x-transformers 1.40.7__py3-none-any.whl → 1.40.8__py3-none-any.whl

x_transformers/x_transformers.py

@@ -1803,7 +1803,10 @@ class AttentionLayers(Module):
 
         skip_hiddens = []
 
+        # for value residuals
+
         first_self_attn_inter = None
+        first_cross_attn_inter = None
 
         # go through the attention and feedforward layers
 
@@ -1856,20 +1859,35 @@ class AttentionLayers(Module):
 
             block = partial(block, **block_forward_kwargs)
 
-            maybe_value_residual = None
-            if self.add_value_residual and exists(first_self_attn_inter):
-                maybe_value_residual = first_self_attn_inter.values
+            # handle maybe value residuals
+
+            maybe_self_attn_value_residual = None
+            maybe_cross_attn_value_residual = None
+
+            if self.add_value_residual:
+                if exists(first_self_attn_inter):
+                    maybe_self_attn_value_residual = first_self_attn_inter.values
+
+                if exists(first_cross_attn_inter):
+                    maybe_cross_attn_value_residual = first_cross_attn_inter.values
+
+            # forward depending on layer type
 
             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, attn_bias = attn_bias, value_residual = maybe_value_residual, 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, value_residual = maybe_self_attn_value_residual, 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)
+                out, inter = block(x, context = context, mask = mask, context_mask = context_mask, prev_attn = prev_cross_attn, cache = next(iter_attn_cache, None), value_residual = maybe_cross_attn_value_residual, return_intermediates = True)
             elif layer_type == 'f':
                 out = block(x)
 
+            # store first self or cross attention intermediate for value residual
+
             if not exists(first_self_attn_inter) and layer_type == 'a':
                 first_self_attn_inter = inter
 
+            if not exists(first_cross_attn_inter) and layer_type == 'c':
+                first_cross_attn_inter = inter
+
             if self.resi_dual:
                 outer_residual = outer_residual + out * self.resi_dual_scale
 

{x_transformers-1.40.7.dist-info → x_transformers-1.40.8.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.40.7
+Version: 1.40.8
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.40.7.dist-info → x_transformers-1.40.8.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=azvG00DKeg4s0tH28QeDhU6X2GaeFjq3-0RMxgVt408,86715
+x_transformers/x_transformers.py,sha256=UyJ3-XoanCHOUZLoyp29jt-yNWbgc7pFqpfjOGFFEPY,87367
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=7S00kCuab4tWQa-vf-z-XfzADjVj48MoFIr7VSIvttg,8575
-x_transformers-1.40.7.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.40.7.dist-info/METADATA,sha256=mhcww_WPNU2q-piUEvoDEXwForJdgk9s_K6iCQVFDqo,661
-x_transformers-1.40.7.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
-x_transformers-1.40.7.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.40.7.dist-info/RECORD,,
+x_transformers-1.40.8.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.40.8.dist-info/METADATA,sha256=epku2YLENdljdFU7ge7hhUDXMujquHyb0jhoDSv9yFk,661
+x_transformers-1.40.8.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
+x_transformers-1.40.8.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.40.8.dist-info/RECORD,,