rxnn 0.1.33__tar.gz → 0.1.34__tar.gz

{rxnn-0.1.33 → rxnn-0.1.34}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rxnn
-Version: 0.1.33
+Version: 0.1.34
 Summary: RxNN: Reactive Neural Networks Platform
 License: Apache-2.0
 Keywords: deep-learning,ai,machine-learning

{rxnn-0.1.33 → rxnn-0.1.34}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "rxnn"
-version = "0.1.33"
+version = "0.1.34"
 description = "RxNN: Reactive Neural Networks Platform"
 
 license = "Apache-2.0"

{rxnn-0.1.33 → rxnn-0.1.34}/src/rxnn/transformers/attention.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from torch.backends.cuda import sdp_kernel, SDPBackend
+from torch.nn.attention import sdpa_kernel, SDPBackend
 import math
 from .positional import RotaryPositionalEmbedding, RelativePositionalEmbedding
 
@@ -102,7 +102,7 @@ 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):
-        with sdp_kernel(backends=[SDPBackend.FLASH_ATTENTION]):
+        with sdpa_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,