From d1ae5ace67c77bedfccc3213c531f2171979fe11 Mon Sep 17 00:00:00 2001
From: Terry Kong <terryk@nvidia.com>
Date: Fri, 25 Apr 2025 00:22:07 -0700
Subject: [PATCH 1/2] fix: ray.sub early exit + fixing more ports

Signed-off-by: Terry Kong <terryk@nvidia.com>
---
 ray.sub | 134 ++++++++++++++++++++++++++++++++++++++++++--------------
 1 file changed, 102 insertions(+), 32 deletions(-)

diff --git a/ray.sub b/ray.sub
index 047d5427b..4baf4460b 100644
--- a/ray.sub
+++ b/ray.sub
@@ -18,19 +18,26 @@ CONTAINER=$CONTAINER
 MOUNTS=$MOUNTS
 COMMAND=${COMMAND:-}  # This is a script relative to the SLURM_SUBMIT_DIR. If left empty, it will leave the cluster idle after it's brought up.
 ########################################################
-# Ray ports
-GCS_SERVER_PORT=${GCS_SERVER_PORT:-6379}
-DASHBOARD_PORT=${DASHBOARD_PORT:-8265}
-OBJECT_MANAGER_PORT=${OBJECT_MANAGER_PORT:-8076}
-NODE_MANAGER_PORT=${NODE_MANAGER_PORT:-8077}
-DASHBOARD_AGENT_PORT=${DASHBOARD_AGENT_PORT:-52365}
-DASHBOARD_AGENT_GRPC_PORT=${DASHBOARD_AGENT_GRPC_PORT:-52366}
-METRICS_PORT=${METRICS_PORT:-9002}
-# On our clusters, the largest port range on an idle worker appeared between 52367-64607
+# Ports for all nodes (should be odd numbers since we place head/worker[0] on the same node) so all workers get the odd ports, but the head will get +1 the ports
+NODE_MANAGER_PORT=${NODE_MANAGER_PORT:-53001}
+OBJECT_MANAGER_PORT=${OBJECT_MANAGER_PORT:-53003}
+RUNTIME_ENV_AGENT_PORT=${RUNTIME_ENV_AGENT_PORT:-53005}
+DASHBOARD_AGENT_GRPC_PORT=${DASHBOARD_AGENT_GRPC_PORT:-53007}
+METRICS_EXPORT_PORT=${METRICS_EXPORT_PORT:-53009}
+
+# Ports for the head node
+PORT=${PORT:-6379}
+RAY_CLIENT_SERVER_PORT=${RAY_CLIENT_SERVER_PORT:-10001}
+#REDIT_SHARD_PORTS=${REDIT_SHARD_PORTS:-"random"} ??
+DASHBOARD_GRPC_PORT=${DASHBOARD_GRPC_PORT:-52367}
+DASHBOARD_PORT=${DASHBOARD_PORT:-8265}  # Also used by debugger
+DASHBOARD_AGENT_LISTEN_PORT=${DASHBOARD_AGENT_LISTEN_PORT:-52365}
+
+# On our clusters, the largest port range on an idle worker appeared between 52369-64607
 # (not including the other ports set by this script). So this range is chosen to be
 # somewhere in the middle
-MIN_WORKER_PORT=${MIN_WORKER_PORT:-53001}
-MAX_WORKER_PORT=${MAX_WORKER_PORT:-53257}
+MIN_WORKER_PORT=${MIN_WORKER_PORT:-54001}
+MAX_WORKER_PORT=${MAX_WORKER_PORT:-54257}
 ########################################################
 
 # Defaults to placing uv cache inside the SLURM_SUBMIT_DIR
@@ -39,7 +46,8 @@ UV_CACHE_DIR="${UV_CACHE_DIR:-$SLURM_SUBMIT_DIR/uv_cache}"
 mkdir -p $UV_CACHE_DIR
 
 # Create logs directory
-LOG_DIR="$SLURM_SUBMIT_DIR/$SLURM_JOB_ID-logs"
+BASE_LOG_DIR=${BASE_LOG_DIR:-$SLURM_SUBMIT_DIR}/
+LOG_DIR="$BASE_LOG_DIR/$SLURM_JOB_ID-logs"
 mkdir -p $LOG_DIR
 
 COMMON_SRUN_ARGS=""
@@ -56,7 +64,7 @@ COMMON_SRUN_ARGS+=" --gres=gpu:8"
 # Number of GPUs per node
 gpus_per_node=8
 
-num_retries=5
+num_retries=3
 
 # Getting the node names and IP addresses in the SLURM allocation
 nodes=$(scontrol show hostnames "$SLURM_JOB_NODELIST")
@@ -72,7 +80,7 @@ done
 head_node=${nodes_array[0]}
 head_node_ip=${ip_addresses_array[0]}
 
-ip_head=$head_node_ip:$GCS_SERVER_PORT
+ip_head=$head_node_ip:$PORT
 
 # First we start the head of the ray cluster on one of the physical nodes
 # Set GPU/CPU resources to 0 to avoid scheduling on the head node
@@ -82,30 +90,60 @@ head_cmd=$(cat <<EOF
 # Overlapping srun commands will check this file to determine if we can overlap a container command
 touch $LOG_DIR/STARTED_RAY_HEAD
 env
+
+exit-dramatically() {
+    # Use SIGTERM to forcefully terminate the srun process
+    pkill -P $$ || true
+    kill -TERM 0 || true
+    # As a last resort, exit with a non-zero code
+    exit 1
+}
+export -f exit-dramatically
+
+# Background process to check for ENDED file
+monitor-sidecar() {
+  set +x
+  while true; do
+    sleep 60
+    if [[ -f "$LOG_DIR/ENDED" ]]; then
+      echo "Detected ENDED file, terminating..."
+      exit-dramatically
+    fi
+  done
+}
+monitor-sidecar &
+
 cat <<EOFINNER | tee /launch-head.sh
 ray start --head \
---disable-usage-stats \
---num-cpus=0 \
---num-gpus=0 \
---node-ip-address="$head_node_ip" \
---port=${GCS_SERVER_PORT} \
---dashboard-port=${DASHBOARD_PORT} \
---object-manager-port=${OBJECT_MANAGER_PORT} \
---node-manager-port=${NODE_MANAGER_PORT} \
---metrics-export-port=${METRICS_PORT} \
---dashboard-agent-grpc-port=${DASHBOARD_AGENT_GRPC_PORT} \
---dashboard-agent-listen-port=${DASHBOARD_AGENT_PORT} \
---block
+    --disable-usage-stats \
+    --num-cpus=0 \
+    --num-gpus=0 \
+    --node-ip-address="$head_node_ip" \
+    --port=${PORT} \
+    --ray-client-server-port=${RAY_CLIENT_SERVER_PORT} \
+    --dashboard-grpc-port=${DASHBOARD_GRPC_PORT} \
+    --dashboard-port=${DASHBOARD_PORT} \
+    \
+    --node-manager-port=$((${NODE_MANAGER_PORT} + 1)) \
+    --object-manager-port=$((${OBJECT_MANAGER_PORT} + 1)) \
+    --runtime-env-agent-port=$((${RUNTIME_ENV_AGENT_PORT} + 1)) \
+    --dashboard-agent-grpc-port=$((${DASHBOARD_AGENT_GRPC_PORT} + 1)) \
+    --dashboard-agent-listen-port=$((${DASHBOARD_AGENT_LISTEN_PORT} + 1)) \
+    --metrics-export-port=$((${METRICS_EXPORT_PORT} + 1)) \
+    \
+    --block
 EOFINNER
 chmod +x /launch-head.sh
 
 count=0
-while true; do
+while [[ \$count -lt $num_retries ]]; do
   bash /launch-head.sh
   count=\$((count+1))
-  echo "Head node failed \$count times, restarting..."
+  echo "Head node failed \$count/$num_retries times, restarting in 5 seconds..."
+  sleep 5
 done
-echo ret_code=\$?
+touch $LOG_DIR/ENDED
+exit 1
 EOF
 )
 srun $COMMON_SRUN_ARGS --container-name=ray-head --nodes=1 --ntasks=1 -w "$head_node" -o $LOG_DIR/ray-head.log bash -x -c "$head_cmd" &
@@ -120,22 +158,54 @@ for ((i = 0; i < SLURM_JOB_NUM_NODES; i++)); do
     
   worker_cmd=$(cat <<EOF
 env
+
+exit-dramatically() {
+    # Use SIGTERM to forcefully terminate the srun process
+    pkill -P $$ || true
+    kill -TERM 0 || true
+    # As a last resort, exit with a non-zero code
+    exit 1
+}
+
+# Background process to check for ENDED file
+monitor-sidecar() {
+  set +x
+  while true; do
+    sleep 60
+    if [[ -f "$LOG_DIR/ENDED" ]]; then
+      echo "Detected ENDED file, terminating..."
+      exit-dramatically
+    fi
+  done
+}
+monitor-sidecar &
+
 cat <<EOFINNER | tee /launch-worker.sh
 ray start --address "$ip_head" \
           --disable-usage-stats \
           --resources="{\"worker_units\": $gpus_per_node}" \
           --min-worker-port=${MIN_WORKER_PORT} \
           --max-worker-port=${MAX_WORKER_PORT} \
+          \
+          --node-manager-port=${NODE_MANAGER_PORT} \
+          --object-manager-port=${OBJECT_MANAGER_PORT} \
+          --runtime-env-agent-port=${RUNTIME_ENV_AGENT_PORT} \
+          --dashboard-agent-grpc-port=${DASHBOARD_AGENT_GRPC_PORT} \
+          --dashboard-agent-listen-port=${DASHBOARD_AGENT_LISTEN_PORT} \
+          --metrics-export-port=${METRICS_EXPORT_PORT} \
+          \
           --block
 EOFINNER
 
 count=0
-while true; do
+while [[ \$count -lt $num_retries ]]; do
   bash /launch-worker.sh
   count=\$((count+1))
-  echo "Worker failed \$count times, restarting..."
+  echo "Worker failed \$count/$num_retries times, restarting in 5 seconds..."
+  sleep 5
 done
-echo ret_code=\$?
+touch $LOG_DIR/ENDED
+exit 1
 EOF
 )
   if [[ $i -eq 0 ]]; then

From e581babf48157de93f976bdd5f389f73f336ec28 Mon Sep 17 00:00:00 2001
From: Terry Kong <terryk@nvidia.com>
Date: Fri, 25 Apr 2025 00:22:53 -0700
Subject: [PATCH 2/2] tiny

Signed-off-by: Terry Kong <terryk@nvidia.com>
---
 ray.sub | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ray.sub b/ray.sub
index 4baf4460b..0fe8cbdcd 100644
--- a/ray.sub
+++ b/ray.sub
@@ -46,7 +46,7 @@ UV_CACHE_DIR="${UV_CACHE_DIR:-$SLURM_SUBMIT_DIR/uv_cache}"
 mkdir -p $UV_CACHE_DIR
 
 # Create logs directory
-BASE_LOG_DIR=${BASE_LOG_DIR:-$SLURM_SUBMIT_DIR}/
+BASE_LOG_DIR=${BASE_LOG_DIR:-$SLURM_SUBMIT_DIR}
 LOG_DIR="$BASE_LOG_DIR/$SLURM_JOB_ID-logs"
 mkdir -p $LOG_DIR