broccoli-ml 15.3.0__tar.gz → 15.4.0__tar.gz

{broccoli_ml-15.3.0 → broccoli_ml-15.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: broccoli-ml
-Version: 15.3.0
+Version: 15.4.0
 Summary: Some useful Pytorch models, circa 2025
 License: MIT
 Author: Nicholas Bailey

{broccoli_ml-15.3.0 → broccoli_ml-15.4.0}/broccoli/vit.py

@@ -520,9 +520,10 @@ class ViT(nn.Module):
                 "SwiGLU": SwiGLU,
             }[transformer_activation]
 
-        # Set alpha and beta according to Microsoft's DeepNorm
-        self.alpha = (2 * transformer_layers) ** 0.25
-        self.beta = (8 * transformer_layers) ** -0.25
+        # Set alpha according to Microsoft's DeepNorm if layers > 50
+        if transformer_layers > 50:
+            self.alpha = (2 * transformer_layers) ** 0.25
+            # beta is not needed as we norm the Q and K vectors in MSA!
 
         self.encoder = ViTEncoder(
             input_size=input_size,

{broccoli_ml-15.3.0 → broccoli_ml-15.4.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "broccoli-ml"
-version = "15.3.0"
+version = "15.4.0"
 description = "Some useful Pytorch models, circa 2025"
 authors = [
     {name = "Nicholas Bailey"}