[bench] Support common prefix len config (for decode-only bench) (#29934)

Signed-off-by: Ming Yang <minos.future@gmail.com>

vllm/benchmarks/datasets.py

@@ -1842,6 +1842,7 @@ def get_samples(args, tokenizer) -> list[SampleRequest]:
                 random_seed=args.seed,
                 dataset_path=args.dataset_path,
                 disable_shuffle=args.disable_shuffle,
+                prefix_len=args.common_prefix_len,
             ).sample(
                 tokenizer=tokenizer,
                 num_requests=args.num_prompts,

vllm/benchmarks/serve.py

@@ -1221,6 +1221,12 @@ def add_cli_args(parser: argparse.ArgumentParser):
         help="Repetition penalty sampling parameter. Only has effect on "
         "openai-compatible backends.",
     )
+    sampling_group.add_argument(
+        "--common-prefix-len",
+        type=int,
+        default=None,
+        help="Common prefix length shared by all prompts (used by random dataset)",
+    )
 
     parser.add_argument(
         "--tokenizer-mode",