broccoli-ml 15.0.0__tar.gz → 15.1.0__tar.gz

{broccoli_ml-15.0.0 → broccoli_ml-15.1.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: broccoli-ml
-Version: 15.0.0
+Version: 15.1.0
 Summary: Some useful Pytorch models, circa 2025
 License: MIT
 Author: Nicholas Bailey

{broccoli_ml-15.0.0 → broccoli_ml-15.1.0}/broccoli/transformer.py

@@ -122,6 +122,9 @@ class MHAttention(nn.Module):
 
         self.head_dim = self.embed_dim // self.n_heads
 
+        self.query_norm = nn.RMSNorm(self.head_dim)
+        self.key_norm = nn.RMSNorm(self.head_dim)
+
         if self.scaling == "sqrtd":
             self.scaling_factor = 1 / math.sqrt(self.head_dim)
         elif self.scaling == "d":
@@ -225,8 +228,9 @@ class MHAttention(nn.Module):
 
         freqs = self.rotary_embedding.get_axial_freqs(*self.source_size)
 
-        q_img = apply_rotary_emb(freqs, q_img)
-        k_img = apply_rotary_emb(freqs, k_img)
+        # norm Qs/Ks to protect axial rope, like https://arxiv.org/abs/2302.05442
+        q_img = apply_rotary_emb(freqs, self.query_norm(q_img))
+        k_img = apply_rotary_emb(freqs, self.key_norm(k_img))
 
         q_img = rearrange(
             q_img,
@@ -416,7 +420,7 @@ class FeedforwardBlock(nn.Module):
                 self.activation,
                 self.inner_dropout,
                 (
-                    nn.LayerNorm(int(ratio * output_features))
+                    nn.RMSNorm(int(ratio * output_features))
                     if normformer
                     else nn.Identity()
                 ),
@@ -474,13 +478,7 @@ class FeedforwardBlock(nn.Module):
 
 
 class EncoderBlock(nn.Module):
-    """
-    Performs LayerNorms first (as in PyTorch Transformers when norm_first=True),
-        which is also what is seen in e.g.
-        https://github.com/karpathy/minGPT/blob/master/mingpt/model.py
-        and is recommended by https://arxiv.org/abs/2002.04745
-
-    """
+    """ """
 
     def __init__(
         self,
@@ -534,16 +532,16 @@ class EncoderBlock(nn.Module):
         self.drop_path = DropPath(drop_prob=identity_probability, scale_by_keep=True)
 
         if self.pre_norm:
-            self.pre_attention_norm = nn.LayerNorm(d_model)
-            self.pre_mlp_norm = nn.LayerNorm(d_model)
+            self.pre_attention_norm = nn.RMSNorm(d_model)
+            self.pre_mlp_norm = nn.RMSNorm(d_model)
 
         if normformer:
-            self.normformer_norm = nn.LayerNorm(d_model)
+            self.normformer_norm = nn.RMSNorm(d_model)
 
         if self.post_norm:
-            self.input_norm = nn.LayerNorm(d_model)
-            self.post_attention_norm = nn.LayerNorm(d_model)
-            self.post_mlp_norm = nn.LayerNorm(d_model)
+            self.input_norm = nn.RMSNorm(d_model)
+            self.post_attention_norm = nn.RMSNorm(d_model)
+            self.post_mlp_norm = nn.RMSNorm(d_model)
 
         if relative_position_embedding:
             max_freq = int(max(source_size) / 2)  # Suggested by Gemini!

{broccoli_ml-15.0.0 → broccoli_ml-15.1.0}/broccoli/vit.py

@@ -403,7 +403,7 @@ class ViTEncoder(nn.Module):
                 checkpoint=transformer_checkpoint_ff,
                 beta=self.beta,
             )
-            self.layer_norm = nn.LayerNorm(transformer_embedding_size)
+            self.layer_norm = nn.RMSNorm(transformer_embedding_size)
         else:
             self.initial_ff = None
 
@@ -417,7 +417,7 @@ class ViTEncoder(nn.Module):
                     f"N C {spatial_dim_names} -> N ({spatial_dim_names}) C"
                 ),
                 self.pooling_channels_padding,
-                nn.LayerNorm(transformer_embedding_size),
+                nn.RMSNorm(transformer_embedding_size),
             ]
         )
 
@@ -522,7 +522,7 @@ class ViT(nn.Module):
 
         # Set alpha and beta according to Microsoft's DeepNorm
         self.alpha = (2 * transformer_layers) ** 0.25
-        self.beta = (8 * transformer_layers) ** 0.25
+        self.beta = (8 * transformer_layers) ** -0.25
 
         self.encoder = ViTEncoder(
             input_size=input_size,

{broccoli_ml-15.0.0 → broccoli_ml-15.1.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "broccoli-ml"
-version = "15.0.0"
+version = "15.1.0"
 description = "Some useful Pytorch models, circa 2025"
 authors = [
     {name = "Nicholas Bailey"}