use enable_gqa for flex attention for the sliding windows branch

Author: lucidrains

native_sparse_attention_pytorch/native_sparse_attention.py

@@ -446,11 +446,11 @@ def forward(
         sk = rotated_k
         sv = v
 
-        sk, sv = tuple(repeat(t, 'b h ... -> b (num_grouped_queries h) ...', num_grouped_queries = self.num_grouped_queries) for t in (sk, sv))
-
         if exists(sliding_window_flex_mask):
-            sliding_window_attn_out = flex_attention(sq, sk, sv, block_mask = sliding_window_flex_mask)
+            sliding_window_attn_out = flex_attention(sq, sk, sv, block_mask = sliding_window_flex_mask, enable_gqa = True)
         else:
+            sk, sv = tuple(repeat(t, 'b h ... -> b (num_grouped_queries h) ...', num_grouped_queries = self.num_grouped_queries) for t in (sk, sv))
+
             sliding_window_attn_out = self.sliding_window(sq, sk, sv)
 
         # combine strategies

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "native-sparse-attention-pytorch"
-version = "0.0.29"
+version = "0.0.30"
 description = "Native Sparse Attention"
 authors = [
     { name = "Phil Wang", email = "[email protected]" }