Raise if fsdp plugin is unset

[alex-jw-brooks]

When we try to run multigpu training, we explicitly set:

trainer.accelerator.state.fsdp_plugin.auto_wrap_policy = fsdp_auto_wrap_policy(model)

before launching the trainer. However, the Accelerator state doesn't always define the fsdp_plugin attribute, so this can crash in a cryptic way if we launch with Torchrun for multigpu with the wrong settings.

The fsdp_plugin of the Accelerator state for multigpu configurations is set here in the Accelerate source. If you don't set ACCELERATE_USE_FSDP, it won't be defined.

This PR explicitly checks to see if the accelerator state has an fsdp_plugin before trying to update the auto wrap policy; if it doesn't, it raises an attribute error with a hint that you probably need to set ACCELERATE_USE_FSDP to run with a multigpu configuration.

[fabianlim]

@alex-jw-brooks @Ssukriti I dont think his change is needed if #53 is merged. This is because we now explicitly check trainer.is_fsdp_enabled, which will be True only if FSDP is enabled.

[alex-jw-brooks]

Hey @fabianlim, that sounds good - I'm not super familiar with the Accelerate codebase around the plugin, but I assume that will handle the current attribute error and that the fsdp_plugin will always be defined.

I know that a lot of default args are set to None in Accelerate, including the fsdp_plugin in the AcceleratorState, so it might be a good idea to take a closer look at some point to make sure that we can't end up in a situation where fsdp_plugin is a defined state attribute that is set to None, since that would break similarly if it were possible. For now, let's close this PR in favor of the other one though!