bash_test_ccl_multinode.sh

#!/bin/bash -l
#SBATCH --job-name=ccl-multi
#SBATCH --account=AIRR-P51-DAWN-GPU
#SBATCH --partition=pvc9
#SBATCH --nodes=2
#SBATCH --gres=gpu:4
#SBATCH --ntasks-per-node=8
#SBATCH --cpus-per-task=12
#SBATCH --time=02:00:00
#SBATCH --output=Logs/test_ccl_multi_%j.out
#SBATCH --error=Logs/test_ccl_multi_%j.err
#SBATCH --exclude=pvc-s-135

# CCL Multi-Node Stress Test: reproduce epoch-boundary DDP hang.
# Single-node test passed all phases — the hang is multi-node specific (OFI transport).
#
# 2 nodes x 4 XPU cards x 2 tiles/card = 16 XPU tiles

cd /rds/project/rds-TWhPgQVLKbA/Code/OmniMorph

. /etc/profile.d/modules.sh
module purge
module load rhel9/default-dawn

source ~/miniconda3/etc/profile.d/conda.sh
conda activate ~/rds/rds-airr-p51-TWhPgQVLKbA/Env/pub_env/pytorch-xpu

# --- CCL/MPI setup ---
export I_MPI_PMI_LIBRARY=/usr/local/software/slurm/current-rhel8/lib/libpmi2.so
export I_MPI_HYDRA_BOOTSTRAP=slurm
export CCL_WORKER_AFFINITY=auto
export CCL_ZE_CACHE_OPEN_IPC_HANDLES_THRESHOLD=100000

# --- XPU memory allocator ---
export PYTORCH_ALLOC_CONF=expandable_segments:True,max_split_size_mb:512

# --- Multi-node setup ---
export MASTER_ADDR=$(scontrol show hostname "$SLURM_NODELIST" | head -n1)
export MASTER_PORT=12355

SRUN_TIMEOUT=3600  # 1h timeout

echo "============================================"
echo "CCL Multi-Node Stress Test"
echo "============================================"
echo "Job ID:        $SLURM_JOB_ID"
echo "Nodes:         $SLURM_NNODES"
echo "Tasks/node:    $SLURM_NTASKS_PER_NODE"
echo "Total tasks:   $SLURM_NTASKS"
echo "CPUs/task:     $SLURM_CPUS_PER_TASK"
echo "Master:        $MASTER_ADDR:$MASTER_PORT"
echo "Node list:     $SLURM_NODELIST"
echo "Walltime:      02:00:00"
echo "Timeout/srun:  ${SRUN_TIMEOUT}s"
echo "CCL_ZE_CACHE:  $CCL_ZE_CACHE_OPEN_IPC_HANDLES_THRESHOLD"
echo "============================================"
echo "Start time:    $(date)"
echo "============================================"

timeout $SRUN_TIMEOUT srun --kill-on-bad-exit=1 bash -c '
export LOCAL_RANK=$SLURM_LOCALID
export RANK=$SLURM_PROCID
export WORLD_SIZE=$SLURM_NTASKS
export MASTER_ADDR='"$MASTER_ADDR"'
export MASTER_PORT='"$MASTER_PORT"'
python tests/test_ccl_stress.py \
    --spatial-size 64 \
    --batch-size 2 \
    --phase1-steps 200 \
    --phase3-steps 10 \
    --phase4-steps 10 \
    --phase5-broadcasts 50
'
EXIT_CODE=$?

echo "============================================"
echo "End time:      $(date)"
echo "Exit code:     $EXIT_CODE"
if [ $EXIT_CODE -eq 0 ]; then
    echo "Result:        ALL PHASES PASSED"
    echo "  Multi-node CCL hang NOT reproduced with simple model."
    echo "  The issue may require the actual OM_net architecture or training loop specifics."
elif [ $EXIT_CODE -eq 124 ]; then
    echo "Result:        TIMEOUT (CCL hang detected)"
    echo "  Check the last 'PHASE N START' in the log to identify the hanging phase."
    echo "  This confirms the hang is in the CCL inter-node (OFI) transport."
else
    echo "Result:        FAILED (exit code $EXIT_CODE)"
fi
echo "============================================"