-
Notifications
You must be signed in to change notification settings - Fork 1
/
time_flies.py
330 lines (247 loc) · 11.3 KB
/
time_flies.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
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
import time
from hardware_spec import HardwareSpec
import json
class BasicOPCalculator(object):
def __init__(self, dtype_size, compute_ability, bandwidth):
self.dtype_size = dtype_size
self.compute_ability = compute_ability
self.bandwidth = bandwidth
def matmul(self, row, k, col):
total_ops = row * k * col
total_compute_time = total_ops / self.compute_ability
# read write
input_read = row * k
weight_read = k * col
output_write = row * col
total_read_write_bytes = input_read + weight_read + output_write
total_read_write_bytes *= self.dtype_size
total_read_write_time = total_read_write_bytes / self.bandwidth
return total_compute_time, total_read_write_time
class TimeFlies(BasicOPCalculator):
def __init__(self,
model_type="llama2",
seq_len = 100,
config_file = "config/Qwen2-VL-2B-Instruct.json",
machine="A10",
dtype="float16",
use_cache=True,
attention_version="v1"):
self.dtype = dtype
if self.dtype == "int8" or self.dtype == "fp8":
self.dtype_size = 1
if self.dtype == "float16":
self.dtype_size = 2
if self.dtype == "float32":
self.dtype_size = 4
machine = machine.upper()
self.compute_ablility = HardwareSpec[machine]["compute_ablility"][dtype]
self.bandwidth = HardwareSpec[machine]["bandwidth"]
super().__init__(self.dtype_size, self.compute_ablility, self.bandwidth)
with open(config_file, "r") as file:
config = json.load(file)
print(config)
self.seq_len = seq_len
self.hidden_size = config["hidden_size"]
self.heads = config["num_attention_heads"]
self.kv_heads = config["num_key_value_heads"]
self.head_size = self.hidden_size // self.heads
self.layer_num = config["num_hidden_layers"]
self.intermediate_size = config["intermediate_size"]
self.vocab_size = config["vocab_size"]
self.exp_inst = 4
self.rsqrt_inst = 4
self.use_cache = use_cache
self.attn_version = attention_version
self.print_params()
def print_params(self):
print("hidden_size is: ", self.hidden_size)
print("heads is : ", self.heads)
print("head size is: ", self.head_size)
print("layer num is: ", self.layer_num)
def get_real_seq_len(self):
if self.use_cache:
return 1
else:
return self.seq_len
def get_bottleneck_time(self, layer_type, cptt, rwt):
if cptt > rwt:
print(f'{layer_type} is compute bottleneck, ', end = '')
bnt = cptt
else:
print(f'{layer_type} is memory bottleneck ', end = '')
bnt = rwt
cptt = cptt * 1000000
rwt = rwt * 1000000
print(f'compute time is {cptt:.6f} us, read write time is {rwt:.6f} us')
return bnt
def LlamaRMSNorm(self):
seq_len = self.get_real_seq_len()
if self.use_cache:
seq_len = 1
else:
seq_len = self.seq_len
# pow
pow_ops = seq_len * self.hidden_size
# mean
mean_ops = seq_len * self.hidden_size * 2
# + eps
eps_ops = seq_len
# sqrt
rsqrt_ops = seq_len * self.rsqrt_inst
# x / rsqrt
norm_ops = seq_len * self.hidden_size
# * weight
mul_weight_ops = seq_len * self.hidden_size
total_compute_ops = pow_ops + mean_ops + eps_ops + rsqrt_ops + norm_ops + mul_weight_ops
print(total_compute_ops)
total_compute_time = total_compute_ops / self.compute_ablility
print(total_compute_time)
input_bytes = seq_len * self.hidden_size * self.dtype_size
weight_bytes = self.hidden_size * self.dtype_size
output_bytes = seq_len * self.hidden_size * self.dtype_size
total_read_write_bytes = input_bytes + weight_bytes + output_bytes
total_read_write_time = total_read_write_bytes / self.bandwidth
bnt = self.get_bottleneck_time("layernorm", total_compute_time, total_read_write_time)
return bnt
def QKV_MatMul(self):
# Matmul input featues = self.hidden_size, output features = self.hidden_size * 3
# compute
seq_len = self.get_real_seq_len()
total_compute_time, total_read_write_time = self.matmul(seq_len, self.hidden_size, self.hidden_size * 3)
bnt = self.get_bottleneck_time("qkv", total_compute_time, total_read_write_time)
return bnt
def ROPE(self):
"""
ignore rotate_half
"""
seq_len = self.get_real_seq_len()
""" 1. compute time """
# cos * x
cos_q_ops = seq_len * self.hidden_size
# sin * x
sin_q_ops = seq_len * self.hidden_size
# add_q
add_q_ops = seq_len + self.hidden_size
# total q
total_rope_q_ops = cos_q_ops + sin_q_ops + add_q_ops
# total k
total_rope_k_ops = total_rope_q_ops
# total
total_ops = total_rope_q_ops + total_rope_k_ops
total_compute_time = total_ops / self.compute_ablility
""" 2. read write time """
q_read_bytes = seq_len * self.hidden_size * self.dtype_size
cos_read_bypes = q_read_bytes
sin_read_bytes = cos_read_bypes
q_write_bytes = q_read_bytes
k_read_bytes = q_read_bytes
k_write_bytes = q_write_bytes
total_read_write_bytes = q_read_bytes + q_write_bytes + k_read_bytes + k_write_bytes + cos_read_bypes + sin_read_bytes
total_read_write_time = total_read_write_bytes / self.bandwidth
bnt = self.get_bottleneck_time("rope", total_compute_time, total_read_write_time)
return bnt
def attention_v1(self):
pass
def attention_v3(self):
pass
def Attention(self):
seq_len = self.get_real_seq_len()
# 1. attn weight
attn_weight_ops = self.heads * seq_len * self.head_size * self.seq_len # q len * (k len + cache_len)
# 2. div sqrt(d)
atten_weight_size = self.heads * seq_len * self.seq_len
scale_ops = atten_weight_size
# 3. softmax
max_ops = atten_weight_size
# minus max
minus_ops = atten_weight_size
# exp
exp_ops = atten_weight_size * self.exp_inst
# sum
sum_ops = atten_weight_size
# div, equal to * 1/sum(exp(x))
div_ops = atten_weight_size
# 3 * value attn_weight * v
value_ops = self.heads * seq_len * self.seq_len * self.head_size
total_ops = attn_weight_ops + scale_ops + max_ops + minus_ops + exp_ops + sum_ops + div_ops + value_ops
total_compute_time = total_ops / self.compute_ablility
# assume use flush attention, only q, k, v, caluse HBM load and store.
q_read_bytes = seq_len * self.hidden_size
k_read_bytes = self.seq_len * self.hidden_size
v_read_bytes = self.seq_len * self.hidden_size
output_write_bytes = q_read_bytes
total_read_write_bytes = q_read_bytes + k_read_bytes + v_read_bytes + output_write_bytes
total_read_write_bytes *= self.dtype_size
total_read_write_time = total_read_write_bytes / self.bandwidth
bnt = self.get_bottleneck_time("attn", total_compute_time, total_read_write_time)
return bnt
def OutputMatmul(self):
seq_len = self.get_real_seq_len()
total_compute_time, total_read_write_time = self.matmul(seq_len, self.hidden_size, self.hidden_size)
bnt = self.get_bottleneck_time("output matmul", total_compute_time, total_read_write_time)
return bnt
def MLPUp(self):
# intermediate size = 4 * self.hidden_size
seq_len = self.get_real_seq_len()
total_compute_time, total_read_write_time = self.matmul(seq_len, self.hidden_size, self.intermediate_size)
bnt = self.get_bottleneck_time("MLP up", total_compute_time, total_read_write_time)
return bnt
def MLPGate(self):
seq_len = self.get_real_seq_len()
# intermediate size = 4 * self.hidden_size
gate_ops = seq_len * self.hidden_size * self.intermediate_size
exp_ops = seq_len * self.intermediate_size * self.exp_inst
add_ops = seq_len * self.intermediate_size
div_ops = seq_len * self.intermediate_size
# mul gate
mul_ops = seq_len * self.intermediate_size
total_ops = gate_ops + exp_ops + add_ops + div_ops + mul_ops
total_compute_time = total_ops / self.compute_ablility
# read write
gate_input_read = seq_len * self.hidden_size * self.dtype_size
weight_read = self.hidden_size * self.intermediate_size * self.dtype_size
up_input_read = seq_len * self.intermediate_size
output_write = seq_len * self.intermediate_size
total_read_write_bytes = gate_input_read + weight_read + up_input_read + output_write
total_read_write_time = total_read_write_bytes / self.bandwidth
bnt = self.get_bottleneck_time("MLPGate", total_compute_time, total_read_write_time)
return bnt
def MLPDown(self):
seq_len = self.get_real_seq_len()
total_compute_time, total_read_write_time = self.matmul(seq_len, self.intermediate_size, self.hidden_size)
bnt = self.get_bottleneck_time("MLPDown", total_compute_time, total_read_write_time)
return bnt
def MLP(self):
up_bnt = self.MLPUp()
gate_bnt = self.MLPGate()
down_bnt = self.MLPDown()
return up_bnt + gate_bnt + down_bnt
def LMHead(self):
seq_len = self.get_real_seq_len()
total_compute_time, total_read_write_time = self.matmul(seq_len, self.hidden_size, self.vocab_size)
bnt = self.get_bottleneck_time("lm head", total_compute_time, total_read_write_time)
return bnt
def TotalTime(self):
in_ln_t = self.LlamaRMSNorm()
qkv_t = self.QKV_MatMul()
rope_t = self.ROPE()
self.attn_t = self.Attention()
output_t = self.OutputMatmul()
post_ln_t = self.LlamaRMSNorm()
mlp_t = self.MLP()
lm_head_t = self.LMHead()
self.per_layer_total_time = in_ln_t + qkv_t + rope_t + self.attn_t + output_t + post_ln_t + mlp_t
self.total_time = self.per_layer_total_time * self.layer_num + lm_head_t
print("attention occupied: ", self.attn_t / self.per_layer_total_time * 100, " %")
print(f"The best total time is { self.per_layer_total_time * 1000000:.6f} us for each layer, {self.layer_num} layers will use {self.total_time * 1000:.6f} ms")
if __name__ == "__main__":
import sys
config_path = sys.argv[1]
time_model = TimeFlies(model_type="llama2",
seq_len = 6000,
config_file = config_path,
machine="H800",
dtype="float16",
use_cache=True,
attention_version="v1")
time_model.TotalTime()