add ability to configure xpos scale base

local_attention/local_attention.py

@@ -63,7 +63,8 @@ def __init__(
         autopad = False,
         exact_windowsize = False,
         scale = None,
-        use_xpos = False
+        use_xpos = False,
+        xpos_scale_base = None
     ):
         super().__init__()
         look_forward = default(look_forward, 0 if causal else 1)
@@ -92,7 +93,12 @@ def __init__(
         if exists(rel_pos_emb_config) or exists(dim):  # backwards compatible with old `rel_pos_emb_config` deprecated argument
             if exists(rel_pos_emb_config):
                 dim = rel_pos_emb_config[0]
-            self.rel_pos = SinusoidalEmbeddings(dim, use_xpos = use_xpos, scale_base = window_size // 2)
+
+            self.rel_pos = SinusoidalEmbeddings(
+                dim,
+                use_xpos = use_xpos,
+                scale_base = default(xpos_scale_base, window_size // 2)
+            )
 
     def forward(self, q, k, v, mask = None, input_mask = None, window_size = None):
         mask = default(mask, input_mask)

local_attention/transformer.py

@@ -48,6 +48,7 @@ def __init__(
         qk_rmsnorm = False,
         qk_scale = 8,
         use_xpos = False,
+        xpos_scale_base = None,
         **kwargs
     ):
         super().__init__()        
@@ -72,6 +73,7 @@ def __init__(
             scale = (qk_scale if qk_rmsnorm else None),
             exact_windowsize = True,
             use_xpos = use_xpos,
+            xpos_scale_base = xpos_scale_base,
             **kwargs
         )
 
@@ -131,6 +133,7 @@ def __init__(
         ff_dropout = 0.,
         ignore_index = -1,
         use_xpos = False,
+        xpos_scale_base = None,
         **kwargs
     ):
         super().__init__()
@@ -142,7 +145,7 @@ def __init__(
 
         for _ in range(depth):
             self.layers.append(nn.ModuleList([
-                LocalMHA(dim = dim, dim_head = dim_head, heads = heads, dropout = attn_dropout, causal = causal, window_size = local_attn_window_size, use_xpos = use_xpos, prenorm = True, **kwargs),
+                LocalMHA(dim = dim, dim_head = dim_head, heads = heads, dropout = attn_dropout, causal = causal, window_size = local_attn_window_size, use_xpos = use_xpos, xpos_scale_base = xpos_scale_base, prenorm = True, **kwargs),
                 FeedForward(dim = dim, mult = ff_mult, dropout = ff_dropout)
             ]))
 

setup.py

@@ -3,7 +3,7 @@
 setup(
   name = 'local-attention',
   packages = find_packages(),
-  version = '1.7.0',
+  version = '1.7.1',
   license='MIT',
   description = 'Local attention, window with lookback, for language modeling',
   long_description_content_type = 'text/markdown',