diff --git a/benchmarks/vllm_cluster_bench.py b/benchmarks/vllm_cluster_bench.py
index 26c2f62..56e4db7 100755
--- a/benchmarks/vllm_cluster_bench.py
+++ b/benchmarks/vllm_cluster_bench.py
@@ -181,6 +181,10 @@ def get_cluster_env():
     env["NCCL_IB_DISABLE"] = "0"
     env["NCCL_NET_GDR_LEVEL"] = "0"
     
+    # Stability for RDMA (Fix for high-throughput models like Gemma 3)
+    env["NCCL_IB_TIMEOUT"] = "23"  # ~32 seconds (default is 18/~1s)
+    env["NCCL_IB_RETRY_CNT"] = "7" # Default is 3, increase for lossy networks
+    
     return env
 
 def get_model_args(model):
diff --git a/scripts/start_vllm_cluster.py b/scripts/start_vllm_cluster.py
index 9fcfb7f..d6bb561 100755
--- a/scripts/start_vllm_cluster.py
+++ b/scripts/start_vllm_cluster.py
@@ -148,6 +148,9 @@ def setup_worker_node(worker_ip, head_ip):
     export RDMA_IFACE=$(ip -o addr show to {subnet} | awk '{{print $2}}' | head -n1)
     export NCCL_SOCKET_IFNAME=$RDMA_IFACE
     export GLOO_SOCKET_IFNAME=$RDMA_IFACE
+    # Stability for RDMA
+    export NCCL_IB_TIMEOUT=23
+    export NCCL_IB_RETRY_CNT=7
     echo "Starting Ray Worker on {worker_ip} connecting to {head_ip}..."
     ray start --address='{head_ip}:6379' --num-gpus=1 --num-cpus=8 --disable-usage-stats
     """
@@ -183,6 +186,9 @@ def setup_head_node(head_ip):
     export RDMA_IFACE=$(ip -o addr show to {subnet} | awk '{{print $2}}' | head -n1)
     export NCCL_SOCKET_IFNAME=$RDMA_IFACE
     export GLOO_SOCKET_IFNAME=$RDMA_IFACE
+    # Stability for RDMA
+    export NCCL_IB_TIMEOUT=23
+    export NCCL_IB_RETRY_CNT=7
     echo "Starting Ray Head on {head_ip}..."
     ray start --head --port=6379 --node-ip-address={head_ip} --num-gpus=1 --num-cpus=8 --disable-usage-stats
     """