diff --git a/serve/benchmarks/benchmark_throughput.py b/serve/benchmarks/benchmark_throughput.py
index 75c585f0cd..466c27ae5b 100644
--- a/serve/benchmarks/benchmark_throughput.py
+++ b/serve/benchmarks/benchmark_throughput.py
@@ -111,6 +111,7 @@ def main(args: argparse.Namespace):
     engine = LocalProcessInferenceEngine(
         model_module,
         max_batched_tokens=args.max_num_batched_tokens,
+        min_decode_steps=args.min_decode_steps,
     )
 
     requests = sample_requests(
@@ -154,6 +155,7 @@ def main(args: argparse.Namespace):
     parser.add_argument("--num-shards", type=int, default=1)
     parser.add_argument("--max-num-batched-tokens", type=int, default=-1)
     parser.add_argument("--max-input-len", type=int, default=-1)
+    parser.add_argument("--min-decode-steps", type=int, default=256)
     parser.add_argument(
         "--num-prompts", type=int, default=1000, help="Number of prompts to process."
     )
diff --git a/serve/mlc_serve/run.py b/serve/mlc_serve/run.py
index ac137a84a4..c28e6d9bfc 100644
--- a/serve/mlc_serve/run.py
+++ b/serve/mlc_serve/run.py
@@ -34,6 +34,7 @@ def parse_args():
     args.add_argument("--num-shards", type=int, default=1)
     args.add_argument("--max-num-batched-tokens", type=int, default=-1)
     args.add_argument("--max-input-len", type=int, default=-1)
+    args.add_argument("--min-decode-steps", type=int, default=256)
     args.add_argument("--debug-logging", action="store_true")
     parsed = args.parse_args()
     parsed.model, parsed.quantization = parsed.local_id.rsplit("-", 1)
@@ -85,6 +86,7 @@ def run_server():
     engine = LocalProcessInferenceEngine(
         model_module,
         max_batched_tokens=args.max_num_batched_tokens,
+        min_decode_steps=args.min_decode_steps,
     )
     connector = AsyncEngineConnector(engine)
     app = create_app(connector)