Unused Variables when using DDP Spawn

[jk-tripathy]

When tuning the base models, some variables are unused ( I presume the classification heads that we replace with our own classification head/seq2seq model). This breaks the implementation, making the code unrunnable.
This can be fixed by using ddp_spawn_find_unused_parameters_true strategy rather than ddp_spawn.
An issue discussing this can be found here

NOTE: this issue does not pop up when freezing the base model, leading more credence to the idea that the unused variables are the original classification head that the base model comes with