x-transformers 2.1.11__tar.gz → 2.1.14__tar.gz

{x_transformers-2.1.11 → x_transformers-2.1.14}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.1.11
+Version: 2.1.14
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.1.11 → x_transformers-2.1.14}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "x-transformers"
-version = "2.1.11"
+version = "2.1.14"
 description = "X-Transformers"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{x_transformers-2.1.11 → x_transformers-2.1.14}/tests/test_x_transformers.py

@@ -695,8 +695,10 @@ def test_lime(
     model(x)
 
 @pytest.mark.parametrize('backward_ar_loss_weight', (1., 0.5))
+@pytest.mark.parametrize('goal_suffix', (False, True))
 def test_belief_state_wrapper(
-    backward_ar_loss_weight
+    backward_ar_loss_weight,
+    goal_suffix
 ):
     from x_transformers.belief_state_wrapper import BeliefStateWrapper
 
@@ -733,5 +735,9 @@ def test_belief_state_wrapper(
     loss = model(seq, backward = False)
     loss.backward()
 
-    sampled = model.generate_with_suffix_token_only(seq[:, :1], 16)
+    suffix = None
+    if goal_suffix:
+        suffix = torch.randint(0, 20000, (2, 2))
+
+    sampled = model.generate_with_suffix_token_only(seq[:, :1], 16, suffix = suffix)
     assert sampled.shape == (2, 16)

{x_transformers-2.1.11 → x_transformers-2.1.14}/x_transformers/belief_state_wrapper.py

@@ -132,11 +132,11 @@ class BeliefStateWrapper(Module):
 
         # get the encoded suffix token once
 
-        if not exists(suffix):
-            suffix = out[:, 0:0]
+        if exists(suffix):
+            if suffix.ndim == 1:
+                suffix = repeat(suffix, 'n -> b n', b = batch)
 
-        if suffix.ndim == 1:
-            suffix = repeat(suffix, 'n -> b n', b = batch)
+            suffix = suffix.flip(1) # reverse autoregressive
 
         suffix_sos_tokens = rearrange(self.suffix_token, 'd -> 1 1 d')
 
@@ -148,6 +148,10 @@ class BeliefStateWrapper(Module):
             return_embeddings = True
         )
 
+        # pick out the last embedding for fill in the model
+
+        suffix_embed = suffix_embed[:, -1:]
+
         # sampling up to seq_len
 
         for _ in range(seq_len):

{x_transformers-2.1.11 → x_transformers-2.1.14}/x_transformers/x_transformers.py

@@ -2898,6 +2898,15 @@ class TransformerWrapper(Module):
         to_logits_kwargs = dict(),
         **kwargs,
     ):
+
+        # if sequence is None, auto create an empty one if `prepend_embeds` was supplied
+
+        if not exists(x):
+            assert exists(prepend_embeds)
+            x = prepend_embeds.new_empty((prepend_embeds.shape[0], 0), dtype = torch.long)
+
+        # shapes and variables
+
         b, n, device, num_mems, has_memory_tokens, emb_frac_gradient, orig_mask = x.shape[0], x.shape[1], x.device, self.num_memory_tokens, self.num_memory_tokens > 0, self.emb_frac_gradient, mask
 
         return_hiddens = return_mems | return_attn | return_intermediates | return_attn_z_loss