x-transformers 1.44.0__py3-none-any.whl → 1.44.2__py3-none-any.whl

x_transformers/x_transformers.py

@@ -7,10 +7,11 @@ from random import random, randrange
 from packaging import version
 
 import torch
+from torch.amp import autocast
 import torch.nn.functional as F
 from torch import nn, einsum, Tensor
+from torch.utils._pytree import tree_flatten
 from torch.nn import Module, ModuleList, ModuleDict
-from torch.amp import autocast
 
 from functools import partial, wraps
 from collections import namedtuple
@@ -1138,6 +1139,7 @@ class Attention(Module):
         selective = False,
         custom_attn_fn: Callable | None = None,
         hybrid_module: Module | None = None,
+        hybrid_mask_kwarg: str | None = None,
         one_kv_head = False,
         kv_heads = None,
         shared_kv = False,
@@ -1341,6 +1343,8 @@ class Attention(Module):
 
         self.hybrid_module = deepcopy(hybrid_module) if exists(hybrid_module) else None
 
+        self.hybrid_mask_kwarg = hybrid_mask_kwarg # for bidirectional, can forward `mask` into the hybrid module and let it handle variable lengths
+
         # output dimension by default same as input, but can be overridden
 
         dim_out = default(dim_out, dim)
@@ -1592,7 +1596,21 @@ class Attention(Module):
         # hybrid module
 
         if exists(self.hybrid_module):
-            hybrid_out, _ = self.hybrid_module(x)
+
+            # hybrid input
+
+            hybrid_forward_kwargs = dict()
+
+            if not self.causal and exists(self.hybrid_mask_kwarg):
+                hybrid_forward_kwargs = {self.hybrid_mask_kwarg: mask}
+
+            # hybrid forward
+
+            hybrid_outputs = self.hybrid_module(x, **hybrid_forward_kwargs)
+
+            # handle hybrid out
+
+            (hybrid_out, *rest_hybrid_outs), _ = tree_flatten(hybrid_outputs)
             out = 0.5 * (out + hybrid_out)
 
         # alphafold2 styled gating of the values

{x_transformers-1.44.0.dist-info → x_transformers-1.44.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.44.0
+Version: 1.44.2
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.44.0.dist-info → x_transformers-1.44.2.dist-info}/RECORD

@@ -6,11 +6,11 @@ x_transformers/dpo.py,sha256=xt4OuOWhU8pN3OKN2LZAaC2NC8iiEnchqqcrPWVqf0o,3521
 x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg,9252
 x_transformers/neo_mlp.py,sha256=XCNnnop9WLarcxap1kGuYc1x8GHvwkZiDRnXOxSl3Po,3452
 x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dCN7fjlzd3K1rLUY,10510
-x_transformers/x_transformers.py,sha256=BI3RU3XFvwSNDZgoQBrFBSJ4SavJr38rOCCVgHZBTx0,101241
+x_transformers/x_transformers.py,sha256=W8qNdv-1CctdU4zwN2rYwu2CrgVMT1WL_3lwTSy5cCg,101862
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=7S00kCuab4tWQa-vf-z-XfzADjVj48MoFIr7VSIvttg,8575
-x_transformers-1.44.0.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.44.0.dist-info/METADATA,sha256=MNVwW_pDeKEIHRVEA1XOUNfzFmL706X7Npoh7xc3wIk,738
-x_transformers-1.44.0.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
-x_transformers-1.44.0.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.44.0.dist-info/RECORD,,
+x_transformers-1.44.2.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.44.2.dist-info/METADATA,sha256=yNNX1JyY8T3KF0Q6vO525au_90esaJZhw3quktk8n7g,738
+x_transformers-1.44.2.dist-info/WHEEL,sha256=A3WOREP4zgxI0fKrHUG8DC8013e3dK3n7a6HDbcEIwE,91
+x_transformers-1.44.2.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.44.2.dist-info/RECORD,,

{x_transformers-1.44.0.dist-info → x_transformers-1.44.2.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (75.6.0)
+Generator: setuptools (75.7.0)
 Root-Is-Purelib: true
 Tag: py3-none-any