patch

default to 4 residual streams, hyper-connections paper from bytedance

x-transformers fix for learned value residual in presence of cross at…

…tention

turn on a new effective technique from x-transformers

make lfq loss positive with softplus

turn on a breakthrough technique for vq

update to latest cfg lib

just pin linear attention version

just pin linear attention version

address #92 (comment) again