File size: 8,977 Bytes
7cd710e
 
 
 
 
 
 
 
 
 
 
 
 
 
39b15f4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7cd710e
 
 
 
39b15f4
 
 
 
7cd710e
 
 
39b15f4
 
 
 
 
7cd710e
 
 
 
39b15f4
7cd710e
 
 
 
39b15f4
7cd710e
 
 
 
 
39b15f4
7cd710e
 
39b15f4
 
 
 
 
 
 
 
 
 
7cd710e
0910acf
39b15f4
7cd710e
39b15f4
 
 
7cd710e
 
39b15f4
 
 
 
 
 
 
 
 
 
 
7cd710e
 
39b15f4
 
 
 
 
 
 
 
7cd710e
39b15f4
 
 
7cd710e
 
39b15f4
 
 
 
 
 
 
 
 
 
 
7cd710e
 
39b15f4
 
 
 
 
 
 
 
7cd710e
39b15f4
 
 
 
 
 
 
 
 
 
 
 
 
 
7cd710e
39b15f4
 
 
 
 
 
0910acf
39b15f4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7cd710e
39b15f4
 
 
 
 
 
 
 
 
 
 
 
7cd710e
 
39b15f4
 
 
 
 
 
 
 
 
 
 
7cd710e
 
39b15f4
 
 
 
 
 
0910acf
7cd710e
 
 
39b15f4
 
7cd710e
39b15f4
 
 
7cd710e
 
 
 
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
import matplotlib
matplotlib.use('Agg')

import functools
import gradio as gr
import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd


FIGURE_PATH = "plt.png"
FIG_DPI = 300


def get_plot(task, gpu, omit_offload):
    # slice the dataframe according to the inputs
    df = pd.read_csv("data.csv")
    df = df[df["task"] == task]
    df = df[df["gpu"] == gpu]
    if omit_offload == "Yes":
        df = df[df["offload"] == 0]

    # combine model name and dtype
    df["model and dtype"] = df['model_name'].str.cat(df[['dtype']], sep=', ')

    # fuse the two columns to be compared (original and assisted generation)
    df = df.melt(
        id_vars=["task", "gpu", "model and dtype", "offload"],
        value_vars=["Greedy", "Assisted"],
        var_name="generation_type",
        value_name="generation_time",
    )

    g = sns.catplot(
        data=df,
        kind="bar",
        x="model and dtype",
        y="generation_time",
        hue="generation_type",
        palette={"Greedy": "blue", "Assisted": "orange"},
        alpha=.9,
    )
    g.despine(left=True)
    g.set_axis_labels("Model size and dtype", "Latency (ms/token)")
    g.set_xticklabels(fontsize=7)
    g.set_yticklabels(fontsize=7)
    g.legend.set_title("Generation Type")
    plt.setp(g._legend.get_texts(), fontsize='7')  # for legend text

    # Add the number to the top of each bar
    ax = g.facet_axis(0, 0)
    for i in ax.containers:
        ax.bar_label(i, fontsize=7)

    plt.savefig(FIGURE_PATH, dpi=FIG_DPI)
    return FIGURE_PATH


demo = gr.Blocks()

with demo:
    gr.Markdown(
        """
        # Assisted Generation Benchmark
        """
    )
    # components shared across tabs
    omit_offload_fn = functools.partial(
        gr.Radio, ["Yes", "No"], value="No", label="Omit cases with memory offload?", interactive=True
    )

    def gpu_selector_fn(gpu_list):
        return gr.Dropdown(
            gpu_list, value=gpu_list[-1], label="GPU", interactive=True
        )

    with gr.Tabs():
        with gr.TabItem("OPT: Open"):
            plot_fn = functools.partial(get_plot, "OPT: Open Text Generation")
            with gr.Row():
                with gr.Column(scale=0.3):
                    gpu_selector = gpu_selector_fn(["3090", "T4", "T4 *2", "A100 (80GB)"])
                    omit_offload = omit_offload_fn()
                    gr.Markdown(
                        """
                        ### Assistant Model
                        - `facebook/opt-125m`

                        ### Model Names:
                        - 1.3B: `facebook/opt-1.3b`
                        - 6.7B: `facebook/opt-6.7b`
                        - 30B: `facebook/opt-30b`
                        - 66B: `facebook/opt-66b`

                        ### Dataset used as input prompt:
                        - C4 (en, validation set)
                        """
                    )
                # Show plot when the gradio app is initialized
                plot = gr.Image(value=plot_fn("A100 (80GB)", "No"))
            # Update plot when any of the inputs change
            plot_inputs = [gpu_selector, omit_offload]
            gpu_selector.change(fn=plot_fn, inputs=plot_inputs, outputs=plot)
            omit_offload.change(fn=plot_fn, inputs=plot_inputs, outputs=plot)
        with gr.TabItem("OPT: Summarization"):
            plot_fn = functools.partial(get_plot, "OPT: Summarization")
            with gr.Row():
                with gr.Column(scale=0.3):
                    gpu_selector = gpu_selector_fn(["3090", "T4", "T4 *2", "A100 (80GB)"])
                    omit_offload = omit_offload_fn()
                    gr.Markdown(
                        """
                        ### Assistant Model
                        - `facebook/opt-125m`

                        ### Model Names:
                        - 1.3B: `facebook/opt-1.3b`
                        - 6.7B: `facebook/opt-6.7b`
                        - 30B: `facebook/opt-30b`
                        - 66B: `facebook/opt-66b`

                        ### Dataset used as input prompt:
                        - CNN Dailymail (3.0.0, validation set)
                        """
                    )
                # Show plot when the gradio app is initialized
                plot = gr.Image(value=plot_fn("A100 (80GB)", "No"))
            # Update plot when any of the inputs change
            plot_inputs = [gpu_selector, omit_offload]
            gpu_selector.change(fn=plot_fn, inputs=plot_inputs, outputs=plot)
            omit_offload.change(fn=plot_fn, inputs=plot_inputs, outputs=plot)
        with gr.TabItem("Whisper: ARS"):
            plot_fn = functools.partial(get_plot, "Whisper: ARS")
            with gr.Row():
                with gr.Column(scale=0.3):
                    gpu_selector = gpu_selector_fn(["3090", "T4"])
                    omit_offload = omit_offload_fn()
                    gr.Markdown(
                        """
                        ### Assistant Model
                        - `openai/whisper-tiny`

                        ### Model Names:
                        - large-v2: `openai/whisper-large-v2`

                        ### Dataset used as input prompt:
                        - Librispeech ARS (clean, validation set)
                        """
                    )
                # Show plot when the gradio app is initialized
                plot = gr.Image(value=plot_fn("T4", "No"))
            # Update plot when any of the inputs change
            plot_inputs = [gpu_selector, omit_offload]
            gpu_selector.change(fn=plot_fn, inputs=plot_inputs, outputs=plot)
            omit_offload.change(fn=plot_fn, inputs=plot_inputs, outputs=plot)
        with gr.TabItem("CodeGen: Code"):
            plot_fn = functools.partial(get_plot, "CodeGen: Code Generation")
            with gr.Row():
                with gr.Column(scale=0.3):
                    gpu_selector = gpu_selector_fn(["3090", "T4", "T4 *2", "A100 (80GB)"])
                    omit_offload = omit_offload_fn()
                    gr.Markdown(
                        """
                        ### Assistant Model
                        - `Salesforce/codegen-350M-mono`

                        ### Model Names:
                        - 2B: `Salesforce/codegen-2B-mono`
                        - 6B: `Salesforce/codegen-6B-mono`
                        - 16B: `Salesforce/codegen-16B-mono`

                        ### Dataset used as input prompt:
                        - The Stack (python)
                        """
                    )
                # Show plot when the gradio app is initialized
                plot = gr.Image(value=plot_fn("A100 (80GB)", "No"))
            # Update plot when any of the inputs change
            plot_inputs = [gpu_selector, omit_offload]
            gpu_selector.change(fn=plot_fn, inputs=plot_inputs, outputs=plot)
            omit_offload.change(fn=plot_fn, inputs=plot_inputs, outputs=plot)
        with gr.TabItem("Flan-T5: Summarization"):
            plot_fn = functools.partial(get_plot, "Flan-T5: Summarization")
            with gr.Row():
                with gr.Column(scale=0.3):
                    gpu_selector = gpu_selector_fn(["3090", "T4", "T4 *2", "A100 (80GB)"])
                    omit_offload = omit_offload_fn()
                    gr.Markdown(
                        """
                        ### Assistant Model
                        - `google/flan-t5-small`

                        ### Model Names:
                        - large: `google/flan-t5-large`
                        - xl: `google/flan-t5-xl`
                        - xxl: `google/flan-t5-xxl`
                        - ul2: `google/flan-ul2`

                        ### Dataset used as input prompt:
                        - CNN Dailymail (3.0.0, validation set)
                        """
                    )
                # Show plot when the gradio app is initialized
                plot = gr.Image(value=plot_fn("A100 (80GB)", "No"))
            # Update plot when any of the inputs change
            plot_inputs = [gpu_selector, omit_offload]
            gpu_selector.change(fn=plot_fn, inputs=plot_inputs, outputs=plot)
            omit_offload.change(fn=plot_fn, inputs=plot_inputs, outputs=plot)
        with gr.TabItem("Benchmark Info"):
            gr.Dataframe(
                headers=["Parameter", "Value"],
                value=[
                    ["Transformers Version", "4.29dev0"],
                    ["Pytorch Version", "2.0.0"],
                    ["OS", "22.04 LTS (3090) / Debian 10 (other GPUs)"],
                    ["CUDA", "11.8 (3090) / 11.3 (others GPUs)"],
                    ["Number of input samples", "20-100 (depending on the model size)"],
                    ["Is there code to reproduce?", "Yes -- https://github.com/gante/huggingface-demos/tree/main/experiments/faster_generation"],
                ],
            )

demo.launch()