titans-pytorch 0.4.2__tar.gz → 0.4.4__tar.gz

{titans_pytorch-0.4.2 → titans_pytorch-0.4.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.4.2
+Version: 0.4.4
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch
@@ -56,7 +56,7 @@ Description-Content-Type: text/markdown
 
 <img src="./fig1.png" width="400px"></img>
 
-## Titans - Pytorch (wip)
+## Titans - Pytorch
 
 Unofficial implementation of [Titans](https://arxiv.org/abs/2501.00663) in Pytorch. Will also contain some explorations into architectures beyond their simple 1-4 layer MLP for the neural memory module, if it works well to any degree.
 

{titans_pytorch-0.4.2 → titans_pytorch-0.4.4}/README.md

@@ -2,7 +2,7 @@
 
 <img src="./fig1.png" width="400px"></img>
 
-## Titans - Pytorch (wip)
+## Titans - Pytorch
 
 Unofficial implementation of [Titans](https://arxiv.org/abs/2501.00663) in Pytorch. Will also contain some explorations into architectures beyond their simple 1-4 layer MLP for the neural memory module, if it works well to any degree.
 

{titans_pytorch-0.4.2 → titans_pytorch-0.4.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.4.2"
+version = "0.4.4"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{titans_pytorch-0.4.2 → titans_pytorch-0.4.4}/tests/test_titans.py

@@ -405,3 +405,23 @@ def test_assoc_scan(
     assert second_half.shape == inputs2.shape
 
     assert torch.allclose(output[:, -1], second_half[:, -1], atol = 1e-5)
+
+def test_mem_state_detach():
+    from titans_pytorch.neural_memory import mem_state_detach
+
+    mem = NeuralMemory(
+        dim = 384,
+        chunk_size = 2,
+        qk_rmsnorm = True,
+        dim_head = 64,
+        heads = 4,
+    )
+
+    seq = torch.randn(4, 64, 384)
+
+    state = None
+
+    for _ in range(2):
+        parallel_retrieved, state = mem(seq, state = state)
+        state = mem_state_detach(state)
+        parallel_retrieved.sum().backward()

{titans_pytorch-0.4.2 → titans_pytorch-0.4.4}/titans_pytorch/__init__.py

@@ -1,5 +1,7 @@
 from titans_pytorch.neural_memory import (
     NeuralMemory,
+    NeuralMemState,
+    mem_state_detach
 )
 
 from titans_pytorch.memory_models import (

{titans_pytorch-0.4.2 → titans_pytorch-0.4.4}/titans_pytorch/neural_memory.py

@@ -7,10 +7,11 @@ from itertools import zip_longest
 from collections import namedtuple
 
 import torch
-from torch import nn, stack, cat, tensor, Tensor
+from torch import nn, stack, cat, is_tensor, tensor, Tensor
 import torch.nn.functional as F
 from torch.nn import Linear, Module, Parameter, ParameterList, ParameterDict
 from torch.func import functional_call, vmap, grad
+from torch.utils._pytree import tree_map, tree_flatten, tree_unflatten
 
 from tensordict import TensorDict
 
@@ -40,6 +41,8 @@ u - key / value updates - allowing a token to emit multiple key / values
 
 LinearNoBias = partial(Linear, bias = False)
 
+# neural mem state related
+
 NeuralMemState = namedtuple('NeuralMemState', [
     'seq_index',
     'weights',
@@ -48,6 +51,13 @@ NeuralMemState = namedtuple('NeuralMemState', [
     'updates',
 ])
 
+def mem_state_detach(
+    state: NeuralMemState
+):
+    assert isinstance(state, NeuralMemState)
+    state = tree_map(lambda t: t.detach() if is_tensor(t) else t, tuple(state))
+    return NeuralMemState(*state)
+
 # functions
 
 def exists(v):
@@ -940,7 +950,7 @@ class NeuralMemory(Module):
 
         # whether to allow network to slowly adjust from initial weight throughout (residual path) to fully updating weights every batch
 
-        surprises = None
+        surprises = (None, None)
         gate = None
 
         if exists(self.transition_gate):
@@ -967,7 +977,7 @@ class NeuralMemory(Module):
 
             updates = accum_updates(updates, next_updates)
 
-            surprises = safe_cat((surprises, chunk_surprises), dim = -1)
+            surprises = tuple(safe_cat(args, dim = -1) for args in zip(surprises, chunk_surprises))
 
             if is_last and not update_after_final_store:
                 continue