Mentions légales du service

Skip to content

add single GPU training support for PyTorch

VESIN Marc requested to merge feature/168-gpu-support into develop

  • implement #168 (closed) single GPU training support for PyTorch.

    • node side: use --gpu, --gpu-num GPU_NUM, --gpu-only tags
    • researcher side: use model_args item 'use_gpu': True
    • no support for: PyTorch single GPU for aggregation (researcher side), PyTorch multi-GPU (not tested, may work with low-level more complex code handling specificities of each node), other training plans (sklearn ignores requests to train on GPU)

  • collateral:

    • fix training plans prototypes to handle model_args
    • fix node error (bad ErrorMessages, nodes crashes when trying to emit an error message)

Merge request reports