Repository navigation
Expand file tree
/
Copy pathswitch_quant_config_preparer.py
More file actions
192 lines (165 loc) · 9.84 KB
/
Copy pathswitch_quant_config_preparer.py
File metadata and controls
192 lines (165 loc) · 9.84 KB
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
import pickle
import torch
import copy
from transformers import SwitchTransformersForConditionalGeneration, HqqConfig
from hqq.core.quantize import *
def compute_expert_order(
model_checkpoint: str,
order_type: str = "router",
zeta: float = 3.0,
):
model = SwitchTransformersForConditionalGeneration.from_pretrained(
model_checkpoint,
device_map='cpu',
)
moe_layers = [1, 3, 5, 7, 9, 11]
encoder_router_order = []
decoder_router_order = []
for i in moe_layers:
en_weight = model.state_dict()[
f"encoder.block.{i}.layer.1.mlp.router.classifier.weight"
]
de_weight = model.state_dict()[
f"decoder.block.{i}.layer.2.mlp.router.classifier.weight"
]
_, en_indices = torch.sort(
torch.linalg.vector_norm(en_weight, ord=2, dim=1)
)
_, de_indices = torch.sort(
torch.linalg.vector_norm(de_weight, ord=2, dim=1)
)
encoder_router_order.append(en_indices.tolist())
decoder_router_order.append(de_indices.tolist())
if order_type == "router":
del model
return encoder_router_order, decoder_router_order
en_var = []
de_var = []
for i in moe_layers:
en_layer_var = []
de_layer_var = []
for j in range(64):
en_w = model.state_dict()[
f"encoder.block.{i}.layer.1.mlp.experts.expert_{j}.wi.weight"
]
de_w = model.state_dict()[
f"decoder.block.{i}.layer.2.mlp.experts.expert_{j}.wi.weight"
]
en_layer_var.append(torch.max(torch.var(en_w, dim=1)))
de_layer_var.append(torch.max(torch.var(de_w, dim=1)))
en_var.append(en_layer_var)
de_var.append(de_layer_var)
if order_type == "variance":
encoder_var_order = [
torch.argsort(torch.tensor(v)).tolist() for v in en_var
]
decoder_var_order = [
torch.argsort(torch.tensor(v)).tolist() for v in de_var
]
del model
return encoder_var_order, decoder_var_order
if order_type == "combined":
encoder_order_var = []
decoder_order_var = []
for i in range(6):
c_l = copy.deepcopy(encoder_router_order[i])
for j in reversed(encoder_router_order[i]):
temp = [
k
for k, val in enumerate([en_var[i][l] for l in c_l])
if (zeta * val) < en_var[i][j]
]
if temp and c_l.index(j) > temp[0]:
y = c_l.pop(c_l.index(j))
c_l.insert(temp[0], y)
encoder_order_var.append(c_l)
c_d = copy.deepcopy(decoder_router_order[i])
for j in reversed(decoder_router_order[i]):
temp_d = [
k_d
for k_d, val_d in enumerate([de_var[i][d] for d in c_d])
if (zeta * val_d) < de_var[i][j]
]
if temp_d and c_d.index(j) > temp_d[0]:
y_d = c_d.pop(c_d.index(j))
c_d.insert(temp_d[0], y_d)
decoder_order_var.append(c_d)
del model
return encoder_order_var, decoder_order_var
raise ValueError(
"order_type must be one of: 'router', 'variance', 'combined'"
)
group_quant_configs_wi = [{'nbits':1,'group_size':None,'axis':1},{'nbits':2,'group_size':None,'axis':1}, {'nbits':3,'group_size':768,'axis':1},
{'nbits':4,'group_size':None,'axis':1}, {'nbits':8,'group_size':None,'axis':1}]
group_quant_configs_wo = [{'nbits':1,'group_size':None,'axis':1},{'nbits':2,'group_size':None,'axis':1}, {'nbits':3,'group_size':3072,'axis':1},
{'nbits':4,'group_size':None,'axis':1}, {'nbits':8,'group_size':None,'axis':1}]
no_of_experts = 64
def create_quant_configs(model_checkpoint, order_type, zeta, encoder_expert_quant_group_sizes, decoder_expert_quant_group_sizes):
encoder_order, decoder_order = compute_expert_order(model_checkpoint, order_type=order_type, zeta=zeta)
module_configs= pickle.load(open('switch_pre_module_configs.pkl','rb'))
for i in [1,3,5,7,9,11]:
for j in range(no_of_experts):
if j<=encoder_expert_quant_group_sizes[(i-1)//2][0]-1:
module_configs.update({'encoder.block.'+str(i)+'.layer.1.mlp.experts.expert_'+
str(encoder_order[(i-1)//2][63-j])+'.wi':group_quant_configs_wi[0]})
module_configs.update({'encoder.block.'+str(i)+'.layer.1.mlp.experts.expert_'+
str(encoder_order[(i-1)//2][63-j])+'.wo':group_quant_configs_wo[0]})
elif j<=encoder_expert_quant_group_sizes[(i-1)//2][0]+encoder_expert_quant_group_sizes[(i-1)//2][1]-1:
module_configs.update({'encoder.block.'+str(i)+'.layer.1.mlp.experts.expert_'+
str(encoder_order[(i-1)//2][63-j])+'.wi':group_quant_configs_wi[1]})
module_configs.update({'encoder.block.'+str(i)+'.layer.1.mlp.experts.expert_'+
str(encoder_order[(i-1)//2][63-j])+'.wo':group_quant_configs_wo[1]})
elif j<=(encoder_expert_quant_group_sizes[(i-1)//2][0]+encoder_expert_quant_group_sizes[(i-1)//2][1]+
encoder_expert_quant_group_sizes[(i-1)//2][2]-1):
module_configs.update({'encoder.block.'+str(i)+'.layer.1.mlp.experts.expert_'+
str(encoder_order[(i-1)//2][63-j])+'.wi':group_quant_configs_wi[2]})
module_configs.update({'encoder.block.'+str(i)+'.layer.1.mlp.experts.expert_'+
str(encoder_order[(i-1)//2][63-j])+'.wo':group_quant_configs_wo[2]})
elif j<=(encoder_expert_quant_group_sizes[(i-1)//2][0]+encoder_expert_quant_group_sizes[(i-1)//2][1]+
encoder_expert_quant_group_sizes[(i-1)//2][2]+encoder_expert_quant_group_sizes[(i-1)//2][3]-1):
module_configs.update({'encoder.block.'+str(i)+'.layer.1.mlp.experts.expert_'+
str(encoder_order[(i-1)//2][63-j])+'.wi':group_quant_configs_wi[3]})
module_configs.update({'encoder.block.'+str(i)+'.layer.1.mlp.experts.expert_'+
str(encoder_order[(i-1)//2][63-j])+'.wo':group_quant_configs_wo[3]})
else:
module_configs.update({'encoder.block.'+str(i)+'.layer.1.mlp.experts.expert_'+
str(encoder_order[(i-1)//2][63-j])+'.wi':group_quant_configs_wi[4]})
module_configs.update({'encoder.block.'+str(i)+'.layer.1.mlp.experts.expert_'+
str(encoder_order[(i-1)//2][63-j])+'.wo':group_quant_configs_wo[4]})
for i in [1,3,5,7,9,11]:
for j in range(no_of_experts):
if j<=decoder_expert_quant_group_sizes[(i-1)//2][0]-1:
module_configs.update({'decoder.block.'+str(i)+'.layer.2.mlp.experts.expert_'+
str(decoder_order[(i-1)//2][63-j])+'.wi':group_quant_configs_wi[0]})
module_configs.update({'decoder.block.'+str(i)+'.layer.2.mlp.experts.expert_'+
str(decoder_order[(i-1)//2][63-j])+'.wo':group_quant_configs_wo[0]})
elif j<=decoder_expert_quant_group_sizes[(i-1)//2][0]+decoder_expert_quant_group_sizes[(i-1)//2][1]-1:
module_configs.update({'decoder.block.'+str(i)+'.layer.2.mlp.experts.expert_'+
str(decoder_order[(i-1)//2][63-j])+'.wi':group_quant_configs_wi[1]})
module_configs.update({'decoder.block.'+str(i)+'.layer.2.mlp.experts.expert_'+
str(decoder_order[(i-1)//2][63-j])+'.wo':group_quant_configs_wo[1]})
elif j<=(decoder_expert_quant_group_sizes[(i-1)//2][0]+decoder_expert_quant_group_sizes[(i-1)//2][1]+
decoder_expert_quant_group_sizes[(i-1)//2][2]-1):
module_configs.update({'decoder.block.'+str(i)+'.layer.2.mlp.experts.expert_'+
str(decoder_order[(i-1)//2][63-j])+'.wi':group_quant_configs_wi[2]})
module_configs.update({'decoder.block.'+str(i)+'.layer.2.mlp.experts.expert_'+
str(decoder_order[(i-1)//2][63-j])+'.wo':group_quant_configs_wo[2]})
elif j<=(decoder_expert_quant_group_sizes[(i-1)//2][0]+decoder_expert_quant_group_sizes[(i-1)//2][1]+
decoder_expert_quant_group_sizes[(i-1)//2][2]+decoder_expert_quant_group_sizes[(i-1)//2][3]-1):
module_configs.update({'decoder.block.'+str(i)+'.layer.2.mlp.experts.expert_'+
str(decoder_order[(i-1)//2][63-j])+'.wi':group_quant_configs_wi[3]})
module_configs.update({'decoder.block.'+str(i)+'.layer.2.mlp.experts.expert_'+
str(decoder_order[(i-1)//2][63-j])+'.wo':group_quant_configs_wo[3]})
else:
module_configs.update({'decoder.block.'+str(i)+'.layer.2.mlp.experts.expert_'+
str(decoder_order[(i-1)//2][63-j])+'.wi':group_quant_configs_wi[4]})
module_configs.update({'decoder.block.'+str(i)+'.layer.2.mlp.experts.expert_'+
str(decoder_order[(i-1)//2][63-j])+'.wo':group_quant_configs_wo[4]})
no_of_blocks = 6
total_bits = 0.0
precisions = [1,2,3,4,8]
for i in range(no_of_blocks):
for j in range(5):
total_bits = total_bits+(encoder_expert_quant_group_sizes[i][j]*precisions[j])+(decoder_expert_quant_group_sizes[i][j]*precisions[j])
avg_bits_per_expert = total_bits/(no_of_experts*no_of_blocks*2)
return avg_bits_per_expert, module_configs