mrq/torchscale
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

rollback LN epsilon in retention

Browse Source 
rollback 2c29de0fb3
This commit is contained in:
Li Dong 2023-09-27 20:40:36 +08:00 committed by GitHub
parent 258eda3308
commit d1fefe9c22
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
torchscale/component/multiscale_retention.py
View File

@ -67,7 +67,7 @@ class MultiScaleRetention(nn.Module):
self.out_proj = MultiwayWrapper(args, nn.Linear(embed_dim * self.factor, embed_dim, bias=True))
self.group_norm = MultiwayWrapper(args, LayerNorm(self.head_dim, eps=1e-6, elementwise_affine=False))
self.group_norm = MultiwayWrapper(args, LayerNorm(self.head_dim, eps=args.layernorm_eps, elementwise_affine=False))
self.reset_parameters()
def reset_parameters(self):