x-transformers 1.42.17__py3-none-any.whl → 1.42.19__py3-none-any.whl

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.dist-info → x_transformers-1.42.19.dist-info}/METADATA

@@ -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.dist-info → x_transformers-1.42.19.dist-info}/RECORD

@@ -6,11 +6,11 @@ x_transformers/dpo.py,sha256=xt4OuOWhU8pN3OKN2LZAaC2NC8iiEnchqqcrPWVqf0o,3521
 x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg,9252
 x_transformers/neo_mlp.py,sha256=XCNnnop9WLarcxap1kGuYc1x8GHvwkZiDRnXOxSl3Po,3452
 x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dCN7fjlzd3K1rLUY,10510
-x_transformers/x_transformers.py,sha256=Wvkw4j_78413LdCnCt_wHgcVFiCbzrC8u4TH2iXhkNU,95181
+x_transformers/x_transformers.py,sha256=pDYtIGhoo-lFn_ULJETnQz1Z0QYuDsD4ReTlPy__jwo,95993
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=7S00kCuab4tWQa-vf-z-XfzADjVj48MoFIr7VSIvttg,8575
-x_transformers-1.42.17.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.42.17.dist-info/METADATA,sha256=T1MDXNdxqdPkqFpGuJVb7vBhniGCbHefm5C-lhb3LJk,717
-x_transformers-1.42.17.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
-x_transformers-1.42.17.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.42.17.dist-info/RECORD,,
+x_transformers-1.42.19.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.42.19.dist-info/METADATA,sha256=pJgi1Jp7FvM1o_x3a7uOaSJ8x0pNgIQnAp4lSI3K__o,739
+x_transformers-1.42.19.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
+x_transformers-1.42.19.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.42.19.dist-info/RECORD,,