Spaces:
Running
Running
File size: 5,177 Bytes
2460b35 dc685a9 2460b35 14d526b 2460b35 14d526b 2460b35 14d526b 2460b35 14d526b 2460b35 dc685a9 08604d0 14d526b dc685a9 14d526b dc685a9 14d526b 08604d0 14d526b 2460b35 14d526b 2460b35 14d526b 2460b35 14d526b 2460b35 14d526b 2460b35 dc685a9 2460b35 dc685a9 2460b35 dc685a9 2460b35 dc685a9 2460b35 dc685a9 2460b35 dc685a9 2460b35 dc685a9 2460b35 dc685a9 14d526b 2460b35 dc685a9 2460b35 dc685a9 2460b35 dc685a9 2460b35 dc685a9 14d526b 2460b35 dc685a9 2460b35 dc685a9 2460b35 dc685a9 2460b35 |
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 |
import gradio as gr
import pandas as pd
import plotly.express as px
QUANT_DATA = [
# open llm
"Model 🤗",
"Arch 🏛️",
"DType 📥",
"Backend 🏭",
"Params (B)",
"Open LLM Score (%)",
# deployment settings
"DType 📥",
"Backend 🏭",
"Optimization 🛠️",
"Quantization 🗜️",
"Optimization 🛠️ Custom Kernel",
"Quantization 🗜️ Custom Kernel",
# primary measurements
"Prefill Latency (s)",
"Prefill Latency (s) Custom Kernel",
"Decode Throughput (tokens/s)",
"Decode Throughput (tokens/s) Custom Kernel",
# speedups
"Prefill Latency Speedup (%)",
"Decode Throughput Speedup (%)",
]
def get_quant_df(llm_perf_df):
copy_df = llm_perf_df.copy()
# seperate vanilla GPTQ experiments from Custom Kernel experiments
vanilla_df = copy_df[
(copy_df["Backend 🏭"] == "pytorch") &
(copy_df["Quantization 🗜️"] == "None") &
(copy_df["Optimization 🛠️"] == "None") &
(copy_df["DType 📥"] == "float16")
]
exllamav1_df = copy_df[(copy_df["Quantization 🗜️"] == "GPTQ.4bit+ExllamaV1")]
exllamav2_df = copy_df[(copy_df["Quantization 🗜️"] == "GPTQ.4bit+ExllamaV2")]
gemm_df = copy_df[(copy_df["Quantization 🗜️"] == "AWQ.4bit+GEMM")]
gemv_df = copy_df[(copy_df["Quantization 🗜️"] == "AWQ.4bit+GEMV")]
# merge the three dataframes
exllamav1_df = pd.merge(
vanilla_df,
exllamav1_df,
on=["Model 🤗"],
suffixes=["", " Custom Kernel"],
)
exllamav2_df = pd.merge(
vanilla_df,
exllamav2_df,
on=["Model 🤗"],
suffixes=["", " Custom Kernel"],
)
gemm_df = pd.merge(
vanilla_df,
gemm_df,
on=["Model 🤗"],
suffixes=["", " Custom Kernel"],
)
gemv_df = pd.merge(
vanilla_df,
gemv_df,
on=["Model 🤗"],
suffixes=["", " Custom Kernel"],
)
# concat the two dataframes row-wise
quant_df = pd.concat([exllamav1_df, exllamav2_df, gemm_df, gemv_df])
# compute speedups
quant_df["Prefill Latency Speedup (%)"] = (
(quant_df["Prefill Latency (s)"] / quant_df["Prefill Latency (s) Custom Kernel"]) * 100
).round(2) - 100
quant_df["Decode Throughput Speedup (%)"] = (
(quant_df["Decode Throughput (tokens/s) Custom Kernel"] / quant_df["Decode Throughput (tokens/s)"]) * 100
).round(2) - 100
# filter speedups > 1000%
quant_df = quant_df[quant_df["Prefill Latency Speedup (%)"] < 1000]
quant_df = quant_df[quant_df["Decode Throughput Speedup (%)"] < 1000]
return quant_df
def get_quant_decode_fig(llm_perf_df):
quant_df = get_quant_df(llm_perf_df)
# plot
decode_fig = px.box(
quant_df,
x="Arch 🏛️",
y="Decode Throughput Speedup (%)",
color_discrete_sequence=px.colors.qualitative.Light24,
custom_data=QUANT_DATA,
color="Quantization 🗜️ Custom Kernel",
points="all",
)
# add hover data
decode_fig.update_traces(
hovertemplate="<br>".join([f"<b>{column}:</b> %{{customdata[{i}]}}" for i, column in enumerate(QUANT_DATA)])
)
# add layout
decode_fig.update_layout(
title={
"text": "Decode Throughput Speedup per Architecture",
"y": 0.95,
"x": 0.5,
"xanchor": "center",
"yanchor": "top",
},
xaxis_title="LLM Architecture",
yaxis_title="Decode Speedup (%)",
legend_title="Quantization Scheme",
width=1200,
height=600,
)
return decode_fig
def get_quant_prefill_fig(llm_perf_df):
quant_df = get_quant_df(llm_perf_df)
# plot
prefill_fig = px.box(
quant_df,
x="Arch 🏛️",
y="Prefill Latency Speedup (%)",
color_discrete_sequence=px.colors.qualitative.Light24,
custom_data=QUANT_DATA,
color="Quantization 🗜️ Custom Kernel",
points="all",
)
# add hover data
prefill_fig.update_traces(
hovertemplate="<br>".join([f"<b>{column}:</b> %{{customdata[{i}]}}" for i, column in enumerate(QUANT_DATA)])
)
# add layout
prefill_fig.update_layout(
title={
"text": "Prefill Latency Speedup per Architecture",
"y": 0.95,
"x": 0.5,
"xanchor": "center",
"yanchor": "top",
},
xaxis_title="LLM Architecture",
yaxis_title="Prefill Speedup (%)",
legend_title="Quantization Scheme",
width=1200,
height=600,
)
return prefill_fig
def create_quant_plots(llm_perf_df):
# descriptive text
gr.HTML("👆 Hover over the points 👆 for additional information.", elem_id="text")
# get figures
prefill_fig = get_quant_prefill_fig(llm_perf_df)
decode_fig = get_quant_decode_fig(llm_perf_df)
# create plots
prefill_plot = gr.components.Plot(value=prefill_fig, elem_id="plot", show_label=False)
decode_plot = gr.components.Plot(value=decode_fig, elem_id="plot", show_label=False)
return prefill_plot, decode_plot
|