File size: 6,410 Bytes
1dd0620 734ca59 28dd0d5 976eb10 1dd0620 2723972 976eb10 2723972 0dfa35f 734ca59 0dfa35f 241d9b0 0dfa35f 241d9b0 0dfa35f 241d9b0 0dfa35f da756b5 0dfa35f 2723972 0dfa35f da756b5 0dfa35f da756b5 0dfa35f da756b5 0dfa35f b1484a9 2723972 b1484a9 2723972 e3f1090 241d9b0 6d9bae3 7aef7b8 979192e 74e572f 979192e 7aef7b8 28dd0d5 b06798e b1484a9 b06798e b1484a9 b06798e b3e1fc6 b1484a9 61212e3 951e56d b1484a9 28dd0d5 7aef7b8 28dd0d5 7c04b7a 9a8c083 7fa6af3 241d9b0 7fa6af3 9a8c083 28dd0d5 7aef7b8 |
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 |
import matplotlib
matplotlib.use('Agg')
import functools
import gradio as gr
import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd
# benchmark order: pytorch, tf eager, tf xla; units = ms
BENCHMARK_DATA = {
"Greedy Search": {
"DistilGPT2": {
"T4": [336.22, 3976.23, 115.84],
"3090": [158.38, 1835.82, 46.56],
"A100": [371.49, 4073.84, 60.94],
},
"GPT2": {
"T4": [607.31, 7140.23, 185.12],
"3090": [297.03, 3308.31, 76.68],
"A100": [691.75, 7323.60, 110.72],
},
"OPT-1.3B": {
"T4": [1303.41, 15939.07, 1488.15],
"3090": [428.33, 7259.43, 468.37],
"A100": [1125.00, 16713.63, 384.52],
},
"GPTJ-6B": {
"T4": [0, 0, 0],
"3090": [0, 0, 0],
"A100": [2664.28, 32783.09, 1440.06],
},
"T5 Small": {
"T4": [99.88, 1527.73, 18.78],
"3090": [55.09, 665.70, 9.25],
"A100": [124.91, 1642.07, 13.72],
},
"T5 Base": {
"T4": [416.56, 6095.05, 106.12],
"3090": [223.00, 2503.28, 46.67],
"A100": [550.76, 6504.11, 64.57],
},
"T5 Large": {
"T4": [645.05, 9587.67, 225.17],
"3090": [377.74, 4216.41, 97.92],
"A100": [944.17, 10572.43, 116.52],
},
"T5 3B": {
"T4": [1493.61, 13629.80, 1494.80],
"3090": [694.75, 6316.79, 489.33],
"A100": [1801.68, 16707.71, 411.93],
},
},
"Sample": {
"DistilGPT2": {
"T4": [],
"3090": [],
"A100": [],
},
"GPT2": {
"T4": [],
"3090": [],
"A100": [],
},
"OPT-1.3B": {
"T4": [],
"3090": [],
"A100": [],
},
"GPTJ-6B": {
"T4": [],
"3090": [],
"A100": [],
},
"T5 Small": {
"T4": [],
"3090": [],
"A100": [],
},
"T5 Base": {
"T4": [],
"3090": [],
"A100": [],
},
"T5 Large": {
"T4": [],
"3090": [],
"A100": [],
},
"T5 3B": {
"T4": [],
"3090": [],
"A100": [],
},
},
"Beam Search": {
"DistilGPT2": {
"T4": [],
"3090": [],
"A100": [],
},
"GPT2": {
"T4": [],
"3090": [],
"A100": [],
},
"OPT-1.3B": {
"T4": [],
"3090": [],
"A100": [],
},
"GPTJ-6B": {
"T4": [],
"3090": [],
"A100": [],
},
"T5 Small": {
"T4": [],
"3090": [],
"A100": [],
},
"T5 Base": {
"T4": [],
"3090": [],
"A100": [],
},
"T5 Large": {
"T4": [],
"3090": [],
"A100": [],
},
"T5 3B": {
"T4": [],
"3090": [],
"A100": [],
},
},
}
def get_plot(model_name, plot_eager, generate_type):
df = pd.DataFrame(BENCHMARK_DATA[generate_type][model_name])
df["framework"] = ["PyTorch", "TF (Eager Execition)", "TF (XLA)"]
df = pd.melt(df, id_vars=["framework"], value_vars=["T4", "3090", "A100"])
if plot_eager == "No":
df = df[df["framework"] != "TF (Eager Execition)"]
g = sns.catplot(
data=df, kind="bar",
x="variable", y="value", hue="framework",
ci="sd", palette="dark", alpha=.6, height=6
)
g.despine(left=True)
g.set_axis_labels("GPU", "Generation time (ms)")
g.legend.set_title("Framework")
return plt.gcf()
demo = gr.Blocks()
with demo:
gr.Markdown(
"""
# TensorFlow XLA Text Generation Benchmark
Pick a tab for the type of generation (or other information), and then select a model from the dropdown menu.
You can also ommit results from TensorFlow Eager Execution, if you wish to better compare the performance of
PyTorch to TensorFlow with XLA.
"""
)
with gr.Tabs():
with gr.TabItem("Greedy Search"):
gr.Markdown(
"""
### Greedy Search benchmark parameters
- `max_new_tokens = 64`;
- `pad_to_multiple_of = 64` for Tensorflow XLA models. Others do not pad (input prompts between 2 and 33 tokens).
"""
)
with gr.Row():
model_selector = gr.Dropdown(
choices=["DistilGPT2", "GPT2", "OPT-1.3B", "GPTJ-6B", "T5 Small", "T5 Base", "T5 Large", "T5 3B"],
value="T5 Small",
label="Model",
interactive=True,
)
eager_enabler = gr.Radio(
["Yes", "No"],
value="Yes",
label="Plot TF Eager Execution?",
interactive=True
)
plot_fn = functools.partial(get_plot, generate_type="Greedy Search")
plot = gr.Plot(value=plot_fn("T5 Small", "Yes")) # Show plot when the gradio app is initialized
model_selector.change(fn=plot_fn, inputs=[model_selector, eager_enabler], outputs=plot)
with gr.TabItem("Sample"):
gr.Button("New Tiger")
with gr.TabItem("Beam Search"):
gr.Button("New Tiger")
with gr.TabItem("Benchmark Information"):
gr.Dataframe(
headers=["Parameter", "Value"],
value=[
["Transformers Version", "4.22.dev0"],
["TensorFlow Version", "2.9.1"],
["Pytorch Version", "1.11.0"],
["OS", "22.04 LTS (3090) / Debian 10 (other GPUs)"],
["CUDA", "11.6 (3090) / 11.3 (others GPUs)"],
["Number of Runs", "100 (the first run was discarded to ignore compilation time)"],
["Is there code to reproduce?", "Yes -- https://gist.github.com/gante/f0017e3f13ac11b0c02e4e4db351f52f"],
],
)
demo.launch()
|