x-transformers 1.31.3__py3-none-any.whl → 1.31.5__py3-none-any.whl

x_transformers/attend.py

@@ -69,7 +69,7 @@ def onnx_create_causal_mask(i, j, device):
 
 # main class
 
-class Attend(nn.Module):
+class Attend(Module):
     def __init__(
         self,
         *,
@@ -81,7 +81,8 @@ class Attend(nn.Module):
         scale = None,
         qk_norm = False,
         flash = False,
-        logit_softclamp_value = None,
+        softclamp_logits = False,
+        logit_softclamp_value = 30.,
         add_zero_kv = False,
         cope = None,
         onnxable = False,
@@ -123,10 +124,11 @@ class Attend(nn.Module):
 
         # soft clamp attention logit value
 
-        if exists(logit_softclamp_value):
+        if softclamp_logits:
             assert not flash, 'flash attention not compatible with logit softclamp value yet'
             assert logit_softclamp_value > 0.
 
+        self.softclamp_logits = softclamp_logits
         self.logit_softclamp_value = logit_softclamp_value
 
         # contextual positional encoding
@@ -308,6 +310,9 @@ class Attend(nn.Module):
         if exists(attn_bias):
             sim = sim + attn_bias
 
+        if self.softclamp_logits:
+            sim = softclamp(sim, self.logit_softclamp_value)
+
         i, j, dtype = *sim.shape[-2:], sim.dtype
 
         mask_value = -torch.finfo(sim.dtype).max
@@ -329,9 +334,6 @@ class Attend(nn.Module):
 
         pre_softmax_attn = sim.clone()
 
-        if exists(self.logit_softclamp_value):
-            sim = softclamp(sim, self.logit_softclamp_value)
-
         attn = self.attn_fn(sim, dim = -1)
         attn = attn.type(dtype)
 

x_transformers/x_transformers.py

@@ -566,7 +566,7 @@ class LayerNorm(Module):
     def __init__(
         self,
         dim,
-        unit_offset = 0.
+        unit_offset = False
     ):
         """
         bias-less layernorm has been shown to be more stable. most newer models have moved towards rmsnorm, also bias-less
@@ -576,9 +576,7 @@ class LayerNorm(Module):
 
         self.ln = nn.LayerNorm(dim, elementwise_affine = False)
         self.gamma = nn.Parameter(torch.ones(dim))
-        nn.init.constant_(self.gamma, 1. - unit_offset)
-
-        self.register_buffer('beta', torch.zeros(dim), persistent = False)
+        nn.init.constant_(self.gamma, 1. - float(unit_offset))
 
     def forward(self, x):
         normed = self.ln(x)
@@ -596,7 +594,6 @@ class AdaptiveLayerNorm(Module):
 
         self.ln = nn.LayerNorm(dim, elementwise_affine = False)
         self.to_gamma = nn.Linear(dim_condition, dim, bias = False)
-
         nn.init.zeros_(self.to_gamma.weight)
 
     def forward(self, x, *, condition):
@@ -608,14 +605,14 @@ class ScaleNorm(Module):
     def __init__(
         self,
         dim,
-        unit_offset = 0.
+        unit_offset = False
     ):
         super().__init__()
         self.unit_offset = unit_offset
         self.scale = dim ** 0.5
 
         self.g = nn.Parameter(torch.zeros(1))
-        nn.init.constant_(self.g, 1. - unit_offset)
+        nn.init.constant_(self.g, 1. - float(unit_offset))
 
     def forward(self, x):
         return F.normalize(x, dim = -1) * self.scale * (self.g + self.unit_offset)
@@ -624,14 +621,14 @@ class RMSNorm(Module):
     def __init__(
         self,
         dim,
-        unit_offset = 0.
+        unit_offset = False
     ):
         super().__init__()
         self.unit_offset = unit_offset
         self.scale = dim ** 0.5
 
         self.g = nn.Parameter(torch.zeros(dim))
-        nn.init.constant_(self.g, 1. - unit_offset)
+        nn.init.constant_(self.g, 1. - float(unit_offset))
 
     def forward(self, x):
         return F.normalize(x, dim = -1) * self.scale * (self.g + self.unit_offset)
@@ -738,14 +735,14 @@ class LayerScale(Module):
         fn: Module,
         dim,
         init_value = 0.,
-        unit_offset = 0.
+        unit_offset = False
     ):
         super().__init__()
         self.unit_offset = unit_offset
 
         self.fn = fn
         self.gamma = nn.Parameter(torch.zeros(dim))
-        nn.init.constant_(self.gamma, init_value - unit_offset)
+        nn.init.constant_(self.gamma, init_value - float(unit_offset))
 
     def forward(self, x, **kwargs):
         out = self.fn(x, **kwargs)
@@ -887,7 +884,8 @@ class Attention(Module):
         cope_max_pos = 16,
         cope_soft_onehot_pos = False,
         cope_talking_heads = False,
-        logit_softclamp_value = None,
+        softclamp_logits = False,
+        logit_softclamp_value = 30.,
         onnxable = False
     ):
         super().__init__()
@@ -990,6 +988,7 @@ class Attention(Module):
             scale = qk_norm_scale if qk_norm else self.scale,
             add_zero_kv = add_zero_kv,
             flash = flash,
+            softclamp_logits = softclamp_logits,
             logit_softclamp_value = logit_softclamp_value,
             cope = cope,
             onnxable = onnxable
@@ -1370,8 +1369,8 @@ class AttentionLayers(Module):
         norm_fn = partial(norm_class, dim)
 
         if not norm_need_condition and norm_add_unit_offset:
-            # researcher Ohad Rubin shares in a blog post by adding an offset to gammas and betas, they can be subjected to weight decay safely
-            norm_fn = partial(norm_fn, unit_offset = 1.)
+            # researcher Ohad Rubin shares in a blog post by adding an offset to gammas, they can be subjected to weight decay safely
+            norm_fn = partial(norm_fn, unit_offset = True)
 
         self.norm_need_condition = norm_need_condition
         self.dim_condition = dim_condition
@@ -1404,7 +1403,7 @@ class AttentionLayers(Module):
         self.post_branch_fn_needs_condition = post_branch_fn_needs_condition
 
         if exists(post_branch_fn) and not post_branch_fn_needs_condition and norm_add_unit_offset:
-            post_branch_fn = partial(post_branch_fn, unit_offset = 1.)
+            post_branch_fn = partial(post_branch_fn, unit_offset = True)
 
         # setup mlp for conditioning
 

{x_transformers-1.31.3.dist-info → x_transformers-1.31.5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.31.3
+Version: 1.31.5
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.31.3.dist-info → x_transformers-1.31.5.dist-info}/RECORD

@@ -1,14 +1,14 @@
 x_transformers/__init__.py,sha256=8LQl-dNL6vj8VHRx5LMSOlRDTXQvYOuM21PDXz8WdiI,703
-x_transformers/attend.py,sha256=ap2QkD-bRadFE9ZFQP84Lo1P2DpLOXPam24Jq9ybpPY,10903
+x_transformers/attend.py,sha256=UWq0bElvJf-_j1N2QbJ2yg28xkWlhnOrLjMJt3If3ao,10956
 x_transformers/autoregressive_wrapper.py,sha256=uX8Mb0zLsQrZECt_9UGt35g7tC05Rk3nPqO6xp2FFCc,9619
 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=D9l3jL1D0RzHynIJcSUpxQec1n-7cHgRQZNDdDoYCFQ,75832
+x_transformers/x_transformers.py,sha256=VL9Dm8L5jnpgyt_V6DWGtLs9MeiTn7ZMCQdcHSFLVo8,75871
 x_transformers/xl_autoregressive_wrapper.py,sha256=DCx4n0_c1tFai4nOqaWVnqx2p9eutsZsDMiMP1ckxNU,4117
 x_transformers/xval.py,sha256=QE1ltYZTR_eGgIHPP2BrMWVWVLqMW-OpDZh87BSmQEg,8563
-x_transformers-1.31.3.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.31.3.dist-info/METADATA,sha256=cS3vVeEi3fSoYqXWDkW6FJhgmX11vrfQr9Lc3kMT9MI,661
-x_transformers-1.31.3.dist-info/WHEEL,sha256=mguMlWGMX-VHnMpKOjjQidIo1ssRlCFu4a4mBpz1s2M,91
-x_transformers-1.31.3.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.31.3.dist-info/RECORD,,
+x_transformers-1.31.5.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.31.5.dist-info/METADATA,sha256=o7z9SmDOf9BAU9FyMRuAbX4LIH8nIACA0acjGHPMz8s,661
+x_transformers-1.31.5.dist-info/WHEEL,sha256=mguMlWGMX-VHnMpKOjjQidIo1ssRlCFu4a4mBpz1s2M,91
+x_transformers-1.31.5.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.31.5.dist-info/RECORD,,