diff --git a/python/sglang/benchmark/one_batch_server.py b/python/sglang/benchmark/one_batch_server.py
index 27a426336..545734b2f 100644
--- a/python/sglang/benchmark/one_batch_server.py
+++ b/python/sglang/benchmark/one_batch_server.py
@@ -124,6 +124,7 @@ class BenchArgs:
     base_url: str = ""
     local_tokenizer_path: str = ""
     skip_warmup: bool = False
+    skip_token_capacity_check: bool = False
     show_report: bool = False
     profile: bool = False
     profile_activities: Tuple[str] = ("CPU", "GPU")
@@ -207,6 +208,11 @@ class BenchArgs:
             ),
         )
         parser.add_argument("--skip-warmup", action="store_true")
+        parser.add_argument(
+            "--skip-token-capacity-check",
+            action="store_true",
+            help="Skip the raw-token capacity check; keep max-running-requests checks.",
+        )
         parser.add_argument("--show-report", action="store_true")
         parser.add_argument("--profile", action="store_true")
         parser.add_argument(
@@ -1230,6 +1236,9 @@ def run_benchmark_internal(
                 "token_capacity", 1000000000
             )
 
+        if bench_args.skip_token_capacity_check:
+            skip_token_capacity_threshold = float("inf")
+
         # Router /get_server_info responses carry "router_manager"; worker
         # responses never do, so its presence confirms a router by design.
         if not internal_states and server_info.get("router_manager"):