GreenRay AIO lab helper — stands in for a physical/UI action, not a real tool

Usage: dataloader-bench [options]

Runs only the input pipeline of the training job configured on this node
(no model) and reports the samples/s it can deliver, compared with what
the GPUs consume. On a real cluster you would do this with your
framework's own benchmark mode (for example a DALI pipeline or a PyTorch
DataLoader loop without the model).

Options:
  -c, --config FILE      job environment file to read (default: the job's own)
  -d, --data DIR         override DATA_DIR from the config
  -w, --workers N        override NUM_WORKERS (loader workers per GPU)
      --gds 0|1          override USE_GDS (read through GPUDirect Storage)
  -h, --help             show this help and exit

Options also accept the --option=value form. Overrides apply to this run
only; edit the config file to change what the job uses.
