titans-pytorch 0.0.38__tar.gz → 0.0.39__tar.gz

{titans_pytorch-0.0.38 → titans_pytorch-0.0.39}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.0.38
+Version: 0.0.39
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch
@@ -43,6 +43,7 @@ Requires-Dist: ninja
 Requires-Dist: rotary-embedding-torch
 Requires-Dist: tensordict
 Requires-Dist: torch>=2.2
+Requires-Dist: x-transformers
 Provides-Extra: examples
 Requires-Dist: local-attention>=1.10.1; extra == 'examples'
 Requires-Dist: taylor-series-linear-attention; extra == 'examples'

{titans_pytorch-0.0.38 → titans_pytorch-0.0.39}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.0.38"
+version = "0.0.39"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -34,6 +34,7 @@ dependencies = [
     "rotary-embedding-torch",
     "tensordict",
     "torch>=2.2",
+    "x-transformers"
 ]
 
 [project.urls]

{titans_pytorch-0.0.38 → titans_pytorch-0.0.39}/titans_pytorch/mac_transformer.py

@@ -7,7 +7,7 @@ from torch import nn, cat
 import torch.nn.functional as F
 from torch.nn import Module, ModuleList, Linear
 
-from einops import repeat, rearrange, pack, unpack
+from einops import einsum, repeat, rearrange, pack, unpack
 from einops.layers.torch import Rearrange
 
 from hyper_connections import get_init_and_expand_reduce_stream_functions
@@ -16,6 +16,7 @@ from hyper_connections import get_init_and_expand_reduce_stream_functions
 
 from axial_positional_embedding import ContinuousAxialPositionalEmbedding
 from rotary_embedding_torch import RotaryEmbedding
+from x_transformers.attend import Attend
 
 # proposed neural memory
 
@@ -93,6 +94,7 @@ class SegmentedAttention(Module):
         num_longterm_mem_tokens = 0,
         dim_head = 64,
         heads = 8,
+        attend_kwargs: dict = dict()
     ):
         super().__init__()
         self.norm = nn.RMSNorm(dim)
@@ -101,6 +103,8 @@ class SegmentedAttention(Module):
 
         self.rotary_emb = RotaryEmbedding(dim_head)
 
+        self.attend = Attend(causal = True, **attend_kwargs)
+
         self.to_qkv = LinearNoBias(dim, dim_inner * 3)
         self.to_out = LinearNoBias(dim_inner, dim)
 
@@ -145,9 +149,9 @@ class SegmentedAttention(Module):
         k = cat((pmk, k), dim = -2)
         v = cat((pmv, v), dim = -2)
 
-        # sdpa
+        # attention
 
-        out = F.scaled_dot_product_attention(q, k, v, is_causal = True)
+        out, _ = self.attend(q, k, v)
 
         out = self.merge_heads(out)