x-transformers 1.30.21__py3-none-any.whl → 1.30.23__py3-none-any.whl

x_transformers/x_transformers.py

@@ -710,7 +710,7 @@ class LayerScale(Module):
         out, *rest = out
         return out * self.gamma, *rest
 
-class ConditionedLayerScale(Module):
+class AdaptiveLayerScale(Module):
     def __init__(self, fn: Module, dim, dim_condition = None, init_bias_value = -2.):
         super().__init__()
         self.fn = fn
@@ -1181,7 +1181,7 @@ class AttentionLayers(Module):
         use_simple_rmsnorm = False,
         use_adaptive_layernorm = False,
         use_adaptive_rmsnorm = False,
-        use_conditioned_layerscale = False, # paired with use_adaptive_layernorm for ada-ln-zero from DiT paper
+        use_adaptive_layerscale = False, # paired with use_adaptive_layernorm for ada-ln-zero from DiT paper
         dim_condition = None,
         adaptive_condition_mlp = False,
         adaptive_condition_mlp_expansion = 4,
@@ -1201,8 +1201,8 @@ class AttentionLayers(Module):
         rotary_xpos_scale_base = 512,
         rotary_base_rescale_factor = 1.,
         weight_tie_layers = False,
-        custom_layers: Tuple[str] | None = None,
-        layers_execute_order: Tuple[int] | None = None,
+        custom_layers: Tuple[str, ...] | None = None,
+        layers_execute_order: Tuple[int, ...] | None = None,
         sandwich_coef = None,
         par_ratio = None,
         residual_attn = False,
@@ -1332,15 +1332,15 @@ class AttentionLayers(Module):
 
         # determine post branch wrapper
 
-        assert at_most_one_of(use_layerscale, use_conditioned_layerscale)
+        assert at_most_one_of(use_layerscale, use_adaptive_layerscale)
 
         post_branch_fn = None
         post_branch_fn_needs_condition = False
 
         if use_layerscale:
             post_branch_fn = partial(LayerScale, dim = dim, init_value = layerscale_init_value)
-        elif use_conditioned_layerscale:
-            post_branch_fn = partial(ConditionedLayerScale, dim = dim, dim_condition = dim_condition * dim_condition_mult)
+        elif use_adaptive_layerscale:
+            post_branch_fn = partial(AdaptiveLayerScale, dim = dim, dim_condition = dim_condition * dim_condition_mult)
             post_branch_fn_needs_condition = True
 
         self.post_branch_fn_needs_condition = post_branch_fn_needs_condition
@@ -1484,7 +1484,8 @@ class AttentionLayers(Module):
         cache_age = 1,
         return_hiddens = False,
         rotary_pos_emb = None,
-        condition = None
+        condition = None,
+        layers_execute_order: Tuple[int, ...] | None = None
     ):
         assert not (self.cross_attend ^ exists(context)), 'context must be passed in if cross_attend is set to True'
         assert not (exists(condition) ^ self.need_condition), 'condition needs to be passed in if using adaptive layernorm or vice versa'
@@ -1576,7 +1577,11 @@ class AttentionLayers(Module):
             self.layer_dropouts
         )
 
-        layer_variables = tuple(tuple(layer_variable[i] for i in self.layers_execute_order) for layer_variable in layer_variables)
+        # able to override the layers execution order on forward, for trying to depth extrapolate
+
+        layers_execute_order = default(layers_execute_order, self.layers_execute_order)
+
+        layer_variables = tuple(tuple(layer_variable[i] for i in layers_execute_order) for layer_variable in layer_variables)
 
         # go through the attention and feedforward layers
 

{x_transformers-1.30.21.dist-info → x_transformers-1.30.23.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.30.21
+Version: 1.30.23
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.30.21.dist-info → x_transformers-1.30.23.dist-info}/RECORD

@@ -4,11 +4,11 @@ x_transformers/autoregressive_wrapper.py,sha256=uX8Mb0zLsQrZECt_9UGt35g7tC05Rk3n
 x_transformers/continuous.py,sha256=WO52n9lFAXv5-SGadi2cApGF8dkouN8QSTEOuC7erj8,6180
 x_transformers/dpo.py,sha256=LjvWgCkqTl-UuehrzQ8nkX5guLr4whYwsmm7SKSwdls,3450
 x_transformers/nonautoregressive_wrapper.py,sha256=ys_p8obc7lTeeodCqvkRKxOXQ1C9T3j5Jwr-JbVgnXk,10432
-x_transformers/x_transformers.py,sha256=YFycvL28Y-Y-sr-_P2umC9qZ0nAfKFIBR_5mI1oKog0,73974
+x_transformers/x_transformers.py,sha256=tZixUvlsaEj3CpB49KLDOJ2BwYSPjdWotDUjB9Rbf7g,74213
 x_transformers/xl_autoregressive_wrapper.py,sha256=DCx4n0_c1tFai4nOqaWVnqx2p9eutsZsDMiMP1ckxNU,4117
 x_transformers/xval.py,sha256=QE1ltYZTR_eGgIHPP2BrMWVWVLqMW-OpDZh87BSmQEg,8563
-x_transformers-1.30.21.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.30.21.dist-info/METADATA,sha256=7HN3_fnSG0wGIGijv85bqfd8w4_K1m10Pv_PmI8kNrk,662
-x_transformers-1.30.21.dist-info/WHEEL,sha256=cpQTJ5IWu9CdaPViMhC9YzF8gZuS5-vlfoFihTBC86A,91
-x_transformers-1.30.21.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.30.21.dist-info/RECORD,,
+x_transformers-1.30.23.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.30.23.dist-info/METADATA,sha256=LM8Y0bkOF259zCn_FE2A-Uw5Yjr8YrqCKNYuW4DqtQY,662
+x_transformers-1.30.23.dist-info/WHEEL,sha256=cpQTJ5IWu9CdaPViMhC9YzF8gZuS5-vlfoFihTBC86A,91
+x_transformers-1.30.23.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.30.23.dist-info/RECORD,,