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 11, 00:00 UTC
Revert "[bench] Support common prefix len config (for decode-only bench)" (#31240)
Signed-off-by: Ming Yang <minos.future@gmail.com>
Ming Yang committed
Dec 23, 2025 at 21:17 UTC
3bb9561928bb4b0fb1047e738e2f160f841b04b6
2 files changed
-7
vllm/benchmarks/datasets.py
-1
@@ -1847,7 +1847,6 @@ def get_samples(args, tokenizer: TokenizerLike) -> list[SampleRequest]:
1847
random_seed=args.seed,
1848
dataset_path=args.dataset_path,
1849
disable_shuffle=args.disable_shuffle,
1850
- prefix_len=args.common_prefix_len,
1850
).sample(
1851
tokenizer=tokenizer,
1852
num_requests=args.num_prompts,
vllm/benchmarks/serve.py
-6
@@ -1281,12 +1281,6 @@ def add_cli_args(parser: argparse.ArgumentParser):
1281
help="Repetition penalty sampling parameter. Only has effect on "
1282
"openai-compatible backends.",
1283
)
1284
- sampling_group.add_argument(
1285
- "--common-prefix-len",
1286
- type=int,
1287
- default=None,
1288
- help="Common prefix length shared by all prompts (used by random dataset)",
1289
- )
1284
1285
parser.add_argument(
1286
"--served-model-name",