Add sample GPT training run (#194)

Fix deepspeed for single GPU (#187)

Add deepspeed support (#186)

Add cosine annealing to LR scheduler options (#182)

Add join run cli command (#177)

Support multi-GPU training with DataParallel (#178)

Support multi-GPU single node training by default.

Note that more advanced techniques (e.g. DistributedDataParallel or DeepSpeed) will gain even more performance.

Add save and load methods to the model (#172)

Add join sweep method (#173)

Allows adding extra agents to a sweep.

Reduce model complexity (#171)

Remove the abstract encoder/decoder classes and just have one interface for the model itself

Fix activation resampler normalization dimension (#155)