DeepSpeed
f26900ec
- Modify normalize kernels to save variance in floating-point for backward
Go
Login via GitHub
Home
Pricing
FAQ
Install
Login
via GitHub
Commit
View On
GitHub
Commit
4 years ago
Modify normalize kernels to save variance in floating-point for backward
References
#1404 - Transformer/fix layer norm
Author
Reza Yazdani
Parents
4017a7c9
Loading