For training jobs, if you specify --accelerators/--accelerators-limits and not --instance-type , it will request
resources:
limits:
nvidia.com/gpu: '0'
requests:
nvidia.com/gpu: '0'
independently of how many accelerators you request. If you specify --instance-type additionally, it will use the correct value from --accelerators/--accelerators-limits for some reason. But: you cannot use --accelerators/--accelerators-limits and --node-count at the same time, it will give a validation error.
Thus: You cannot submit multi-node jobs with manually specified resources.