x-transformers 1.40.7__py3-none-any.whl → 1.40.9__py3-none-any.whl

x_transformers/x_transformers.py

@@ -944,6 +944,8 @@ class Attention(Module):
         cope_talking_heads = False,
         softclamp_logits = False,
         logit_softclamp_value = 50.,
+        neutreno_value_residual = False, # Nguyen et al. https://arxiv.org/abs/2312.00751
+        neutreno_alpha = 0.4,
         onnxable = False
     ):
         super().__init__()
@@ -982,6 +984,11 @@ class Attention(Module):
 
         self.to_r = LinearNoBias(dim, v_dim) if tensor_product else None
 
+        # the value residual used by Nguyen et al. in https://arxiv.org/abs/2312.00751 for countering oversmoothing
+
+        self.neutreno_value_residual = neutreno_value_residual
+        self.neutreno_alpha = neutreno_alpha
+
         # add GLU gating for aggregated values, from alphafold2
 
         self.to_v_gate = None
@@ -1244,11 +1251,15 @@ 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 previous values passed in for residual, either invoke resformer or neutreno
 
         if exists(value_residual):
-            v = v + value_residual
+            if self.neutreno_value_residual:
+                diff_values = (value_residual - v) * self.neutreno_alpha
+                diff_values = repeat(diff_values, 'b h n d -> b (r h) n d', r = h // kv_h)
+            else:
+                # https://arxiv.org/abs/2410.17897v1
+                v = v + value_residual
 
         # attention is all we need
 
@@ -1259,10 +1270,13 @@ class Attention(Module):
             prev_attn = prev_attn
         )
 
-        # store the values for resformer from Zhou et al. https://arxiv.org/abs/2410.17897v1
+        # store the values for resformer or Neutreno
 
         intermediates.values = v
 
+        if exists(value_residual) and self.neutreno_value_residual:
+            out = out + diff_values
+
         # https://arxiv.org/abs/2208.06061 proposes to add a residual for better gradients
 
         if exists(r):
@@ -1365,7 +1379,7 @@ 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
+        add_value_residual = False, # resformer from Zhou et al - https://arxiv.org/abs/2410.17897v1
         **kwargs
     ):
         super().__init__()
@@ -1378,6 +1392,7 @@ class AttentionLayers(Module):
         assert len(kwargs) == 0, f'unrecognized kwargs passed in {kwargs.keys()}'
 
         dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD)
+        add_value_residual |= attn_kwargs.get('neutreno_value_residual', False)
 
         self.dim = dim
         self.causal = causal
@@ -1803,7 +1818,10 @@ 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 +1874,35 @@ 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-1.40.9.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.40.7
+Version: 1.40.9
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.40.7.dist-info → x_transformers-1.40.9.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=NoWBoiz1t8_QytYo1T2YBFk-7H9s38k2t-EksxqUkMU,88072
 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-1.40.9.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.40.9.dist-info/METADATA,sha256=xSxqFkhGfr5dU2xI0xo3UzlPMSuaaR4Rd2TrDpEyxcE,661
+x_transformers-1.40.9.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
+x_transformers-1.40.9.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.40.9.dist-info/RECORD,,