BlinkDL
|
50587bd65f
|
fix for jit modules
|
3 years ago |
BlinkDL
|
2b4539cd08
|
faster (use torch 1.12.1+cu116 or newer)
|
3 years ago |
BlinkDL
|
2f33901c10
|
no message
|
3 years ago |
BlinkDL
|
09c76b185a
|
no message
|
3 years ago |
BlinkDL
|
2815260d83
|
better
|
3 years ago |
BlinkDL
|
dc7e0802d0
|
faster
|
3 years ago |
BlinkDL
|
c43a17cfb3
|
10% faster training
|
3 years ago |
BlinkDL
|
73b96705d7
|
+ fp32 mode (slow but good for verification)
|
3 years ago |
BlinkDL
|
6299c087a4
|
fixed VRAM consumpition
|
3 years ago |
BlinkDL
|
68c486ad10
|
supports RWKV-4 pile models
|
3 years ago |
BlinkDL
|
61b7c429df
|
no message
|
3 years ago |
BlinkDL
|
13bb641007
|
no message
|
3 years ago |
BlinkDL
|
f79137b524
|
supports megatron bin+idx format
|
3 years ago |
BlinkDL
|
083f9504c6
|
+ bf16 mode (more stable)
|
3 years ago |
BlinkDL
|
6667ad18c2
|
more training tips
|
3 years ago |
BlinkDL
|
165dfd1b9e
|
RWKV-4 with DeepSpeed & FP16 & Better CUDA Kernel
|
3 years ago |