DDP, MPIEnv and numdevices

[timocafe]

Hello,

I am confused by the API and the different scenario using the MPI plugin. Let's consider a single node 2 GPUs (XPU I am working for intel). I want to run on the two GPUs only.

Using DDP setting and basic training, if I do python main.py:

trainer = L.Trainer(
  devices=2, 
  strategy=ddp,
  accelerator='xpu')

The trainer will spawn 1 additional process and do parallel training.

Now if I plug MPIEnvironment(), and start my job mpirun -n 2 python main.py.
Do I still need to specify devices = 2 in the trainer ?

trainer = L.Trainer(
  plugins=[MPIEnvironment()],
  # devices=2, needed or not ? 
  strategy=ddp,
  accelerator='xpu')