[Bharath1234]

Since the Transformer decoder cannot be parallelized during inference, how can it be cost effective at scale?!