x-transformers 1.42.16__tar.gz → 1.42.18__tar.gz

{x_transformers-1.42.16/x_transformers.egg-info → x_transformers-1.42.18}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.42.16
+Version: 1.42.18
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang
@@ -14,7 +14,8 @@ Classifier: License :: OSI Approved :: MIT License
 Classifier: Programming Language :: Python :: 3.6
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: torch>=2.0
 Requires-Dist: einx>=0.3.0
 Requires-Dist: einops>=0.8.0
+Requires-Dist: loguru
 Requires-Dist: packaging>=21.0
+Requires-Dist: torch>=2.0

{x_transformers-1.42.16 → x_transformers-1.42.18}/setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.42.16',
+  version = '1.42.18',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',
@@ -16,10 +16,11 @@ setup(
     'transformers'
   ],
   install_requires=[
-    'torch>=2.0',
     'einx>=0.3.0',
     'einops>=0.8.0',
+    'loguru',
     'packaging>=21.0',
+    'torch>=2.0',
   ],
   setup_requires=[
     'pytest-runner',

{x_transformers-1.42.16 → x_transformers-1.42.18}/tests/test_x_transformers.py

@@ -352,7 +352,10 @@ def test_value_residual(
 
     model(x)
 
-def test_forgetting_transformer():
+@pytest.mark.parametrize('has_num_mem_kv', (False, True))
+def test_forgetting_transformer(
+    has_num_mem_kv: bool
+):
 
     model = TransformerWrapper(
         num_tokens = 20000,
@@ -361,7 +364,8 @@ def test_forgetting_transformer():
             dim = 128,
             depth = 6,
             heads = 8,
-            attn_data_dependent_alibi = False
+            attn_num_mem_kv = 1 if has_num_mem_kv else 0,
+            attn_data_dependent_alibi = True
         )
     )
 

{x_transformers-1.42.16 → x_transformers-1.42.18}/x_transformers/x_transformers.py

@@ -20,6 +20,8 @@ import einx
 from einops.layers.torch import Rearrange
 from einops import rearrange, repeat, reduce, pack, unpack
 
+from loguru import logger
+
 from x_transformers.attend import Attend, Intermediates
 from x_transformers.autoregressive_wrapper import AutoregressiveWrapper
 
@@ -1428,13 +1430,15 @@ class Attention(Module):
             else:
                 attn_bias = rel_pos(i, j)
 
-            attn_bias = pad_at_dim(attn_bias, (num_mem_kv, 0), value = 0.) # handle memory key / values
+            attn_bias = pad_at_dim(attn_bias, (num_mem_kv, 0)) # handle memory key / values
 
         # prepare data dependent alibi from forgetting transformers paper, if needed
 
         if exists(self.data_dependent_alibi):
             attn_bias = self.data_dependent_alibi(x)
 
+            attn_bias = pad_at_dim(attn_bias, (num_mem_kv, 0))
+
         # attention is all we need
 
         out, intermediates = self.attend(
@@ -1578,7 +1582,12 @@ class AttentionLayers(Module):
 
         self.disable_abs_pos_emb = default(disable_abs_pos_emb, (rel_pos_bias or rotary_pos_emb))
 
-        rotary_emb_dim = max(default(rotary_emb_dim, dim_head // 2), 32)
+        rotary_emb_dim = default(rotary_emb_dim, dim_head // 2)
+
+        assert rotary_emb_dim <= dim_head, f'rotary emb dim {rotary_emb_dim} must be less than or equal to attention head dimension {dim_head}'
+
+        if rotary_emb_dim < 32:
+            logger.warning('when training language model, rotary embedding dimension should be at least 32')
 
         assert not (rotary_xpos and not causal), 'rotary xpos is not compatible with bidirectional attention'
         self.rotary_pos_emb = RotaryEmbedding(rotary_emb_dim, use_xpos = rotary_xpos, scale_base = rotary_xpos_scale_base, interpolation_factor = rotary_interpolation_factor, base_rescale_factor = rotary_base_rescale_factor) if rotary_pos_emb else None

{x_transformers-1.42.16 → x_transformers-1.42.18/x_transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.42.16
+Version: 1.42.18
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang
@@ -14,7 +14,8 @@ Classifier: License :: OSI Approved :: MIT License
 Classifier: Programming Language :: Python :: 3.6
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: torch>=2.0
 Requires-Dist: einx>=0.3.0
 Requires-Dist: einops>=0.8.0
+Requires-Dist: loguru
 Requires-Dist: packaging>=21.0
+Requires-Dist: torch>=2.0

{x_transformers-1.42.16 → x_transformers-1.42.18}/x_transformers.egg-info/requires.txt

@@ -1,4 +1,5 @@
-torch>=2.0
 einx>=0.3.0
 einops>=0.8.0
+loguru
 packaging>=21.0
+torch>=2.0