Merge pull request #172 from VectorInstitute/f/random-ports

Fix: Randomize port search to avoid port collision during model loading

Author: XkunW

vec_inf/config/models.yaml

@@ -1059,12 +1059,11 @@ models:
     model_family: gpt-oss
     model_variant: 120b
     model_type: LLM
-    gpus_per_node: 4
-    num_nodes: 2
+    gpus_per_node: 2
+    num_nodes: 1
     vocab_size: 201088
     time: 08:00:00
     resource_type: l40s
     vllm_args:
-      --tensor-parallel-size: 4
-      --pipeline-parallel-size: 2
-      --max-model-len: 40960
+      --tensor-parallel-size: 2
+      --max-model-len: 32768

vec_inf/find_port.sh

@@ -28,7 +28,16 @@ find_available_port() {
     local base_port=$2
     local max_port=$3
 
-    for ((port=base_port; port<=max_port; port++)); do
+    # Generate shuffled list of ports; fallback to sequential if shuf not present
+    if command -v shuf >/dev/null 2>&1; then
+        local port_list
+        port_list=$(shuf -i "${base_port}-${max_port}")
+    else
+        local port_list
+        port_list=$(seq $base_port $max_port)
+    fi
+
+    for port in $port_list; do
         if is_port_available $ip $port; then
             echo $port
             return