titans-pytorch 0.1.14__py3-none-any.whl → 0.1.15__py3-none-any.whl

titans_pytorch/titans.py

@@ -301,6 +301,45 @@ class MemoryAttention(Module):
 
         return out
 
+# associative scan wrapper
+
+class AssocScan(Module):
+    def __init__(
+        self,
+        use_accelerated = False
+    ):
+        super().__init__()
+        self.use_accelerated = use_accelerated
+
+    def forward(self, gates, inputs):
+
+        if not self.use_accelerated:
+            _, outputs = associative_scan(binary_operator, (gates, inputs))
+            return outputs
+
+        from accelerated_scan.triton import scan as triton_scan
+        from accelerated_scan.warp import scan as warp_scan
+
+        scan = triton_scan if gates.is_cuda else warp_scan
+
+        def accelerate_scan_fn(gates, inputs):
+            gates = gates.expand_as(inputs)
+            gates, inputs = tuple(rearrange(t, 'b n d -> b d n') for t in (gates, inputs))
+
+            seq_len = gates.shape[-1]
+            next_power_two_seq_len = 2 ** max(5, int(math.ceil(math.log2(seq_len))))
+
+            gates = F.pad(gates, (0, next_power_two_seq_len - seq_len))
+            inputs = F.pad(inputs, (0, next_power_two_seq_len - seq_len))
+
+            outputs = scan(gates.contiguous(), inputs.contiguous())
+
+            outputs = outputs[..., :seq_len]
+            outputs = rearrange(outputs, 'b d n -> b n d')
+            return outputs
+
+        return accelerate_scan_fn(gates, inputs)
+
 # main neural memory
 
 def default_adaptive_step_transform(adaptive_step, max_lr = 1e-2):
@@ -339,6 +378,10 @@ class NeuralMemory(Module):
 
         self.retrieve_chunk_size, self.store_chunk_size = pair(chunk_size)
 
+        # associative scan
+
+        self.assoc_scan = AssocScan(use_accelerated = use_accelerated_scan)
+
         # norms
 
         self.retrieve_norm = nn.RMSNorm(dim) if pre_rmsnorm else nn.Identity()
@@ -564,38 +607,6 @@ class NeuralMemory(Module):
 
         surprises = grads.apply(lambda t: -t)
 
-        # determine scan function
-
-        def default_associative_scan(gates, inputs):
-            _, outputs = associative_scan(binary_operator, (gates, inputs))
-            return outputs
-
-        if self.use_accelerated_scan:
-            from accelerated_scan.triton import scan as triton_scan
-            from accelerated_scan.warp import scan as warp_scan
-
-            scan = triton_scan if seq.is_cuda else warp_scan
-
-            def accelerate_scan_fn(gates, inputs):
-                gates = gates.expand_as(inputs)
-                gates, inputs = tuple(rearrange(t, 'b n d -> b d n') for t in (gates, inputs))
-
-                seq_len = gates.shape[-1]
-                next_power_two_seq_len = 2 ** max(5, int(math.ceil(math.log2(seq_len))))
-
-                gates = F.pad(gates, (0, next_power_two_seq_len - seq_len))
-                inputs = F.pad(inputs, (0, next_power_two_seq_len - seq_len))
-
-                outputs = scan(gates.contiguous(), inputs.contiguous())
-
-                outputs = outputs[..., :seq_len]
-                outputs = rearrange(outputs, 'b d n -> b n d')
-                return outputs
-
-            scan_fn = accelerate_scan_fn
-        else:
-            scan_fn = default_associative_scan
-
         # momentum + weight decay - momentum is the new contribution, as most linear RNNs have learned forgetting gates
 
         next_momentum = TensorDict() if has_momentum else None
@@ -610,12 +621,12 @@ class NeuralMemory(Module):
             # derive momentum with associative scan - eq (10)
 
             if has_momentum:
-                update = scan_fn(adaptive_momentum, surprise) # momentum is S / surprise in the paper
+                update = self.assoc_scan(adaptive_momentum, surprise) # momentum is S / surprise in the paper
                 momentum = update
 
             # use associative scan again for learned forgetting (weight decay) - eq (13)
 
-            update = scan_fn(1. - decay_factor, update)
+            update = self.assoc_scan(1. - decay_factor, update)
 
             updates[param_name] = inverse_pack(update)
 

{titans_pytorch-0.1.14.dist-info → titans_pytorch-0.1.15.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.1.14
+Version: 0.1.15
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch
@@ -35,7 +35,7 @@ Classifier: Programming Language :: Python :: 3.9
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.9
 Requires-Dist: accelerated-scan>=0.2.0
-Requires-Dist: axial-positional-embedding>=0.3.7
+Requires-Dist: axial-positional-embedding>=0.3.9
 Requires-Dist: einops>=0.8.0
 Requires-Dist: einx>=0.3.0
 Requires-Dist: hyper-connections>=0.1.8

titans_pytorch-0.1.15.dist-info/RECORD

@@ -0,0 +1,8 @@
+titans_pytorch/__init__.py,sha256=u0tta_KqhOdfzCEDWT9P4_jejJEK2q1XxhsEzB5MnQU,223
+titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
+titans_pytorch/mac_transformer.py,sha256=zxknstaI_Uz47Y8WvZ3S7geJ-TNdqKV5Rvj0Jlw8njs,19271
+titans_pytorch/titans.py,sha256=R0e25ly2uTHkHSZEb-9Eqb0DqtFq8wFBB8iH1T6bYVg,22240
+titans_pytorch-0.1.15.dist-info/METADATA,sha256=SnNsoK4obeOAWFPhQypYJfJWZ_abXKr7WCvLMqFdyg0,6340
+titans_pytorch-0.1.15.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+titans_pytorch-0.1.15.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+titans_pytorch-0.1.15.dist-info/RECORD,,

titans_pytorch-0.1.14.dist-info/RECORD

@@ -1,8 +0,0 @@
-titans_pytorch/__init__.py,sha256=u0tta_KqhOdfzCEDWT9P4_jejJEK2q1XxhsEzB5MnQU,223
-titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
-titans_pytorch/mac_transformer.py,sha256=zxknstaI_Uz47Y8WvZ3S7geJ-TNdqKV5Rvj0Jlw8njs,19271
-titans_pytorch/titans.py,sha256=J7UbhhL0YKoQjrKPCgwUYwJd8-YfCZrHTSVjIFvvRRw,22077
-titans_pytorch-0.1.14.dist-info/METADATA,sha256=lM_x9obTibKKMXXdsW0U5GcQLTR4aXCtemEKPzYPqYo,6340
-titans_pytorch-0.1.14.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-titans_pytorch-0.1.14.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-titans_pytorch-0.1.14.dist-info/RECORD,,