accelerate
Accelerate is a library that enables the same PyTorch code to be run across any distributed configuration.
TLDR
Print environment information
$ accelerate env
Interactively create a configuration file
$ accelerate config
Print the estimated GPU memory cost of running a huggingface model with different data types
$ accelerate estimate-memory [name/model]
Test an Accelerate configuration file
$ accelerate test --config_file [path/to/config.yaml]
Run a model on CPU with Accelerate
$ accelerate launch [path/to/script.py] [--cpu]
Run a model on multi-GPU with Accelerate, with 2 machines
$ accelerate launch [path/to/script.py] --multi_gpu --num_machines [2]