titans-pytorch 0.1.15__tar.gz → 0.1.18__tar.gz

{titans_pytorch-0.1.15 → titans_pytorch-0.1.18}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.1.15
+Version: 0.1.18
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch
@@ -38,7 +38,7 @@ Requires-Dist: accelerated-scan>=0.2.0
 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
+Requires-Dist: hyper-connections>=0.1.9
 Requires-Dist: ninja
 Requires-Dist: rotary-embedding-torch
 Requires-Dist: tensordict

{titans_pytorch-0.1.15 → titans_pytorch-0.1.18}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.1.15"
+version = "0.1.18"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -29,7 +29,7 @@ dependencies = [
     "axial_positional_embedding>=0.3.9",
     "einops>=0.8.0",
     "einx>=0.3.0",
-    "hyper-connections>=0.1.8",
+    "hyper-connections>=0.1.9",
     "Ninja",
     "rotary-embedding-torch",
     "tensordict",

{titans_pytorch-0.1.15 → titans_pytorch-0.1.18}/tests/test_titans.py

@@ -3,7 +3,7 @@ from torch import nn
 
 import pytest
 from titans_pytorch import NeuralMemory
-from titans_pytorch.mac_transformer import flex_attention, SegmentedAttention
+from titans_pytorch.mac_transformer import flex_attention, SegmentedAttention, MemoryAsContextTransformer
 
 def exists(v):
     return v is not None
@@ -92,8 +92,6 @@ def test_mac(
     num_longterm_mem_tokens,
     neural_mem_gate_attn_output
 ):
-    from titans_pytorch.mac_transformer import MemoryAsContextTransformer
-
     transformer = MemoryAsContextTransformer(
         num_tokens = 256,
         dim = 256,
@@ -109,6 +107,25 @@ def test_mac(
     logits = transformer(x)
     assert logits.shape == (1, seq_len, 256)
 
+def test_mac_sampling():
+    transformer = MemoryAsContextTransformer(
+        num_tokens = 256,
+        dim = 256,
+        depth = 2,
+        segment_len = 32,
+        num_persist_mem_tokens = 4,
+        num_longterm_mem_tokens = 16,
+    )
+
+    ids = torch.randint(0, 256, (1, 1023))
+
+    # after much training
+
+    sampled = transformer.sample(ids[:, :4], 53, use_cache = False, temperature = 0.)
+    sampled_with_cache = transformer.sample(ids[:, :4], 53, use_cache = True, temperature = 0.)
+
+    assert torch.allclose(sampled, sampled_with_cache)
+
 @pytest.mark.parametrize('seq_len', (1023, 17))
 @pytest.mark.parametrize('sliding', (True, False))
 def test_flex(

{titans_pytorch-0.1.15 → titans_pytorch-0.1.18}/titans_pytorch/mac_transformer.py

@@ -217,7 +217,8 @@ class SegmentedAttention(Module):
         self,
         seq,
         value_residual = None,
-        flex_attn_fn: Callable | None = None
+        flex_attn_fn: Callable | None = None,
+        output_gating = None
     ):
 
         assert not (exists(value_residual) ^ exists(self.to_learned_v_mix))
@@ -267,6 +268,9 @@ class SegmentedAttention(Module):
 
         out = self.to_out(out)
 
+        if exists(output_gating):
+            out = out * output_gating
+
         return out, orig_v
 
     def forward(
@@ -274,10 +278,11 @@ class SegmentedAttention(Module):
         seq,
         value_residual = None,
         flex_attn_fn: Callable | None = None,
-        disable_flex_attn = False
+        disable_flex_attn = False,
+        output_gating = None
     ):
         if seq.is_cuda and self.use_flex_attn and not disable_flex_attn:
-            return self.forward_flex(seq, value_residual, flex_attn_fn)
+            return self.forward_flex(seq, value_residual, flex_attn_fn, output_gating = output_gating)
 
         assert not (exists(value_residual) ^ exists(self.to_learned_v_mix))
 
@@ -361,50 +366,10 @@ class SegmentedAttention(Module):
 
         out = inverse_segment(out)
 
-        return out, orig_v
-
-# Attention + Neural Memory gating configuration, as depicted in Figure 2
-
-class NeuralMemoryGatingWrapper(Module):
-    def __init__(
-        self,
-        dim,
-        attn: SegmentedAttention,
-        neural_mem: NeuralMemory | None = None,
-        gate_attn_output = True
-    ):
-        super().__init__()
-        self.attn = attn
-        self.neural_mem = neural_mem
-        self.gate_attn_output = gate_attn_output
-
-    def forward(
-        self,
-        seq,
-        *args,
-        **kwargs
-    ):
-        batch, seq_len = seq.shape[:2]
-        mem = self.neural_mem
-
-        if not exists(mem):
-            return self.attn(seq, *args, **kwargs), 0.
-
-        # initial retrieve, still should store first, it doesn't make sense not to, unless if all layers share the same neural memory
-
-        retrieved, kv_aux_loss = mem(seq, return_aux_kv_loss = True)
-
-        if not self.gate_attn_output:
-            seq = seq + retrieved
-
-        # attention
-
-        attn_out, values = self.attn(seq, *args, **kwargs)
-
-        if self.gate_attn_output:
-            attn_out = attn_out * retrieved.sigmoid()
+        if exists(output_gating):
+            out = out * output_gating
 
-        return (attn_out, values), kv_aux_loss
+        return out, orig_v
 
 # MAC transformer
 
@@ -494,16 +459,10 @@ class MemoryAsContextTransformer(Module):
                     **neural_memory_kwargs
                 )
 
-            attn = NeuralMemoryGatingWrapper(
-                dim,
-                attn = attn,
-                neural_mem = mem,
-                gate_attn_output = neural_mem_gate_attn_output
-            )
-
             ff = FeedForward(dim = dim, mult = ff_mult)
 
             self.layers.append(ModuleList([
+                init_hyper_conn(dim = dim, branch = mem, add_branch_out_to_residual = not neural_mem_gate_attn_output) if exists(mem) else None,
                 init_hyper_conn(dim = dim, branch = attn),
                 init_hyper_conn(dim = dim, branch = ff)
             ]))
@@ -512,6 +471,10 @@ class MemoryAsContextTransformer(Module):
 
         self.to_logits = LinearNoBias(dim, num_tokens)
 
+        # whether to gate the attention output with the retrieved memories
+
+        self.gate_attn_output = neural_mem_gate_attn_output
+
         # auxiliary loss on kv recon
 
         self.has_aux_kv_recon_loss = aux_kv_recon_loss_weight > 0.
@@ -537,7 +500,8 @@ class MemoryAsContextTransformer(Module):
         filter_kwargs: dict = dict(
             min_p = 0.1,
         ),
-        show_progress = True
+        show_progress = True,
+        use_cache = False
     ):
         was_training = self.training
         self.eval()
@@ -547,8 +511,37 @@ class MemoryAsContextTransformer(Module):
 
         iter_wrap = tqdm.tqdm if show_progress else identity
 
+        # cache for axial pos, attention, and neural memory
+
+        cache = None
+        factorized_pos_emb = None
+
+        # precompute factorized pos emb
+
+        if use_cache:
+            round_up_seq_len = round_up_multiple(seq_len, self.segment_len)
+            longterm_mem_lens = (round_up_seq_len // self.segment_len) * self.num_longterm_mem_tokens
+            seq_len_with_mem = round_up_seq_len + longterm_mem_lens
+
+            axial_dims = self.axial_pos_emb.maybe_derive_outer_dim(seq_len_with_mem, (self.neural_memory_segment_len,))
+
+            factorized_pos_emb = self.axial_pos_emb(axial_dims, return_factorized = True)
+
+        # sample
+
         for _ in iter_wrap(range(sample_num_times)):
-            logits = self.forward(out, disable_flex_attn = True)
+
+            logits, next_cache = self.forward(
+                out,
+                disable_flex_attn = True,
+                cache = cache,
+                return_cache = True,
+                factorized_pos_emb = factorized_pos_emb
+            )
+
+            if use_cache:
+                cache = next_cache
+
             logits = logits[:, -1]
 
             logits = filter_fn(logits, **filter_kwargs)
@@ -565,7 +558,10 @@ class MemoryAsContextTransformer(Module):
         x,
         return_loss = False,
         return_loss_breakdown = False,
-        disable_flex_attn = False
+        disable_flex_attn = False,
+        cache = None,
+        return_cache = False,
+        factorized_pos_emb = None
     ):
 
         if return_loss:
@@ -593,7 +589,7 @@ class MemoryAsContextTransformer(Module):
         # apply axial positional embedding
         # so intra and inter segment can be more easily discerned by the network
 
-        pos_emb = self.axial_pos_emb.forward_with_seq_len(seq_len_with_mem, (neural_mem_segment_len,))
+        pos_emb = self.axial_pos_emb.forward_with_seq_len(seq_len_with_mem, (neural_mem_segment_len,), factorized = factorized_pos_emb)
 
         x = x + pos_emb
 
@@ -619,19 +615,34 @@ class MemoryAsContextTransformer(Module):
 
         x = self.expand_streams(x)
 
-        for attn, ff in self.layers:
+        for mem, attn, ff in self.layers:
+
+            retrieved = None
+            attn_out_gates = None
+
+            if exists(mem):
+                retrieved, mem_kv_aux_loss = mem(x, return_aux_kv_loss = True)
+                kv_recon_losses = kv_recon_losses + mem_kv_aux_loss
 
-            (x, values), maybe_mem_kv_aux_loss = attn(
+                if self.gate_attn_output:
+                    attn_out_gates = retrieved.sigmoid()
+                else:
+                    seq = retrieved
+
+            # attention
+
+            x, values = attn(
                 x,
                 value_residual = value_residual,
                 disable_flex_attn = disable_flex_attn,
-                flex_attn_fn = flex_attn_fn
+                flex_attn_fn = flex_attn_fn,
+                output_gating = attn_out_gates
             )
 
-            kv_recon_losses = kv_recon_losses + maybe_mem_kv_aux_loss
-
             value_residual = default(value_residual, values)
 
+            # feedforward
+
             x = ff(x)
 
         x = self.reduce_streams(x)
@@ -651,7 +662,10 @@ class MemoryAsContextTransformer(Module):
         logits = self.to_logits(x)
 
         if not return_loss:
-            return logits
+            if not return_cache:
+                return logits
+
+            return logits, cache
 
         ar_loss = F.cross_entropy(rearrange(logits, 'b n l -> b l n'), labels)