broccoli-ml 15.3.0__py3-none-any.whl → 15.4.1__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.
broccoli/vit.py CHANGED
@@ -520,9 +520,13 @@ class ViT(nn.Module):
520
520
  "SwiGLU": SwiGLU,
521
521
  }[transformer_activation]
522
522
 
523
- # Set alpha and beta according to Microsoft's DeepNorm
524
- self.alpha = (2 * transformer_layers) ** 0.25
525
- self.beta = (8 * transformer_layers) ** -0.25
523
+ # Set alpha according to Microsoft's DeepNorm if layers > 50
524
+ if transformer_layers > 50:
525
+ self.alpha = (2 * transformer_layers) ** 0.25
526
+ self.beta = 1.0 # beta is not needed as we norm the Q and K vectors in MSA!
527
+ else:
528
+ self.alpha = 1.0
529
+ self.beta = 1.0 # beta is not needed as we norm the Q and K vectors in MSA!
526
530
 
527
531
  self.encoder = ViTEncoder(
528
532
  input_size=input_size,
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.3
2
2
  Name: broccoli-ml
3
- Version: 15.3.0
3
+ Version: 15.4.1
4
4
  Summary: Some useful Pytorch models, circa 2025
5
5
  License: MIT
6
6
  Author: Nicholas Bailey
@@ -6,8 +6,8 @@ broccoli/rope.py,sha256=GRqApBNmYCFaDak0WL1xE_BC5CTTYKQU_PBdeTcQcjc,12557
6
6
  broccoli/tensor.py,sha256=um8mrxkYbvNDo-QvHlmJm8Aw6qcngOlUZPoAk_PMReA,4480
7
7
  broccoli/transformer.py,sha256=XCr3fkDrYvNEA3m5rdVDHU5qbf1IYJQS1Cjgmn5RX-o,28177
8
8
  broccoli/utils.py,sha256=oOWzn6dJ5nC_9r4zq0emmfmaYACJXJNFS48AOpW2jqc,358
9
- broccoli/vit.py,sha256=9z1MJce38dppX35OGJvmnNIOnt-euMVV4n160W_aQYU,22799
10
- broccoli_ml-15.3.0.dist-info/LICENSE,sha256=0BAzJE5BqQ7Iixp_AFdB2W1uO-HCRX-Qfun8PHt6yVM,1073
11
- broccoli_ml-15.3.0.dist-info/METADATA,sha256=StOW7vZd1c4xNddKcA8XJ-zbXgUnAsrQlg9G97wT-2w,1369
12
- broccoli_ml-15.3.0.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
13
- broccoli_ml-15.3.0.dist-info/RECORD,,
9
+ broccoli/vit.py,sha256=lhWOcVg4GfTHJp3ABflKjQQzWGKMqrOU6QKqRWU11wo,23012
10
+ broccoli_ml-15.4.1.dist-info/LICENSE,sha256=0BAzJE5BqQ7Iixp_AFdB2W1uO-HCRX-Qfun8PHt6yVM,1073
11
+ broccoli_ml-15.4.1.dist-info/METADATA,sha256=Vg2KdoHSlbV1IDYuNk8h07sVT1UTW69bt49JLiEr0os,1369
12
+ broccoli_ml-15.4.1.dist-info/WHEEL,sha256=b4K_helf-jlQoXBBETfwnf4B04YC67LOev0jo4fX5m8,88
13
+ broccoli_ml-15.4.1.dist-info/RECORD,,