rxnn 0.1.32__py3-none-any.whl → 0.1.33__py3-none-any.whl

rxnn/transformers/attention.py

@@ -1,6 +1,7 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from torch.backends.cuda import sdp_kernel, SDPBackend
 import math
 from .positional import RotaryPositionalEmbedding, RelativePositionalEmbedding
 
@@ -17,7 +18,7 @@ class MultiHeadAttention(nn.Module):
             rope_only_for_query: bool = False,
             use_relative_embeddings: bool = False,
             max_seq_len: int = 1024,
-            use_flash_attention: bool = False,
+            use_flash_attention: bool = True,
             is_causal: bool = False,
             use_bias: bool = False,
             *args,
@@ -101,13 +102,14 @@ class MultiHeadAttention(nn.Module):
 
     def _flash_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, b: int, t: int, d: int,
                          mask: torch.Tensor = None, enable_gqa: bool = False):
-        attn_output = F.scaled_dot_product_attention(
-            q, k, v,
-            attn_mask=mask if not self.is_causal else None,
-            dropout_p=self.dropout.p if self.training else 0.0,
-            is_causal=self.is_causal,
-            enable_gqa=enable_gqa,
-        )
+        with sdp_kernel(backends=[SDPBackend.FLASH_ATTENTION]):
+            attn_output = F.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask=mask if not self.is_causal else None,
+                dropout_p=self.dropout.p if self.training else 0.0,
+                is_causal=self.is_causal,
+                enable_gqa=enable_gqa,
+            )
         return self._transpose_output(attn_output, b, t, d)
 
     def _calculate_flash_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, b: int, t: int, d: int,

{rxnn-0.1.32.dist-info → rxnn-0.1.33.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rxnn
-Version: 0.1.32
+Version: 0.1.33
 Summary: RxNN: Reactive Neural Networks Platform
 License: Apache-2.0
 Keywords: deep-learning,ai,machine-learning

{rxnn-0.1.32.dist-info → rxnn-0.1.33.dist-info}/RECORD

@@ -16,7 +16,7 @@ rxnn/training/dataset.py,sha256=vQ5mDF3bA0HXya474n4D4iL8Mn3AEpJukgzFNVkxjGU,5106
 rxnn/training/scheduler.py,sha256=ow6oALzWjWQmHSpcJEjv6tg4g4CDMvr73TypxfcefMc,712
 rxnn/training/tokenizer.py,sha256=4Y41f07uo2KPA_7bp3FCcwGKbXoS2hsckOoXUsXfQxY,8052
 rxnn/transformers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-rxnn/transformers/attention.py,sha256=FHATZVf_kt3OHnG02zEeG9QdUXLncKDjrhyT28Pk0E4,14185
+rxnn/transformers/attention.py,sha256=Nox986BH9qq4rDYLiYmfj1DeMeULF3akexIl99MPccM,14331
 rxnn/transformers/ff.py,sha256=jJnuBDsnnX5uYC_WZH8cXAYrMnz0P-iX7MwcPivjRtI,2533
 rxnn/transformers/layers.py,sha256=n_jZTqEF_vLkF31AkB5XGErfm2sQFd9CRqJUHKRFkKI,6956
 rxnn/transformers/mask.py,sha256=J0cfLVLt3SzS2ra3KcY4khrkhI975Dw4CjpUi3Sn25s,419
@@ -25,7 +25,7 @@ rxnn/transformers/moe.py,sha256=6Cffyo0QjmEWc4rK1ncOmLRCQbY0OpQJ4D7xH_4nTN4,4738
 rxnn/transformers/positional.py,sha256=2l38RS0Dini3f6Z3LUHr3XwWzg1UK7fO2C6wazWDAYU,4292
 rxnn/transformers/sampler.py,sha256=poWBpxg1iuK5gEJtxHkk5VVfS9V48hs2Olqdhy_Gw8c,6548
 rxnn/utils.py,sha256=d5U8i5ukovgDyqiycc2AoxObTz_eF_bgo2MKvdtJ98s,467
-rxnn-0.1.32.dist-info/LICENSE,sha256=C8coDFIUYuOcke4JLPwTqahQUCyXyGq6WOaigOkx8tY,11275
-rxnn-0.1.32.dist-info/METADATA,sha256=oGoeF0F8LvsoUc8qGSd6n0oCIZmoj59bj42QwkaWDuM,16627
-rxnn-0.1.32.dist-info/WHEEL,sha256=fGIA9gx4Qxk2KDKeNJCbOEwSrmLtjWCwzBz351GyrPQ,88
-rxnn-0.1.32.dist-info/RECORD,,
+rxnn-0.1.33.dist-info/LICENSE,sha256=C8coDFIUYuOcke4JLPwTqahQUCyXyGq6WOaigOkx8tY,11275
+rxnn-0.1.33.dist-info/METADATA,sha256=m3DWDnTu7Lx1kHYPIAQCdKU8t4QZBdqG0QcSIFvB924,16627
+rxnn-0.1.33.dist-info/WHEEL,sha256=fGIA9gx4Qxk2KDKeNJCbOEwSrmLtjWCwzBz351GyrPQ,88
+rxnn-0.1.33.dist-info/RECORD,,