x-transformers 1.42.4__py3-none-any.whl → 1.42.6__py3-none-any.whl

x_transformers/x_transformers.py

@@ -512,12 +512,15 @@ class DataDependentAlibi(Module):
         self,
         dim,
         heads,
+        causal = True,
         bias_init = 5.,
-        post_log_scale = 1.
+        post_log_scale = 1.,
     ):
         super().__init__()
 
-        linear = nn.Linear(dim, heads)
+        self.causal = causal
+
+        linear = nn.Linear(dim, heads * (1 if causal else 2))
 
         self.to_forget_gates = nn.Sequential(
             linear,
@@ -529,9 +532,21 @@ class DataDependentAlibi(Module):
         self.post_log_scale = post_log_scale
 
     def forward(self, x):
+        bidirectional = not self.causal
+
         forget_gates = self.to_forget_gates(x) * self.post_log_scale
+
         forget_gates = forget_gates.cumsum(dim = -1)
+
+        if bidirectional:
+            forget_gates, forget_gates_reversed = forget_gates.chunk(2, dim = 1)
+
         forget_gates = einx.subtract('b h i, b h j -> b h i j', forget_gates, forget_gates)
+
+        if bidirectional:
+            forget_gates_reversed = einx.subtract('b h j, b h i -> b h i j', forget_gates_reversed, forget_gates_reversed)
+            forget_gates = forget_gates.tril() + forget_gates_reversed.triu()
+
         return forget_gates
 
 class PerRowDataDependentAlibi(Module):
@@ -541,10 +556,13 @@ class PerRowDataDependentAlibi(Module):
         self,
         dim,
         heads,
+        causal = True,
         dim_head = 8,
         post_log_scale = 1.
     ):
         super().__init__()
+        assert causal, 'bidirectional not supported yet'
+
         self.scale = dim_head ** -0.5
 
         linear = nn.Linear(dim, heads * dim_head * 2, bias = False)
@@ -1037,7 +1055,12 @@ class Attention(Module):
         logit_softclamp_value = 50.,
         neutreno_value_residual = False, # Nguyen et al. https://arxiv.org/abs/2312.00751
         neutreno_alpha = 0.4,
-        onnxable = False
+        onnxable = False,
+        attend_sdp_kwargs: dict = dict(
+            enable_flash = True,
+            enable_math = True,
+            enable_mem_efficient = True
+        )
     ):
         super().__init__()
         dim_kv = default(dim_context, dim)
@@ -1138,10 +1161,9 @@ class Attention(Module):
         self.data_dependent_alibi = None
 
         if data_dependent_alibi:
-            assert causal, 'data dependent alibi only works for autoregressive for now until further research'
 
             dda_klass = DataDependentAlibi if not data_dependent_alibi_per_row else PerRowDataDependentAlibi
-            dda_kwargs = dict(dim = dim, heads = heads)
+            dda_kwargs = dict(dim = dim, heads = heads, causal = causal)
 
             if data_dependent_alibi_per_row:
                 dda_kwargs.update(dim_head = data_dependent_alibi_per_row_dim_head)
@@ -1171,7 +1193,8 @@ class Attention(Module):
             softclamp_logits = softclamp_logits,
             logit_softclamp_value = logit_softclamp_value,
             cope = cope,
-            onnxable = onnxable
+            onnxable = onnxable,
+            sdp_kwargs = attend_sdp_kwargs
         )
 
         # head scaling

{x_transformers-1.42.4.dist-info → x_transformers-1.42.6.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.42.4
+Version: 1.42.6
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.42.4.dist-info → x_transformers-1.42.6.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=o_Rm-v1XJyIYU_zDcXWxbHN6whFcK8VKRHvlqTNaQTc,92062
+x_transformers/x_transformers.py,sha256=cPsSl1s14_c9fMdn9cZwe6Eg3aDbcRyCTsoXUJusWUg,92706
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=7S00kCuab4tWQa-vf-z-XfzADjVj48MoFIr7VSIvttg,8575
-x_transformers-1.42.4.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.42.4.dist-info/METADATA,sha256=9KikrLFEmmDn92O9ne5Qd6pEuiztY21vkXg5KWiChhw,689
-x_transformers-1.42.4.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-x_transformers-1.42.4.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.42.4.dist-info/RECORD,,
+x_transformers-1.42.6.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.42.6.dist-info/METADATA,sha256=OANeMK9I504gC7iErAdYMTGBUEl6FOcEwm97o4OyC1k,689
+x_transformers-1.42.6.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+x_transformers-1.42.6.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.42.6.dist-info/RECORD,,