--- up.sh
+++ up.sh
@@ -17,5 +17,6 @@
   q=$(squeue -u "$USER" -h -n "llm-$NAME" -o '%i %T %N' | tr '\n' ' ')
   [ -n "$q" ] && { echo "already queued/running: $q"; exit 0; }
 fi
-sbatch --parsable -J "llm-$NAME" -p "$PARTITION" -G "$GPUS" -c "$CPUS" --mem "$MEM" -t "$TIME" \
+# One vLLM process uses local tensor-parallel GPUs; never spread them across nodes.
+sbatch --parsable --nodes=1 -J "llm-$NAME" -p "$PARTITION" -G "$GPUS" -c "$CPUS" --mem "$MEM" -t "$TIME" \
   -o "$B/logs/$NAME.%j.out" ${AFTER:+--dependency=afterany:$AFTER} "$B/serve.sh" "$NAME"