Using DDPPlugin changes accelerator to ddp

[Rizhiy]

🐛 Bug

If DDPPlugin is passed to pl.Trainer

Please reproduce using the BoringModel

https://colab.research.google.com/drive/1XsSQoQghHhIxA-9al9dnsjbBJOxzo1Ml?usp=sharing

To Reproduce

As can be seen in output of "Run Test", ddp is initialized even though dp is chosen as accelerator.

Expected behavior

Either ignore the plugin or raise an error about incompatability.

Environment

• CUDA:
  • GPU:
    • Tesla T4
  • available: True
  • version: 10.1
• Packages:
  • numpy: 1.19.5
  • pyTorch_debug: False
  • pyTorch_version: 1.8.1+cu101
  • pytorch-lightning: 1.3.3
  • tqdm: 4.41.1
• System:
  • OS: Linux
  • architecture:
    • 64bit
  • processor: x86_64
  • python: 3.7.10
  • version: Proposal for help #1 SMP Tue Apr 20 19:55:43 PDT 2021

Additional context

[Rizhiy]

As a consequence, DDPPlugin(find_unused_parameters=False) cannot be used with ddp_spawn.

[awaelchli]

Directly related to #6090, we want to simplify how the accelerator and plugins are selected. This will also enable better checking of mismatching combinations.

[kaushikb11]

@Rizhiy @awaelchli Picking up #6090 next week, it would simplify things.

[awaelchli]

Sounds great!

[Rizhiy]

Decided to revisit this issue and found out that to use find_unused_parameters with ddp_spawn I have to use DDPSpawnPlugin instead of DDPPlugin.
Probably would be a good idea to add it to documentation.