Question: Different environment variables set when using srun or sbatch

[Sohn123]

Hello,
I ran a container interactively using srun to do some development. Then I wanted to start a job using sbatch based on that container. I now have the problem, that the environment variables inside the container differ when running the container using sbatch instead of srun.

This is the srun command I used:

#!/bin/bash -eux
srun \
        --account students \
        --container-image /data/pytorch.sqsh \
        --container-mounts /data:/data \
        --container-name test_container1 \
        --container-writable \
        --partition gpu-interactive \
        --cpus-per-task 64 \
        --mem 64gb \
        --gpus 1 \
        --time 8:00:00 \
        --nodelist gx01 \
       python training.py

the file pytorch.sqsh was created by running: enroot import -o pytorch.sqsh 'docker://nvcr.io#nvidia/pytorch:25.05-py3'

I created the following sbatch file based on the srun script:

#!/bin/bash -eux
#SBATCH --job-name=finetuning
#SBATCH --output=finetuning_%j.log
#SBATCH --error=finetuning_%j.err
#SBATCH --account=students
#SBATCH --container-image=/data/pytorch.sqsh
#SBATCH --container-mounts=/data:/data
#SBATCH --container-name=test_container1
#SBATCH --container-writable
#SBATCH --partition=gpu
#SBATCH --cpus-per-task=64
#SBATCH --mem=64gb
#SBATCH --gpus=1
#SBATCH --time=48:00:00
#SBATCH --nodelist=gx01

python training.py

When running training.py pytorch complains, that the environment variable WORLD_SIZE is not set. By running env in both scenarios I could verify that WORLD_SIZE is indeed only set when using srun. What could be the reason for the different environment variables? The only thing that differs is the partition I am running the job on. Could that be the reason for a difference in environment variables?