diff --git a/fastvideo/attention/backends/sageattn/api.py b/fastvideo/attention/backends/sageattn/api.py
index 28b11f4f1..def4f50fe 100644
--- a/fastvideo/attention/backends/sageattn/api.py
+++ b/fastvideo/attention/backends/sageattn/api.py
@@ -24,8 +24,8 @@ import fp4quant_cuda
 
 # Centralized block size configuration for sageattn_blackwell kernels
 # These should match the values in fastvideo/attention/backends/sageattn/blackwell/block_config.h
-BLOCK_M = 128  # Block size for M dimension (query sequence length)
-BLOCK_N = 128  # Block size for N dimension (key/value sequence length)
+BLOCK_M = 64  # Block size for M dimension (query sequence length)
+BLOCK_N = 64  # Block size for N dimension (key/value sequence length)
 
 
 @triton.jit
diff --git a/fastvideo/attention/backends/sageattn/blackwell/block_config.h b/fastvideo/attention/backends/sageattn/blackwell/block_config.h
index e4ecfda77..631e605ae 100644
--- a/fastvideo/attention/backends/sageattn/blackwell/block_config.h
+++ b/fastvideo/attention/backends/sageattn/blackwell/block_config.h
@@ -20,9 +20,9 @@
 // Block sizes for M and N dimensions
 namespace flash {
     // Block size for M dimension (query sequence length)
-    static constexpr int BLOCK_M = 128;
+    static constexpr int BLOCK_M = 64;
     
     // Block size for N dimension (key/value sequence length)
-    static constexpr int BLOCK_N = 128;
+    static constexpr int BLOCK_N = 64;
 }