Spaces:
Paused
Paused
File size: 4,032 Bytes
0084610 |
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 |
import math
import torch
from torch import Tensor, IntTensor, BoolTensor
from torch.nn.attention.flex_attention import BlockMask, _mask_mod_signature
from einops import rearrange
def exist(item):
return item is not None
def freeze(model):
for p in model.parameters():
p.requires_grad = False
return model
@torch.autocast(device_type="cuda", enabled=False)
def get_freqs(dim, max_period=10000.0):
freqs = torch.exp(
-math.log(max_period)
* torch.arange(start=0, end=dim, dtype=torch.float32)
/ dim
)
return freqs
def fractal_flatten(x, rope, shape, block_mask=False):
if block_mask:
pixel_size = 8
x = local_patching(x, shape, (1, pixel_size, pixel_size), dim=0)
rope = local_patching(rope, shape, (1, pixel_size, pixel_size), dim=0)
x = x.flatten(0, 1)
rope = rope.flatten(0, 1)
else:
x = x.flatten(0, 2)
rope = rope.flatten(0, 2)
return x, rope
def fractal_unflatten(x, shape, block_mask=False):
if block_mask:
pixel_size = 8
x = x.reshape(-1, pixel_size**2, *x.shape[1:])
x = local_merge(x, shape, (1, pixel_size, pixel_size), dim=0)
else:
x = x.reshape(*shape, *x.shape[1:])
return x
def local_patching(x, shape, group_size, dim=0):
duration, height, width = shape
g1, g2, g3 = group_size
x = x.reshape(
*x.shape[:dim],
duration // g1,
g1,
height // g2,
g2,
width // g3,
g3,
*x.shape[dim + 3 :]
)
x = x.permute(
*range(len(x.shape[:dim])),
dim,
dim + 2,
dim + 4,
dim + 1,
dim + 3,
dim + 5,
*range(dim + 6, len(x.shape))
)
x = x.flatten(dim, dim + 2).flatten(dim + 1, dim + 3)
return x
def local_merge(x, shape, group_size, dim=0):
duration, height, width = shape
g1, g2, g3 = group_size
x = x.reshape(
*x.shape[:dim],
duration // g1,
height // g2,
width // g3,
g1,
g2,
g3,
*x.shape[dim + 2 :]
)
x = x.permute(
*range(len(x.shape[:dim])),
dim,
dim + 3,
dim + 1,
dim + 4,
dim + 2,
dim + 5,
*range(dim + 6, len(x.shape))
)
x = x.flatten(dim, dim + 1).flatten(dim + 1, dim + 2).flatten(dim + 2, dim + 3)
return x
def fast_sta_nabla(
T: int, H: int, W: int, wT: int = 3, wH: int = 3, wW: int = 3, device="cuda"
) -> Tensor:
l = torch.Tensor([T, H, W]).amax()
r = torch.arange(0, l, 1, dtype=torch.int16, device=device)
mat = (r.unsqueeze(1) - r.unsqueeze(0)).abs()
sta_t, sta_h, sta_w = (
mat[:T, :T].flatten(),
mat[:H, :H].flatten(),
mat[:W, :W].flatten(),
)
sta_t = sta_t <= wT // 2
sta_h = sta_h <= wH // 2
sta_w = sta_w <= wW // 2
sta_hw = (
(sta_h.unsqueeze(1) * sta_w.unsqueeze(0))
.reshape(H, H, W, W)
.transpose(1, 2)
.flatten()
)
sta = (
(sta_t.unsqueeze(1) * sta_hw.unsqueeze(0))
.reshape(T, T, H * W, H * W)
.transpose(1, 2)
)
return sta.reshape(T * H * W, T * H * W)
def nablaT_v2(
q: Tensor,
k: Tensor,
sta: Tensor,
thr: float = 0.9,
) -> BlockMask:
# Map estimation
B, h, S, D = q.shape
s1 = S // 64
qa = q.reshape(B, h, s1, 64, D).mean(-2)
ka = k.reshape(B, h, s1, 64, D).mean(-2).transpose(-2, -1)
map = qa @ ka
map = torch.softmax(map / math.sqrt(D), dim=-1)
# Map binarization
vals, inds = map.sort(-1)
cvals = vals.cumsum_(-1)
mask = (cvals >= 1 - thr).int()
mask = mask.gather(-1, inds.argsort(-1))
mask = torch.logical_or(mask, sta)
# BlockMask creation
kv_nb = mask.sum(-1).to(torch.int32)
kv_inds = mask.argsort(dim=-1, descending=True).to(torch.int32)
return BlockMask.from_kv_blocks(
torch.zeros_like(kv_nb), kv_inds, kv_nb, kv_inds, BLOCK_SIZE=64, mask_mod=None
)
|