titans-pytorch 0.1.15__tar.gz → 0.1.17__tar.gz

{titans_pytorch-0.1.15 → titans_pytorch-0.1.17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.1.15
+Version: 0.1.17
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

{titans_pytorch-0.1.15 → titans_pytorch-0.1.17}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.1.15"
+version = "0.1.17"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{titans_pytorch-0.1.15 → titans_pytorch-0.1.17}/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-0.1.17}/titans_pytorch/mac_transformer.py

@@ -537,7 +537,8 @@ 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 +548,37 @@ 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 +595,10 @@ 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 +626,7 @@ 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
 
@@ -651,7 +684,10 @@ 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)