conv2d(
in_channels, inner_dim, kernel_size=1, stride=1, padding=0
)
else:
self.proj_in = nn.Linear(in_channels, inner_dim)
self.transformer_blocks = nn.ModuleList(
[
BasicTransformerBlock(
inner_dim,
n_heads,
d_head,
dropout=dropout,
context_dim=context_dim[d],
add_context_dim=add_context_dim,
disable_self_attn=disable_self_attn,
attn_mode=attn_type,
checkpoint=use_checkpoint,
sdp_backend=sdp_backend,
)
for d in range(depth)
]
)
if not use_linear:
self.proj_out = zero_module(
nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
)
else:
# self.proj_out = zero_module(nn.Linear(in_channels, inner_dim))
self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
self.use_linear = use_linear
def forward(self, x, context=None, add_context=None):
# note: if no context is given, cross-attention defaults to self-attention
if not isinstance(context, list):
context = [context]
b, c, h, w = x.shape
x_in = x
x = self.norm(x)
if not self.use_linear:
x = self.proj_in(x)
x = rearrange(x, "b c h w -> b (h w) c").contiguous()
if self.use_linear:
x = self.proj_in(x)
for i, block in enumerate(self.transformer_blocks):
if i > 0 and len(context) == 1:
i = 0 # use same context for each block
x = block(x, context=context[i], add_context=add_context)
if self.use_linear:
x = self.proj_out(x)
x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w).contiguous()
if not self.use_linear:
x = self.proj_out(x)
return x + x_in
def benchmark_attn():
# Lets define a helpful benchmarking function:
# https://pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html
device = "cuda" if torch.cuda.is_available() else "cpu"
import torch.nn.functional as F
import torch.utils.benchmark as benchmark
def benchmark_torch_function_in_microseconds(f, *args, **kwargs):
t0 = benchmark.Timer(
stmt="f(*args, **kwargs)", globals={"args": args, "kwargs": kwargs, "f": f}
)
return t0.blocked_autorange().mean * 1e6
# Lets define the hyper-parameters of our input
batch_size = 32
max_sequence_len = 1024
num_heads = 32
embed_dimension = 32
dtype = torch.float16
query = torch.rand(
batch_size,
num_heads,
max_sequence_len,
embed_dimension,
device=device,
dtype=dtype,
)
key = torch.rand(
batch_size,
num_heads,
max_sequence_len,
embed_dimension,
device=device,
dtype=dtype,
)
value = torch.rand(
batch_size,
num_heads,
max_sequence_len,
embed_dimension,
device=device,
dtype=dtype,
)
print(f"q/k/v shape:", query.shape, key.shape, value.shape)
# Lets explore the speed of each of the 3 implementations
from torch.backends.cuda import SDPBackend, sdp_kernel
# Helpful arguments mapper
backend_map = {
SDPBackend.MATH: {
"enable_math": True,
"enable_flash": False,
"enable_mem_efficient": False,
},
SDPBackend.FLASH_ATTENTION: {
"enable_math": False,
"enable_flash": True,
"enable_mem_efficient": False,
},
SDPBackend.EFFICIENT_ATTENTION: {
"enable_math": False,
"enable_flash": False,
"enable_mem_efficient": True,
},
}
from torch.profiler import ProfilerActivity, profile, record_function
activities = [ProfilerActivity.CPU, ProfilerActivity.CUDA]
print(
f"The default implementation runs in {benchmark_torch_function_in_microseconds(F.scaled_dot_product_attention, query, key, value):.3f} microseconds"
)
with profile(
activities=activities, record_shapes=False, profile_memory=True
) as prof:
with record_function("Default detailed stats"):
for _ in range(25):
o = F.scaled_dot_product_attention(query, key, value)
print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
print(
f"The math implementation runs in {benchmark_torch_function_in_microseconds(F.scaled_dot_product_attention, query, key, value):.3f} microseconds"
)
with sdp_kernel(**backend_map[SDPBackend.MATH]):
with profile(
activities=activities, record_shapes=False, profile_memory=True
) as prof:
with record_function("Math implmentation stats"):
for _ in range(25):
o = F.scaled_dot_product_attention(query, key, value)
print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
with sdp_kernel(**backend_map[SDPBackend.FLASH_ATTENTION]):
try:
print(
f"The flash attention implementation runs in {benchmark_torch_function_in_microseconds(F.scaled_dot_product_attention, query, key, value):.3f} microseconds"
)
except RuntimeError:
print("FlashAttention is not supported. See warnings for reasons.")
with profile(
activities=activities, record_shapes=False, profile_memory=True
) as prof:
with record_function("FlashAttention stats"):
for _ in range(25):
o = F.scaled_dot_product_attention(query, key, value)
print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
with sdp_kernel(**backend_map[SDPBackend.EFFICIENT_ATTENTION]):
try:
print(
f"The memory efficient implementation runs in {benchmark_torch_function_in_microseconds(F.scaled_dot_product_attention, query, key, value):.3f} microseconds"
)
except RuntimeError:
print("EfficientAttention is not supported. See warnings for reasons.")
with profile(
activities=activities, record_shapes=False, profile_memory=True
) as prof:
with record_function("EfficientAttention stats"):
for _ in range(25):
o = F.scaled_dot_product_attention(query, key, value)
print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
def run_model(model, x, context):
return model(x, context)
def benchmark_transformer_blocks():
device = "cuda" if torch.cuda.is_available() else "cpu"
import torch.utils.benchmark as benchmark
def benchmark_torch_function_in_microseconds(f, *args, **kwargs):
t0 = benchmark.Timer(
stmt="f(*args, **kwargs)", globals={"args": args, "kwargs": kwargs, "f": f}
)
return t0.blocked_autorange().mean * 1e6
checkpoint = True
compile = False
batch_size = 32
h, w = 64, 64
context_len = 77
embed_dimension = 1024
context_dim = 1024
d_head = 64
transformer_depth = 4
n_heads = embed_dimension // d_head
dtype = torch.float16
model_native = SpatialTransformer(
embed_dimension,
n_heads,
d_head,
context_dim=context_dim,
use_linear=True,
use_checkpoint=checkpoint,
attn_type="softmax",
depth=transformer_depth,
sdp_backend=SDPBackend.FLASH_ATTENTION,
).to(device)
model_efficient_attn = SpatialTransformer(
embed_dimension,
n_heads,
d_head,
context_dim=context_dim,
use_linear=True,
depth=transformer_depth,
use_checkpoint=checkpoint,
attn_type="softmax-xformers",
).to(device)
if not checkpoint and compile:
print("compiling models")
model_native = torch.compile(model_native)
model_efficient_attn = torch.compile(model_efficient_attn)
x = torch.rand(batch_size, embed_dimension, h, w, device=device, dtype=dtype)
c = torch.rand(batch_size, context_len, context_dim, device=device, dtype=dtype)
from torch.profiler import ProfilerActivity, profile, record_function
activities = [ProfilerActivity.CPU, ProfilerActivity.CUDA]
with torch.autocast("cuda"):
print(
f"The native model runs in {benchmark_torch_function_in_microseconds(model_native.forward, x, c):.3f} microseconds"
)
print(
f"The efficientattn model runs in {benchmark_torch_function_in_microseconds(model_efficient_attn.forward, x, c):.3f} microseconds"
)
print(75 * "+")
print("NATIVE")
print(75 * "+")
torch.cuda.reset_peak_memory_stats()
with profile(
activities=activities, record_shapes=False, profile_memory=True
) as prof:
with record_function("NativeAttention stats"):
for _ in range(25):
model_native(x, c)
print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
print(torch.cuda.max_memory_allocated() * 1e-9, "GB used by native block")
print(75 * "+")
print("Xformers")
print(75 * "+")
torch.cuda.reset_peak_memory_stats()
with profile(
activities=activities, record_shapes=False, profile_memory=True
) as prof:
with record_function("xformers stats"):
for _ in range(25):
model_efficient_attn(x, c)
print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))
print(torch.cuda.max_memory_allocated() * 1e-9, "GB used by xformers block")
def test01():
# conv1x1 vs linear
from ..util import count_params
conv = nn.Conv2d(3, 32, kernel_size=1).cuda()
print(count_params(conv))
linear = torch.nn.Linear(3, 32).cuda()
print(count_params(linear))
print(conv.weight.shape)
# use same initialization
linear.weight = torch.nn.Parameter(conv.weight.squeeze(-1).squeeze(-1))
linear.bias = torch.nn.Parameter(conv.bias)
print(linear.weight.shape)
x = torch.randn(11, 3, 64, 64).cuda()
xr = rearrange(x, "b c h w -> b (h w) c").contiguous()
print(xr.shape)
out_linear = linear(xr)
print(out_linear.mean(), out_linear.shape)
out_conv = conv(x)
print(out_conv.mean(), out_conv.shape)
print("done with test01.\n")
def test02():
# try cosine flash attention
import time
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cudnn.benchmark = True
print("testing cosine flash attention...")
DIM = 1024
SEQLEN = 4096
BS = 16
print(" softmax (vanilla) first...")
model = BasicTransformerBlock(
dim=DIM,
n_heads=16,
d_head=64,
dropout=0.0,
context_dim=None,
attn_mode="softmax",
).cuda()
try:
x = torch.randn(BS, SEQLEN, DIM).cuda()
tic = time.time()
y = model(x)
toc = time.time()
print(y.shape, toc - tic)
except RuntimeError as e:
# likely oom
print(str(e))
print("\n now flash-cosine...")
model = BasicTransformerBlock(
dim=DIM,
n_heads=16,
d_head=64,
dropout=0.0,
context_dim=None,
attn_mode="flash-cosine",
).cuda()
x = torch.randn(BS, SEQLEN, DIM).cuda()
tic = time.time()
y = model(x)
toc = time.time()
print(y.shape, toc - tic)
print("done with test02.\n")
if __name__ == "__main__":
# test01()
# test02()
# test03()
# benchmark_attn()
benchmark_transformer_blocks()
print("done.")
|