File size: 10,232 Bytes
f7cde70 9ff97e7 f7cde70 9ff97e7 f7cde70 9ff97e7 f7cde70 9ff97e7 f7cde70 b5adb0e f7cde70 c6dc17f 8a3223a 51f2901 8a3223a 51f2901 8a3223a 1a1c2cb c6dc17f 8a3223a f7cde70 9ff97e7 f7cde70 9ff97e7 8f99c19 f7cde70 9ff97e7 f7cde70 9ff97e7 f7cde70 9ff97e7 f7cde70 9ff97e7 f7cde70 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 |
from contextlib import ExitStack
from dataclasses import dataclass
from typing import List
import click
import gradio as gr
import pandas as pd
from parse_results import build_results
@dataclass
class PlotConfig:
x_title: str
y_title: str
title: str
percentiles: List[float] = None
def run(from_results_dir, datasource, port):
css = '''
.summary span {
font-size: 10px;
padding-top:0;
padding-bottom:0;
}
'''
summary_desc = '''
## Summary
This table shows the average of the metrics for each model and QPS rate.
The metrics are:
* Inter token latency: Time to generate a new output token for each user querying the system.
It translates as the “speed” perceived by the end-user. We aim for at least 300 words per minute (average reading speed), so ITL<150ms
* Time to First Token: Time the user has to wait before seeing the first token of its answer.
Lower waiting time are essential for real-time interactions, less so for offline workloads.
* End-to-end latency: The overall time the system took to generate the full response to the user.
* Throughput: The number of tokens per second the system can generate across all requests
* Successful requests: The number of requests the system was able to honor in the benchmark timeframe
* Error rate: The percentage of requests that ended up in error, as the system could not process them in time or failed to process them.
'''
df_bench = pd.DataFrame()
line_plots_bench = []
column_mappings = {'inter_token_latency_ms_p90': 'ITL P90 (ms)', 'time_to_first_token_ms_p90': 'TTFT P90 (ms)',
'e2e_latency_ms_p90': 'E2E P90 (ms)', 'token_throughput_secs': 'Throughput (tokens/s)',
'successful_requests': 'Successful requests', 'error_rate': 'Error rate (%)', 'model': 'Model',
'rate': 'QPS', 'run_id': 'Run ID'}
default_df = pd.DataFrame.from_dict(
{"rate": [1, 2], "inter_token_latency_ms_p90": [10, 20],
"version": ["default", "default"],
"model": ["default", "default"]})
def load_demo(model_bench, percentiles):
return update_bench(model_bench, percentiles)
def update_bench(model, percentiles):
res = []
for plot in line_plots_bench:
if plot['config'].percentiles:
k = plot['metric'] + '_' + str(percentiles)
df_bench[plot['metric']] = df_bench[k] if k in df_bench.columns else 0
res.append(df_bench[(df_bench['model'] == model)])
return res + [summary_table()]
def summary_table() -> pd.DataFrame:
data = df_bench.groupby(['model', 'run_id', 'rate']).agg(
{'inter_token_latency_ms_p90': 'mean', 'time_to_first_token_ms_p90': 'mean',
'e2e_latency_ms_p90': 'mean', 'token_throughput_secs': 'mean',
'successful_requests': 'mean', 'error_rate': 'mean'}).reset_index()
data = data[
['run_id', 'model', 'rate', 'inter_token_latency_ms_p90', 'time_to_first_token_ms_p90',
'e2e_latency_ms_p90',
'token_throughput_secs']]
#for metric in ['inter_token_latency_ms_p90', 'time_to_first_token_ms_p90', 'e2e_latency_ms_p90',
# 'token_throughput_secs']:
# data[metric] = data[metric].apply(lambda x: f"{x:.2f}")
data = data.rename(
columns=column_mappings)
# uncomment the following line if you want to return the raw DataFrame without formatting
# return data
# Ensure numeric columns are properly typed for sorting
numeric_cols_to_ensure = [
'ITL P90 (ms)', 'TTFT P90 (ms)', 'E2E P90 (ms)',
'Throughput (tokens/s)', 'QPS'
]
for col in numeric_cols_to_ensure:
if col in data.columns:
data[col] = pd.to_numeric(data[col], errors='coerce')
# Round numeric columns to reduce decimal places while maintaining sorting
rounding_rules = {
'ITL P90 (ms)': 2,
'TTFT P90 (ms)': 2,
'E2E P90 (ms)': 2,
'Throughput (tokens/s)': 2,
'QPS': 0 # Round to integers
}
for col, decimals in rounding_rules.items():
if col in data.columns:
data[col] = data[col].round(decimals)
return data
def load_bench_results(source) -> pd.DataFrame:
data = pd.read_parquet(source)
# remove warmup and throughput
data = data[(data['id'] != 'warmup') & (data['id'] != 'throughput')]
# only keep constant rate
data = data[data['executor_type'] == 'ConstantArrivalRate']
return data
def select_region(selection: gr.SelectData, model):
min_w, max_w = selection.index
data = df_bench[(df_bench['model'] == model) & (df_bench['rate'] >= min_w) & (
df_bench['rate'] <= max_w)]
res = []
for plot in line_plots_bench:
# find the y values for the selected region
metric = plot["metric"]
y_min = data[metric].min()
y_max = data[metric].max()
res.append(gr.LinePlot(x_lim=[min_w, max_w], y_lim=[y_min, y_max]))
return res
def reset_region():
res = []
for _ in line_plots_bench:
res.append(gr.LinePlot(x_lim=None, y_lim=None))
return res
def load_datasource(datasource, fn):
if datasource.startswith('file://'):
return fn(datasource)
elif datasource.startswith('s3://'):
return fn(datasource)
else:
raise ValueError(f"Unknown datasource: {datasource}")
if from_results_dir is not None:
build_results(from_results_dir, 'benchmarks.parquet', None)
# Load data
df_bench = load_datasource(datasource, load_bench_results)
# Define metrics
metrics = {
"inter_token_latency_ms": PlotConfig(title="Inter Token Latency (lower is better)", x_title="QPS",
y_title="Time (ms)", percentiles=[0.5, 0.6, 0.7, 0.8, 0.9, 0.95, 0.99]),
"time_to_first_token_ms": PlotConfig(title="TTFT (lower is better)", x_title="QPS",
y_title="Time (ms)", percentiles=[0.5, 0.6, 0.7, 0.8, 0.9, 0.95, 0.99]),
"e2e_latency_ms": PlotConfig(title="End to End Latency (lower is better)", x_title="QPS",
y_title="Time (ms)", percentiles=[0.5, 0.6, 0.7, 0.8, 0.9, 0.95, 0.99]),
"token_throughput_secs": PlotConfig(title="Request Output Throughput (higher is better)", x_title="QPS",
y_title="Tokens/s"),
"successful_requests": PlotConfig(title="Successful requests (higher is better)", x_title="QPS",
y_title="Count"),
"error_rate": PlotConfig(title="Error rate", x_title="QPS", y_title="%"),
"prompt_tokens": PlotConfig(title="Prompt tokens", x_title="QPS", y_title="Count"),
"decoded_tokens": PlotConfig(title="Decoded tokens", x_title="QPS", y_title="Count")
}
models = df_bench["model"].unique()
run_ids = df_bench["run_id"].unique()
# get all available percentiles
percentiles = set()
for k, v in metrics.items():
if v.percentiles:
percentiles.update(v.percentiles)
percentiles = map(lambda p: f'p{int(float(p) * 100)}', percentiles)
percentiles = sorted(list(percentiles))
percentiles.append('avg')
with gr.Blocks(css=css, title="Inference Benchmarker") as demo:
with gr.Row():
gr.Markdown("# Inference-benchmarker 🤗\n## Benchmarks results")
with gr.Row():
gr.Markdown(summary_desc)
with gr.Row():
table = gr.DataFrame(
pd.DataFrame(),
elem_classes=["summary"],
)
with gr.Row():
details_desc = gr.Markdown("## Details")
with gr.Row():
model = gr.Dropdown(list(models), label="Select model", value=models[0])
with gr.Row():
percentiles_bench = gr.Radio(percentiles, label="", value="avg")
i = 0
with ExitStack() as stack:
for k, v in metrics.items():
if i % 2 == 0:
stack.close()
gs = stack.enter_context(gr.Row())
line_plots_bench.append(
{"component": gr.LinePlot(default_df, label=f'{v.title}', x="rate", y=k,
y_title=v.y_title, x_title=v.x_title,
color="run_id"
),
"model": model.value,
"metric": k,
"config": v
},
)
i += 1
for component in [model, percentiles_bench]:
component.change(update_bench, [model, percentiles_bench],
[item["component"] for item in line_plots_bench] + [table])
gr.on([plot["component"].select for plot in line_plots_bench], select_region, [model],
outputs=[item["component"] for item in line_plots_bench])
gr.on([plot["component"].double_click for plot in line_plots_bench], reset_region, None,
outputs=[item["component"] for item in line_plots_bench])
demo.load(load_demo, [model, percentiles_bench],
[item["component"] for item in line_plots_bench] + [table])
demo.launch(server_port=port, server_name="0.0.0.0")
@click.command()
@click.option('--from-results-dir', default=None, help='Load inference-benchmarker results from a directory')
@click.option('--datasource', default='file://benchmarks.parquet', help='Load a Parquet file already generated')
@click.option('--port', default=7860, help='Port to run the dashboard')
def main(from_results_dir, datasource, port):
run(from_results_dir, datasource, port)
if __name__ == '__main__':
main(auto_envvar_prefix='DASHBOARD')
|