Distributed Training

Chutes SN64 says Parallax trained without blocking GPUs

Chutes says Parallax trained a recurrent model across distributed GPUs with a stated 0.6 percent quality gap versus centralized training.