Not working with ddp_cpu

[lvoegtlin]

Describe the bug
If we want to run the framework with ddp_cpu as accelerator it wont work as it has a working directory problem.

To Reproduce
python run.py trainer.accelerator='ddp_cpu' trainer.precision=32

Expected behavior
We can use ddp_cpu to debug our system

Additional context
To avoid this problem at the moment we can just use the full path to the run.py file ($PWD/run.py).

Checklist

• Add a warning if ddp_cpu and not presicion=32

[lvoegtlin]

When I run debug trainer with Pytorch lightning 1.3.8 I do not get an error! python run.py trainer=debug_trainer to run it. CAn you test that too @powl7 ?

[lvoegtlin]

We can add a user warning if he uses the ddp_cpu accelerator and the precision is not 32

[powl7]

Both commands work for me. No error:
python run.py trainer.accelerator='ddp_cpu' trainer.precision=32
python run.py trainer=debug_trainer