x-transformers 1.40.2__py3-none-any.whl → 1.40.4__py3-none-any.whl

x_transformers/x_transformers.py

@@ -442,10 +442,10 @@ class DynamicPositionBias(Module):
         return bias
 
 class AlibiPositionalBias(Module):
-    def __init__(self, heads, total_heads, **kwargs):
+    def __init__(self, heads, total_heads = None, **kwargs):
         super().__init__()
         self.heads = heads
-        self.total_heads = total_heads
+        self.total_heads = default(total_heads, heads)
 
         slopes = Tensor(self._get_slopes(heads))
         slopes = rearrange(slopes, 'h -> h 1 1')
@@ -1353,6 +1353,7 @@ class AttentionLayers(Module):
         use_layerscale = False,
         layerscale_init_value = 0.,
         unet_skips = False,
+        reinject_input = True, # seen first in DEQ paper https://arxiv.org/abs/1909.01377, but later used in a number of papers trying to achieve depthwise generalization https://arxiv.org/abs/2410.03020v1
         **kwargs
     ):
         super().__init__()
@@ -1582,6 +1583,11 @@ class AttentionLayers(Module):
 
         self.skip_combines = ModuleList([])
 
+        # whether there is reinjection of input at every layer
+
+        self.reinject_input = reinject_input
+        self.reinject_input_proj = nn.Linear(dim, dim, bias = False) if reinject_input else None
+
         # iterate and construct layers
 
         for ind, (layer_type, layer_shift_tokens) in enumerate(zip(self.layer_types, shift_tokens)):
@@ -1665,6 +1671,7 @@ class AttentionLayers(Module):
         cache_age = 1,
         return_hiddens = False,
         rotary_pos_emb = None,
+        attn_bias = None,
         condition = None,
         layers_execute_order: tuple[int, ...] | None = None
     ):
@@ -1765,6 +1772,11 @@ class AttentionLayers(Module):
 
         layer_variables = tuple(tuple(layer_variable[i] for i in layers_execute_order) for layer_variable in layer_variables)
 
+        # derived input for reinjection if needed
+
+        if self.reinject_input:
+            inp_inject = self.reinject_input_proj(x)
+
         # store all hiddens for skips
 
         skip_hiddens = []
@@ -1809,6 +1821,9 @@ class AttentionLayers(Module):
                 post_branch_norm = maybe(partial)(post_branch_norm, **norm_kwargs)
                 post_main_norm = maybe(partial)(post_main_norm, **norm_kwargs)
 
+            if self.reinject_input:
+                x = x + inp_inject
+
             if exists(pre_norm):
                 x = pre_norm(x)
 
@@ -1818,7 +1833,7 @@ class AttentionLayers(Module):
             block = partial(block, **block_forward_kwargs)
 
             if layer_type == 'a':
-                out, inter = block(x, mask = mask, context_mask = self_attn_kv_mask, attn_mask = attn_mask, rel_pos = self.rel_pos, rotary_pos_emb = rotary_pos_emb, prev_attn = prev_attn, cache = next(iter_attn_cache, None), mem = layer_mem, mem_mask = layer_mem_mask, return_intermediates = True)
+                out, inter = block(x, mask = mask, context_mask = self_attn_kv_mask, attn_mask = attn_mask, rel_pos = self.rel_pos, rotary_pos_emb = rotary_pos_emb, prev_attn = prev_attn, cache = next(iter_attn_cache, None), mem = layer_mem, mem_mask = layer_mem_mask, attn_bias = attn_bias, return_intermediates = True)
             elif layer_type == 'c':
                 out, inter = block(x, context = context, mask = mask, context_mask = context_mask, prev_attn = prev_cross_attn, cache = next(iter_attn_cache, None), return_intermediates = True)
             elif layer_type == 'f':

{x_transformers-1.40.2.dist-info → x_transformers-1.40.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.40.2
+Version: 1.40.4
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.40.2.dist-info → x_transformers-1.40.4.dist-info}/RECORD

@@ -5,11 +5,11 @@ x_transformers/continuous.py,sha256=cIVEdhfei258__ziV7kQBrJMxCel54bExBTDrO9rfCI,
 x_transformers/dpo.py,sha256=LjvWgCkqTl-UuehrzQ8nkX5guLr4whYwsmm7SKSwdls,3450
 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=SfM0ql3wK7t8KzBXRNnGTdcyq3tQVmHB4VIcfg5sSv4,84604
+x_transformers/x_transformers.py,sha256=g7cGYUcq344QZejqdB_PPHHz1O_zqmdplxgQA07hqks,85298
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=QE1ltYZTR_eGgIHPP2BrMWVWVLqMW-OpDZh87BSmQEg,8563
-x_transformers-1.40.2.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.40.2.dist-info/METADATA,sha256=G1LWuKpy25e1rXV7MFRT5r2F4bHjvyMUrgxTgJIQLic,661
-x_transformers-1.40.2.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
-x_transformers-1.40.2.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.40.2.dist-info/RECORD,,
+x_transformers-1.40.4.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.40.4.dist-info/METADATA,sha256=K7tcPd4ZY7qO773XErYQbik-XDEfLbFQ0lFYSfjAFvY,661
+x_transformers-1.40.4.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
+x_transformers-1.40.4.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.40.4.dist-info/RECORD,,