ESPnet3 Multi-GPU And Cluster Examples

Masao SomekiLess than 1 minute

ESPnet3 Multi-GPU And Cluster Examples

Runner-based workloads can scale through the parallel config.

Local multi-GPU

parallel:
  env: local_gpu
  n_workers: 2
  options: {}

This uses dask_cuda.LocalCUDACluster when available.

SLURM example

parallel:
  env: slurm
  n_workers: 4
  options:
    queue: gpu
    cores: 8
    processes: 1
    memory: 32GB
    walltime: "04:00:00"
    job_extra_directives:
      - "--gres=gpu:1"

The selected options are forwarded to the corresponding Dask cluster class.