diff --git a/examples/disaggregated/slurm/benchmark/submit.py b/examples/disaggregated/slurm/benchmark/submit.py
index f1cf53f87996..60c51c81626f 100644
--- a/examples/disaggregated/slurm/benchmark/submit.py
+++ b/examples/disaggregated/slurm/benchmark/submit.py
@@ -291,15 +291,17 @@ def build_worker_environment(worker_config, env_config, role, benchmark_mode,
         upsert_env_config(env_config, 'worker_env_var',
                           'TRTLLM_DISAGG_BENCHMARK_GEN_ONLY',
                           'TRTLLM_DISAGG_BENCHMARK_GEN_ONLY=1')
-    if benchmark_mode == "gen_only":
-        upsert_env_config(env_config, 'worker_env_var',
+    if benchmark_mode == "gen_only" and role == "GEN":
+        # GEN worker only: skipping transfer-state polling helps the generation
+        # worker, but the same flag on the CTX worker has been seen to hang
+        # gen_only runs with KV blocks never released.
+        upsert_env_config(env_config, 'gen_worker_env_var',
                           'TRTLLM_DISABLE_KV_CACHE_TRANSFER_OVERLAP',
                           'TRTLLM_DISABLE_KV_CACHE_TRANSFER_OVERLAP=1')
-        if role == "GEN":
-            concurrency = _parse_positive_concurrency(concurrency)
-            upsert_env_config(env_config, 'gen_worker_env_var',
-                              'TLLM_BENCHMARK_REQ_QUEUES_SIZE',
-                              f'TLLM_BENCHMARK_REQ_QUEUES_SIZE={concurrency}')
+        concurrency = _parse_positive_concurrency(concurrency)
+        upsert_env_config(env_config, 'gen_worker_env_var',
+                          'TLLM_BENCHMARK_REQ_QUEUES_SIZE',
+                          f'TLLM_BENCHMARK_REQ_QUEUES_SIZE={concurrency}')
 
     # 2. Add profiling env vars to env_config (conditional)
     if nsys_on:
diff --git a/jenkins/scripts/perf/local/submit.py b/jenkins/scripts/perf/local/submit.py
index 8e571eb2f890..e359b81bfff7 100755
--- a/jenkins/scripts/perf/local/submit.py
+++ b/jenkins/scripts/perf/local/submit.py
@@ -901,9 +901,8 @@ def main():
             srun_args_lines.append("--container-env=TRTLLM_DISAGG_BENCHMARK_GEN_ONLY")
         elif "gen_only" in bm_config.get("mode", ""):
             concurrency = bm_config.get("concurrency", 1)
-            ctx_worker_env_vars = (
-                f"TRTLLM_DISABLE_KV_CACHE_TRANSFER_OVERLAP=1 {ctx_worker_env_vars}"
-            )
+            # GEN worker only: the same flag on the CTX worker has been seen to
+            # hang gen_only runs with KV blocks never released.
             gen_worker_env_vars = (
                 f"TRTLLM_DISABLE_KV_CACHE_TRANSFER_OVERLAP=1 "
                 f"TLLM_BENCHMARK_REQ_QUEUES_SIZE={concurrency} {gen_worker_env_vars}"
diff --git a/jenkins/scripts/perf/submit.py b/jenkins/scripts/perf/submit.py
index f1c54fb13690..dad770df9654 100755
--- a/jenkins/scripts/perf/submit.py
+++ b/jenkins/scripts/perf/submit.py
@@ -796,9 +796,8 @@ def main():
             srun_args_lines.append("--container-env=TRTLLM_DISAGG_BENCHMARK_GEN_ONLY")
         elif benchmark_mode == "gen_only":
             concurrency = benchmark_config.get("concurrency", 1)
-            ctx_worker_env_vars = (
-                f"TRTLLM_DISABLE_KV_CACHE_TRANSFER_OVERLAP=1 {ctx_worker_env_vars}"
-            )
+            # GEN worker only: the same flag on the CTX worker has been seen to
+            # hang gen_only runs with KV blocks never released.
             gen_worker_env_vars = (
                 f"TRTLLM_DISABLE_KV_CACHE_TRANSFER_OVERLAP=1 "
                 f"TLLM_BENCHMARK_REQ_QUEUES_SIZE={concurrency} {gen_worker_env_vars}"