x-transformers 1.40.6__py3-none-any.whl → 1.40.8__py3-none-any.whl

x_transformers/attend.py

@@ -22,6 +22,7 @@ class Intermediates:
     qk_similarities:    Tensor | None = None
     pre_softmax_attn:   Tensor | None = None
     post_softmax_attn:  Tensor | None = None
+    values:             Tensor | None = None
     cached_kv:          Tuple[Tensor, Tensor] | None = None
     layer_type:         str | None = None
 

x_transformers/x_transformers.py

@@ -1114,6 +1114,7 @@ class Attention(Module):
         mem_mask = None,
         return_intermediates = False,
         cache: Intermediates | None = None,
+        value_residual = None
     ):
         b, n, h, kv_h, head_scale, num_mem_kv, device, has_context = x.shape[0], x.shape[1], self.heads, self.kv_heads, self.head_scale, self.num_mem_kv, x.device, exists(context)
 
@@ -1243,6 +1244,12 @@ class Attention(Module):
             attn_bias = rel_pos(i, j)
             attn_bias = pad_at_dim(attn_bias, (num_mem_kv, 0), value = 0.) # handle memory key / values
 
+        # previous values passed in
+        # https://arxiv.org/abs/2410.17897v1
+
+        if exists(value_residual):
+            v = v + value_residual
+
         # attention is all we need
 
         out, intermediates = self.attend(
@@ -1252,6 +1259,10 @@ class Attention(Module):
             prev_attn = prev_attn
         )
 
+        # store the values for resformer from Zhou et al. https://arxiv.org/abs/2410.17897v1
+
+        intermediates.values = v
+
         # https://arxiv.org/abs/2208.06061 proposes to add a residual for better gradients
 
         if exists(r):
@@ -1354,6 +1365,7 @@ class AttentionLayers(Module):
         layerscale_init_value = 0.,
         unet_skips = False,
         reinject_input = False, # 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
+        add_value_residual = False, # resformer from Zhou et al - https://arxiv.org/abs/2410.17897v1 | TODO: also add NeuTRENO from Nguyen et al. https://arxiv.org/abs/2312.00751
         **kwargs
     ):
         super().__init__()
@@ -1588,6 +1600,10 @@ class AttentionLayers(Module):
         self.reinject_input = reinject_input
         self.reinject_input_proj = nn.Linear(dim, dim, bias = False) if reinject_input else None
 
+        # add the value from the first self attention block to all latter projected self attention values as a residual
+
+        self.add_value_residual = add_value_residual
+
         # iterate and construct layers
 
         for ind, (layer_type, layer_shift_tokens) in enumerate(zip(self.layer_types, shift_tokens)):
@@ -1787,6 +1803,11 @@ 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
 
         for ind, (layer_type, skip_combine, (norm, block, residual_fn), layer_dropout) in enumerate(zip(*layer_variables)):
@@ -1838,13 +1859,35 @@ class AttentionLayers(Module):
 
             block = partial(block, **block_forward_kwargs)
 
+            # 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, 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-1.40.6.dist-info → x_transformers-1.40.8.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.40.6
+Version: 1.40.8
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.40.6.dist-info → x_transformers-1.40.8.dist-info}/RECORD

@@ -1,15 +1,15 @@
 x_transformers/__init__.py,sha256=-MkQrSc37cTVDX7AOykxunYnqVtFlQ7lb0Cse5dsGWU,793
-x_transformers/attend.py,sha256=VbB0fi-ETgAF4dc2a_Meaqvt14LMaRVIjZ8NexUX8F0,17239
+x_transformers/attend.py,sha256=SdWlV8Vp5DtpsOzAd0LRhm4VGrJf0lJCGiV2_j_CtoA,17284
 x_transformers/autoregressive_wrapper.py,sha256=DOJJCMMDOqDYKWy_IaG5IyKsXD3AW6amzfUgdAADOLY,10500
 x_transformers/continuous.py,sha256=cIVEdhfei258__ziV7kQBrJMxCel54bExBTDrO9rfCI,6450
 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=FKHaJFQuMNiFMrjDF13OE3vk-iYf_qwogBNxVpiQSc4,85671
+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-1.40.6.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.40.6.dist-info/METADATA,sha256=HsoNUu71hkonsBhThVN46rakFnIAGOav3pHDpYnX9t8,661
-x_transformers-1.40.6.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
-x_transformers-1.40.6.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.40.6.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,,