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 9, 03:45 UTC
Change defaults for vllm bench startup (#33489)
Luka Govedič committed
Feb 1, 2026 at 02:46 UTC
672023877bfd17a8720ae05ed198456d9b454363
1 file changed
+3
-3
vllm/benchmarks/startup.py
+3
-3
@@ -159,19 +159,19 @@ def add_cli_args(parser: argparse.ArgumentParser):
159
parser.add_argument(
160
"--num-iters-cold",
161
type=int,
162
- default=5,
162
+ default=3,
163
help="Number of cold startup iterations.",
164
)
165
parser.add_argument(
166
"--num-iters-warmup",
167
type=int,
168
- default=3,
168
+ default=1,
169
help="Number of warmup iterations before benchmarking warm startups.",
170
)
171
parser.add_argument(
172
"--num-iters-warm",
173
type=int,
174
- default=5,
174
+ default=3,
175
help="Number of warm startup iterations.",
176
)
177
parser.add_argument(