start using RMSNorm, used in Gopher and AlphaCode, and as a way to go complete bias-less (purportedly more stable according to PaLM)

dalle2_pytorch/dalle2_pytorch.py

@@ -44,10 +44,21 @@ def freeze_model_and_make_eval_(model):
 
 # diffusion prior
 
+class RMSNorm(nn.Module):
+    def __init__(self, dim, eps = 1e-8):
+        super().__init__()
+        self.scale = dim ** -0.5
+        self.eps = eps
+        self.g = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        norm = torch.norm(x, dim = -1, keepdim = True) * self.scale
+        return x / norm.clamp(min = self.eps) * self.g
+
 def FeedForward(dim, mult = 4, dropout = 0.):
     inner_dim = int(mult * dim)
     return nn.Sequential(
-        nn.LayerNorm(dim),
+        RMSNorm(dim),
         nn.Linear(dim, inner_dim, bias = False),
         nn.GELU(),
         nn.Dropout(dropout),
@@ -67,7 +78,7 @@ class Attention(nn.Module):
         self.scale = dim_head ** -0.5
         inner_dim = dim_head * heads
 
-        self.norm = nn.LayerNorm(dim)
+        self.norm = RMSNorm(dim)
         self.dropout = nn.Dropout(dropout)
 
         self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
@@ -124,7 +135,7 @@ class Transformer(nn.Module):
                 FeedForward(dim = dim, mult = ff_mult, dropout = ff_dropout)
             ]))
 
-        self.norm = nn.LayerNorm(dim) if norm_out else nn.Identity()  # unclear in paper whether they projected after the classic layer norm for the final denoised image embedding, or just had the transformer output it directly: plan on offering both options
+        self.norm = RMSNorm(dim) if norm_out else nn.Identity()  # unclear in paper whether they projected after the classic layer norm for the final denoised image embedding, or just had the transformer output it directly: plan on offering both options
 
     def forward(
         self,