@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 6fa718a460

Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 15, 06:00 UTC

[Misc] Modularize CLI Argument Parsing in Benchmark Scripts (#19593)

Signed-off-by: reidliu41 <reid201711@gmail.com> Co-authored-by: reidliu41 <reid201711@gmail.com>

Reid committed Jun 14, 2025 at 16:54 UTC 6fa718a46007ae97e98a5bb3bcfb506748a2473e

7 files changed +47 -8

benchmarks/benchmark_latency.py

+7 -1

@@ -123,7 +123,7 @@ def main(args: argparse.Namespace):

123 save_to_pytorch_benchmark_format(args, results)

124

125

126 -if __name__ == "__main__":

126 +def create_argument_parser():

127 parser = FlexibleArgumentParser(

128 description="Benchmark the latency of processing a single batch of "

129 "requests till completion."

@@ -171,6 +171,12 @@ if __name__ == "__main__":

171 # V1 enables prefix caching by default which skews the latency

172 # numbers. We need to disable prefix caching by default.

173 parser.set_defaults(enable_prefix_caching=False)

174 +

175 + return parser

176 +

177 +

178 +if __name__ == "__main__":

179 + parser = create_argument_parser()

180 args = parser.parse_args()

181 if args.profile and not envs.VLLM_TORCH_PROFILER_DIR:

182 raise OSError(

benchmarks/benchmark_long_document_qa_throughput.py

+7 -1

@@ -142,7 +142,7 @@ def main(args):

142 )

143

144

145 -if __name__ == "__main__":

145 +def create_argument_parser():

146 parser = FlexibleArgumentParser(

147 description="Benchmark the performance with or "

148 "without automatic prefix caching."

@@ -192,5 +192,11 @@ if __name__ == "__main__":

192 )

193

194 parser = EngineArgs.add_cli_args(parser)

195 +

196 + return parser

197 +

198 +

199 +if __name__ == "__main__":

200 + parser = create_argument_parser()

201 args = parser.parse_args()

202 main(args)

benchmarks/benchmark_prefix_caching.py

+7 -1

@@ -218,7 +218,7 @@ def main(args):

218 )

219

220

221 -if __name__ == "__main__":

221 +def create_argument_parser():

222 parser = FlexibleArgumentParser(

223 description="Benchmark the performance with or without "

224 "automatic prefix caching."

@@ -268,5 +268,11 @@ if __name__ == "__main__":

268 )

269

270 parser = EngineArgs.add_cli_args(parser)

271 +

272 + return parser

273 +

274 +

275 +if __name__ == "__main__":

276 + parser = create_argument_parser()

277 args = parser.parse_args()

278 main(args)

benchmarks/benchmark_prioritization.py

+7 -1

@@ -161,7 +161,7 @@ def main(args: argparse.Namespace):

161 json.dump(results, f, indent=4)

162

163

164 -if __name__ == "__main__":

164 +def create_argument_parser():

165 parser = FlexibleArgumentParser(description="Benchmark the throughput.")

166 parser.add_argument(

167 "--backend", type=str, choices=["vllm", "hf", "mii"], default="vllm"

@@ -204,6 +204,12 @@ if __name__ == "__main__":

204 )

205

206 parser = EngineArgs.add_cli_args(parser)

207 +

208 + return parser

209 +

210 +

211 +if __name__ == "__main__":

212 + parser = create_argument_parser()

213 args = parser.parse_args()

214 if args.tokenizer is None:

215 args.tokenizer = args.model

benchmarks/benchmark_serving.py

+6 -2

@@ -875,7 +875,7 @@ def main(args: argparse.Namespace):

875 save_to_pytorch_benchmark_format(args, result_json, file_name)

876

877

878 -if __name__ == "__main__":

878 +def create_argument_parser():

879 parser = FlexibleArgumentParser(

880 description="Benchmark the online serving throughput."

881 )

@@ -1225,6 +1225,10 @@ if __name__ == "__main__":

1225 "script chooses a LoRA module at random.",

1226 )

1227

1228 - args = parser.parse_args()

1228 + return parser

1229 +

1230

1231 +if __name__ == "__main__":

1232 + parser = create_argument_parser()

1233 + args = parser.parse_args()

1234 main(args)

benchmarks/benchmark_serving_structured_output.py

+6 -1

@@ -850,7 +850,7 @@ def main(args: argparse.Namespace):

850 json.dump(results, outfile, indent=4)

851

852

853 -if __name__ == "__main__":

853 +def create_argument_parser():

854 parser = FlexibleArgumentParser(

855 description="Benchmark the online serving throughput."

856 )

@@ -1034,5 +1034,10 @@ if __name__ == "__main__":

1034 help="Ratio of Structured Outputs requests",

1035 )

1036

1037 + return parser

1038 +

1039 +

1040 +if __name__ == "__main__":

1041 + parser = create_argument_parser()

1042 args = parser.parse_args()

1043 main(args)

benchmarks/benchmark_throughput.py

+7 -1

@@ -595,7 +595,7 @@ def validate_args(args):

595 )

596

597

598 -if __name__ == "__main__":

598 +def create_argument_parser():

599 parser = FlexibleArgumentParser(description="Benchmark the throughput.")

600 parser.add_argument(

601 "--backend",

@@ -717,6 +717,12 @@ if __name__ == "__main__":

717 )

718

719 parser = AsyncEngineArgs.add_cli_args(parser)

720 +

721 + return parser

722 +

723 +

724 +if __name__ == "__main__":

725 + parser = create_argument_parser()

726 args = parser.parse_args()

727 if args.tokenizer is None:

728 args.tokenizer = args.model