transformersprofiling/parallel_batch_decoding.py at main · YangZhou08/transformersprofiling · GitHub

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
import numpy as np
import matplotlib.pyplot as plt

import matplotlib as mpl

mpl.rcParams['xtick.labelsize'] = 22
mpl.rcParams['ytick.labelsize'] = 22
mpl.rcParams['axes.titlesize'] = 22
mpl.rcParams['axes.labelsize'] = 22
mpl.rcParams['legend.fontsize'] = 22

x = list(range(1, 21, 2))
# latency_measurement = [16.540902376174927, 19.05730414390564, 19.612385272979736, 20.869054794311523, 20.67660117149353, 21.932927131652832, 22.34955334663391, 22.983564138412476, 23.834572792053223, 24.18628692626953]
latency_measurement = [9.712086200714111, 10.404566049575806, 10.668877124786377, 10.944726943969727, 11.201996803283691, 11.497334480285645, 11.744471311569214, 12.02213454246521, 12.955510139465332, 12.561285972595215]
latency_bf16 = [9.770136833190918, 10.448176383972168, 10.71483826637268, 10.992271661758423, 11.24536418914795, 11.538933515548706, 11.78537654876709, 12.074774265289307, 12.992205142974854, 12.607782125473022]
x_ticks = [str(i) for i in x]

fig, ax = plt.subplots()
bar_width = 0.6

# bars = ax.bar([i - bar_width/2 for i in x], latency_measurement, color = "blue", width = 0.6, label = "FP32")
bars = ax.bar([i - bar_width/2 for i in x], latency_bf16, color = "orange", width = 0.6, label = "BF16")
# bar2 = ax.bar([i + bar_width/2 for i in x], latency_bf16, color = "orange", width = 0.6, label = "BF16")
bar2 = ax.bar([i + bar_width/2 for i in x], latency_measurement, color = "blue", width = 0.6, label = "FP16")
for i in range(len(bars)):
    bar = bars[i]
    yval = bar.get_height()
    plt.text(bar.get_x() + bar.get_width()/2.0, yval, "{:.2f}".format(yval), va='bottom', ha='center')
    b2 = bar2[i]
    yval2 = b2.get_height()
    plt.text(b2.get_x() + b2.get_width()/2.0, yval2, "{:.2f}".format(yval2), va='bottom', ha='center')

ax.set_xticks(x, x_ticks)
ax.set_xlabel("Batch Size")
ax.set_ylabel("Latency (s)")
ax.legend()
ax.grid()

plt.show()