x-transformers 2.9.2__py3-none-any.whl → 2.10.1__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Potentially problematic release.
This version of x-transformers might be problematic. Click here for more details.
- x_transformers/attend.py +20 -0
- x_transformers/x_transformers.py +2 -0
- {x_transformers-2.9.2.dist-info → x_transformers-2.10.1.dist-info}/METADATA +13 -1
- {x_transformers-2.9.2.dist-info → x_transformers-2.10.1.dist-info}/RECORD +6 -6
- {x_transformers-2.9.2.dist-info → x_transformers-2.10.1.dist-info}/WHEEL +0 -0
- {x_transformers-2.9.2.dist-info → x_transformers-2.10.1.dist-info}/licenses/LICENSE +0 -0
x_transformers/attend.py
CHANGED
|
@@ -183,6 +183,7 @@ class Attend(Module):
|
|
|
183
183
|
gumbel_softmax = False,
|
|
184
184
|
gumbel_softmax_temp = 1.,
|
|
185
185
|
gumbel_softmax_hard = True,
|
|
186
|
+
cog_signed = False,
|
|
186
187
|
custom_attn_fn: Callable | None = None,
|
|
187
188
|
flash = False,
|
|
188
189
|
softclamp_logits = False,
|
|
@@ -260,6 +261,12 @@ class Attend(Module):
|
|
|
260
261
|
assert not (selective and not causal), 'selective attention is designed for autoregressive'
|
|
261
262
|
self.selective = selective
|
|
262
263
|
|
|
264
|
+
# cog attention - negative weights for expressiveness
|
|
265
|
+
# https://openreview.net/forum?id=ezRrwwbxd0
|
|
266
|
+
|
|
267
|
+
assert not (flash and cog_signed), 'cog attention not available for flash'
|
|
268
|
+
self.cog_signed = cog_signed
|
|
269
|
+
|
|
263
270
|
# l2 distance attention
|
|
264
271
|
|
|
265
272
|
self.l2_distance = l2_distance
|
|
@@ -509,6 +516,14 @@ class Attend(Module):
|
|
|
509
516
|
if self.softclamp_logits:
|
|
510
517
|
sim = softclamp(sim, self.logit_softclamp_value)
|
|
511
518
|
|
|
519
|
+
# pre-masking - handle cog by storing sign
|
|
520
|
+
|
|
521
|
+
if self.cog_signed:
|
|
522
|
+
sim_sign = sim.sign()
|
|
523
|
+
sim = sim.abs()
|
|
524
|
+
|
|
525
|
+
# masking
|
|
526
|
+
|
|
512
527
|
i, j, dtype = *sim.shape[-2:], sim.dtype
|
|
513
528
|
|
|
514
529
|
mask_value = -torch.finfo(sim.dtype).max
|
|
@@ -542,6 +557,11 @@ class Attend(Module):
|
|
|
542
557
|
|
|
543
558
|
attn = attn.type(dtype)
|
|
544
559
|
|
|
560
|
+
# add back the sign
|
|
561
|
+
|
|
562
|
+
if self.cog_signed:
|
|
563
|
+
attn = attn * sim_sign
|
|
564
|
+
|
|
545
565
|
post_softmax_attn = attn
|
|
546
566
|
|
|
547
567
|
if self.head_learned_sink:
|
x_transformers/x_transformers.py
CHANGED
|
@@ -1340,6 +1340,7 @@ class Attention(Module):
|
|
|
1340
1340
|
gumbel_softmax_temp = 1.,
|
|
1341
1341
|
gumbel_softmax_hard = True,
|
|
1342
1342
|
selective = False,
|
|
1343
|
+
cog_signed = False,
|
|
1343
1344
|
custom_attn_fn: Callable | None = None,
|
|
1344
1345
|
hybrid_module: Module | None = None,
|
|
1345
1346
|
hybrid_mask_kwarg: str | None = None,
|
|
@@ -1548,6 +1549,7 @@ class Attention(Module):
|
|
|
1548
1549
|
gumbel_softmax_temp = gumbel_softmax_temp,
|
|
1549
1550
|
gumbel_softmax_hard = gumbel_softmax_hard,
|
|
1550
1551
|
selective = selective,
|
|
1552
|
+
cog_signed = cog_signed,
|
|
1551
1553
|
custom_attn_fn = custom_attn_fn,
|
|
1552
1554
|
add_zero_kv = add_zero_kv,
|
|
1553
1555
|
head_learned_sink = head_learned_sink,
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: x-transformers
|
|
3
|
-
Version: 2.
|
|
3
|
+
Version: 2.10.1
|
|
4
4
|
Summary: X-Transformers
|
|
5
5
|
Project-URL: Homepage, https://pypi.org/project/x-transformers/
|
|
6
6
|
Project-URL: Repository, https://github.com/lucidrains/x-transformers
|
|
@@ -2586,4 +2586,16 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
|
|
|
2586
2586
|
}
|
|
2587
2587
|
```
|
|
2588
2588
|
|
|
2589
|
+
```bibtex
|
|
2590
|
+
@misc{lv2025expressiveattentionnegativeweights,
|
|
2591
|
+
title = {More Expressive Attention with Negative Weights},
|
|
2592
|
+
author = {Ang Lv and Ruobing Xie and Shuaipeng Li and Jiayi Liao and Xingwu Sun and Zhanhui Kang and Di Wang and Rui Yan},
|
|
2593
|
+
year = {2025},
|
|
2594
|
+
eprint = {2411.07176},
|
|
2595
|
+
archivePrefix = {arXiv},
|
|
2596
|
+
primaryClass = {cs.CL},
|
|
2597
|
+
url = {https://arxiv.org/abs/2411.07176},
|
|
2598
|
+
}
|
|
2599
|
+
```
|
|
2600
|
+
|
|
2589
2601
|
*solve intelligence... then use that to solve everything else.* - Demis Hassabis
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
x_transformers/__init__.py,sha256=aVuhUU0572TJHW88BVc4yA2tla0Zb8l3NH7W4RZ1AEs,1005
|
|
2
|
-
x_transformers/attend.py,sha256=
|
|
2
|
+
x_transformers/attend.py,sha256=uu4lIEfiwzZLeuBY2dJLG9709DZbWK8-on4ds8SCCJ0,19207
|
|
3
3
|
x_transformers/autoregressive_wrapper.py,sha256=BsGO9xfVYkvynqbU1__tu_S_cxl7gss0YwnkhIa2baY,18401
|
|
4
4
|
x_transformers/belief_state_wrapper.py,sha256=YLUMk6t2MhFBEw5lHDDHJHcoCxTIkHvxTNY__GGZEKU,13374
|
|
5
5
|
x_transformers/continuous.py,sha256=WwpQCjyVY4PtuEAOFY68zqgklbF9I7AL5w6874YlDe8,13249
|
|
@@ -10,10 +10,10 @@ x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg
|
|
|
10
10
|
x_transformers/neo_mlp.py,sha256=XCNnnop9WLarcxap1kGuYc1x8GHvwkZiDRnXOxSl3Po,3452
|
|
11
11
|
x_transformers/nonautoregressive_wrapper.py,sha256=hMQqNimGtchNIe13cR5LZule1V7I1qM5LmY8VQfVdnA,11698
|
|
12
12
|
x_transformers/up_wrapper.py,sha256=YC2LN14_7Xx9Wtiek2rtEJ_qHqdfSmKlh3d7Cgxwd80,7073
|
|
13
|
-
x_transformers/x_transformers.py,sha256=
|
|
13
|
+
x_transformers/x_transformers.py,sha256=ADr83Fz2cehj_F7N1bMwxhAg-r48fGhlaZqw3hxoxMQ,125765
|
|
14
14
|
x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
|
|
15
15
|
x_transformers/xval.py,sha256=AwwYUm8yDAtKQyKJDIhYMsiLTJ_skh3scUFMjp5sda8,8597
|
|
16
|
-
x_transformers-2.
|
|
17
|
-
x_transformers-2.
|
|
18
|
-
x_transformers-2.
|
|
19
|
-
x_transformers-2.
|
|
16
|
+
x_transformers-2.10.1.dist-info/METADATA,sha256=sEfcxJr3l0W4Yga0NLHq1sMk90Zr5-Lpr-9fIlmG9H4,95799
|
|
17
|
+
x_transformers-2.10.1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
|
|
18
|
+
x_transformers-2.10.1.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
|
|
19
|
+
x_transformers-2.10.1.dist-info/RECORD,,
|
|
File without changes
|
|
File without changes
|