x-transformers 2.3.22__tar.gz → 2.3.23__tar.gz

{x_transformers-2.3.22 → x_transformers-2.3.23}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.3.22
+Version: 2.3.23
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.3.22 → x_transformers-2.3.23}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "x-transformers"
-version = "2.3.22"
+version = "2.3.23"
 description = "X-Transformers"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{x_transformers-2.3.22 → x_transformers-2.3.23}/x_transformers/attend.py

@@ -276,21 +276,22 @@ class Attend(Module):
 
         # torch 2.3 uses new backend and context manager
 
-        if torch_version >= version.parse('2.3'):
-            from torch.nn.attention import SDPBackend
-
-            str_to_backend = dict(
-                enable_flash = SDPBackend.FLASH_ATTENTION,
-                enable_mem_efficient = SDPBackend.EFFICIENT_ATTENTION,
-                enable_math = SDPBackend.MATH,
-                enable_cudnn = SDPBackend.CUDNN_ATTENTION
-            )
+        if self.flash:
+            if torch_version >= version.parse('2.3'):
+                from torch.nn.attention import SDPBackend
 
-            sdpa_backends = [str_to_backend[enable_str] for enable_str, enable in sdp_kwargs.items() if enable]
+                str_to_backend = dict(
+                    enable_flash = SDPBackend.FLASH_ATTENTION,
+                    enable_mem_efficient = SDPBackend.EFFICIENT_ATTENTION,
+                    enable_math = SDPBackend.MATH,
+                    enable_cudnn = SDPBackend.CUDNN_ATTENTION
+                )
 
-            self.sdp_context_manager = partial(torch.nn.attention.sdpa_kernel, sdpa_backends)
-        else:
-            self.sdp_context_manager = partial(torch.backends.cuda.sdp_kernel, **sdp_kwargs)
+                sdpa_backends = [str_to_backend[enable_str] for enable_str, enable in sdp_kwargs.items() if enable]
+
+                self.sdp_context_manager = partial(torch.nn.attention.sdpa_kernel, sdpa_backends)
+            else:
+                self.sdp_context_manager = partial(torch.backends.cuda.sdp_kernel, **sdp_kwargs)
 
     def flash_attn(
         self,