Katz/plot_end2end_latency.py at main · modelscope/Katz · 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
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
import os

import numpy as np

import matplotlib
import matplotlib.pyplot as plt
matplotlib.rcParams['pdf.fonttype'] = 42
matplotlib.rcParams['ps.fonttype'] = 42

one_controlnet_switch_latency = 0.28

def get_latency(foldername):
    latencies = []
    with open(f"{foldername}/sdxl_pipeline.log", "r") as f:
        for line in f:
            if "End2End inference latency:" in line:
                latencies.append(float(line.split(":")[-1].strip()))
    latencies = latencies[1:-1]
    return latencies

foldernames = {
    "Diffusers": {
        "0C/0L": "./logs/diffusers-0c-0l",
        "1C/0L": "./logs/diffusers-1c-0l",
        "0C/1L": "./logs/diffusers-0c-1l",
        "1C/1L": "./logs/diffusers-1c-1l",
        "3C/0L": "./logs/diffusers-3c-0l",
        "2C/2L": "./logs/diffusers-2c-2l",
        "3C/2L": "./logs/diffusers-3c-2l",
    },
    "Nirvana-10": {
        "0C/0L": "./logs/nirvana-0c-0l-skip10",
        "1C/0L": "./logs/nirvana-1c-0l-skip10",
        "0C/1L": "./logs/nirvana-0c-1l-skip10",
        "1C/1L": "./logs/nirvana-1c-1l-skip10",
        "3C/0L": "./logs/nirvana-3c-0l-skip10",
        "2C/2L": "./logs/nirvana-2c-2l-skip10",
        "3C/2L": "./logs/nirvana-3c-2l-skip10",
    },
    "Nirvana-20": {
        "0C/0L": "./logs/nirvana-0c-0l-skip20",
        "1C/0L": "./logs/nirvana-1c-0l-skip20",
        "0C/1L": "./logs/nirvana-0c-1l-skip20",
        "1C/1L": "./logs/nirvana-1c-1l-skip20",
        "3C/0L": "./logs/nirvana-3c-0l-skip20",
        "2C/2L": "./logs/nirvana-2c-2l-skip20",
        "3C/2L": "./logs/nirvana-3c-2l-skip20",
    },
    "DistriFusion": {
        "0C/0L": 0.0,
        "1C/0L":0.0,
        "0C/1L": 0.0,
        "1C/1L": 0.0,
        "3C/0L": 0.0,
        "2C/2L": 0.0,
        "3C/2L": 0.0,
    },
    "Katz": {
        "0C/0L": "./logs/katz-0c-0l",
        "1C/0L": "./logs/katz-1c-0l",
        "0C/1L": "./logs/katz-0c-1l",
        "1C/1L": "./logs/katz-1c-1l",
        "3C/0L": "./logs/katz-3c-0l",
        "2C/2L": "./logs/katz-2c-2l",
        "3C/2L": "./logs/katz-3c-2l",
    },
}

latencies = {}
for baseline, foldername in foldernames.items():
    if baseline not in latencies:
        latencies[baseline] = {}
    if baseline != "DistriFusion":
        for setting, folder in foldername.items():
            if setting not in latencies[baseline]:
                latencies[baseline][setting] = 0.0
            latencies[baseline][setting] = np.mean(get_latency(folder))
    else:
        # parse the latencies from the DistriFusion
        distriFusionFolder = "./logs/distrifuser_benchmark_logs"
        for setting in foldernames[baseline]:
            # print(f"Setting: {setting}")
            setting_filename = setting.replace("/", "_")
            with open(f"{distriFusionFolder}/distrifuser_{setting_filename}.log", "r") as fr:
                # print(f"{distriFusionFolder}/distrifuser_{ setting_filename }.log")
                end2end_inference_times = []
                for line in fr:
                    if "End2End inference latency" in line:
                        end2end_inference_times.append(float(line.split(":")[-1]))

                latencies[baseline][setting] = np.mean(end2end_inference_times[:-1])

fig, ax = plt.subplots(figsize=(14, 5))
fontsize = 32

barWidth = 0.18
bars = []
hatches = ['-', '\\', "/", "o", "x"]
# matplotlib default colors
bar_colors = ["tab:blue", "tab:orange", "tab:green", "tab:purple", "tab:red"]

for i, (baseline, latencies) in enumerate(latencies.items()):
    x = np.arange(len(latencies))
    plot_latencies = np.array(list(latencies.values()))

    num_controlnets = np.array( [ int(item.split("/")[0][0]) for item in list(latencies.keys()) ] )
    controlnet_switch_latency = num_controlnets * one_controlnet_switch_latency
    if "Katz" not in baseline:
        plot_latencies += controlnet_switch_latency
    if "Nirvana" in baseline:
        plot_latencies += 0.35  #

    print(baseline, plot_latencies)
    bars.append(ax.bar(x + i * barWidth - 1.5*barWidth, plot_latencies, barWidth, \
                       label=baseline, hatch=hatches[i], color=bar_colors[i] , linewidth=1.5))

    # add bar height
    if baseline == "Diffusers" or baseline == "Katz":
        for j, latency in enumerate(plot_latencies):
            ax.text(x[j] + 1.05 * i * barWidth - 1.5*barWidth, latency + 0.3, f"{latency:.1f}", fontsize=fontsize-12, ha='center')

ax.grid(axis='y')
ax.set_ylim([1, 17.8])

ax.set_xlabel("Adapter settings", fontsize=fontsize)
ax.set_ylabel('Latency (s)', fontsize=fontsize)
ax.set_xticks(np.arange(len(latencies)))
ax.set_xticklabels(latencies.keys(), fontsize=fontsize)
# set y tick labels
ax.yaxis.set_ticks([2,5,8,11,14,17])

ax.tick_params(axis='y', labelsize=fontsize)

ax.legend(loc='upper left', fontsize=fontsize, frameon=False, borderaxespad=0.1, labelspacing=0.1, ncol=2, columnspacing=0.2, handletextpad=0.1, handlelength=2)
fig.tight_layout()

figure_path = "./figures"
if not os.path.exists(figure_path):
    os.makedirs(figure_path)
latency_figure = f"{figure_path}/end2end_latency.pdf"
plt.savefig(latency_figure, format='pdf', bbox_inches='tight', pad_inches=0.03)
print(f"Saved figure to {latency_figure}")