x-transformers 1.43.1__tar.gz → 1.43.4__tar.gz

{x_transformers-1.43.1/x_transformers.egg-info → x_transformers-1.43.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.43.1
+Version: 1.43.4
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.43.1 → x_transformers-1.43.4}/setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.43.1',
+  version = '1.43.4',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',

{x_transformers-1.43.1 → x_transformers-1.43.4}/tests/test_x_transformers.py

@@ -409,7 +409,8 @@ def test_custom_alibi(flash: bool):
 
     logits = model(x, pos = pos)
 
-def test_custom_rotary_pos_emb():
+@pytest.mark.parametrize('rotary_xpos', (True, False))
+def test_custom_rotary_pos_emb(rotary_xpos):
     from einops import repeat
 
     model = TransformerWrapper(
@@ -419,7 +420,8 @@ def test_custom_rotary_pos_emb():
             dim = 512,
             depth = 2,
             heads = 8,
-            rotary_pos_emb = True
+            rotary_pos_emb = True,
+            rotary_xpos = rotary_xpos
         )
     )
 

{x_transformers-1.43.1 → x_transformers-1.43.4}/x_transformers/x_transformers.py

@@ -666,7 +666,7 @@ class RotaryEmbedding(Module):
             return freqs, 1.
 
         power = (t - (max_pos // 2)) / self.scale_base
-        scale = self.scale ** rearrange(power, 'n -> n 1')
+        scale = self.scale ** rearrange(power, '... n -> ... n 1')
         scale = torch.stack((scale, scale), dim = -1)
         scale = rearrange(scale, '... d r -> ... (d r)')
 
@@ -2270,16 +2270,16 @@ class AttentionLayers(Module):
         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
 

{x_transformers-1.43.1 → x_transformers-1.43.4/x_transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.43.1
+Version: 1.43.4
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang