BlinkDL
|
710d3e34b7
|
better init for RWKV
|
4 years ago |
|
BlinkDL
|
619ed00e4b
|
misc improvement
|
4 years ago |
|
BlinkDL
|
3329161ed7
|
rapid convergence using ZERO initialization
|
4 years ago |
|
BlinkDL
|
7f391c5758
|
+ RWKV tiny-attn and now it's great for ctx 1024 or 2048
|
4 years ago |
|
BlinkDL
|
4ffd8f1b76
|
+ new comparison
|
4 years ago |
|
BlinkDL
|
ad627311f4
|
clean init code
|
4 years ago |
|
BlinkDL
|
c675b47705
|
misc improvements
|
4 years ago |
|
BlinkDL
|
ef29f4b9e8
|
fixed nan loss
|
4 years ago |
|
BlinkDL
|
4fd8716976
|
improve RWKV time_w initialization
|
4 years ago |
|
BlinkDL
|
a31a3b2e92
|
+ MHA_shift
|
4 years ago |
|
BlinkDL
|
3b9005ea11
|
RWKV: now faster and less params
|
4 years ago |
|
BlinkDL
|
546114c6a5
|
still use layernorm for everything
|
4 years ago |
|
BlinkDL
|
fd098b1d2e
|
small update
|
4 years ago |
|
BlinkDL
|
3b60c5b266
|
add wandb, and rename variables
|
4 years ago |
|
BlinkDL
|
440bebff1a
|
fixed nan in large models
|
4 years ago |
|
BlinkDL
|
62e2cb06d6
|
fixing nan in large models
|
4 years ago |
|
BlinkDL
|
d699a69169
|
misc improvements
|
4 years ago |
|
BlinkDL
|
6266f481da
|
minor changes
|
4 years ago |
|
BlinkDL
|
89eab46e60
|
+ info
|
4 years ago |
|
BlinkDL
|
e9fbd9bf70
|
remove layernorm -> better RWKV
|
4 years ago |
|
BlinkDL
|
447eae5841
|
add MHA-plus model
|
4 years ago |
|
BlinkDL
|
aa4e2a68f4
|
first commit
|
4 years ago |