x-transformers 2.6.3__py3-none-any.whl → 2.6.5__py3-none-any.whl

x_transformers/attend.py

@@ -4,8 +4,8 @@ from functools import partial
 from typing import Tuple, Callable
 
 import torch
-from torch.nn import Module
-from torch import nn, einsum, Tensor
+from torch.nn import Module, Parameter
+from torch import cat, nn, einsum, Tensor
 import torch.nn.functional as F
 
 from collections import namedtuple
@@ -176,6 +176,7 @@ class Attend(Module):
         softclamp_logits = False,
         logit_softclamp_value = 50.,
         add_zero_kv = False,
+        head_learned_sinks = 0,
         selective = False,
         hard = False,
         cope = None,
@@ -254,6 +255,13 @@ class Attend(Module):
 
         self.add_zero_kv = add_zero_kv
 
+        # learned sink concatted pre-softmax, working solution from gpt-oss
+
+        self.has_head_learned_sinks = head_learned_sinks > 0
+        assert not (self.has_head_learned_sinks and flash), f'not supported for flash attention yet'
+
+        self.head_attn_sinks = Parameter(torch.zeros(heads, head_learned_sinks)) if self.has_head_learned_sinks else None
+
         # soft clamp attention logit value
 
         if softclamp_logits:
@@ -315,10 +323,10 @@ class Attend(Module):
         if self.l2_distance:
             k_norm_sq = k.norm(dim = -1, keepdim = True) ** 2
             k = F.pad(k, (0, 1), value = -1.)
-            k = torch.cat((k, k_norm_sq), dim = -1)
+            k = cat((k, k_norm_sq), dim = -1)
 
             q_norm_sq = q.norm(dim = -1, keepdim = True) ** 2
-            q = torch.cat((2 * q, q_norm_sq), dim = -1)
+            q = cat((2 * q, q_norm_sq), dim = -1)
             q = F.pad(q, (0, 1), value = -1.)
 
         # handle scale - by default they scale by dim_head ** -0.5, but need to take care if using cosine sim attention
@@ -509,6 +517,12 @@ class Attend(Module):
         if self.selective:
             sim = selective_attn(sim)
 
+        if self.has_head_learned_sinks:
+            # add learned attention sink
+            num_sinks = self.head_attn_sinks.shape[-1]
+            attn_sink = repeat(self.head_attn_sinks, 'h sinks -> b h i sinks', b = sim.shape[0], i = sim.shape[2])
+            sim = cat((attn_sink, sim), dim = -1)
+
         pre_softmax_attn = sim
 
         attn = self.attn_fn(sim)
@@ -517,6 +531,10 @@ class Attend(Module):
 
         post_softmax_attn = attn
 
+        if self.has_head_learned_sinks:
+            # remove attention sink
+            attn = attn[..., num_sinks:]
+
         attn = self.attn_dropout(attn)
 
         if exists(self.post_softmax_talking_heads):

x_transformers/x_transformers.py

@@ -1319,6 +1319,7 @@ class Attention(Module):
         value_dim_head = None,
         dim_out = None,
         add_zero_kv = False,         # same as add_zero_attn in pytorch
+        head_learned_sinks = 0,
         rotate_num_heads = None,
         data_dependent_alibi = False,
         data_dependent_alibi_per_row = False,
@@ -1515,6 +1516,7 @@ class Attention(Module):
             selective = selective,
             custom_attn_fn = custom_attn_fn,
             add_zero_kv = add_zero_kv,
+            head_learned_sinks = head_learned_sinks,
             flash = flash,
             softclamp_logits = softclamp_logits,
             logit_softclamp_value = logit_softclamp_value,

{x_transformers-2.6.3.dist-info → x_transformers-2.6.5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.6.3
+Version: 2.6.5
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers
@@ -2507,4 +2507,14 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@misc{openai_gpt_oss,
+  author       = {OpenAI},
+  title        = {Introducing gpt-oss},
+  howpublished = {https://openai.com/index/introducing-gpt-oss},
+  month        = {August},
+  year         = {2025}
+}
+```
+
 *solve intelligence... then use that to solve everything else.* - Demis Hassabis

{x_transformers-2.6.3.dist-info → x_transformers-2.6.5.dist-info}/RECORD

@@ -1,5 +1,5 @@
 x_transformers/__init__.py,sha256=aVuhUU0572TJHW88BVc4yA2tla0Zb8l3NH7W4RZ1AEs,1005
-x_transformers/attend.py,sha256=Ax34Rw56xXAWfFPqtZ_B8iKEW2EfQdbVoc9uFjfeNjA,17404
+x_transformers/attend.py,sha256=DX_qrDkz98Db0eNapbzciJbVp5dsWIFWdpv2LUfebJs,18223
 x_transformers/autoregressive_wrapper.py,sha256=BsGO9xfVYkvynqbU1__tu_S_cxl7gss0YwnkhIa2baY,18401
 x_transformers/belief_state_wrapper.py,sha256=YLUMk6t2MhFBEw5lHDDHJHcoCxTIkHvxTNY__GGZEKU,13374
 x_transformers/continuous.py,sha256=hpb1sSbt3k2LNzzjrjSd8F5xOIbKj7IluV9MBEAFLkw,13031
@@ -9,10 +9,10 @@ x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg
 x_transformers/neo_mlp.py,sha256=XCNnnop9WLarcxap1kGuYc1x8GHvwkZiDRnXOxSl3Po,3452
 x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dCN7fjlzd3K1rLUY,10510
 x_transformers/up_wrapper.py,sha256=YC2LN14_7Xx9Wtiek2rtEJ_qHqdfSmKlh3d7Cgxwd80,7073
-x_transformers/x_transformers.py,sha256=B7dv_LuzODwCrTsfDnp28g-_lMnirQE3gteQwSGyW5k,122100
+x_transformers/x_transformers.py,sha256=F_ZR9jysYmkbqKvsZmzXqOP3VznVeivXVOstAwKIdPU,122185
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=AwwYUm8yDAtKQyKJDIhYMsiLTJ_skh3scUFMjp5sda8,8597
-x_transformers-2.6.3.dist-info/METADATA,sha256=DaTrEChlXc_zpUXv-Jw3A4ca4aon0Ons7wl4-wj1XzY,90223
-x_transformers-2.6.3.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-x_transformers-2.6.3.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-2.6.3.dist-info/RECORD,,
+x_transformers-2.6.5.dist-info/METADATA,sha256=yMl0MlBbo7D9dOu_cBQz38iJQ3a6F8PlaPCo5RQXrSA,90445
+x_transformers-2.6.5.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+x_transformers-2.6.5.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-2.6.5.dist-info/RECORD,,