titans-pytorch 0.1.33__tar.gz → 0.1.34__tar.gz

{titans_pytorch-0.1.33 → titans_pytorch-0.1.34}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.1.33
+Version: 0.1.34
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

{titans_pytorch-0.1.33 → titans_pytorch-0.1.34}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.1.33"
+version = "0.1.34"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{titans_pytorch-0.1.33 → titans_pytorch-0.1.34}/tests/test_titans.py

@@ -86,25 +86,32 @@ def test_retrieve_store_diff_seq():
     assert retrieve_seq.shape == retrieved.shape
 
 def test_weight_tied_mlp_neural_mem():
+    from titans_pytorch import MemoryMLP
+
+    mlp = MemoryMLP(64, depth = 2)
+
     mem  = NeuralMemory(
         dim = 384,
         dim_head = 64,
         heads = 2,
-        chunk_size = 2
+        chunk_size = 2,
+        model = mlp
     )
 
     mem2 = NeuralMemory(
         dim = 384,
         dim_head = 64,
         heads = 2,
-        chunk_size = 2
+        chunk_size = 2,
+        model = mlp
     )
 
     mem3 = NeuralMemory(
         dim = 384,
         dim_head = 64,
         heads = 2,
-        chunk_size = 2
+        chunk_size = 2,
+        model = mlp
     )
 
     seq = torch.randn(2, 128, 384)
@@ -113,6 +120,31 @@ def test_weight_tied_mlp_neural_mem():
     seq, cache2 = mem2(seq, prev_layer_updates = cache.updates)
     seq, cache3 = mem3(seq, prev_layer_updates = cache2.updates)
 
+def test_mac_with_weight_tied_neural_mem():
+    from titans_pytorch import MemoryMLP, MemoryAsContextTransformer
+
+    transformer = MemoryAsContextTransformer(
+        num_tokens = 256,
+        dim = 256,
+        depth = 2,
+        segment_len = 2,
+        num_persist_mem_tokens = 0,
+        num_longterm_mem_tokens = 2,
+        neural_memory_segment_len = 2,
+        sliding_window_attn = True,
+        neural_memory_layers = (1, 2),
+        neural_memory_model = MemoryMLP(256, depth = 1),
+        num_residual_streams = 4,
+        weight_tie_memory_model = True,
+        neural_mem_gate_attn_output = True,
+    )
+
+
+    ids = torch.randint(0, 256, (1, 1023))
+    logits = transformer(ids)
+
+    assert logits.shape == (1, 1023, 256)
+
 def test_overriding_chunk_size():
     mem = NeuralMemory(
         dim = 384,

{titans_pytorch-0.1.33 → titans_pytorch-0.1.34}/titans_pytorch/mac_transformer.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 from typing import Callable
 
 from math import ceil
+from copy import deepcopy
 from functools import partial
 from collections import namedtuple
 
@@ -485,11 +486,13 @@ class MemoryAsContextTransformer(Module):
         heads = 8,
         ff_mult = 4,
         num_residual_streams = 4,
+        neural_memory_model: Module | None = None,
         neural_memory_kwargs: dict = dict(),
         neural_memory_layers: tuple[int, ...] | None = None,
         aux_kv_recon_loss_weight = 0.,
         use_flex_attn = False,
-        sliding_window_attn = False
+        sliding_window_attn = False,
+        weight_tie_memory_model = False
     ):
         super().__init__()
 
@@ -523,6 +526,15 @@ class MemoryAsContextTransformer(Module):
 
         neural_memory_layers = default(neural_memory_layers, layers)
 
+        # weight tying neural memory model
+
+        maybe_copy = deepcopy if not weight_tie_memory_model else identity
+
+        if weight_tie_memory_model:
+            assert exists(neural_memory_model), '`neural_memory_model` must be explicitly set'
+
+        self.weight_tie_memory_model = weight_tie_memory_model
+
         # mem, attn, and feedforward layers
 
         for layer in layers:
@@ -551,6 +563,7 @@ class MemoryAsContextTransformer(Module):
                 mem = NeuralMemory(
                     dim = dim,
                     chunk_size = self.neural_memory_segment_len,
+                    model = maybe_copy(neural_memory_model),
                     **neural_memory_kwargs
                 )
 
@@ -683,7 +696,7 @@ class MemoryAsContextTransformer(Module):
 
         # math
 
-        batch, seq_len, neural_mem_segment_len, segment_len, num_longterm_mem_tokens, attn_window_size = *x.shape, self.neural_memory_segment_len, self.segment_len, self.num_longterm_mem_tokens, self.attn_window_size
+        batch, seq_len, neural_mem_segment_len, segment_len, num_longterm_mem_tokens, attn_window_size, weight_tie_memory_model = *x.shape, self.neural_memory_segment_len, self.segment_len, self.num_longterm_mem_tokens, self.attn_window_size, self.weight_tie_memory_model
 
         seq_len_with_mem = self.seq_len_with_longterm_mem(seq_len)
 
@@ -736,6 +749,10 @@ class MemoryAsContextTransformer(Module):
         next_kv_caches = []
         next_neural_mem_caches = []
 
+        # weight tied neural memory
+
+        neural_memory_updates = None
+
         # value residual
 
         value_residual = None
@@ -769,7 +786,8 @@ class MemoryAsContextTransformer(Module):
                 if not is_inferencing:
                     (retrieved, next_neural_mem_cache), mem_kv_aux_loss = mem(
                         mem_input,
-                        return_aux_kv_loss = True
+                        return_aux_kv_loss = True,
+                        prev_layer_updates = neural_memory_updates
                     )
 
                     kv_recon_losses = kv_recon_losses + mem_kv_aux_loss
@@ -777,9 +795,13 @@ class MemoryAsContextTransformer(Module):
                 else:
                     retrieved, next_neural_mem_cache = mem.forward_inference(
                         mem_input,
-                        state = next(neural_mem_caches, None)
+                        state = next(neural_mem_caches, None),
+                        prev_layer_updates = neural_memory_updates
                     )
 
+                if weight_tie_memory_model:
+                    neural_memory_updates = next_neural_mem_cache.updates
+
                 if self.gate_attn_output:
                     attn_out_gates = retrieved.sigmoid()
                 else:

{titans_pytorch-0.1.33 → titans_pytorch-0.1.34}/titans_pytorch/neural_memory.py

@@ -822,6 +822,7 @@ class NeuralMemory(Module):
         self,
         token: Tensor,
         state = None,
+        prev_layer_updates: dict[str, Tensor] | None = None
     ):
 
         # unpack previous state
@@ -863,12 +864,17 @@ class NeuralMemory(Module):
         else:
             updates = updates.apply(lambda t: t[:, -1:])
 
+        if exists(prev_layer_updates):
+            prev_layer_updates = TensorDict(prev_layer_updates)
+            prev_layer_updates = prev_layer_updates.apply(lambda t: t[:, -1:])
+
         if store_seq_cache_len == self.chunk_size:
 
             next_updates, next_states, _ = self.store_memories(
                 cache_store_seq,
                 weights,
-                past_state = past_states
+                past_state = past_states,
+                prev_layer_updates = prev_layer_updates,
             )
 
             updates = next_updates