x-transformers 1.42.17__tar.gz → 1.42.19__tar.gz

{x_transformers-1.42.17/x_transformers.egg-info → x_transformers-1.42.19}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.42.17
+Version: 1.42.19
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang
@@ -14,7 +14,8 @@ Classifier: License :: OSI Approved :: MIT License
 Classifier: Programming Language :: Python :: 3.6
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: torch>=2.0
 Requires-Dist: einx>=0.3.0
 Requires-Dist: einops>=0.8.0
+Requires-Dist: loguru
 Requires-Dist: packaging>=21.0
+Requires-Dist: torch>=2.0

{x_transformers-1.42.17 → x_transformers-1.42.19}/README.md

@@ -2352,4 +2352,13 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@inproceedings{Duvvuri2024LASERAW,
+    title   = {LASER: Attention with Exponential Transformation},
+    author  = {Sai Surya Duvvuri and Inderjit S. Dhillon},
+    year    = {2024},
+    url     = {https://api.semanticscholar.org/CorpusID:273849947}
+}
+```
+
 *solve intelligence... then use that to solve everything else.* - Demis Hassabis

{x_transformers-1.42.17 → x_transformers-1.42.19}/setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.42.17',
+  version = '1.42.19',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',
@@ -16,10 +16,11 @@ setup(
     'transformers'
   ],
   install_requires=[
-    'torch>=2.0',
     'einx>=0.3.0',
     'einops>=0.8.0',
+    'loguru',
     'packaging>=21.0',
+    'torch>=2.0',
   ],
   setup_requires=[
     'pytest-runner',

{x_transformers-1.42.17 → x_transformers-1.42.19}/tests/test_x_transformers.py

@@ -516,3 +516,19 @@ def test_to_logits(to_logits):
     output = model(x, to_logits_kwargs=to_logits_kwargs)
 
     assert output.shape == (2, 1024, 20000)
+
+def test_laser():
+    model = TransformerWrapper(
+        num_tokens = 20000,
+        max_seq_len = 1024,
+        attn_layers = Decoder(
+            dim = 128,
+            depth = 6,
+            heads = 8,
+            attn_laser = True
+        )
+    )
+
+    x = torch.randint(0, 20000, (2, 1024))
+
+    model(x)

{x_transformers-1.42.17 → x_transformers-1.42.19}/x_transformers/x_transformers.py

@@ -20,6 +20,8 @@ import einx
 from einops.layers.torch import Rearrange
 from einops import rearrange, repeat, reduce, pack, unpack
 
+from loguru import logger
+
 from x_transformers.attend import Attend, Intermediates
 from x_transformers.autoregressive_wrapper import AutoregressiveWrapper
 
@@ -1073,6 +1075,7 @@ class Attention(Module):
         neutreno_value_residual = False, # Nguyen et al. https://arxiv.org/abs/2312.00751
         neutreno_alpha = 0.4,
         learned_value_residual_mix = False,
+        laser = False, # https://arxiv.org/abs/2411.03493v1
         onnxable = False,
         attend_sdp_kwargs: dict = dict(
             enable_flash = True,
@@ -1112,6 +1115,11 @@ class Attention(Module):
         assert not (shared_kv and value_dim_head != dim_head), 'key and value head dimensions must be equal for shared key / values'
         self.to_v = LinearNoBias(dim_kv, v_dim) if not shared_kv else None
 
+        # enhancing gradients to attention through exponentiated values
+        # todo - compare it to `attn = attn * large_value + attn.detach() * (1. - large_value)`
+
+        self.laser = laser
+
         # relations projection from tp-attention
 
         self.to_r = LinearNoBias(dim, v_dim) if tensor_product else None
@@ -1437,6 +1445,11 @@ class Attention(Module):
 
             attn_bias = pad_at_dim(attn_bias, (num_mem_kv, 0))
 
+        if self.laser:
+            values_max = v.amax(dim = -2, keepdim = True).detach() # numerical stability
+            v = v - values_max
+            v = v.exp()
+
         # attention is all we need
 
         out, intermediates = self.attend(
@@ -1446,6 +1459,11 @@ class Attention(Module):
             prev_attn = prev_attn
         )
 
+        # laser
+
+        if self.laser:
+            out = out.log() + values_max
+
         # store the values for resformer or Neutreno
 
         intermediates.values = orig_values
@@ -1580,7 +1598,12 @@ class AttentionLayers(Module):
 
         self.disable_abs_pos_emb = default(disable_abs_pos_emb, (rel_pos_bias or rotary_pos_emb))
 
-        rotary_emb_dim = max(default(rotary_emb_dim, dim_head // 2), 32)
+        rotary_emb_dim = default(rotary_emb_dim, dim_head // 2)
+
+        assert rotary_emb_dim <= dim_head, f'rotary emb dim {rotary_emb_dim} must be less than or equal to attention head dimension {dim_head}'
+
+        if rotary_emb_dim < 32:
+            logger.warning('when training language model, rotary embedding dimension should be at least 32')
 
         assert not (rotary_xpos and not causal), 'rotary xpos is not compatible with bidirectional attention'
         self.rotary_pos_emb = RotaryEmbedding(rotary_emb_dim, use_xpos = rotary_xpos, scale_base = rotary_xpos_scale_base, interpolation_factor = rotary_interpolation_factor, base_rescale_factor = rotary_base_rescale_factor) if rotary_pos_emb else None

{x_transformers-1.42.17 → x_transformers-1.42.19/x_transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.42.17
+Version: 1.42.19
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang
@@ -14,7 +14,8 @@ Classifier: License :: OSI Approved :: MIT License
 Classifier: Programming Language :: Python :: 3.6
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: torch>=2.0
 Requires-Dist: einx>=0.3.0
 Requires-Dist: einops>=0.8.0
+Requires-Dist: loguru
 Requires-Dist: packaging>=21.0
+Requires-Dist: torch>=2.0

{x_transformers-1.42.17 → x_transformers-1.42.19}/x_transformers.egg-info/requires.txt

@@ -1,4 +1,5 @@
-torch>=2.0
 einx>=0.3.0
 einops>=0.8.0
+loguru
 packaging>=21.0
+torch>=2.0