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": [617.40, 6078.81, 221.65], "3090": [310.37, 2843.73, 85.44], "A100": [729.05, 7140.05, 121.83], }, "GPT2": { "T4": [1205.34, 12256.98, 378.69], "3090": [577.12, 5637.11, 160.02], "A100": [1377.68, 15605.72, 234.47], }, "OPT-1.3B": { "T4": [2166.72, 19126.25, 2341.32], "3090": [706.50, 9616.97, 731.58], "A100": [2019.70, 28621.09, 690.36], }, "GPTJ-6B": { "T4": [0, 0, 0], "3090": [0, 0, 0], "A100": [5150.35, 70554.07, 2744.49], }, "T5 Small": { "T4": [235.93, 3599.47, 41.07], "3090": [100.41, 1093.33, 23.24], "A100": [267.42, 3366.73, 28.53], }, "T5 Base": { "T4": [812.59, 7966.73, 196.85], "3090": [407.81, 4904.54, 97.56], "A100": [1033.05, 11521.97, 123.93], }, "T5 Large": { "T4": [1114.22, 16433.31, 424.91], "3090": [647.61, 7184.71, 160.97], "A100": [1668.73, 19962.78, 200.75], }, "T5 3B": { "T4": [2282.56, 20891.22, 2196.02], "3090": [1011.32, 9735.97, 734.40], "A100": [2769.64, 26440.65, 612.98], }, }, "Beam Search": { "DistilGPT2": { "T4": [2407.89, 19442.60, 3313.92], "3090": [998.52, 8286.03, 900.28], "A100": [2237.41, 21771.40, 760.47], }, "GPT2": { "T4": [3767.43, 34813.93, 5559.42], "3090": [1633.04, 14606.93, 1533.55], "A100": [3705.43, 34586.23, 1295.87], }, "OPT-1.3B": { "T4": [16649.82, 78500.33, 21894.31], "3090": [508518, 32822.81, 5762.46], "A100": [5967.32, 78334.56, 4096.38], }, "GPTJ-6B": { "T4": [0, 0, 0], "3090": [0, 0, 0], "A100": [15119.10, 134000.40, 10214.17], }, "T5 Small": { "T4": [283.64, 25089.12, 1391.66], "3090": [137.38, 10680.28, 486.96], "A100": [329.28, 24747.38, 513.99], }, "T5 Base": { "T4": [1383.21, 44809.14, 3920.40], "3090": [723.11, 18657.48, 1258.60], "A100": [2360.85, 45085.07, 1107.58], }, "T5 Large": { "T4": [1663.50, 81902.41, 9551.29], "3090": [922.53, 35524.30, 2838.86], "A100": [2168.22, 86890.00, 2373.04], }, "T5 3B": { "T4": [0, 0, 0], "3090": [1521.05, 35337.30, 8282.09], "A100": [3162.54, 88453.65, 5585.20], }, }, } 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={"PyTorch": "blue", "TF (Eager Execition)": "orange", "TF (XLA)": "red"}, 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 omit 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) eager_enabler.change(fn=plot_fn, inputs=[model_selector, eager_enabler], outputs=plot) with gr.TabItem("Sample"): gr.Markdown( """ ### Sample benchmark parameters - `max_new_tokens = 128`; - `temperature = 2.0`; - `top_k = 50`; - `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="Sample") 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) eager_enabler.change(fn=plot_fn, inputs=[model_selector, eager_enabler], outputs=plot) with gr.TabItem("Beam Search"): gr.Markdown( """ ### Beam Search benchmark parameters - `max_new_tokens = 256`; - `num_beams = 16`; - `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="Beam 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) eager_enabler.change(fn=plot_fn, inputs=[model_selector, eager_enabler], outputs=plot) 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()