-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathkv_cache_qwen.py
executable file
·225 lines (190 loc) · 9.28 KB
/
kv_cache_qwen.py
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
import torch
import random
def slice2d(x, start, end):
return x[:, :, start:end, ...]
def slice3d(x, start, end):
return x[:, :, :, start:end, ...]
def slice1d(x, start, end):
return x[:, start:end, ...]
import torch.nn.functional as F
DIM_TO_SLICE = {
1: slice1d,
2: slice2d,
3: slice3d,
}
class ElasticCache:
def __init__(
self,
start_size=4,
recent_size=512,
k_seq_dim=2,
v_seq_dim=2,
ratio=0.,
distance=-25,
layer_num=40,
):
self.start_size = start_size
self.recent_size = recent_size
self.cache_size = start_size + recent_size
self.k_seq_dim = k_seq_dim
self.v_seq_dim = v_seq_dim
self.k_slice = DIM_TO_SLICE[k_seq_dim]
self.v_slice = DIM_TO_SLICE[v_seq_dim]
self.score_sum = torch.zeros(layer_num, self.cache_size + 1).cuda()
self.ratio = ratio
self.protect_size = 1
self.lazy_size = 0
self.flag = True
self.distance = distance
self.layer_num = layer_num
self.selected_idx = 0
def __call__(self, past_key_values, num_of_token=None, attentions=None):
if past_key_values is None:
return None
attn_score = [attention for attention in attentions]
seq_len = past_key_values[0][0].size(self.k_seq_dim)
# update attn score
attn_score = torch.cat(attn_score, dim=0)
attn_score = attn_score.mean(dim=1, keepdim=False)
if attn_score.shape[-2] > 1:
assert self.flag is True # only use for the first time
for idx in range(attn_score.shape[-1]):
cur_score = attn_score[:, idx, :idx+1]
self.score_sum[:, :(cur_score.shape[-1])] += cur_score
else:
pass
forget_num = int(seq_len - num_of_token * (1 - self.ratio))
if forget_num <= 0:
return past_key_values
else:
if forget_num > 1:
assert self.flag is True
self.flag = False
selected_idx_all = []
merge_idx_all = []
throw_idx_all = []
for idx in range(self.layer_num):
selected_idx = torch.where(torch.argsort(self.score_sum[idx, self.start_size:(seq_len - self.protect_size)]) > forget_num)[0] + self.start_size
throw_idx = torch.where(torch.argsort(self.score_sum[idx, self.start_size:(seq_len - self.protect_size)]) <= forget_num)[0]
merge_idx = []
for i in range(len(throw_idx)):
merge_idx.append(selected_idx[torch.abs((selected_idx - throw_idx[i])).argmin()].unsqueeze(0))
merge_idx = torch.cat(merge_idx)
selected_idx = torch.cat([torch.arange(self.start_size).cuda(), selected_idx, torch.tensor([seq_len - self.protect_size]).cuda()], dim=0) # the last token is always kept
selected_idx_all.append(selected_idx)
merge_idx_all.append(merge_idx)
throw_idx_all.append(throw_idx)
if self.distance > 0:
self.selected_idx = self.distance
else:
self.selected_idx = seq_len - forget_num + self.distance
past_key_values_return = []
for idx, (k, v) in enumerate(past_key_values):
selected_idx = selected_idx_all[idx]
merge_idx = merge_idx_all[idx]
throw_idx = throw_idx_all[idx]
k_forget = k.gather(dim=1, index=throw_idx.view(1,-1,1,1).expand(k.shape[0], -1, k.shape[2], k.shape[-1]))
v_forget = v.gather(dim=1, index=throw_idx.view(1,-1,1,1).expand(v.shape[0], -1, v.shape[2], v.shape[-1]))
k = k.scatter_reduce(1, merge_idx.view(1,-1,1,1).expand(k.shape[0], -1, k.shape[2], k.shape[-1]), k_forget, 'mean')
v = v.scatter_reduce(1, merge_idx.view(1,-1,1,1).expand(v.shape[0], -1, v.shape[2], v.shape[-1]), v_forget, 'mean')
k_new = k.gather(dim=1, index=selected_idx.view(1,-1,1,1).expand(k.shape[0], -1, k.shape[2] ,k.shape[-1]))
v_new = v.gather(dim=1, index=selected_idx.view(1,-1,1,1).expand(v.shape[0], -1, v.shape[2] ,v.shape[-1]))
past_key_values_return.append([k_new, v_new])
return past_key_values_return
else:
selected_idx = self.selected_idx
return [[torch.cat([self.k_slice(k, 0, selected_idx), self.k_slice(k, (selected_idx+1), seq_len),],
dim=self.k_seq_dim,),
torch.cat([self.v_slice(v, 0, selected_idx), self.v_slice(v, (selected_idx+1), seq_len),],
dim=self.v_seq_dim,)]
for k, v in past_key_values]
class LocalCache:
def __init__(
self,
start_size=4,
recent_size=512,
k_seq_dim=2,
v_seq_dim=2,
ratio=0.
):
self.start_size = start_size
self.recent_size = recent_size
self.cache_size = start_size + recent_size
self.k_seq_dim = k_seq_dim
self.v_seq_dim = v_seq_dim
self.k_slice = DIM_TO_SLICE[k_seq_dim]
self.v_slice = DIM_TO_SLICE[v_seq_dim]
self.ratio = ratio
def __call__(self, past_key_values, num_of_token=None, attentions=None):
if past_key_values is None:
return None
seq_len = past_key_values[0][0].size(self.k_seq_dim)
forget_num = int(seq_len - num_of_token * (1 - self.ratio))
if forget_num <= 0:
return past_key_values
else:
return [[torch.cat([self.k_slice(k, 0, self.start_size), self.k_slice(k, forget_num + self.start_size, seq_len),],
dim=self.k_seq_dim,),
torch.cat([self.v_slice(v, 0, self.start_size), self.v_slice(v, forget_num + self.start_size, seq_len),],
dim=self.v_seq_dim,),]
for k, v in past_key_values]
class H2OCache:
def __init__(
self,
start_size=4,
recent_size=512,
k_seq_dim=2,
v_seq_dim=2,
ratio=0.
):
self.start_size = start_size
self.recent_size = recent_size
self.cache_size = start_size + recent_size
self.k_seq_dim = k_seq_dim
self.v_seq_dim = v_seq_dim
self.k_slice = DIM_TO_SLICE[k_seq_dim]
self.v_slice = DIM_TO_SLICE[v_seq_dim]
self.score_sum = torch.zeros(self.cache_size + 1).cuda()
self.ratio = ratio
self.protect_size = 1
self.flag = True
def __call__(self, past_key_values, num_of_token=None, attentions=None):
if past_key_values is None:
return None
attn_score = [attention for attention in attentions]
past_key_values_new = tuple(x for x in past_key_values)
seq_len = past_key_values_new[0][0].size(self.k_seq_dim)
# update attn score
attn_score = torch.cat(attn_score, dim=0)
attn_score = attn_score.mean(dim=1, keepdim=False).mean(dim=0, keepdim=False)
if attn_score.shape[-2] > 1:
assert self.flag is True # only use for the first time
for idx in range(attn_score.shape[-1]):
cur_score = attn_score[idx][:idx+1]
self.score_sum[:len(cur_score)] += cur_score
else:
attn_score = attn_score.squeeze(0)
self.score_sum[:seq_len] += attn_score
forget_num = int(seq_len - num_of_token * (1 - self.ratio))
if forget_num <= 0:
return past_key_values_new
else:
if forget_num > 1:
assert self.flag is True
self.flag = False
selected_idx = torch.where(torch.argsort(self.score_sum[:(seq_len - self.protect_size)]) > forget_num)[0]
selected_idx = torch.cat([selected_idx, torch.arange(seq_len - self.protect_size, seq_len).cuda()], dim=0)
past_key_values_return = []
for k, v in past_key_values_new:
k_new = k.gather(dim=1, index=selected_idx.view(1,-1,1,1).expand(k.shape[0], -1, k.shape[2], k.shape[-1]))
v_new = v.gather(dim=1, index=selected_idx.view(1,-1,1,1).expand(v.shape[0], -1, v.shape[2], v.shape[-1]))
past_key_values_return.append([k_new, v_new])
return past_key_values_return
else:
selected_idx = self.score_sum[self.start_size:(seq_len - self.protect_size)].argmin() + self.start_size
self.score_sum[(selected_idx):-1] = self.score_sum[(selected_idx+1):].clone()
return [[torch.cat([self.k_slice(k, 0, selected_idx), self.k_slice(k, (selected_idx+1), seq_len),],
dim=self.k_seq_dim,),
torch.cat([self.v_slice(v, 0, selected_idx), self.v_slice(v, (selected_idx+1), seq_len),],
dim=self.v_seq_dim,)]
for k, v in past_key_values_new]