fix after changing importance score to compressed attention values

Author: lucidrains

native_sparse_attention_pytorch/native_sparse_attention.py

@@ -225,7 +225,7 @@ def forward(
         if self.use_diff_topk:
             gates = selected_importance_values + (1. - selected_importance_values).detach()
 
-        fmask = selected_importance_values > mask_value
+        fmask = selected_importance_values > 1e-10
 
         fq = q
         fk = k

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "native-sparse-attention-pytorch"
-version = "0.0.5"
+version = "0.0.6"
 description = "Native Sparse Attention"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }