[Jack000]

note that the fairseq implementation requires 2x 32GB v100 gpus for inference.

[visarga]

And that's why we won't see this deployed in production anywhere.