diff --git a/tests/integration/defs/perf/test_perf_sanity.py b/tests/integration/defs/perf/test_perf_sanity.py
index 4119913b1fa4..1a0cac76f959 100644
--- a/tests/integration/defs/perf/test_perf_sanity.py
+++ b/tests/integration/defs/perf/test_perf_sanity.py
@@ -872,6 +872,11 @@ def __init__(
         self.model_path = ""
         self.dataset_file = client_config_data.get("dataset_file", "")
         self.use_nv_sa_benchmark = client_config_data.get("use_nv_sa_benchmark", False)
+        # Derived from lane identity only (gen_only + concurrency == 1); do not
+        # set this from lane YAML. warmup is intentionally not a baseline match
+        # key, which is only sound while its value stays fully determined by
+        # benchmark_mode and concurrency.
+        self.warmup = client_config_data.get("warmup", False)
         self.env_vars = env_vars
         # spec_decoding flag is retained for DB matching (b_eos column). --ignore-eos
         # is now always passed; output-length stability with spec decoding comes from
@@ -950,11 +955,15 @@ def _to_default_benchmark_cmd(self) -> List[str]:
             str(self.concurrency * self.iterations),
             "--max-concurrency",
             str(self.concurrency),
-            "--no-test-input",
             "--percentile-metrics",
             "ttft,tpot,itl,e2el",
             "--ignore-eos",
         ]
+        # benchmark_serving's initial single-prompt test run (excluded from
+        # metrics) doubles as a warmup request; keep it disabled unless the
+        # lane requests one.
+        if not self.warmup:
+            benchmark_cmd.append("--no-test-input")
         if dataset_path:
             benchmark_cmd.append("--dataset-name")
             benchmark_cmd.append("trtllm_custom")
@@ -1012,6 +1021,7 @@ def to_db_data(self) -> dict:
             "b_trust_remote_code": self.trust_remote_code,
             "b_use_nv_sa_benchmark": self.use_nv_sa_benchmark,
             "b_eos": self.spec_decoding,
+            "b_warmup": self.warmup,
             "s_client_log_link": "",
             "s_client_env_vars": self.env_vars,
         }
@@ -1978,6 +1988,15 @@ def _parse_disagg_config_file(self, config_file_path: str, config_file: str):
                 "use_nv_sa_benchmark": use_nv_sa_benchmark,
                 "accuracy_config": accuracy_data,
                 "only_run_accuracy": only_run_accuracy,
+                # gen_only measures a single round (iterations forced to 1
+                # above), so one-time costs like the cache transceiver's lazy
+                # connection setup would otherwise land entirely on the
+                # measured TTFT. Scoped to concurrency == 1: the gen executor's
+                # fill gate (TLLM_BENCHMARK_REQ_QUEUES_SIZE) only opens once
+                # `concurrency` requests are queued, so a lone warmup request
+                # would deadlock higher-concurrency lanes — which amortize the
+                # cold start anyway.
+                "warmup": benchmark_mode == "gen_only" and concurrency == 1,
             }
             client_config = ClientConfig(
                 client_config_data,