Decoder Mask

[Muennighoff]

2 quick q's on the decoder:

• Because of the neat cumsum, are we treating the decoder masks the same as the encoder masks (i can't find any different treatment in thecode except for the context route map) ?
• the cumsum at the same time seems to be memory intensive; for an enc-dec / dec-only ~4096 toks text summarizer would you go with the performer (if yes enc-dec or discard the enc) ?

& awesome repo as always!