我想测试多机时开启容器节点间的通信是否通畅，结果报错：

[W1117 10:11:23.318584088 socket.cpp:204] [c10d] The hostname of the client socket cannot be retrieved. err=-3

[W1117 10:11:23.141307645 socket.cpp:204] [c10d] The hostname of the client socket cannot be retrieved. err=-3

[W1117 10:11:33.324091075 socket.cpp:204] [c10d] The hostname of the client socket cannot be retrieved. err=-3

[W1117 10:11:43.334863788 socket.cpp:204] [c10d] The hostname of the client socket cannot be retrieved. err=-3

容器脚本:

#!/bin/bash -l
#SBATCH --comment=xxx
#SBATCH --error=%j.err
#SBATCH --output=%j.out
#SBATCH --partition=g078t
#SBATCH --job-name=megatron-gpt-mt
#SBATCH --exclusive
##SBATCH --qos=boost_qos_dbg # if you need to run debug, max 30 mins 2 nodes
#SBATCH --time=00:30:00
#SBATCH --nodes=2
#SBATCH --nodelist=gpu9,gpu10
#SBATCH --exclusive
#SBATCH --gres=gpu:8
#SBATCH --mem=0
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=64

nodes=( $( scontrol show hostnames $SLURM_JOB_NODELIST ) )
nodes_array=($nodes)
head_node=${nodes_array[0]}
head_node_ip=$(srun --nodes=1 --ntasks=1 -w "$head_node" hostname --ip-address)

export MASTER_ADDR=${head_node_ip}
export MASTER_PORT=6000
export WORLD_SIZE=$((SLURM_NNODES * SLURM_GPUS_PER_NODE))

export NCCL_SOCKET_IFNAME=ib0
export NCCL_IB_DISABLE=0
export NCCL_IB_CUDA_SUPPORT=1
export NCCL_DEBUG=INFO

srun singularity exec \
    --nv --pwd /home/u2025110710/repos/megatron -B /opt/app/nvidia/535.154.05/lib/,/opt/app/cuda/12.1/,/home/u2025110710/.triton,/home/u2025110710/.cache,/home/u2025110710/.local,/home/u2025110710/.nv,/home/u2025110710/.pki,/home/u2025110710/models/,/home/u2025110710/datasets/,/home/u2025110710/repos/megatron,/home/u2025110710/containers/megatron/py312_torch260:/py312_torch260 megatron \
    bash -c "
        export SLURM_NNODES=$SLURM_NNODES
        export SLURM_GPUS_PER_NODE=8
        export MASTER_ADDR=$MASTER_ADDR
        export MASTER_PORT=$MASTER_PORT
        export WORLD_SIZE=$WORLD_SIZE
        source /set_env.sh
        bash /run_comm_test.sh --tensor-size=5000000 --test-rounds=8
    "

torchrun --nnodes=2 --nproc_per_node=8 --master_addr=192.168.0.11 --master_port=6000 dist_comm_test.py --tensor-size=1024000 --dtype=float32 --backend=nccl --warmup=3 --test-rounds=5