titans-pytorch 0.0.57__py3-none-any.whl → 0.0.61__py3-none-any.whl

titans_pytorch/mac_transformer.py

@@ -24,8 +24,8 @@ def create_mac_block_mask(seq_len, window_size, persist_mem_len):
 
     def create_mac_mask(b, h, q_idx, kv_idx):
         is_persist_mem = kv_idx < persist_mem_len
-        causal_mask = q_idx >= (kv_idx - is_persist_mem)
-        block_diagonal = (q_idx // window_size) == ((kv_idx - is_persist_mem) // window_size)
+        causal_mask = q_idx >= (kv_idx - persist_mem_len)
+        block_diagonal = (q_idx // window_size) == ((kv_idx - persist_mem_len) // window_size)
         return is_persist_mem | (~is_persist_mem & (causal_mask & block_diagonal))
 
     block_mask = create_block_mask(create_mac_mask, B = None, H = None, Q_LEN = seq_len, KV_LEN = seq_len + persist_mem_len, _compile = True)
@@ -292,12 +292,13 @@ class NeuralMemoryGatingWrapper(Module):
         self,
         dim,
         attn: SegmentedAttention,
-        neural_mem: NeuralMemory | None = None
+        neural_mem: NeuralMemory | None = None,
+        gate_attn_output = True
     ):
         super().__init__()
         self.attn = attn
         self.neural_mem = neural_mem
-        self.to_gates = nn.Linear(dim, dim) if exists(neural_mem) else None
+        self.gate_attn_output = gate_attn_output
 
     def forward(
         self,
@@ -313,21 +314,19 @@ class NeuralMemoryGatingWrapper(Module):
 
         # initial retrieve, still should store first, it doesn't make sense not to, unless if all layers share the same neural memory
 
-        retrieved, first_kv_aux_loss = mem(seq, return_aux_kv_loss = True)
+        retrieved, kv_aux_loss = mem(seq, return_aux_kv_loss = True)
 
-        seq = seq + retrieved
+        if not self.gate_attn_output:
+            seq = seq + retrieved
 
         # attention
 
         attn_out, values = self.attn(seq, *args, **kwargs)
 
-        # another retrieve, but this time gate the attention output
+        if self.gate_attn_output:
+            attn_out = attn_out * retrieved.sigmoid()
 
-        retrieved, second_kv_aux_loss = mem(attn_out, return_aux_kv_loss = True)
-
-        attn_out = attn_out * self.to_gates(retrieved).sigmoid()
-
-        return (attn_out, values), (first_kv_aux_loss + second_kv_aux_loss)
+        return (attn_out, values), kv_aux_loss
 
 # MAC transformer
 
@@ -340,6 +339,7 @@ class MemoryAsContextTransformer(Module):
         depth,
         segment_len,
         neural_memory_segment_len = None,
+        neural_mem_gate_attn_output = True,
         num_longterm_mem_tokens = 0,
         num_persist_mem_tokens = 0,
         dim_head = 64,
@@ -414,6 +414,7 @@ class MemoryAsContextTransformer(Module):
                 dim,
                 attn = attn,
                 neural_mem = mem,
+                gate_attn_output = neural_mem_gate_attn_output
             )
 
             ff = FeedForward(dim = dim, mult = ff_mult)
@@ -488,7 +489,7 @@ class MemoryAsContextTransformer(Module):
         flex_attn_fn = None
 
         if use_flex_attn:
-            block_mask = create_mac_block_mask(seq_len_with_mem, self.segment_len, self.num_persist_mem_tokens)
+            block_mask = create_mac_block_mask(seq_len_with_mem, segment_len + num_longterm_mem_tokens, self.num_persist_mem_tokens)
             flex_attn_fn = partial(flex_attention, block_mask = block_mask)
 
         # value residual

titans_pytorch/titans.py

@@ -289,6 +289,8 @@ class NeuralMemory(Module):
 
         self.use_accelerated_scan = use_accelerated_scan
 
+        self.register_buffer('zero', torch.tensor(0.), persistent = False)
+
     def init_weights_and_momentum(self):
         params = TensorDict(dict(self.memory_model.named_parameters()))
 
@@ -306,6 +308,13 @@ class NeuralMemory(Module):
         past_state: tuple[dict[str, Tensor], dict[str, Tensor]],
         return_aux_kv_loss = False
     ):
+        seq_len = seq.shape[-2]
+
+        # handle edge case
+
+        if seq_len < self.chunk_size:
+            past_weight, _ = past_state
+            return TensorDict(past_weight).clone().zero_(), self.zero
 
         seq = self.store_norm(seq)
 
@@ -425,12 +434,10 @@ class NeuralMemory(Module):
 
         last_update = updates.apply(lambda t: t[:, -1])
 
-        next_state = (curr_weights + last_update, next_momentum)
-
         if not return_aux_kv_loss:
-            return updates, next_state
+            return updates
 
-        return updates, next_state, aux_kv_recon_loss.mean()
+        return updates, aux_kv_recon_loss.mean()
 
     def retrieve_memories(
         self,
@@ -442,7 +449,8 @@ class NeuralMemory(Module):
 
         seq = self.retrieve_norm(seq)
 
-        assert seq_len >= chunk_size
+        if seq_len < self.chunk_size:
+            return self.init_empty_memory_embed(batch, seq_len)
 
         seq = seq[:, (chunk_size - 1):]
         curtailed_seq_len = seq.shape[-2]
@@ -524,7 +532,7 @@ class NeuralMemory(Module):
 
         store_seq = default(store_seq, seq)
 
-        updates, next_memories, aux_kv_recon_loss = self.store_memories(store_seq, past_state, return_aux_kv_loss = True)
+        updates, aux_kv_recon_loss = self.store_memories(store_seq, past_state, return_aux_kv_loss = True)
 
         past_weights, _ = past_state
 

{titans_pytorch-0.0.57.dist-info → titans_pytorch-0.0.61.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.0.57
+Version: 0.0.61
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

titans_pytorch-0.0.61.dist-info/RECORD

@@ -0,0 +1,8 @@
+titans_pytorch/__init__.py,sha256=Q0MQA3RS8vqzs-KzSGZkvLR7afQ6ZW9uMOq1MeNuFoY,170
+titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
+titans_pytorch/mac_transformer.py,sha256=R9Xu-BjqAe9ZY60IGk4aNXBx_L8THsjJ4QrkbTnRNHo,15346
+titans_pytorch/titans.py,sha256=5wuAoDULbgXTM8Nbq8bXrW3Fd2nsn22kpERRfJOwZiU,16367
+titans_pytorch-0.0.61.dist-info/METADATA,sha256=Cfhqnse_9nnFNqVGo9p_kxO_LVawwv4uuZOx4anqhf0,4457
+titans_pytorch-0.0.61.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+titans_pytorch-0.0.61.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+titans_pytorch-0.0.61.dist-info/RECORD,,

titans_pytorch-0.0.57.dist-info/RECORD

@@ -1,8 +0,0 @@
-titans_pytorch/__init__.py,sha256=Q0MQA3RS8vqzs-KzSGZkvLR7afQ6ZW9uMOq1MeNuFoY,170
-titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
-titans_pytorch/mac_transformer.py,sha256=-WS_SI--_5f1whMaJOH-mYCU37EjYU_iZTurGfs8zgI,15331
-titans_pytorch/titans.py,sha256=ZKm-LnVKh1Cxs2tSxr4CcY37KroOOmYtTFM2F3Zb8Xg,16122
-titans_pytorch-0.0.57.dist-info/METADATA,sha256=rwLIRndtBo22oJt0Xm9xK9zqOYV50Jfo6g7oVrKq7CQ,4457
-titans_pytorch-0.0.57.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-titans_pytorch-0.0.57.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-titans_pytorch-0.0.57.dist-info/RECORD,,