Hello Team,
I want to ask your help to understand/fix the following error.
I’m setting up a batch_connect app to configure and start a vLLM + OpenHands session in our local cluster.
The error happens during the slurm script execution on the compute node. I can see the host portion of the `connection.yml` is populated, and then the permission denied showing in the output.log:
Script starting...
Generating connection YAML file...
/var/spool/slurm/d/job33111772/slurm_script: line 144: /gpfs/home/myuser/ondemand/data/sys/dashboard/batch_connect/dev/ood-vllm/output/0a37d2a6-6208-4c3a-95a8-7ca95943b3dc/script.sh: Permission denied
When I open the `connection.yml` file I can see only the host is there:
```
host: hpcgpu020
port:
password:
```
The template script.sh uses Python socket library to find a free port for vLLM. I’m guessing that is the problem, but I cannot understand/explain exactly why.
Here is how that script looks:
#!/usr/bin/env bash
set -euo pipefail
cd "${HOME}"
CONDA_BASE_FALLBACK="/gpfs/hpc_fs/scratch/conda"
VLLM_CONDA_ENV="vllm-serve"
OPENHANDS_CONDA_ENV="openhands-ai"
# Shared HuggingFace model cache on GPFS.
HF_HOME_SHARED="/gpfs/hpc_fs/scratch/hf-cache"
# Outbound HTTP proxy used by compute nodes.
PROXY_HOST="hpclogin01"
PROXY_PORT="3128"
# =============================================================================
# 1. ENVIRONMENT
# =============================================================================
module purge 2>/dev/null || true
module load miniforge3 2>/dev/null || true
CONDA_BASE=$(conda info --base 2>/dev/null || echo "${CONDA_BASE_FALLBACK}")
echo "[conda] base environment: ${CONDA_BASE}"
_HOST=$(hostname)
_HOST_IP=$(hostname -i 2>/dev/null | awk '{print $1}' || echo "127.0.0.1")
# =============================================================================
# 2. PORT DISCOVERY
# Two ports: one for vLLM, one reserved for OpenHands.
# =============================================================================
_free_port() {
python3 -c \
'import socket; s=socket.socket(); s.bind(("",0)); p=s.getsockname()[1]; s.close(); print(p)'
}
VLLM_PORT=$(_free_port)
OPENHANDS_PORT=$(_free_port)
# =============================================================================
# 3. WRITE connection.yml
# OOD reads this YAML file and makes every key available as a local variable
# in view.html.erb. Starting with vllm_ready=false; the health-check loop
# flips that flag to "true" once /health responds.
# =============================================================================
STAGED_ROOT="/gpfs/hpc_fs/home/hpcuser/ondemand/data/sys/dashboard/batch_connect/dev/ood-vllm/output/0a37d2a6-6208-4c3a-95a8-7ca95943b3dc"
CONNECTION_YML="${STAGED_ROOT}/connection.yml"
cat > "${CONNECTION_YML}" << EOL
---
host: ${_HOST}
port: ${VLLM_PORT}
vllm_port: ${VLLM_PORT}
openhands_port: ${OPENHANDS_PORT}
model_repo_id: "${MODEL_REPO_ID}"
model_served_name: "${MODEL_SERVED_NAME}"
hardware_pool: "${HARDWARE_POOL}"
gpu_count: "${GPU_COUNT}"
enable_openhands: "${ENABLE_OPENHANDS}"
vllm_ready: "false"
EOL
echo "[ood] connection.yml initialised at ${CONNECTION_YML}"
