File size: 29,607 Bytes
3698d0a
dd4f101
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3698d0a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dd4f101
3698d0a
dd4f101
 
 
 
 
 
 
3698d0a
 
dd4f101
 
3698d0a
 
 
dd4f101
 
 
 
3698d0a
 
 
 
dd4f101
 
 
 
 
 
 
 
 
 
 
 
3698d0a
5f0df3a
3698d0a
 
 
dd4f101
 
 
 
 
 
 
 
 
 
 
 
3698d0a
5f0df3a
3698d0a
 
 
dd4f101
 
 
 
 
 
 
 
 
 
 
 
3698d0a
5f0df3a
3698d0a
 
 
 
5f0df3a
 
dd4f101
3698d0a
 
 
 
 
 
 
 
 
5f0df3a
3698d0a
 
 
dd4f101
 
 
 
 
 
 
 
 
 
 
 
3698d0a
 
 
 
 
 
 
dd4f101
 
 
 
 
 
 
3698d0a
 
 
 
dd4f101
 
 
 
 
 
 
 
 
 
 
 
 
3698d0a
 
dd4f101
3698d0a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dd4f101
3698d0a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dd4f101
3698d0a
 
 
 
 
 
 
 
 
 
 
dd4f101
3698d0a
 
 
 
 
 
 
 
dd4f101
3698d0a
 
 
 
 
 
 
5f0df3a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ed50ee5
 
5f0df3a
 
 
 
 
 
 
 
 
 
dd4f101
 
 
5f0df3a
 
 
 
 
 
 
 
 
 
 
 
 
ed50ee5
5f0df3a
 
 
 
dd4f101
5f0df3a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ed50ee5
 
 
 
 
 
 
 
 
 
dd4f101
ed50ee5
 
 
 
 
 
 
 
 
 
 
dd4f101
ed50ee5
 
 
 
 
 
 
 
dd4f101
ed50ee5
 
 
5f0df3a
 
ed50ee5
 
5f0df3a
 
 
 
 
 
 
 
c93009d
 
 
 
 
 
 
 
 
 
 
 
 
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
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
import numpy as np
from collections import defaultdict
from functools import partial
from typing import List
from model_util import get_module_tensors_matched

def calc_model_size_from_model(model_config, inference_config):
    get_module_tensors_matched_partial = partial(get_module_tensors_matched, module_classes_dict = model_config['module_classes'])

    parameter_count = defaultdict(float)
    parameter_count['word_embedding'] = sum([v.numel() for v in get_module_tensors_matched_partial(lambda x: 'embed' in x and 'pos' not in x)])
    parameter_count['positional_embedding'] = sum([v.numel() for v in get_module_tensors_matched_partial(lambda x: 'embed' in x and 'pos' in x)])

    parameter_count['attention_Q'] = sum([v.numel() for v in get_module_tensors_matched_partial(lambda x: 'att' in x and 'q' in x)])
    parameter_count['attention_K'] = sum([v.numel() for v in get_module_tensors_matched_partial(lambda x: 'att' in x and 'k' in x)])
    parameter_count['attention_V'] = sum([v.numel() for v in get_module_tensors_matched_partial(lambda x: 'att' in x and 'v' in x)])
    parameter_count['attention_out'] = sum([v.numel() for v in get_module_tensors_matched_partial(lambda x: 'att' in x and ('out_' in x or 'o_' in x))])
    
    parameter_count['layernorm'] = sum([v.numel() for v in get_module_tensors_matched_partial(lambda x: 'norm' in x)])
    parameter_count['mlp_weights'] = sum([v.numel() for v in get_module_tensors_matched_partial(lambda x: 'fc' in x or 'mlp' in x)])
       
    parameter_count['embedding_weights'] = parameter_count['word_embedding'] + parameter_count['positional_embedding']
    parameter_count['attention_weights'] = parameter_count['attention_out'] + parameter_count['attention_Q'] + parameter_count['attention_K'] + parameter_count['attention_V']
        
    return parameter_count

def model_size_estimate(model_config, inference_config):
    parameter_count = {}
    parameter_count['word_embedding'] = model_config['vocab_size']*model_config['hidden_size']
    parameter_count['positional_embedding'] = model_config['max_position_embeddings']*model_config['hidden_size']
    
    parameter_count['attention_Q']   = model_config['num_hidden_layers']*model_config['hidden_size']*model_config['hidden_size']/model_config['num_attention_heads']*model_config['num_attention_heads']
    parameter_count['attention_K']   = model_config['num_hidden_layers']*model_config['hidden_size']*model_config['hidden_size']/model_config['num_attention_heads']*model_config['num_attention_heads']
    parameter_count['attention_V']   = model_config['num_hidden_layers']*model_config['hidden_size']*model_config['hidden_size']/model_config['num_attention_heads']*model_config['num_attention_heads']
    parameter_count['attention_out'] = model_config['num_hidden_layers']*model_config['hidden_size']*model_config['hidden_size']/model_config['num_attention_heads']*model_config['num_attention_heads']
    
    parameter_count['layernorm'] = 2*model_config['layernorm_operation']*model_config['num_hidden_layers']*model_config['hidden_size']
    parameter_count['mlp1'] = model_config['num_hidden_layers']*model_config['hidden_size']*model_config['intermediate_size']
    parameter_count['mlp2'] = model_config['num_hidden_layers']*model_config['hidden_size']*model_config['intermediate_size']
    parameter_count['embedding_weights'] = parameter_count['word_embedding'] + parameter_count['positional_embedding']
    parameter_count['attention_weights'] = parameter_count['attention_out'] + parameter_count['attention_Q'] + parameter_count['attention_K'] + parameter_count['attention_V']
    parameter_count['mlp_weights'] = parameter_count['mlp1'] + parameter_count['mlp2']
    
    return parameter_count

def multiplication_in_int64(array):
    return np.cumprod(np.array(array, dtype=np.int64))[-1]

def matrix_operation(shapeA, shapeB):
    assert(shapeA[-1] == shapeB[0])
    op = np.cumprod(np.array(shapeA[:-1], np.float64))    
    return multiplication_in_int64([2, op[-1], shapeA[-1], shapeB[-1]])

def word_embedding_operation(model_config, inference_config):
    #Given:
    #\begin{itemize}
    #    \item Matrix \( X \) of size \( B \times s \) (representing the batch size and sequence length respectively).
    #    \item Embedding matrix \( W_e \) of size \( n_{vocab} \times d_{model} \).
    #\end{itemize}
    
    #The resultant matrix after the multiplication will be of size \( B \times s \times d_{model} \).
    #For each element in this resultant matrix, the number of FLOPs required is \( 2 \times n_{vocab} \). This is because for a single element in the output matrix, we have \( 2N \) FLOPs (with \( N \) being the common dimension), leading to the matrix multiplication FLOP count as:
    #\begin{equation}
    #2 \times B \times s \times n_{v ocab} \times d_{model}
    #\end{equation}
    if model_config['module_classes']:
        modules = get_module_tensors_matched(lambda x: 'embed' in x and 'pos' not in x, model_config['module_classes'])
        A = [inference_config['batchsize'], inference_config['input_seq_length'], modules[0][0]]
        B = modules[0]
        op_count = matrix_operation(A, B)
        return op_count

    A = [inference_config['batchsize'], inference_config['input_seq_length'], model_config['vocab_size']]
    B = [model_config['vocab_size'], model_config['hidden_size']]
    op_count = matrix_operation(A, B)
    return op_count


def positional_embedding_operation(model_config, inference_config):
    if model_config['module_classes']:
        modules = get_module_tensors_matched(lambda x: 'embed' in x and 'pos' in x, model_config['module_classes'])
        return multiplication_in_int64([inference_config['batchsize'], inference_config['input_seq_length'], modules[0][-1]])

    return multiplication_in_int64([inference_config['batchsize'], inference_config['input_seq_length'], model_config['hidden_size']])

### Below three are the same
def attention_K_operation(model_config, inference_config, seq_length):
    if model_config['module_classes']:
        modules = get_module_tensors_matched(lambda x: 'att' in x and 'k' in x , model_config['module_classes'])
        total = 0
        for module in modules:
            if len(module) > 1:
                A = [inference_config['batchsize'], seq_length, model_config['hidden_size']]
                B = [model_config['hidden_size'], model_config['hidden_size_per_head']]
                total += model_config['num_attention_heads']*matrix_operation(A, B)
            else:
                total += model_config['hidden_size']
        return total

    A = [inference_config['batchsize'], seq_length, model_config['hidden_size']]
    B = [model_config['hidden_size'], model_config['hidden_size_per_head']]
    return model_config['num_hidden_layers'] * model_config['num_attention_heads'] * matrix_operation(A, B)

def attention_Q_operation(model_config, inference_config, seq_length):
    if model_config['module_classes']:
        modules = get_module_tensors_matched(lambda x: 'att' in x and 'q' in x , model_config['module_classes'])
        total = 0
        for module in modules:
            if len(module) > 1:
                A = [inference_config['batchsize'], seq_length, model_config['hidden_size']]
                B = [model_config['hidden_size'], model_config['hidden_size_per_head']]
                total += model_config['num_attention_heads']*matrix_operation(A, B)
            else:
                total += model_config['hidden_size']
        return total
     
    A = [inference_config['batchsize'], seq_length, model_config['hidden_size']]
    B = [model_config['hidden_size'], model_config['hidden_size_per_head']]
    return model_config['num_hidden_layers'] * model_config['num_attention_heads'] * matrix_operation(A, B)

def attention_V_operation(model_config, inference_config, seq_length):
    if model_config['module_classes']:
        modules = get_module_tensors_matched(lambda x: 'att' in x and 'v' in x , model_config['module_classes'])
        total = 0
        for module in modules:
            if len(module) > 1:
                A = [inference_config['batchsize'], seq_length, model_config['hidden_size']]
                B = [model_config['hidden_size'], model_config['hidden_size_per_head']]
                total += model_config['num_attention_heads']*matrix_operation(A, B)
            else:
                total += model_config['hidden_size']
        return total
     
    A = [inference_config['batchsize'], seq_length, model_config['hidden_size']]
    B = [model_config['hidden_size'], model_config['hidden_size_per_head']]
    return model_config['num_hidden_layers'] * model_config['num_attention_heads'] * matrix_operation(A, B)

## 
def attention_QK_operation(model_config, inference_config, seq_length_Q, seq_length_K):
    A = [inference_config['batchsize'], seq_length_Q, model_config['hidden_size_per_head']]
    B = [model_config['hidden_size_per_head'], seq_length_K]
    return model_config['num_hidden_layers'] * model_config['num_attention_heads'] * matrix_operation(A, B)

def attention_softmax_operation(model_config, inference_config,seq_length):
    # Ref: Ouyang, A. (2023). Understanding the Performance of Transformer Inference (Doctoral dissertation, Massachusetts Institute of Technology).
    # 3 is a modeled value
    softmax_operation = (3*inference_config['batchsize']*seq_length*seq_length)
    return model_config['num_hidden_layers'] * model_config['num_attention_heads'] * softmax_operation

def attention_multV_operation(model_config, inference_config, seq_length_Q, seq_length_V):
    A = [inference_config['batchsize'], seq_length_Q, seq_length_V]
    B = [seq_length_V, model_config['hidden_size_per_head']]
    return model_config['num_hidden_layers'] * model_config['num_attention_heads']* matrix_operation(A, B)

def attention_out_operation(model_config, inference_config, seq_length):
    if model_config['module_classes']:
        modules = get_module_tensors_matched(lambda x: 'att' in x and 'k' in x , model_config['module_classes'])
        total = 0
        for module in modules:
            if len(module) > 1:
                A = [inference_config['batchsize'], seq_length, model_config['hidden_size']]
                B = [model_config['hidden_size'], model_config['hidden_size']]
                total += matrix_operation(A, B)
            else:
                total += model_config['hidden_size']
        return total
    
    A = [inference_config['batchsize'], seq_length, model_config['hidden_size']]
    B = [model_config['hidden_size'], model_config['hidden_size']]
    return model_config['num_hidden_layers'] * matrix_operation(A, B)

def layernorm_operation(model_config, inference_config, seq_length):
    # Ref: Ouyang, A. (2023). Understanding the Performance of Transformer Inference (Doctoral dissertation, Massachusetts Institute of Technology).
    # 5 is a modeled value
    if model_config['module_classes']:
        modules = get_module_tensors_matched(lambda x: 'norm' in x, model_config['module_classes'])
        total = 0
        for module in modules:
            total += model_config['hidden_size']
        return 5*total
    
    layernorm_operation = (5*inference_config['batchsize']*seq_length*model_config['hidden_size'])
    return model_config['num_hidden_layers'] * model_config['layernorm_operation'] * layernorm_operation


def mlp_operation(model_config, inference_config, seq_length):
    if model_config['module_classes']:
        modules = get_module_tensors_matched(lambda x: 'fc' in x or 'mlp' in x, model_config['module_classes'])
        total = 0
        for module in modules:
            if len(module) > 1:
                A = [inference_config['batchsize'], seq_length, module[1]]
                B = [module[1], module[0]]
                total += matrix_operation(A, B)
            else:
                total += modules[-1][0]
        return total
    
    A = [inference_config['batchsize'], seq_length, model_config['hidden_size']]
    B = [model_config['hidden_size'], model_config['intermediate_size']]
    return model_config['num_hidden_layers'] * (2*matrix_operation(A, B))


def prefilling_operation(model_config, inference_config):
    prefilling_operation_count = {}
    prefilling_operation_count['word_embedding'] = word_embedding_operation(model_config, inference_config)
    prefilling_operation_count['positional_embedding'] = positional_embedding_operation(model_config, inference_config)
    
    prefilling_operation_count['attention_Q'] = attention_Q_operation(model_config, inference_config, inference_config['input_seq_length'])
    prefilling_operation_count['attention_K'] = attention_K_operation(model_config, inference_config, inference_config['input_seq_length'])
    prefilling_operation_count['attention_V'] = attention_V_operation(model_config, inference_config, inference_config['input_seq_length'])
    prefilling_operation_count['attention_QK'] = attention_QK_operation(model_config, inference_config, inference_config['input_seq_length'], inference_config['input_seq_length'])
    prefilling_operation_count['attention_softmax'] = attention_softmax_operation(model_config, inference_config, inference_config['input_seq_length'])
    prefilling_operation_count['attention_multV'] = attention_multV_operation(model_config, inference_config, inference_config['input_seq_length'], inference_config['input_seq_length'])
    prefilling_operation_count['attention_out'] = attention_out_operation(model_config, inference_config, inference_config['input_seq_length'])

    prefilling_operation_count['layernorm'] =layernorm_operation(model_config, inference_config, inference_config['input_seq_length'])

    prefilling_operation_count['mlp'] = mlp_operation(model_config, inference_config, inference_config['input_seq_length'])
    
    prefilling_operation_count['embeddings'] = prefilling_operation_count['word_embedding'] + prefilling_operation_count['positional_embedding']
    prefilling_operation_count['attention'] = sum([v for k,v in prefilling_operation_count.items() if 'attention' in k])
    prefilling_operation_count['total'] = (prefilling_operation_count['embeddings'] + prefilling_operation_count['attention'] + prefilling_operation_count['mlp'] + prefilling_operation_count['layernorm'])
    
    return prefilling_operation_count

def generation_operation(model_config, inference_config):
    generation_operation_count = {}
    generation_operation_count['word_embedding'] = 0
    generation_operation_count['positional_embedding'] = 0
    generation_operation_count['attention_K'] = 0
    generation_operation_count['attention_V'] = 0
    generation_operation_count['attention_Q'] = 0
    generation_operation_count['attention_QK'] = 0
    generation_operation_count['attention_softmax'] = 0
    generation_operation_count['attention_multV'] = 0
    generation_operation_count['attention_out'] = 0
    generation_operation_count['mlp'] = 0
    generation_operation_count['layernorm'] = 0

    for t in range(inference_config['output_seq_length']):
        if inference_config['KV_cache']:
            generation_operation_count['attention_K'] += attention_K_operation(model_config, inference_config, 1)
            generation_operation_count['attention_V'] += attention_V_operation(model_config, inference_config, 1)
            generation_operation_count['attention_Q'] += attention_Q_operation(model_config, inference_config, 1)
            generation_operation_count['attention_QK'] += attention_QK_operation(model_config, inference_config, seq_length_Q=1, seq_length_K=(t+1)+inference_config['input_seq_length'])
            generation_operation_count['attention_softmax'] += attention_softmax_operation(model_config, inference_config, 1)
            generation_operation_count['attention_multV'] += attention_multV_operation(model_config, inference_config, seq_length_Q=1, seq_length_V=(t+1)+inference_config['input_seq_length'])
            generation_operation_count['attention_out'] += attention_out_operation(model_config, inference_config, 1)
            generation_operation_count['mlp'] += mlp_operation(model_config, inference_config, 1)
        else:
            generation_operation_count['attention_K'] += attention_K_operation(model_config, inference_config, (t+1)+inference_config['input_seq_length'])
            generation_operation_count['attention_V'] += attention_V_operation(model_config, inference_config, (t+1)+inference_config['input_seq_length'])
            generation_operation_count['attention_Q'] += attention_Q_operation(model_config, inference_config, (t+1)+inference_config['input_seq_length'])
            generation_operation_count['attention_QK'] += attention_QK_operation(model_config, inference_config, seq_length_Q=(t+1)+inference_config['input_seq_length'], seq_length_K=(t+1)+inference_config['input_seq_length'])
            generation_operation_count['attention_softmax'] += attention_softmax_operation(model_config, inference_config, (t+1)+inference_config['input_seq_length'])
            generation_operation_count['attention_multV'] += attention_multV_operation(model_config, inference_config, seq_length_Q=(t+1)+inference_config['input_seq_length'], seq_length_V=(t+1)+inference_config['input_seq_length'])
            generation_operation_count['attention_out'] += attention_out_operation(model_config, inference_config, (t+1)+inference_config['input_seq_length'])
            generation_operation_count['mlp'] += mlp_operation(model_config, inference_config, (t+1)+inference_config['input_seq_length'])

        generation_operation_count['layernorm'] += layernorm_operation(model_config, inference_config, (t+1)+inference_config['input_seq_length'])

    generation_operation_count['embeddings'] = generation_operation_count['word_embedding'] + generation_operation_count['positional_embedding'] 
    generation_operation_count['attention'] = sum([v for k,v in generation_operation_count.items() if 'attention' in k])
    generation_operation_count['total'] = (generation_operation_count['attention'] + generation_operation_count['mlp'] + generation_operation_count['layernorm'])

    return generation_operation_count


def word_embedding_activation_memory(model_config, inference_config, seq_length):
    return inference_config['batchsize'] * seq_length * (model_config['vocab_size'] + model_config['hidden_size'])

def positional_embedding_activation_memory(model_config, inference_config, seq_length):
    return 2 * inference_config['batchsize'] * seq_length * model_config['hidden_size']

def attention_K_activation_memory(model_config, inference_config, seq_length):
    per_head_per_layer = inference_config['batchsize'] * seq_length * (model_config['hidden_size'] + model_config['hidden_size_per_head'])
    return model_config['num_hidden_layers'] * model_config['num_attention_heads'] * per_head_per_layer

def attention_V_activation_memory(model_config, inference_config, seq_length):
    per_head_per_layer = inference_config['batchsize'] * seq_length * (model_config['hidden_size'] + model_config['hidden_size_per_head'])
    return model_config['num_hidden_layers'] * model_config['num_attention_heads'] * per_head_per_layer

def attention_Q_activation_memory(model_config, inference_config, seq_length):
    per_head_per_layer = inference_config['batchsize'] * seq_length * (model_config['hidden_size'] + model_config['hidden_size_per_head'])
    return model_config['num_hidden_layers'] * model_config['num_attention_heads'] * per_head_per_layer

def attention_QK_activation_memory(model_config, inference_config, seq_length_Q, seq_length_K):
    inputs_Q = inference_config['batchsize'] * seq_length_Q * model_config['hidden_size_per_head']
    inputs_K = inference_config['batchsize'] * seq_length_K * model_config['hidden_size_per_head']
    outputs =  inference_config['batchsize'] * seq_length_Q * seq_length_K
    return model_config['num_hidden_layers'] * model_config['num_attention_heads'] * (inputs_Q + inputs_K + outputs)

def attention_softmax_activation_memory(model_config, inference_config, seq_length):
    per_head_per_layer = (2 * inference_config['batchsize'] * seq_length * seq_length)
    return model_config['num_hidden_layers'] * model_config['num_attention_heads'] * per_head_per_layer

def attention_multV_activation_memory(model_config, inference_config, seq_length_Q, seq_length_V):
    per_head_per_layer = inference_config['batchsize'] * seq_length_Q * seq_length_V + inference_config['batchsize'] * seq_length_Q * model_config['hidden_size_per_head'] + inference_config['batchsize'] * seq_length_V * model_config['hidden_size_per_head']
    return model_config['num_hidden_layers'] * model_config['num_attention_heads'] * per_head_per_layer

def attention_out_activation_memory(model_config, inference_config, seq_length):
    per_head_per_layer = 2 * inference_config['batchsize'] * seq_length * model_config['hidden_size']
    return model_config['num_hidden_layers'] * model_config['num_attention_heads'] * per_head_per_layer

def layernorm_activation_memory(model_config, inference_config, seq_length):
    per_layernorm_per_layer = 2 * inference_config['batchsize'] * seq_length * model_config['hidden_size']
    return model_config['num_hidden_layers'] * model_config['layernorm_operation'] * per_layernorm_per_layer

def mlp_activation_memory(model_config, inference_config, seq_length):
    # two mlp layer
    per_layer = 2 * inference_config['batchsize'] * seq_length * (model_config['hidden_size'] + model_config['intermediate_size'])
    return model_config['num_hidden_layers'] * per_layer

def prefilling_activation_memory(model_config, inference_config):
    activation_memory = {}
    
    activation_memory['word_embedding'] = word_embedding_activation_memory(model_config, inference_config, inference_config['input_seq_length'])
    activation_memory['positional_embedding'] = positional_embedding_activation_memory(model_config, inference_config, inference_config['input_seq_length'])
    
    activation_memory['attention_Q'] = attention_Q_activation_memory(model_config, inference_config, inference_config['input_seq_length'])
    activation_memory['attention_K'] = attention_K_activation_memory(model_config, inference_config, inference_config['input_seq_length'])
    activation_memory['attention_V'] = attention_V_activation_memory(model_config, inference_config, inference_config['input_seq_length'])
    activation_memory['attention_QK'] = attention_QK_activation_memory(model_config, inference_config, inference_config['input_seq_length'], inference_config['input_seq_length'])
    activation_memory['attention_softmax'] = attention_softmax_activation_memory(model_config, inference_config, inference_config['input_seq_length'])
    activation_memory['attention_multV'] = attention_multV_activation_memory(model_config, inference_config, inference_config['input_seq_length'], inference_config['input_seq_length'])
    activation_memory['attention_out'] = attention_out_activation_memory(model_config, inference_config, inference_config['input_seq_length'])

    activation_memory['layernorm'] = layernorm_activation_memory(model_config, inference_config, inference_config['input_seq_length'])

    activation_memory['mlp'] = mlp_activation_memory(model_config, inference_config, inference_config['input_seq_length'])
    
    activation_memory['embeddings'] = activation_memory['word_embedding'] + activation_memory['positional_embedding']
    activation_memory['attention'] = (
        activation_memory['attention_Q'] + activation_memory['attention_K'] +
        activation_memory['attention_V'] + activation_memory['attention_QK'] +
        activation_memory['attention_softmax'] + activation_memory['attention_multV'] +
        activation_memory['attention_out']
    )
    activation_memory['total'] = (
        activation_memory['embeddings'] + activation_memory['attention'] +
        activation_memory['mlp'] + activation_memory['layernorm']
    )
    
    activation_memory['embeddings'] = activation_memory['word_embedding'] + activation_memory['positional_embedding'] 
    activation_memory['attention'] = sum([v for k,v in activation_memory.items() if 'attention' in k])
    activation_memory['total'] = (activation_memory['attention'] + activation_memory['mlp'] + activation_memory['layernorm'])

    return activation_memory

def generation_activation_memory(model_config, inference_config):
    activation_memory = {}

    activation_memory['word_embedding'] = 0
    activation_memory['positional_embedding'] = 0
    activation_memory['attention_K'] = 0
    activation_memory['attention_V'] = 0
    activation_memory['attention_Q'] = 0
    activation_memory['attention_QK'] = 0
    activation_memory['attention_softmax'] = 0
    activation_memory['attention_multV'] = 0
    activation_memory['attention_out'] = 0
    activation_memory['mlp'] = 0
    activation_memory['layernorm'] = 0

    for t in range(inference_config['output_seq_length']):
        if inference_config['KV_cache']:
            activation_memory['attention_K'] += attention_K_activation_memory(model_config, inference_config, 1)
            activation_memory['attention_V'] += attention_V_activation_memory(model_config, inference_config, 1)
            activation_memory['attention_Q'] += attention_Q_activation_memory(model_config, inference_config, 1)
            activation_memory['attention_QK'] += attention_QK_activation_memory(model_config, inference_config, seq_length_Q=1, seq_length_K=(t+1)+inference_config['input_seq_length'])
            activation_memory['attention_softmax'] += attention_softmax_activation_memory(model_config, inference_config, 1)
            activation_memory['attention_multV'] += attention_multV_activation_memory(model_config, inference_config, seq_length_Q=1, seq_length_V=(t+1)+inference_config['input_seq_length'])
            activation_memory['attention_out'] += attention_out_activation_memory(model_config, inference_config, 1)
            activation_memory['mlp'] += mlp_activation_memory(model_config, inference_config, 1)
        else:
            activation_memory['attention_K'] += attention_K_activation_memory(model_config, inference_config, (t+1)+inference_config['input_seq_length'])
            activation_memory['attention_V'] += attention_V_activation_memory(model_config, inference_config, (t+1)+inference_config['input_seq_length'])
            activation_memory['attention_Q'] += attention_Q_activation_memory(model_config, inference_config, (t+1)+inference_config['input_seq_length'])
            activation_memory['attention_QK'] += attention_QK_activation_memory(model_config, inference_config, seq_length_Q=(t+1)+inference_config['input_seq_length'], seq_length_K=(t+1)+inference_config['input_seq_length'])
            activation_memory['attention_softmax'] += attention_softmax_activation_memory(model_config, inference_config, (t+1)+inference_config['input_seq_length'])
            activation_memory['attention_multV'] += attention_multV_activation_memory(model_config, inference_config, seq_length_Q=(t+1)+inference_config['input_seq_length'], seq_length_V=(t+1)+inference_config['input_seq_length'])
            activation_memory['attention_out'] += attention_out_activation_memory(model_config, inference_config, (t+1)+inference_config['input_seq_length'])
            activation_memory['mlp'] += mlp_activation_memory(model_config, inference_config, (t+1)+inference_config['input_seq_length'])

        activation_memory['layernorm'] += layernorm_activation_memory(model_config, inference_config, (t+1)+inference_config['input_seq_length'])

    activation_memory['embeddings'] = activation_memory['word_embedding'] + activation_memory['positional_embedding']
    activation_memory['attention'] = (
        activation_memory['attention_K'] + activation_memory['attention_V'] +
        activation_memory['attention_Q'] + activation_memory['attention_QK'] +
        activation_memory['attention_softmax'] + activation_memory['attention_multV'] +
        activation_memory['attention_out']
    )
    activation_memory['total'] = (
        activation_memory['embeddings'] + activation_memory['attention'] +
        activation_memory['mlp'] + activation_memory['layernorm']
    )

    return activation_memory


def calc_prefilling_throughput(model_config, inference_config, inference_info):
    inference_info['prefilling_throughput'] = inference_config['input_seq_length']*inference_config['batchsize'] / max([inference_info['inference_prefilling_time'], inference_info['prefilling_memory_latency']])
    inference_info['prefilling_bound_type'] = "memory" if inference_info['inference_prefilling_time'] < inference_info['prefilling_memory_latency'] else "arithmetic"

def calc_generation_throughput(model_config, inference_config, inference_info):
    inference_info['generation_throughput'] = inference_config['input_seq_length']*inference_config['batchsize'] / max([inference_info['inference_generation_time'], inference_info['generation_memory_latency']])
    inference_info['generation_bound_type'] = "memory" if inference_info['inference_generation_time'] < inference_info['generation_memory_latency'] else "arithmetic"
    
    total_time = max([inference_info['inference_prefilling_time'], inference_info['prefilling_memory_latency']]) + max([inference_info['inference_generation_time'], inference_info['generation_memory_latency']])
    inference_info['client_generation_throughput'] = inference_config['output_seq_length']*inference_config['batchsize'] / total_time