Spaces:
Running
on
Zero
Running
on
Zero
File size: 4,170 Bytes
04b20ec 6b88e7a |
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 |
import os
import torch
import torch.nn.functional as F
scaled_dot_product_attention = F.scaled_dot_product_attention
if os.environ.get('CA_USE_SAGEATTN', '0') == '1':
try:
from sageattention import sageattn
except ImportError:
raise ImportError('Please install the package "sageattention" to use this USE_SAGEATTN.')
scaled_dot_product_attention = sageattn
class CrossAttentionProcessor:
def __call__(self, attn, q, k, v):
out = scaled_dot_product_attention(q, k, v)
return out
class FlashVDMCrossAttentionProcessor:
def __init__(self, topk=None):
self.topk = topk
def __call__(self, attn, q, k, v):
if k.shape[-2] == 3072:
topk = 1024
elif k.shape[-2] == 512:
topk = 256
else:
topk = k.shape[-2] // 3
if self.topk is True:
q1 = q[:, :, ::100, :]
sim = q1 @ k.transpose(-1, -2)
sim = torch.mean(sim, -2)
topk_ind = torch.topk(sim, dim=-1, k=topk).indices.squeeze(-2).unsqueeze(-1)
topk_ind = topk_ind.expand(-1, -1, -1, v.shape[-1])
v0 = torch.gather(v, dim=-2, index=topk_ind)
k0 = torch.gather(k, dim=-2, index=topk_ind)
out = scaled_dot_product_attention(q, k0, v0)
elif self.topk is False:
out = scaled_dot_product_attention(q, k, v)
else:
idx, counts = self.topk
start = 0
outs = []
for grid_coord, count in zip(idx, counts):
end = start + count
q_chunk = q[:, :, start:end, :]
q1 = q_chunk[:, :, ::50, :]
sim = q1 @ k.transpose(-1, -2)
sim = torch.mean(sim, -2)
topk_ind = torch.topk(sim, dim=-1, k=topk).indices.squeeze(-2).unsqueeze(-1)
topk_ind = topk_ind.expand(-1, -1, -1, v.shape[-1])
v0 = torch.gather(v, dim=-2, index=topk_ind)
k0 = torch.gather(k, dim=-2, index=topk_ind)
out = scaled_dot_product_attention(q_chunk, k0, v0)
outs.append(out)
start += count
out = torch.cat(outs, dim=-2)
self.topk = False
return out
class FlashVDMTopMCrossAttentionProcessor:
def __init__(self, topk=None):
self.topk = topk
def __call__(self, attn, q, k, v):
if k.shape[-2] == 3072:
topk = 1024
elif k.shape[-2] == 512:
topk = 256
else:
topk = k.shape[-2] // 3
if self.topk is True:
q1 = q[:, :, ::100, :]
sim = q1 @ k.transpose(-1, -2)
sim = torch.mean(sim, -2)
topk_ind = torch.topk(sim, dim=-1, k=topk).indices.squeeze(-2).unsqueeze(-1)
topk_ind = topk_ind.expand(-1, -1, -1, v.shape[-1])
v0 = torch.gather(v, dim=-2, index=topk_ind)
k0 = torch.gather(k, dim=-2, index=topk_ind)
out = scaled_dot_product_attention(q, k0, v0)
elif self.topk is False:
out = scaled_dot_product_attention(q, k, v)
else:
idx, counts = self.topk
start = 0
outs = []
for grid_coord, count in zip(idx, counts):
end = start + count
q_chunk = q[:, :, start:end, :]
q1 = q_chunk[:, :, ::30, :]
sim = q1 @ k.transpose(-1, -2)
# sim = sim.to(torch.float32)
sim = sim.softmax(-1)
sim = torch.mean(sim, 1)
activated_token = torch.where(sim > 1e-6)[2]
index = torch.unique(activated_token, return_counts=True)[0].unsqueeze(0).unsqueeze(0).unsqueeze(-1)
index = index.expand(-1, v.shape[1], -1, v.shape[-1])
v0 = torch.gather(v, dim=-2, index=index)
k0 = torch.gather(k, dim=-2, index=index)
out = scaled_dot_product_attention(q_chunk, k0, v0) # bhnc
outs.append(out)
start += count
out = torch.cat(outs, dim=-2)
self.topk = False
return out
|