merge hyper connection streams before final norm, to avoid edge case …

…with adaptive layernorm
lucidrains · Dec 24, 2024 · e1be411 · e1be411
1 parent 003275c
commit e1be411
Show file tree

Hide file tree

Showing 2 changed files with 6 additions and 6 deletions.
diff --git a/setup.py b/setup.py
@@ -3,7 +3,7 @@
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.43.2',
+  version = '1.43.4',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',

diff --git a/x_transformers/x_transformers.py b/x_transformers/x_transformers.py
@@ -2270,16 +2270,16 @@ def forward(
         if self.need_condition:
             final_norm = maybe(partial)(final_norm, **norm_kwargs)
 
-        if self.resi_dual:
-            x = x + final_norm(outer_residual)
-        else:
-            x = final_norm(x)
-
         # take care of multistreams if needed, use sum for now
 
         if is_multistream:
             x = reduce(x, '(b s) n d -> b n d', 'sum', s = streams)
 
+        if self.resi_dual:
+            x = x + final_norm(outer_residual)
+        else:
+            x = final_norm(x)
+
         if not return_hiddens:
             return x