x-transformers 1.41.0__py3-none-any.whl → 1.41.2__py3-none-any.whl

x_transformers/x_transformers.py

@@ -101,12 +101,6 @@ def log(t, eps = 1e-20):
 def max_neg_value(tensor):
     return -torch.finfo(tensor.dtype).max
 
-def reverse_cumsum(t, dim = -1):
-    t = t.flip(dims = (dim,))
-    t = t.cumsum(dim = dim)
-    t = t.flip(dims = (dim,))
-    return t
-
 def l2norm(t, groups = 1):
     t = rearrange(t, '... (g d) -> ... g d', g = groups)
     t = F.normalize(t, p = 2, dim = -1)
@@ -514,26 +508,54 @@ class DataDependentAlibi(Module):
         self.to_forget_gates = nn.Sequential(
             linear,
             Rearrange('b n h -> b h n'),
-            nn.Sigmoid()
+            nn.LogSigmoid()
         )
 
         nn.init.constant_(linear.bias, 5.)
 
     def forward(self, x):
-        seq = x.shape[-2]
+        forget_gates = self.to_forget_gates(x)
+        forget_gates = forget_gates.cumsum(dim = -1)
+        forget_gates = einx.subtract('b h i, b h j -> b h i j', forget_gates, forget_gates)
+        return forget_gates
 
-        forget_gates = self.to_forget_gates(x).log()
-        forget_gates = repeat(forget_gates, 'b h j -> b h i j', i = seq)
+class PerRowDataDependentAlibi(Module):
+    """ same as data dependent alibi from forgetting transformer, but the forgetting gates are also derived by a queris and keys with a small head dimension """
 
-        # causal mask out, including diagonal (so token to itself attention is never masked out)
+    def __init__(
+        self,
+        dim,
+        heads,
+        dim_head = 8
+    ):
+        super().__init__()
+        self.scale = dim_head ** -0.5
 
-        causal_mask = torch.ones((seq, seq), dtype = torch.bool, device = x.device).triu()
+        linear = nn.Linear(dim, heads * dim_head * 2, bias = False)
+
+        self.to_forget_gates = nn.Sequential(
+            linear,
+            Rearrange('b n (kv h d) -> kv b h n d', kv = 2, d = dim_head)
+        )
+
+    def forward(self, x):
+        q, k = self.to_forget_gates(x)
+        forget_gates = einsum('... i d, ... j d -> ... i j', q, k) * self.scale
+
+        forget_gates = F.logsigmoid(forget_gates)
+
+        # mask out upper triangle + diagonal
+
+        n = x.shape[-2]
+        causal_mask = torch.ones((n, n), dtype = torch.bool, device = x.device).triu()
 
         forget_gates = forget_gates.masked_fill(causal_mask, 0.)
 
-        # reverse cumulative sum in log space (equivalent to cumprod)
+        # reverse cumsum
 
-        forget_gates = reverse_cumsum(forget_gates)
+        forget_gates = forget_gates.flip(dims = (-1,))
+        forget_gates = forget_gates.cumsum(dim = -1)
+        forget_gates = forget_gates.flip(dims = (-1,))
 
         return forget_gates
 
@@ -986,6 +1008,8 @@ class Attention(Module):
         add_zero_kv = False,         # same as add_zero_attn in pytorch
         rotary_embed_values = False,
         data_dependent_alibi = False,
+        data_dependent_alibi_per_row = False,
+        data_dependent_alibi_per_row_dim_head = 8,
         use_cope = False,
         cope_max_pos = 16,
         cope_soft_onehot_pos = False,
@@ -1097,10 +1121,13 @@ class Attention(Module):
         if data_dependent_alibi:
             assert causal, 'data dependent alibi only works for autoregressive for now until further research'
 
-            self.data_dependent_alibi = DataDependentAlibi(
-                dim,
-                heads = heads
-            )
+            dda_klass = DataDependentAlibi if not data_dependent_alibi_per_row else PerRowDataDependentAlibi
+            dda_kwargs = dict(dim = dim, heads = heads)
+
+            if data_dependent_alibi_per_row:
+                dda_kwargs.update(dim_head = data_dependent_alibi_per_row_dim_head)
+
+            self.data_dependent_alibi = dda_klass(**dda_kwargs)
 
         # attend class - includes core attention algorithm + talking heads
 

{x_transformers-1.41.0.dist-info → x_transformers-1.41.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.41.0
+Version: 1.41.2
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.41.0.dist-info → x_transformers-1.41.2.dist-info}/RECORD

@@ -5,11 +5,11 @@ x_transformers/continuous.py,sha256=cIVEdhfei258__ziV7kQBrJMxCel54bExBTDrO9rfCI,
 x_transformers/dpo.py,sha256=xt4OuOWhU8pN3OKN2LZAaC2NC8iiEnchqqcrPWVqf0o,3521
 x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg,9252
 x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dCN7fjlzd3K1rLUY,10510
-x_transformers/x_transformers.py,sha256=fz71zW2IQ3NQU_csHbzCwFzGHNwrdIF9rZTLhUjmM_Q,90260
+x_transformers/x_transformers.py,sha256=5yoN96mTpw3qygZg2plVed5lTOs9p0REtt50lOlllHU,91334
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=7S00kCuab4tWQa-vf-z-XfzADjVj48MoFIr7VSIvttg,8575
-x_transformers-1.41.0.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.41.0.dist-info/METADATA,sha256=832V1ChJ77viLmQO7d9RU7R9SV_bspLgVl9vtdRdq5Q,689
-x_transformers-1.41.0.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-x_transformers-1.41.0.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.41.0.dist-info/RECORD,,
+x_transformers-1.41.2.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.41.2.dist-info/METADATA,sha256=84WHjWdVHIxkHdHxMEKK5OEssLU80cdj4aGTJQ7SpO4,689
+x_transformers-1.41.2.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+x_transformers-1.41.2.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.41.2.dist-info/RECORD,,