Megatron-DeepSpeed
7145f6df
- I think bias starts to diverge first
Go
Login via GitHub
Home
Pricing
FAQ
Install
Login
via GitHub
Commit
View On
GitHub
Commit
3 years ago
I think bias starts to diverge first
References
#270 - Test different layer norm
#271 - Sync layer norm
#274 - Sync 4 layer norms - bf16, fp32, optimizer states on restart
Author
thomasw21
Parents
a5e32958
Loading