x-transformers 2.6.2__tar.gz → 2.6.3__tar.gz

{x_transformers-2.6.2 → x_transformers-2.6.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.6.2
+Version: 2.6.3
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.6.2 → x_transformers-2.6.3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "x-transformers"
-version = "2.6.2"
+version = "2.6.3"
 description = "X-Transformers"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{x_transformers-2.6.2 → x_transformers-2.6.3}/tests/test_x_transformers.py

@@ -1228,8 +1228,8 @@ def test_external_key_values():
     seq = torch.randint(0, 20000, (3, 1024))
 
     key_values = [
-        (torch.randn(3, 8, 32, 16), torch.randn(3, 8, 32, 16)),
-        (torch.randn(3, 8, 32, 16), torch.randn(3, 8, 32, 16)),
+        (torch.randn(3, 2, 32, 16), torch.randn(3, 2, 32, 16)),
+        (torch.randn(3, 2, 32, 16), torch.randn(3, 2, 32, 16)),
     ]
 
     additional_kv_mask = torch.randint(0, 2, (3, 32)).bool()

{x_transformers-2.6.2 → x_transformers-2.6.3}/x_transformers/x_transformers.py

@@ -1795,6 +1795,13 @@ class Attention(Module):
             seq_len = k.shape[-2]
 
             added_k, added_v = additional_key_values
+            added_kv_heads, added_kv_len = added_k.shape[1], added_k.shape[-2]
+
+            # take care of expanding to query heads if mismatch between key / value heads with the ones coming from vlm
+
+            if added_kv_heads != kv_h:
+                assert divisible_by(h, added_kv_heads)
+                k, v, added_k, added_v = tuple(repeat(t, 'b h ... -> b (r h) ...', r = h // t.shape[1]) for t in (k, v, added_k, added_v))
 
             k = cat((added_k, k), dim = -2)
             v = cat((added_v, v), dim = -2)
@@ -1802,7 +1809,6 @@ class Attention(Module):
             if (exists(input_mask) or exists(additional_key_value_mask)):
 
                 if not exists(additional_key_value_mask):
-                    added_kv_len = added_k.shape[-2]
                     input_mask = pad_at_dim(input_mask, (added_kv_len, 0), dim = -1, value = True)
                 elif not exists(input_mask):
                     input_mask = pad_at_dim(additional_key_value_mask, (0, seq_len), dim = -1, value = True)